diff --git a/.github/workflows/build-tarball.yml b/.github/workflows/build-tarball.yml
index a784b43cfc6268..0c73938764165b 100644
--- a/.github/workflows/build-tarball.yml
+++ b/.github/workflows/build-tarball.yml
@@ -96,7 +96,7 @@ jobs:
           echo "TAR_DIR=$RUNNER_TEMP/`basename tarballs/*.tar.gz .tar.gz`" >> $GITHUB_ENV
       - name: Copy directories needed for testing
         run: |
-          cp -r tools/eslint $TAR_DIR/tools
+          cp -r tools/node_modules $TAR_DIR/tools
           cp -r tools/eslint-rules $TAR_DIR/tools
       - name: Build
         run: |
diff --git a/.github/workflows/build-windows.yml b/.github/workflows/build-windows.yml
deleted file mode 100644
index 37e1d789bb3df6..00000000000000
--- a/.github/workflows/build-windows.yml
+++ /dev/null
@@ -1,53 +0,0 @@
-name: Build Windows
-
-on:
-  pull_request:
-    paths-ignore:
-      - README.md
-      - .github/**
-      - '!.github/workflows/build-windows.yml'
-    types: [opened, synchronize, reopened, ready_for_review]
-  push:
-    branches:
-      - main
-      - canary
-      - v[0-9]+.x-staging
-      - v[0-9]+.x
-    paths-ignore:
-      - README.md
-      - .github/**
-      - '!.github/workflows/build-windows.yml'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
-  cancel-in-progress: true
-
-env:
-  PYTHON_VERSION: '3.12'
-  FLAKY_TESTS: keep_retrying
-
-permissions:
-  contents: read
-
-jobs:
-  build-windows:
-    if: github.event.pull_request.draft == false
-    strategy:
-      matrix:
-        windows: [windows-2019, windows-2022]
-      fail-fast: false
-    runs-on: ${{ matrix.windows }}
-    steps:
-      - uses: actions/checkout@692973e3d937129bcbf40652eb9f2f61becf3332  # v4.1.7
-        with:
-          persist-credentials: false
-      - name: Set up Python ${{ env.PYTHON_VERSION }}
-        uses: actions/setup-python@f677139bbe7f9c59b41e40162b753c062f5d49a3  # v5.2.0
-        with:
-          python-version: ${{ env.PYTHON_VERSION }}
-      - name: Install deps
-        run: choco install nasm
-      - name: Environment Information
-        run: npx envinfo
-      - name: Build
-        run: ./vcbuild.bat
diff --git a/.github/workflows/test-linux.yml b/.github/workflows/test-linux.yml
index 2266d0da067cdd..6e8ff1fd6492c0 100644
--- a/.github/workflows/test-linux.yml
+++ b/.github/workflows/test-linux.yml
@@ -35,16 +35,11 @@ permissions:
 jobs:
   test-linux:
     if: github.event.pull_request.draft == false
-    runs-on: ${{ matrix.os }}
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ubuntu-24.04, ubuntu-24.04-arm]
+    runs-on: ubuntu-24.04
     steps:
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683  # v4.2.2
         with:
           persist-credentials: false
-          path: node
       - name: Set up Python ${{ env.PYTHON_VERSION }}
         uses: actions/setup-python@8d9ed9ac5c53483de85588cdf95a591a75ab9f55  # v5.5.0
         with:
@@ -56,13 +51,6 @@ jobs:
       - name: Environment Information
         run: npx envinfo
       - name: Build
-        run: make -C node build-ci -j4 V=1 CONFIG_FLAGS="--error-on-warn"
+        run: make build-ci -j4 V=1 CONFIG_FLAGS="--error-on-warn"
       - name: Test
-        run: make -C node run-ci -j4 V=1 TEST_CI_ARGS="-p actions --measure-flakiness 9"
-      - name: Re-run test in a folder whose name contains unusual chars
-        run: |
-          mv node "$DIR"
-          cd "$DIR"
-          ./tools/test.py --flaky-tests keep_retrying -p actions -j 4
-        env:
-          DIR: dir%20with $unusual"chars?'åß∂ƒ©∆¬…`
+        run: make run-ci -j4 V=1 TEST_CI_ARGS="-p actions --measure-flakiness 9"
diff --git a/.github/workflows/test-macos.yml b/.github/workflows/test-macos.yml
index c45799a022b035..285d67bdb8580d 100644
--- a/.github/workflows/test-macos.yml
+++ b/.github/workflows/test-macos.yml
@@ -58,7 +58,7 @@ jobs:
       - name: Set up sccache
         uses: Mozilla-Actions/sccache-action@7d986dd989559c6ecdb630a3fd2557667be217ad  # v0.0.9
         with:
-          version: v0.8.1
+          version: v0.10.0
       - name: Environment Information
         run: npx envinfo
       # The `npm ci` for this step fails a lot as part of the Test step. Run it
diff --git a/CHANGELOG.md b/CHANGELOG.md
index f0d7f9bfa8675a..d72f6492205aa2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -35,7 +35,8 @@ release.
 </tr>
 <tr>
   <td valign="top">
-<b><a href="doc/changelogs/CHANGELOG_V20.md#20.19.2">20.19.2</a></b><br/>
+<b><a href="doc/changelogs/CHANGELOG_V20.md#20.19.3">20.19.3</a></b><br/>
+<a href="doc/changelogs/CHANGELOG_V20.md#20.19.2">20.19.2</a><br/>
 <a href="doc/changelogs/CHANGELOG_V20.md#20.19.1">20.19.1</a><br/>
 <a href="doc/changelogs/CHANGELOG_V20.md#20.19.0">20.19.0</a><br/>
 <a href="doc/changelogs/CHANGELOG_V20.md#20.18.3">20.18.3</a><br/>
diff --git a/Makefile b/Makefile
index 6af722038efcc5..9b7293a78c9e38 100644
--- a/Makefile
+++ b/Makefile
@@ -171,8 +171,7 @@ with-code-cache test-code-cache:
 	$(warning '$@' target is a noop)
 
 out/Makefile: config.gypi common.gypi node.gyp \
-	deps/uv/uv.gyp deps/llhttp/llhttp.gyp deps/zlib/zlib.gyp \
-	deps/simdutf/simdutf.gyp deps/ada/ada.gyp \
+	deps/*/*.gyp \
 	tools/v8_gypfiles/toolchain.gypi tools/v8_gypfiles/features.gypi \
 	tools/v8_gypfiles/inspector.gypi tools/v8_gypfiles/v8.gyp
 	$(PYTHON) tools/gyp_node.py -f make
diff --git a/README.md b/README.md
index 82ce19db742f46..4475e203792140 100644
--- a/README.md
+++ b/README.md
@@ -317,6 +317,8 @@ For information about the governance of the Node.js project, see
   **Kohei Ueno** <<kohei.ueno119@gmail.com>> (he/him)
 * [daeyeon](https://github.com/daeyeon) -
   **Daeyeon Jeong** <<daeyeon.dev@gmail.com>> (he/him)
+* [dario-piotrowicz](https://github.com/dario-piotrowicz) -
+  **Dario Piotrowicz** <<dario.piotrowicz@gmail.com>> (he/him)
 * [debadree25](https://github.com/debadree25) -
   **Debadree Chatterjee** <<debadree333@gmail.com>> (he/him)
 * [deokjinkim](https://github.com/deokjinkim) -
@@ -745,6 +747,8 @@ maintaining the Node.js project.
   **Feng Yu** <<F3n67u@outlook.com>> (he/him)
 * [gireeshpunathil](https://github.com/gireeshpunathil) -
   **Gireesh Punathil** <<gpunathi@in.ibm.com>> (he/him)
+* [gurgunday](https://github.com/gurgunday) -
+  **Gürgün Dayıoğlu** <<hey@gurgun.day>>
 * [iam-frankqiu](https://github.com/iam-frankqiu) -
   **Frank Qiu** <<iam.frankqiu@gmail.com>> (he/him)
 * [KevinEady](https://github.com/KevinEady) -
diff --git a/SECURITY.md b/SECURITY.md
index b8f54307d5ed5b..3193db0d4dd0b2 100644
--- a/SECURITY.md
+++ b/SECURITY.md
@@ -55,6 +55,39 @@ Here is the security disclosure policy for Node.js
   possible; however, we must follow the release process above to ensure that we
   handle disclosure consistently.
 
+## Code of Conduct and Vulnerability Reporting Guidelines
+
+When reporting security vulnerabilities, reporters must adhere to the following guidelines:
+
+1. **Code of Conduct Compliance**: All security reports must comply with our
+   [Code of Conduct](CODE_OF_CONDUCT.md). Reports that violate our code of conduct
+   will not be considered and may result in being banned from future participation.
+
+2. **No Harmful Actions**: Security research and vulnerability reporting must not:
+   * Cause damage to running systems or production environments.
+   * Disrupt Node.js development or infrastructure.
+   * Affect other users' applications or systems.
+   * Include actual exploits that could harm users.
+   * Involve social engineering or phishing attempts.
+
+3. **Responsible Testing**: When testing potential vulnerabilities:
+   * Use isolated, controlled environments.
+   * Do not test on production systems.
+   * Do not attempt to access or modify other users' data.
+   * Immediately stop testing if unauthorized access is gained accidentally.
+
+4. **Report Quality**
+   * Provide clear, detailed steps to reproduce the vulnerability.
+   * Include only the minimum proof of concept required to demonstrate the issue.
+   * Remove any malicious payloads or components that could cause harm.
+
+Failure to follow these guidelines may result in:
+
+* Rejection of the vulnerability report.
+* Forfeiture of any potential bug bounty.
+* Temporary or permanent ban from the bug bounty program.
+* Legal action in cases of malicious intent.
+
 ## The Node.js threat model
 
 In the Node.js threat model, there are trusted elements such as the
diff --git a/common.gypi b/common.gypi
index d3c17d47bc141b..888b8b94ca1abd 100644
--- a/common.gypi
+++ b/common.gypi
@@ -36,7 +36,7 @@
 
     # Reset this number to 0 on major V8 upgrades.
     # Increment by one for each non-official patch applied to deps/v8.
-    'v8_embedder_string': '-node.26',
+    'v8_embedder_string': '-node.29',
 
     ##### V8 defaults for Node.js #####
 
@@ -268,6 +268,9 @@
     # Defines these mostly for node-gyp to pickup.
     'defines': [
       '_GLIBCXX_USE_CXX11_ABI=1',
+      # This help forks when building Node.js on a 32-bit arch as
+      # libuv is always compiled with _FILE_OFFSET_BITS=64
+      '_FILE_OFFSET_BITS=64'
     ],
 
     # Forcibly disable -Werror.  We support a wide range of compilers, it's
diff --git a/deps/corepack/CHANGELOG.md b/deps/corepack/CHANGELOG.md
index 88363683a9d5f6..ad0bba51c6b83f 100644
--- a/deps/corepack/CHANGELOG.md
+++ b/deps/corepack/CHANGELOG.md
@@ -1,5 +1,23 @@
 # Changelog
 
+## [0.32.0](https://github.com/nodejs/corepack/compare/v0.31.0...v0.32.0) (2025-02-28)
+
+
+### Features
+
+* add limited support for `devEngines` ([#643](https://github.com/nodejs/corepack/issues/643)) ([b456268](https://github.com/nodejs/corepack/commit/b4562688513f23e37e37b0d69a0daff33ca84c8d))
+* add more informative error when fetching latest stable fails ([#644](https://github.com/nodejs/corepack/issues/644)) ([53b1fe7](https://github.com/nodejs/corepack/commit/53b1fe75c47c06bd72a8b8f8bb699a47c9ca32fb))
+* add support for `.corepack.env` ([#642](https://github.com/nodejs/corepack/issues/642)) ([9b95b46](https://github.com/nodejs/corepack/commit/9b95b46f05e50fe1c60f05309c210ba8fe4e23c5))
+* update package manager versions ([#617](https://github.com/nodejs/corepack/issues/617)) ([b83bb5e](https://github.com/nodejs/corepack/commit/b83bb5ec150980c998b9c7053dff307d912cb508))
+
+
+### Bug Fixes
+
+* do not resolve fallback descriptor when `packageManager` is defined ([#632](https://github.com/nodejs/corepack/issues/632)) ([12e77e5](https://github.com/nodejs/corepack/commit/12e77e506946d42a0de9ce8e68d75af8454d6776))
+* **doc:** fix link to proxy library ([#636](https://github.com/nodejs/corepack/issues/636)) ([bae0839](https://github.com/nodejs/corepack/commit/bae08397943d4b99437389b4286546361091f4b3))
+* replace explicit with specify as verb ([#665](https://github.com/nodejs/corepack/issues/665)) ([351d86c](https://github.com/nodejs/corepack/commit/351d86c20226a8c18bfe212be27401f2908b1595))
+* **use:** do not throw on invalid `packageManager` ([#663](https://github.com/nodejs/corepack/issues/663)) ([4be72f6](https://github.com/nodejs/corepack/commit/4be72f6941afa0c9b2b7d26635016bb7b560df8a))
+
 ## [0.31.0](https://github.com/nodejs/corepack/compare/v0.30.0...v0.31.0) (2025-01-27)
 
 
diff --git a/deps/corepack/README.md b/deps/corepack/README.md
index 66bfbc3fb6aae3..67d396d5ca9642 100644
--- a/deps/corepack/README.md
+++ b/deps/corepack/README.md
@@ -41,6 +41,25 @@ is distributed along with Node.js itself.
 
 </details>
 
+<details><summary>Update Corepack using npm</summary>
+
+To install the latest version of Corepack, use:
+
+```shell
+npm install -g corepack@latest
+```
+
+If Corepack was installed on your system using a Node.js Windows Installer
+`.msi` package then you might need to remove it before attempting to install a
+different version of Corepack using npm. You can select the Modify option of the
+Node.js app settings to access the Windows Installer feature selection, and on
+the "corepack manager" feature of the Node.js `.msi` package by selecting
+"Entire feature will be unavailable". See
+[Repair apps and programs in Windows](https://support.microsoft.com/en-us/windows/repair-apps-and-programs-in-windows-e90eefe4-d0a2-7c1b-dd59-949a9030f317)
+for instructions on accessing the Windows apps page to modify settings.
+
+</details>
+
 <details><summary>Install Corepack from source</summary>
 
 See [`CONTRIBUTING.md`](./CONTRIBUTING.md).
@@ -94,6 +113,35 @@ use in the archive).
 }
 ```
 
+#### `devEngines.packageManager`
+
+When a `devEngines.packageManager` field is defined, and is an object containing
+a `"name"` field (can also optionally contain `version` and `onFail` fields),
+Corepack will use it to validate you're using a compatible package manager.
+
+Depending on the value of `devEngines.packageManager.onFail`:
+
+- if set to `ignore`, Corepack won't print any warning or error.
+- if unset or set to `error`, Corepack will throw an error in case of a mismatch.
+- if set to `warn` or some other value, Corepack will print a warning in case
+  of mismatch.
+
+If the top-level `packageManager` field is missing, Corepack will use the
+package manager defined in `devEngines.packageManager` – in which case you must
+provide a specific version in `devEngines.packageManager.version`, ideally with
+a hash, as explained in the previous section:
+
+```json
+{
+  "devEngines":{
+    "packageManager": {
+      "name": "yarn",
+      "version": "3.2.3+sha224.953c8233f7a92884eee2de69a1b92d1f2ec1655e66d08071ba9a02fa"
+    }
+  }
+}
+```
+
 ## Known Good Releases
 
 When running Corepack within projects that don't list a supported package
@@ -227,6 +275,7 @@ it.
 
 Unlike `corepack use` this command doesn't take a package manager name nor a
 version range, as it will always select the latest available version from the
+range specified in `devEngines.packageManager.version`, or fallback to the
 same major line. Should you need to upgrade to a new major, use an explicit
 `corepack use {name}@latest` call (or simply `corepack use {name}`).
 
@@ -248,6 +297,7 @@ same major line. Should you need to upgrade to a new major, use an explicit
   set to `1` to have the URL shown. By default, when Corepack is called
   explicitly (e.g. `corepack pnpm …`), it is set to `0`; when Corepack is called
   implicitly (e.g. `pnpm …`), it is set to `1`.
+  The default value cannot be overridden in a `.corepack.env` file.
   When standard input is a TTY and no CI environment is detected, Corepack will
   ask for user input before starting the download.
 
@@ -273,6 +323,14 @@ same major line. Should you need to upgrade to a new major, use an explicit
   project. This means that it will always use the system-wide package manager
   regardless of what is being specified in the project's `packageManager` field.
 
+- `COREPACK_ENV_FILE` can be set to `0` to request Corepack to not attempt to
+  load `.corepack.env`; it can be set to a path to specify a different env file.
+  Only keys that start with `COREPACK_` and are not in the exception list
+  (`COREPACK_ENABLE_DOWNLOAD_PROMPT` and `COREPACK_ENV_FILE` are ignored)
+  will be taken into account.
+  For Node.js 18.x users, this setting has no effect as that version doesn't
+  support parsing of `.env` files.
+
 - `COREPACK_HOME` can be set in order to define where Corepack should install
   the package managers. By default it is set to `%LOCALAPPDATA%\node\corepack`
   on Windows, and to `$HOME/.cache/node/corepack` everywhere else.
@@ -294,7 +352,7 @@ same major line. Should you need to upgrade to a new major, use an explicit
   empty password, explicitly set `COREPACK_NPM_PASSWORD` to an empty string.
 
 - `HTTP_PROXY`, `HTTPS_PROXY`, and `NO_PROXY` are supported through
-  [`node-proxy-agent`](https://github.com/TooTallNate/node-proxy-agent).
+  [`proxy-from-env`](https://github.com/Rob--W/proxy-from-env).
 
 - `COREPACK_INTEGRITY_KEYS` can be set to an empty string or `0` to
   instruct Corepack to skip integrity checks, or to a JSON string containing
diff --git a/deps/corepack/dist/lib/corepack.cjs b/deps/corepack/dist/lib/corepack.cjs
index 7a92f3334f7687..7a098cbe250aab 100644
--- a/deps/corepack/dist/lib/corepack.cjs
+++ b/deps/corepack/dist/lib/corepack.cjs
@@ -1090,18 +1090,18 @@ var require_node = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/debug.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/debug.js
 var require_debug = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/debug.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/debug.js"(exports2, module2) {
     var debug2 = typeof process === "object" && process.env && process.env.NODE_DEBUG && /\bsemver\b/i.test(process.env.NODE_DEBUG) ? (...args) => console.error("SEMVER", ...args) : () => {
     };
     module2.exports = debug2;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/constants.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/constants.js
 var require_constants = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/constants.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/constants.js"(exports2, module2) {
     var SEMVER_SPEC_VERSION = "2.0.0";
     var MAX_LENGTH = 256;
     var MAX_SAFE_INTEGER = Number.MAX_SAFE_INTEGER || /* istanbul ignore next */
@@ -1130,9 +1130,9 @@ var require_constants = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/re.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/re.js
 var require_re = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/re.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/re.js"(exports2, module2) {
     var {
       MAX_SAFE_COMPONENT_LENGTH,
       MAX_SAFE_BUILD_LENGTH,
@@ -1143,6 +1143,7 @@ var require_re = __commonJS({
     var re = exports2.re = [];
     var safeRe = exports2.safeRe = [];
     var src = exports2.src = [];
+    var safeSrc = exports2.safeSrc = [];
     var t = exports2.t = {};
     var R = 0;
     var LETTERDASHNUMBER = "[a-zA-Z0-9-]";
@@ -1163,6 +1164,7 @@ var require_re = __commonJS({
       debug2(name2, index, value);
       t[name2] = index;
       src[index] = value;
+      safeSrc[index] = safe;
       re[index] = new RegExp(value, isGlobal ? "g" : void 0);
       safeRe[index] = new RegExp(safe, isGlobal ? "g" : void 0);
     };
@@ -1215,9 +1217,9 @@ var require_re = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/parse-options.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/parse-options.js
 var require_parse_options = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/parse-options.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/parse-options.js"(exports2, module2) {
     var looseOption = Object.freeze({ loose: true });
     var emptyOpts = Object.freeze({});
     var parseOptions = (options) => {
@@ -1233,9 +1235,9 @@ var require_parse_options = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/identifiers.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/identifiers.js
 var require_identifiers = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/identifiers.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/identifiers.js"(exports2, module2) {
     var numeric = /^[0-9]+$/;
     var compareIdentifiers = (a, b) => {
       const anum = numeric.test(a);
@@ -1254,12 +1256,12 @@ var require_identifiers = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/semver.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/semver.js
 var require_semver = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/semver.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/semver.js"(exports2, module2) {
     var debug2 = require_debug();
     var { MAX_LENGTH, MAX_SAFE_INTEGER } = require_constants();
-    var { safeRe: re, t } = require_re();
+    var { safeRe: re, safeSrc: src, t } = require_re();
     var parseOptions = require_parse_options();
     var { compareIdentifiers } = require_identifiers();
     var SemVer3 = class _SemVer {
@@ -1399,6 +1401,18 @@ var require_semver = __commonJS({
       // preminor will bump the version up to the next minor release, and immediately
       // down to pre-release. premajor and prepatch work the same way.
       inc(release, identifier, identifierBase) {
+        if (release.startsWith("pre")) {
+          if (!identifier && identifierBase === false) {
+            throw new Error("invalid increment argument: identifier is empty");
+          }
+          if (identifier) {
+            const r = new RegExp(`^${this.options.loose ? src[t.PRERELEASELOOSE] : src[t.PRERELEASE]}$`);
+            const match = `-${identifier}`.match(r);
+            if (!match || match[1] !== identifier) {
+              throw new Error(`invalid identifier: ${identifier}`);
+            }
+          }
+        }
         switch (release) {
           case "premajor":
             this.prerelease.length = 0;
@@ -1418,12 +1432,20 @@ var require_semver = __commonJS({
             this.inc("patch", identifier, identifierBase);
             this.inc("pre", identifier, identifierBase);
             break;
+          // If the input is a non-prerelease version, this acts the same as
+          // prepatch.
           case "prerelease":
             if (this.prerelease.length === 0) {
               this.inc("patch", identifier, identifierBase);
             }
             this.inc("pre", identifier, identifierBase);
             break;
+          case "release":
+            if (this.prerelease.length === 0) {
+              throw new Error(`version ${this.raw} is not a prerelease`);
+            }
+            this.prerelease.length = 0;
+            break;
           case "major":
             if (this.minor !== 0 || this.patch !== 0 || this.prerelease.length === 0) {
               this.major++;
@@ -1445,11 +1467,10 @@ var require_semver = __commonJS({
             }
             this.prerelease = [];
             break;
+          // This probably shouldn't be used publicly.
+          // 1.0.0 'pre' would become 1.0.0-0 which is the wrong direction.
           case "pre": {
             const base = Number(identifierBase) ? 1 : 0;
-            if (!identifier && identifierBase === false) {
-              throw new Error("invalid increment argument: identifier is empty");
-            }
             if (this.prerelease.length === 0) {
               this.prerelease = [base];
             } else {
@@ -1496,27 +1517,27 @@ var require_semver = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/compare.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/compare.js
 var require_compare = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/compare.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/compare.js"(exports2, module2) {
     var SemVer3 = require_semver();
     var compare = (a, b, loose) => new SemVer3(a, loose).compare(new SemVer3(b, loose));
     module2.exports = compare;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/rcompare.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/rcompare.js
 var require_rcompare = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/rcompare.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/rcompare.js"(exports2, module2) {
     var compare = require_compare();
     var rcompare = (a, b, loose) => compare(b, a, loose);
     module2.exports = rcompare;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/parse.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/parse.js
 var require_parse = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/parse.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/parse.js"(exports2, module2) {
     var SemVer3 = require_semver();
     var parse5 = (version3, options, throwErrors = false) => {
       if (version3 instanceof SemVer3) {
@@ -1535,9 +1556,9 @@ var require_parse = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/valid.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/valid.js
 var require_valid = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/valid.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/valid.js"(exports2, module2) {
     var parse5 = require_parse();
     var valid = (version3, options) => {
       const v = parse5(version3, options);
@@ -1547,9 +1568,9 @@ var require_valid = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/lrucache.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/lrucache.js
 var require_lrucache = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/internal/lrucache.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/internal/lrucache.js"(exports2, module2) {
     var LRUCache = class {
       constructor() {
         this.max = 1e3;
@@ -1584,63 +1605,63 @@ var require_lrucache = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/eq.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/eq.js
 var require_eq = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/eq.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/eq.js"(exports2, module2) {
     var compare = require_compare();
     var eq = (a, b, loose) => compare(a, b, loose) === 0;
     module2.exports = eq;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/neq.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/neq.js
 var require_neq = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/neq.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/neq.js"(exports2, module2) {
     var compare = require_compare();
     var neq = (a, b, loose) => compare(a, b, loose) !== 0;
     module2.exports = neq;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/gt.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/gt.js
 var require_gt = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/gt.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/gt.js"(exports2, module2) {
     var compare = require_compare();
     var gt = (a, b, loose) => compare(a, b, loose) > 0;
     module2.exports = gt;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/gte.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/gte.js
 var require_gte = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/gte.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/gte.js"(exports2, module2) {
     var compare = require_compare();
     var gte = (a, b, loose) => compare(a, b, loose) >= 0;
     module2.exports = gte;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/lt.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/lt.js
 var require_lt = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/lt.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/lt.js"(exports2, module2) {
     var compare = require_compare();
     var lt = (a, b, loose) => compare(a, b, loose) < 0;
     module2.exports = lt;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/lte.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/lte.js
 var require_lte = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/lte.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/lte.js"(exports2, module2) {
     var compare = require_compare();
     var lte = (a, b, loose) => compare(a, b, loose) <= 0;
     module2.exports = lte;
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/cmp.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/cmp.js
 var require_cmp = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/cmp.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/cmp.js"(exports2, module2) {
     var eq = require_eq();
     var neq = require_neq();
     var gt = require_gt();
@@ -1687,9 +1708,9 @@ var require_cmp = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/comparator.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/comparator.js
 var require_comparator = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/comparator.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/comparator.js"(exports2, module2) {
     var ANY = Symbol("SemVer ANY");
     var Comparator = class _Comparator {
       static get ANY() {
@@ -1799,9 +1820,9 @@ var require_comparator = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/range.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/range.js
 var require_range = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/classes/range.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/classes/range.js"(exports2, module2) {
     var SPACE_CHARACTERS = /\s+/g;
     var Range3 = class _Range {
       constructor(range, options) {
@@ -2174,9 +2195,9 @@ var require_range = __commonJS({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/ranges/valid.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/ranges/valid.js
 var require_valid2 = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/ranges/valid.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/ranges/valid.js"(exports2, module2) {
     var Range3 = require_range();
     var validRange = (range, options) => {
       try {
@@ -2189,9 +2210,9 @@ var require_valid2 = __commonJS({
   }
 });
 
-// .yarn/cache/ms-npm-2.1.2-ec0c1512ff-a437714e2f.zip/node_modules/ms/index.js
+// .yarn/cache/ms-npm-2.1.3-81ff3cfac1-d924b57e73.zip/node_modules/ms/index.js
 var require_ms = __commonJS({
-  ".yarn/cache/ms-npm-2.1.2-ec0c1512ff-a437714e2f.zip/node_modules/ms/index.js"(exports2, module2) {
+  ".yarn/cache/ms-npm-2.1.3-81ff3cfac1-d924b57e73.zip/node_modules/ms/index.js"(exports2, module2) {
     var s = 1e3;
     var m = s * 60;
     var h = m * 60;
@@ -2305,9 +2326,9 @@ var require_ms = __commonJS({
   }
 });
 
-// .yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/common.js
+// .yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/common.js
 var require_common = __commonJS({
-  ".yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/common.js"(exports2, module2) {
+  ".yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/common.js"(exports2, module2) {
     function setup(env2) {
       createDebug.debug = createDebug;
       createDebug.default = createDebug;
@@ -2408,50 +2429,64 @@ var require_common = __commonJS({
         createDebug.namespaces = namespaces;
         createDebug.names = [];
         createDebug.skips = [];
-        let i;
-        const split = (typeof namespaces === "string" ? namespaces : "").split(/[\s,]+/);
-        const len = split.length;
-        for (i = 0; i < len; i++) {
-          if (!split[i]) {
-            continue;
+        const split = (typeof namespaces === "string" ? namespaces : "").trim().replace(" ", ",").split(",").filter(Boolean);
+        for (const ns of split) {
+          if (ns[0] === "-") {
+            createDebug.skips.push(ns.slice(1));
+          } else {
+            createDebug.names.push(ns);
           }
-          namespaces = split[i].replace(/\*/g, ".*?");
-          if (namespaces[0] === "-") {
-            createDebug.skips.push(new RegExp("^" + namespaces.slice(1) + "$"));
+        }
+      }
+      function matchesTemplate(search, template) {
+        let searchIndex = 0;
+        let templateIndex = 0;
+        let starIndex = -1;
+        let matchIndex = 0;
+        while (searchIndex < search.length) {
+          if (templateIndex < template.length && (template[templateIndex] === search[searchIndex] || template[templateIndex] === "*")) {
+            if (template[templateIndex] === "*") {
+              starIndex = templateIndex;
+              matchIndex = searchIndex;
+              templateIndex++;
+            } else {
+              searchIndex++;
+              templateIndex++;
+            }
+          } else if (starIndex !== -1) {
+            templateIndex = starIndex + 1;
+            matchIndex++;
+            searchIndex = matchIndex;
           } else {
-            createDebug.names.push(new RegExp("^" + namespaces + "$"));
+            return false;
           }
         }
+        while (templateIndex < template.length && template[templateIndex] === "*") {
+          templateIndex++;
+        }
+        return templateIndex === template.length;
       }
       function disable() {
         const namespaces = [
-          ...createDebug.names.map(toNamespace),
-          ...createDebug.skips.map(toNamespace).map((namespace) => "-" + namespace)
+          ...createDebug.names,
+          ...createDebug.skips.map((namespace) => "-" + namespace)
         ].join(",");
         createDebug.enable("");
         return namespaces;
       }
       function enabled(name2) {
-        if (name2[name2.length - 1] === "*") {
-          return true;
-        }
-        let i;
-        let len;
-        for (i = 0, len = createDebug.skips.length; i < len; i++) {
-          if (createDebug.skips[i].test(name2)) {
+        for (const skip of createDebug.skips) {
+          if (matchesTemplate(name2, skip)) {
             return false;
           }
         }
-        for (i = 0, len = createDebug.names.length; i < len; i++) {
-          if (createDebug.names[i].test(name2)) {
+        for (const ns of createDebug.names) {
+          if (matchesTemplate(name2, ns)) {
             return true;
           }
         }
         return false;
       }
-      function toNamespace(regexp) {
-        return regexp.toString().substring(2, regexp.toString().length - 2).replace(/\.\*\?$/, "*");
-      }
       function coerce(val) {
         if (val instanceof Error) {
           return val.stack || val.message;
@@ -2468,9 +2503,9 @@ var require_common = __commonJS({
   }
 });
 
-// .yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/browser.js
+// .yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/browser.js
 var require_browser = __commonJS({
-  ".yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/browser.js"(exports2, module2) {
+  ".yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/browser.js"(exports2, module2) {
     exports2.formatArgs = formatArgs;
     exports2.save = save;
     exports2.load = load;
@@ -2570,10 +2605,11 @@ var require_browser = __commonJS({
       if (typeof navigator !== "undefined" && navigator.userAgent && navigator.userAgent.toLowerCase().match(/(edge|trident)\/(\d+)/)) {
         return false;
       }
+      let m;
       return typeof document !== "undefined" && document.documentElement && document.documentElement.style && document.documentElement.style.WebkitAppearance || // Is firebug? http://stackoverflow.com/a/398120/376773
       typeof window !== "undefined" && window.console && (window.console.firebug || window.console.exception && window.console.table) || // Is firefox >= v31?
       // https://developer.mozilla.org/en-US/docs/Tools/Web_Console#Styling_messages
-      typeof navigator !== "undefined" && navigator.userAgent && navigator.userAgent.toLowerCase().match(/firefox\/(\d+)/) && parseInt(RegExp.$1, 10) >= 31 || // Double check webkit in userAgent just in case we are in a worker
+      typeof navigator !== "undefined" && navigator.userAgent && (m = navigator.userAgent.toLowerCase().match(/firefox\/(\d+)/)) && parseInt(m[1], 10) >= 31 || // Double check webkit in userAgent just in case we are in a worker
       typeof navigator !== "undefined" && navigator.userAgent && navigator.userAgent.toLowerCase().match(/applewebkit\/(\d+)/);
     }
     function formatArgs(args) {
@@ -2637,7 +2673,7 @@ var require_browser = __commonJS({
   }
 });
 
-// .yarn/cache/supports-color-npm-9.4.0-a415f39758-6c24e6b2b6.zip/node_modules/supports-color/index.js
+// .yarn/cache/supports-color-npm-10.0.0-6cd1bb42a6-0e7884dfd0.zip/node_modules/supports-color/index.js
 var supports_color_exports = {};
 __export(supports_color_exports, {
   createSupportsColor: () => createSupportsColor,
@@ -2650,15 +2686,23 @@ function hasFlag(flag, argv = globalThis.Deno ? globalThis.Deno.args : import_no
   return position !== -1 && (terminatorPosition === -1 || position < terminatorPosition);
 }
 function envForceColor() {
-  if ("FORCE_COLOR" in env) {
-    if (env.FORCE_COLOR === "true") {
-      return 1;
-    }
-    if (env.FORCE_COLOR === "false") {
-      return 0;
-    }
-    return env.FORCE_COLOR.length === 0 ? 1 : Math.min(Number.parseInt(env.FORCE_COLOR, 10), 3);
+  if (!("FORCE_COLOR" in env)) {
+    return;
+  }
+  if (env.FORCE_COLOR === "true") {
+    return 1;
+  }
+  if (env.FORCE_COLOR === "false") {
+    return 0;
   }
+  if (env.FORCE_COLOR.length === 0) {
+    return 1;
+  }
+  const level = Math.min(Number.parseInt(env.FORCE_COLOR, 10), 3);
+  if (![0, 1, 2, 3].includes(level)) {
+    return;
+  }
+  return level;
 }
 function translateLevel(level) {
   if (level === 0) {
@@ -2706,10 +2750,10 @@ function _supportsColor(haveStream, { streamIsTTY, sniffFlags = true } = {}) {
     return 1;
   }
   if ("CI" in env) {
-    if ("GITHUB_ACTIONS" in env || "GITEA_ACTIONS" in env) {
+    if (["GITHUB_ACTIONS", "GITEA_ACTIONS", "CIRCLECI"].some((key) => key in env)) {
       return 3;
     }
-    if (["TRAVIS", "CIRCLECI", "APPVEYOR", "GITLAB_CI", "BUILDKITE", "DRONE"].some((sign) => sign in env) || env.CI_NAME === "codeship") {
+    if (["TRAVIS", "APPVEYOR", "GITLAB_CI", "BUILDKITE", "DRONE"].some((sign) => sign in env) || env.CI_NAME === "codeship") {
       return 1;
     }
     return min;
@@ -2754,7 +2798,7 @@ function createSupportsColor(stream, options = {}) {
 }
 var import_node_process, import_node_os, import_node_tty, env, flagForceColor, supportsColor, supports_color_default;
 var init_supports_color = __esm({
-  ".yarn/cache/supports-color-npm-9.4.0-a415f39758-6c24e6b2b6.zip/node_modules/supports-color/index.js"() {
+  ".yarn/cache/supports-color-npm-10.0.0-6cd1bb42a6-0e7884dfd0.zip/node_modules/supports-color/index.js"() {
     import_node_process = __toESM(require("node:process"), 1);
     import_node_os = __toESM(require("node:os"), 1);
     import_node_tty = __toESM(require("node:tty"), 1);
@@ -2772,9 +2816,9 @@ var init_supports_color = __esm({
   }
 });
 
-// .yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/node.js
+// .yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/node.js
 var require_node2 = __commonJS({
-  ".yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/node.js"(exports2, module2) {
+  ".yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/node.js"(exports2, module2) {
     var tty2 = require("tty");
     var util = require("util");
     exports2.init = init;
@@ -2946,9 +2990,9 @@ var require_node2 = __commonJS({
   }
 });
 
-// .yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/index.js
+// .yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/index.js
 var require_src = __commonJS({
-  ".yarn/__virtual__/debug-virtual-710203f68e/0/cache/debug-npm-4.3.5-b5001f59b7-082c375a2b.zip/node_modules/debug/src/index.js"(exports2, module2) {
+  ".yarn/__virtual__/debug-virtual-f48feae4da/0/cache/debug-npm-4.4.0-f6efe76023-db94f1a182.zip/node_modules/debug/src/index.js"(exports2, module2) {
     if (typeof process === "undefined" || process.type === "renderer" || process.browser === true || process.__nwjs) {
       module2.exports = require_browser();
     } else {
@@ -3027,9 +3071,9 @@ var require_proxy_from_env = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/errors.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/errors.js
 var require_errors = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/errors.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/errors.js"(exports2, module2) {
     "use strict";
     var UndiciError = class extends Error {
       constructor(message) {
@@ -3205,6 +3249,17 @@ var require_errors = __commonJS({
         this.headers = headers;
       }
     };
+    var ResponseError = class extends UndiciError {
+      constructor(message, code2, { headers, data }) {
+        super(message);
+        this.name = "ResponseError";
+        this.message = message || "Response error";
+        this.code = "UND_ERR_RESPONSE";
+        this.statusCode = code2;
+        this.data = data;
+        this.headers = headers;
+      }
+    };
     var SecureProxyConnectionError = class extends UndiciError {
       constructor(cause, message, options) {
         super(message, { cause, ...options ?? {} });
@@ -3236,14 +3291,15 @@ var require_errors = __commonJS({
       BalancedPoolMissingUpstreamError,
       ResponseExceededMaxSizeError,
       RequestRetryError,
+      ResponseError,
       SecureProxyConnectionError
     };
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/symbols.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/symbols.js
 var require_symbols = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/symbols.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/symbols.js"(exports2, module2) {
     module2.exports = {
       kClose: Symbol("close"),
       kDestroy: Symbol("destroy"),
@@ -3314,9 +3370,9 @@ var require_symbols = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/constants.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/constants.js
 var require_constants2 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/constants.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/constants.js"(exports2, module2) {
     "use strict";
     var headerNameLowerCasedRecord = {};
     var wellknownHeaderNames = [
@@ -3429,9 +3485,9 @@ var require_constants2 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/tree.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/tree.js
 var require_tree = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/tree.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/tree.js"(exports2, module2) {
     "use strict";
     var {
       wellknownHeaderNames,
@@ -3569,9 +3625,9 @@ var require_tree = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/util.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/util.js
 var require_util = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/util.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/util.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { kDestroyed, kBodyUsed, kListeners, kBody } = require_symbols();
@@ -3720,7 +3776,7 @@ var require_util = __commonJS({
       if (!host) {
         return null;
       }
-      assert5.strictEqual(typeof host, "string");
+      assert5(typeof host === "string");
       const servername = getHostname(host);
       if (net.isIP(servername)) {
         return "";
@@ -4075,9 +4131,9 @@ var require_util = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/readable.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/readable.js
 var require_readable = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/readable.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/readable.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { Readable: Readable2 } = require("node:stream");
@@ -4170,6 +4226,10 @@ var require_readable = __commonJS({
       async blob() {
         return consume(this, "blob");
       }
+      // https://fetch.spec.whatwg.org/#dom-body-bytes
+      async bytes() {
+        return consume(this, "bytes");
+      }
       // https://fetch.spec.whatwg.org/#dom-body-arraybuffer
       async arrayBuffer() {
         return consume(this, "arrayBuffer");
@@ -4305,6 +4365,22 @@ var require_readable = __commonJS({
       const start = bufferLength > 2 && buffer[0] === 239 && buffer[1] === 187 && buffer[2] === 191 ? 3 : 0;
       return buffer.utf8Slice(start, bufferLength);
     }
+    function chunksConcat(chunks, length) {
+      if (chunks.length === 0 || length === 0) {
+        return new Uint8Array(0);
+      }
+      if (chunks.length === 1) {
+        return new Uint8Array(chunks[0]);
+      }
+      const buffer = new Uint8Array(Buffer.allocUnsafeSlow(length).buffer);
+      let offset = 0;
+      for (let i = 0; i < chunks.length; ++i) {
+        const chunk = chunks[i];
+        buffer.set(chunk, offset);
+        offset += chunk.length;
+      }
+      return buffer;
+    }
     function consumeEnd(consume2) {
       const { type, body, resolve: resolve2, stream, length } = consume2;
       try {
@@ -4313,15 +4389,11 @@ var require_readable = __commonJS({
         } else if (type === "json") {
           resolve2(JSON.parse(chunksDecode(body, length)));
         } else if (type === "arrayBuffer") {
-          const dst = new Uint8Array(length);
-          let pos2 = 0;
-          for (const buf of body) {
-            dst.set(buf, pos2);
-            pos2 += buf.byteLength;
-          }
-          resolve2(dst.buffer);
+          resolve2(chunksConcat(body, length).buffer);
         } else if (type === "blob") {
           resolve2(new Blob(body, { type: stream[kContentType] }));
+        } else if (type === "bytes") {
+          resolve2(chunksConcat(body, length));
         }
         consumeFinish(consume2);
       } catch (err) {
@@ -4352,9 +4424,9 @@ var require_readable = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/util.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/util.js
 var require_util2 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/util.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/util.js"(exports2, module2) {
     var assert5 = require("node:assert");
     var {
       ResponseStatusCodeError
@@ -4413,9 +4485,9 @@ var require_util2 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-request.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-request.js
 var require_api_request = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-request.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-request.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { Readable: Readable2 } = require_readable();
@@ -4479,7 +4551,7 @@ var require_api_request = __commonJS({
             this.removeAbortListener = util.addAbortListener(this.signal, () => {
               this.reason = this.signal.reason ?? new RequestAbortedError();
               if (this.res) {
-                util.destroy(this.res, this.reason);
+                util.destroy(this.res.on("error", util.nop), this.reason);
               } else if (this.abort) {
                 this.abort(this.reason);
               }
@@ -4599,9 +4671,9 @@ var require_api_request = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/abort-signal.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/abort-signal.js
 var require_abort_signal = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/abort-signal.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/abort-signal.js"(exports2, module2) {
     var { addAbortListener } = require_util();
     var { RequestAbortedError } = require_errors();
     var kListener = Symbol("kListener");
@@ -4650,9 +4722,9 @@ var require_abort_signal = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-stream.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-stream.js
 var require_api_stream = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-stream.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-stream.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { finished, PassThrough } = require("node:stream");
@@ -4823,9 +4895,9 @@ var require_api_stream = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-pipeline.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-pipeline.js
 var require_api_pipeline = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-pipeline.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-pipeline.js"(exports2, module2) {
     "use strict";
     var {
       Readable: Readable2,
@@ -5023,9 +5095,9 @@ var require_api_pipeline = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-upgrade.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-upgrade.js
 var require_api_upgrade = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-upgrade.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-upgrade.js"(exports2, module2) {
     "use strict";
     var { InvalidArgumentError, SocketError } = require_errors();
     var { AsyncResource } = require("node:async_hooks");
@@ -5065,8 +5137,8 @@ var require_api_upgrade = __commonJS({
         throw new SocketError("bad upgrade", null);
       }
       onUpgrade(statusCode, rawHeaders, socket) {
+        assert5(statusCode === 101);
         const { callback, opaque, context } = this;
-        assert5.strictEqual(statusCode, 101);
         removeSignal(this);
         this.callback = null;
         const headers = this.responseHeaders === "raw" ? util.parseRawHeaders(rawHeaders) : util.parseHeaders(rawHeaders);
@@ -5115,9 +5187,9 @@ var require_api_upgrade = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-connect.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-connect.js
 var require_api_connect = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/api-connect.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/api-connect.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { AsyncResource } = require("node:async_hooks");
@@ -5205,9 +5277,9 @@ var require_api_connect = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/index.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/index.js
 var require_api = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/api/index.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/api/index.js"(exports2, module2) {
     "use strict";
     module2.exports.request = require_api_request();
     module2.exports.stream = require_api_stream();
@@ -5217,9 +5289,9 @@ var require_api = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/dispatcher.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/dispatcher.js
 var require_dispatcher = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/dispatcher.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/dispatcher.js"(exports2, module2) {
     "use strict";
     var EventEmitter2 = require("node:events");
     var Dispatcher = class extends EventEmitter2 {
@@ -5272,9 +5344,9 @@ var require_dispatcher = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/dispatcher-base.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/dispatcher-base.js
 var require_dispatcher_base = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/dispatcher-base.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/dispatcher-base.js"(exports2, module2) {
     "use strict";
     var Dispatcher = require_dispatcher();
     var {
@@ -5433,9 +5505,9 @@ var require_dispatcher_base = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/fixed-queue.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/fixed-queue.js
 var require_fixed_queue = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/fixed-queue.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/fixed-queue.js"(exports2, module2) {
     "use strict";
     var kSize = 2048;
     var kMask = kSize - 1;
@@ -5490,9 +5562,9 @@ var require_fixed_queue = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool-stats.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool-stats.js
 var require_pool_stats = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool-stats.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool-stats.js"(exports2, module2) {
     var { kFree, kConnected, kPending, kQueued, kRunning, kSize } = require_symbols();
     var kPool = Symbol("pool");
     var PoolStats = class {
@@ -5522,9 +5594,9 @@ var require_pool_stats = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool-base.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool-base.js
 var require_pool_base = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool-base.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool-base.js"(exports2, module2) {
     "use strict";
     var DispatcherBase = require_dispatcher_base();
     var FixedQueue = require_fixed_queue();
@@ -5615,9 +5687,9 @@ var require_pool_base = __commonJS({
       }
       async [kClose]() {
         if (this[kQueue].isEmpty()) {
-          return Promise.all(this[kClients].map((c) => c.close()));
+          await Promise.all(this[kClients].map((c) => c.close()));
         } else {
-          return new Promise((resolve2) => {
+          await new Promise((resolve2) => {
             this[kClosedResolve] = resolve2;
           });
         }
@@ -5630,7 +5702,7 @@ var require_pool_base = __commonJS({
           }
           item.handler.onError(err);
         }
-        return Promise.all(this[kClients].map((c) => c.destroy(err)));
+        await Promise.all(this[kClients].map((c) => c.destroy(err)));
       }
       [kDispatch](opts, handler) {
         const dispatcher = this[kGetDispatcher]();
@@ -5677,9 +5749,9 @@ var require_pool_base = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/diagnostics.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/diagnostics.js
 var require_diagnostics = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/diagnostics.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/diagnostics.js"(exports2, module2) {
     "use strict";
     var diagnosticsChannel = require("node:diagnostics_channel");
     var util = require("node:util");
@@ -5862,9 +5934,9 @@ var require_diagnostics = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/request.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/request.js
 var require_request = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/request.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/request.js"(exports2, module2) {
     "use strict";
     var {
       InvalidArgumentError,
@@ -6186,14 +6258,248 @@ var require_request = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/connect.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/util/timers.js
+var require_timers = __commonJS({
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/util/timers.js"(exports2, module2) {
+    "use strict";
+    var fastNow = 0;
+    var RESOLUTION_MS = 1e3;
+    var TICK_MS = (RESOLUTION_MS >> 1) - 1;
+    var fastNowTimeout;
+    var kFastTimer = Symbol("kFastTimer");
+    var fastTimers = [];
+    var NOT_IN_LIST = -2;
+    var TO_BE_CLEARED = -1;
+    var PENDING2 = 0;
+    var ACTIVE = 1;
+    function onTick() {
+      fastNow += TICK_MS;
+      let idx = 0;
+      let len = fastTimers.length;
+      while (idx < len) {
+        const timer = fastTimers[idx];
+        if (timer._state === PENDING2) {
+          timer._idleStart = fastNow - TICK_MS;
+          timer._state = ACTIVE;
+        } else if (timer._state === ACTIVE && fastNow >= timer._idleStart + timer._idleTimeout) {
+          timer._state = TO_BE_CLEARED;
+          timer._idleStart = -1;
+          timer._onTimeout(timer._timerArg);
+        }
+        if (timer._state === TO_BE_CLEARED) {
+          timer._state = NOT_IN_LIST;
+          if (--len !== 0) {
+            fastTimers[idx] = fastTimers[len];
+          }
+        } else {
+          ++idx;
+        }
+      }
+      fastTimers.length = len;
+      if (fastTimers.length !== 0) {
+        refreshTimeout();
+      }
+    }
+    function refreshTimeout() {
+      if (fastNowTimeout) {
+        fastNowTimeout.refresh();
+      } else {
+        clearTimeout(fastNowTimeout);
+        fastNowTimeout = setTimeout(onTick, TICK_MS);
+        if (fastNowTimeout.unref) {
+          fastNowTimeout.unref();
+        }
+      }
+    }
+    var FastTimer = class {
+      [kFastTimer] = true;
+      /**
+       * The state of the timer, which can be one of the following:
+       * - NOT_IN_LIST (-2)
+       * - TO_BE_CLEARED (-1)
+       * - PENDING (0)
+       * - ACTIVE (1)
+       *
+       * @type {-2|-1|0|1}
+       * @private
+       */
+      _state = NOT_IN_LIST;
+      /**
+       * The number of milliseconds to wait before calling the callback.
+       *
+       * @type {number}
+       * @private
+       */
+      _idleTimeout = -1;
+      /**
+       * The time in milliseconds when the timer was started. This value is used to
+       * calculate when the timer should expire.
+       *
+       * @type {number}
+       * @default -1
+       * @private
+       */
+      _idleStart = -1;
+      /**
+       * The function to be executed when the timer expires.
+       * @type {Function}
+       * @private
+       */
+      _onTimeout;
+      /**
+       * The argument to be passed to the callback when the timer expires.
+       *
+       * @type {*}
+       * @private
+       */
+      _timerArg;
+      /**
+       * @constructor
+       * @param {Function} callback A function to be executed after the timer
+       * expires.
+       * @param {number} delay The time, in milliseconds that the timer should wait
+       * before the specified function or code is executed.
+       * @param {*} arg
+       */
+      constructor(callback, delay, arg) {
+        this._onTimeout = callback;
+        this._idleTimeout = delay;
+        this._timerArg = arg;
+        this.refresh();
+      }
+      /**
+       * Sets the timer's start time to the current time, and reschedules the timer
+       * to call its callback at the previously specified duration adjusted to the
+       * current time.
+       * Using this on a timer that has already called its callback will reactivate
+       * the timer.
+       *
+       * @returns {void}
+       */
+      refresh() {
+        if (this._state === NOT_IN_LIST) {
+          fastTimers.push(this);
+        }
+        if (!fastNowTimeout || fastTimers.length === 1) {
+          refreshTimeout();
+        }
+        this._state = PENDING2;
+      }
+      /**
+       * The `clear` method cancels the timer, preventing it from executing.
+       *
+       * @returns {void}
+       * @private
+       */
+      clear() {
+        this._state = TO_BE_CLEARED;
+        this._idleStart = -1;
+      }
+    };
+    module2.exports = {
+      /**
+       * The setTimeout() method sets a timer which executes a function once the
+       * timer expires.
+       * @param {Function} callback A function to be executed after the timer
+       * expires.
+       * @param {number} delay The time, in milliseconds that the timer should
+       * wait before the specified function or code is executed.
+       * @param {*} [arg] An optional argument to be passed to the callback function
+       * when the timer expires.
+       * @returns {NodeJS.Timeout|FastTimer}
+       */
+      setTimeout(callback, delay, arg) {
+        return delay <= RESOLUTION_MS ? setTimeout(callback, delay, arg) : new FastTimer(callback, delay, arg);
+      },
+      /**
+       * The clearTimeout method cancels an instantiated Timer previously created
+       * by calling setTimeout.
+       *
+       * @param {NodeJS.Timeout|FastTimer} timeout
+       */
+      clearTimeout(timeout) {
+        if (timeout[kFastTimer]) {
+          timeout.clear();
+        } else {
+          clearTimeout(timeout);
+        }
+      },
+      /**
+       * The setFastTimeout() method sets a fastTimer which executes a function once
+       * the timer expires.
+       * @param {Function} callback A function to be executed after the timer
+       * expires.
+       * @param {number} delay The time, in milliseconds that the timer should
+       * wait before the specified function or code is executed.
+       * @param {*} [arg] An optional argument to be passed to the callback function
+       * when the timer expires.
+       * @returns {FastTimer}
+       */
+      setFastTimeout(callback, delay, arg) {
+        return new FastTimer(callback, delay, arg);
+      },
+      /**
+       * The clearTimeout method cancels an instantiated FastTimer previously
+       * created by calling setFastTimeout.
+       *
+       * @param {FastTimer} timeout
+       */
+      clearFastTimeout(timeout) {
+        timeout.clear();
+      },
+      /**
+       * The now method returns the value of the internal fast timer clock.
+       *
+       * @returns {number}
+       */
+      now() {
+        return fastNow;
+      },
+      /**
+       * Trigger the onTick function to process the fastTimers array.
+       * Exported for testing purposes only.
+       * Marking as deprecated to discourage any use outside of testing.
+       * @deprecated
+       * @param {number} [delay=0] The delay in milliseconds to add to the now value.
+       */
+      tick(delay = 0) {
+        fastNow += delay - RESOLUTION_MS + 1;
+        onTick();
+        onTick();
+      },
+      /**
+       * Reset FastTimers.
+       * Exported for testing purposes only.
+       * Marking as deprecated to discourage any use outside of testing.
+       * @deprecated
+       */
+      reset() {
+        fastNow = 0;
+        fastTimers.length = 0;
+        clearTimeout(fastNowTimeout);
+        fastNowTimeout = null;
+      },
+      /**
+       * Exporting for testing purposes only.
+       * Marking as deprecated to discourage any use outside of testing.
+       * @deprecated
+       */
+      kFastTimer
+    };
+  }
+});
+
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/connect.js
 var require_connect = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/core/connect.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/core/connect.js"(exports2, module2) {
     "use strict";
     var net = require("node:net");
     var assert5 = require("node:assert");
     var util = require_util();
     var { InvalidArgumentError, ConnectTimeoutError } = require_errors();
+    var timers = require_timers();
+    function noop2() {
+    }
     var tls;
     var SessionCache;
     if (global.FinalizationRegistry && !(process.env.NODE_V8_COVERAGE || process.env.UNDICI_NO_FG)) {
@@ -6260,8 +6566,9 @@ var require_connect = __commonJS({
           }
           servername = servername || options.servername || util.getServerName(host) || null;
           const sessionKey = servername || hostname;
-          const session = customSession || sessionCache.get(sessionKey) || null;
           assert5(sessionKey);
+          const session = customSession || sessionCache.get(sessionKey) || null;
+          port = port || 443;
           socket = tls.connect({
             highWaterMark: 16384,
             // TLS in node can't have bigger HWM anyway...
@@ -6273,7 +6580,7 @@ var require_connect = __commonJS({
             ALPNProtocols: allowH2 ? ["http/1.1", "h2"] : ["http/1.1"],
             socket: httpSocket,
             // upgrade socket connection
-            port: port || 443,
+            port,
             host: hostname
           });
           socket.on("session", function(session2) {
@@ -6281,12 +6588,13 @@ var require_connect = __commonJS({
           });
         } else {
           assert5(!httpSocket, "httpSocket can only be sent on TLS update");
+          port = port || 80;
           socket = net.connect({
             highWaterMark: 64 * 1024,
             // Same as nodejs fs streams.
             ...options,
             localAddress,
-            port: port || 80,
+            port,
             host: hostname
           });
         }
@@ -6294,16 +6602,16 @@ var require_connect = __commonJS({
           const keepAliveInitialDelay = options.keepAliveInitialDelay === void 0 ? 6e4 : options.keepAliveInitialDelay;
           socket.setKeepAlive(true, keepAliveInitialDelay);
         }
-        const cancelTimeout = setupTimeout(() => onConnectTimeout(socket), timeout);
+        const clearConnectTimeout = setupConnectTimeout(new WeakRef(socket), { timeout, hostname, port });
         socket.setNoDelay(true).once(protocol === "https:" ? "secureConnect" : "connect", function() {
-          cancelTimeout();
+          queueMicrotask(clearConnectTimeout);
           if (callback) {
             const cb = callback;
             callback = null;
             cb(null, this);
           }
         }).on("error", function(err) {
-          cancelTimeout();
+          queueMicrotask(clearConnectTimeout);
           if (callback) {
             const cb = callback;
             callback = null;
@@ -6313,125 +6621,57 @@ var require_connect = __commonJS({
         return socket;
       };
     }
-    function setupTimeout(onConnectTimeout2, timeout) {
-      if (!timeout) {
-        return () => {
-        };
+    var setupConnectTimeout = process.platform === "win32" ? (socketWeakRef, opts) => {
+      if (!opts.timeout) {
+        return noop2;
       }
       let s1 = null;
       let s2 = null;
-      const timeoutId = setTimeout(() => {
+      const fastTimer = timers.setFastTimeout(() => {
         s1 = setImmediate(() => {
-          if (process.platform === "win32") {
-            s2 = setImmediate(() => onConnectTimeout2());
-          } else {
-            onConnectTimeout2();
-          }
+          s2 = setImmediate(() => onConnectTimeout(socketWeakRef.deref(), opts));
         });
-      }, timeout);
+      }, opts.timeout);
       return () => {
-        clearTimeout(timeoutId);
+        timers.clearFastTimeout(fastTimer);
         clearImmediate(s1);
         clearImmediate(s2);
       };
-    }
-    function onConnectTimeout(socket) {
+    } : (socketWeakRef, opts) => {
+      if (!opts.timeout) {
+        return noop2;
+      }
+      let s1 = null;
+      const fastTimer = timers.setFastTimeout(() => {
+        s1 = setImmediate(() => {
+          onConnectTimeout(socketWeakRef.deref(), opts);
+        });
+      }, opts.timeout);
+      return () => {
+        timers.clearFastTimeout(fastTimer);
+        clearImmediate(s1);
+      };
+    };
+    function onConnectTimeout(socket, opts) {
+      if (socket == null) {
+        return;
+      }
       let message = "Connect Timeout Error";
       if (Array.isArray(socket.autoSelectFamilyAttemptedAddresses)) {
-        message += ` (attempted addresses: ${socket.autoSelectFamilyAttemptedAddresses.join(", ")})`;
+        message += ` (attempted addresses: ${socket.autoSelectFamilyAttemptedAddresses.join(", ")},`;
+      } else {
+        message += ` (attempted address: ${opts.hostname}:${opts.port},`;
       }
+      message += ` timeout: ${opts.timeout}ms)`;
       util.destroy(socket, new ConnectTimeoutError(message));
     }
     module2.exports = buildConnector;
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/util/timers.js
-var require_timers = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/util/timers.js"(exports2, module2) {
-    "use strict";
-    var TICK_MS = 499;
-    var fastNow = Date.now();
-    var fastNowTimeout;
-    var fastTimers = [];
-    function onTimeout() {
-      fastNow = Date.now();
-      let len = fastTimers.length;
-      let idx = 0;
-      while (idx < len) {
-        const timer = fastTimers[idx];
-        if (timer.state === 0) {
-          timer.state = fastNow + timer.delay - TICK_MS;
-        } else if (timer.state > 0 && fastNow >= timer.state) {
-          timer.state = -1;
-          timer.callback(timer.opaque);
-        }
-        if (timer.state === -1) {
-          timer.state = -2;
-          if (idx !== len - 1) {
-            fastTimers[idx] = fastTimers.pop();
-          } else {
-            fastTimers.pop();
-          }
-          len -= 1;
-        } else {
-          idx += 1;
-        }
-      }
-      if (fastTimers.length > 0) {
-        refreshTimeout();
-      }
-    }
-    function refreshTimeout() {
-      if (fastNowTimeout?.refresh) {
-        fastNowTimeout.refresh();
-      } else {
-        clearTimeout(fastNowTimeout);
-        fastNowTimeout = setTimeout(onTimeout, TICK_MS);
-        if (fastNowTimeout.unref) {
-          fastNowTimeout.unref();
-        }
-      }
-    }
-    var Timeout = class {
-      constructor(callback, delay, opaque) {
-        this.callback = callback;
-        this.delay = delay;
-        this.opaque = opaque;
-        this.state = -2;
-        this.refresh();
-      }
-      refresh() {
-        if (this.state === -2) {
-          fastTimers.push(this);
-          if (!fastNowTimeout || fastTimers.length === 1) {
-            refreshTimeout();
-          }
-        }
-        this.state = 0;
-      }
-      clear() {
-        this.state = -1;
-      }
-    };
-    module2.exports = {
-      setTimeout(callback, delay, opaque) {
-        return delay <= 1e3 ? setTimeout(callback, delay, opaque) : new Timeout(callback, delay, opaque);
-      },
-      clearTimeout(timeout) {
-        if (timeout instanceof Timeout) {
-          timeout.clear();
-        } else {
-          clearTimeout(timeout);
-        }
-      }
-    };
-  }
-});
-
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/utils.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/utils.js
 var require_utils = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/utils.js"(exports2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/utils.js"(exports2) {
     "use strict";
     Object.defineProperty(exports2, "__esModule", { value: true });
     exports2.enumToMap = void 0;
@@ -6449,9 +6689,9 @@ var require_utils = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/constants.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/constants.js
 var require_constants3 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/constants.js"(exports2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/constants.js"(exports2) {
     "use strict";
     Object.defineProperty(exports2, "__esModule", { value: true });
     exports2.SPECIAL_HEADERS = exports2.HEADER_STATE = exports2.MINOR = exports2.MAJOR = exports2.CONNECTION_TOKEN_CHARS = exports2.HEADER_CHARS = exports2.TOKEN = exports2.STRICT_TOKEN = exports2.HEX = exports2.URL_CHAR = exports2.STRICT_URL_CHAR = exports2.USERINFO_CHARS = exports2.MARK = exports2.ALPHANUM = exports2.NUM = exports2.HEX_MAP = exports2.NUM_MAP = exports2.ALPHA = exports2.FINISH = exports2.H_METHOD_MAP = exports2.METHOD_MAP = exports2.METHODS_RTSP = exports2.METHODS_ICE = exports2.METHODS_HTTP = exports2.METHODS = exports2.LENIENT_FLAGS = exports2.FLAGS = exports2.TYPE = exports2.ERROR = void 0;
@@ -6770,173 +7010,215 @@ var require_constants3 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/llhttp-wasm.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/llhttp-wasm.js
 var require_llhttp_wasm = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/llhttp-wasm.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/llhttp-wasm.js"(exports2, module2) {
     "use strict";
     var { Buffer: Buffer3 } = require("node:buffer");
     module2.exports = Buffer3.from("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", "base64");
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/llhttp_simd-wasm.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/llhttp_simd-wasm.js
 var require_llhttp_simd_wasm = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/llhttp/llhttp_simd-wasm.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/llhttp/llhttp_simd-wasm.js"(exports2, module2) {
     "use strict";
     var { Buffer: Buffer3 } = require("node:buffer");
     module2.exports = Buffer3.from("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", "base64");
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/constants.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/constants.js
 var require_constants4 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/constants.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/constants.js"(exports2, module2) {
     "use strict";
-    var corsSafeListedMethods = ["GET", "HEAD", "POST"];
+    var corsSafeListedMethods = (
+      /** @type {const} */
+      ["GET", "HEAD", "POST"]
+    );
     var corsSafeListedMethodsSet = new Set(corsSafeListedMethods);
-    var nullBodyStatus = [101, 204, 205, 304];
-    var redirectStatus = [301, 302, 303, 307, 308];
+    var nullBodyStatus = (
+      /** @type {const} */
+      [101, 204, 205, 304]
+    );
+    var redirectStatus = (
+      /** @type {const} */
+      [301, 302, 303, 307, 308]
+    );
     var redirectStatusSet = new Set(redirectStatus);
-    var badPorts = [
-      "1",
-      "7",
-      "9",
-      "11",
-      "13",
-      "15",
-      "17",
-      "19",
-      "20",
-      "21",
-      "22",
-      "23",
-      "25",
-      "37",
-      "42",
-      "43",
-      "53",
-      "69",
-      "77",
-      "79",
-      "87",
-      "95",
-      "101",
-      "102",
-      "103",
-      "104",
-      "109",
-      "110",
-      "111",
-      "113",
-      "115",
-      "117",
-      "119",
-      "123",
-      "135",
-      "137",
-      "139",
-      "143",
-      "161",
-      "179",
-      "389",
-      "427",
-      "465",
-      "512",
-      "513",
-      "514",
-      "515",
-      "526",
-      "530",
-      "531",
-      "532",
-      "540",
-      "548",
-      "554",
-      "556",
-      "563",
-      "587",
-      "601",
-      "636",
-      "989",
-      "990",
-      "993",
-      "995",
-      "1719",
-      "1720",
-      "1723",
-      "2049",
-      "3659",
-      "4045",
-      "4190",
-      "5060",
-      "5061",
-      "6000",
-      "6566",
-      "6665",
-      "6666",
-      "6667",
-      "6668",
-      "6669",
-      "6679",
-      "6697",
-      "10080"
-    ];
+    var badPorts = (
+      /** @type {const} */
+      [
+        "1",
+        "7",
+        "9",
+        "11",
+        "13",
+        "15",
+        "17",
+        "19",
+        "20",
+        "21",
+        "22",
+        "23",
+        "25",
+        "37",
+        "42",
+        "43",
+        "53",
+        "69",
+        "77",
+        "79",
+        "87",
+        "95",
+        "101",
+        "102",
+        "103",
+        "104",
+        "109",
+        "110",
+        "111",
+        "113",
+        "115",
+        "117",
+        "119",
+        "123",
+        "135",
+        "137",
+        "139",
+        "143",
+        "161",
+        "179",
+        "389",
+        "427",
+        "465",
+        "512",
+        "513",
+        "514",
+        "515",
+        "526",
+        "530",
+        "531",
+        "532",
+        "540",
+        "548",
+        "554",
+        "556",
+        "563",
+        "587",
+        "601",
+        "636",
+        "989",
+        "990",
+        "993",
+        "995",
+        "1719",
+        "1720",
+        "1723",
+        "2049",
+        "3659",
+        "4045",
+        "4190",
+        "5060",
+        "5061",
+        "6000",
+        "6566",
+        "6665",
+        "6666",
+        "6667",
+        "6668",
+        "6669",
+        "6679",
+        "6697",
+        "10080"
+      ]
+    );
     var badPortsSet = new Set(badPorts);
-    var referrerPolicy = [
-      "",
-      "no-referrer",
-      "no-referrer-when-downgrade",
-      "same-origin",
-      "origin",
-      "strict-origin",
-      "origin-when-cross-origin",
-      "strict-origin-when-cross-origin",
-      "unsafe-url"
-    ];
+    var referrerPolicy = (
+      /** @type {const} */
+      [
+        "",
+        "no-referrer",
+        "no-referrer-when-downgrade",
+        "same-origin",
+        "origin",
+        "strict-origin",
+        "origin-when-cross-origin",
+        "strict-origin-when-cross-origin",
+        "unsafe-url"
+      ]
+    );
     var referrerPolicySet = new Set(referrerPolicy);
-    var requestRedirect = ["follow", "manual", "error"];
-    var safeMethods = ["GET", "HEAD", "OPTIONS", "TRACE"];
+    var requestRedirect = (
+      /** @type {const} */
+      ["follow", "manual", "error"]
+    );
+    var safeMethods = (
+      /** @type {const} */
+      ["GET", "HEAD", "OPTIONS", "TRACE"]
+    );
     var safeMethodsSet = new Set(safeMethods);
-    var requestMode = ["navigate", "same-origin", "no-cors", "cors"];
-    var requestCredentials = ["omit", "same-origin", "include"];
-    var requestCache = [
-      "default",
-      "no-store",
-      "reload",
-      "no-cache",
-      "force-cache",
-      "only-if-cached"
-    ];
-    var requestBodyHeader = [
-      "content-encoding",
-      "content-language",
-      "content-location",
-      "content-type",
-      // See https://github.com/nodejs/undici/issues/2021
-      // 'Content-Length' is a forbidden header name, which is typically
-      // removed in the Headers implementation. However, undici doesn't
-      // filter out headers, so we add it here.
-      "content-length"
-    ];
-    var requestDuplex = [
-      "half"
-    ];
-    var forbiddenMethods = ["CONNECT", "TRACE", "TRACK"];
+    var requestMode = (
+      /** @type {const} */
+      ["navigate", "same-origin", "no-cors", "cors"]
+    );
+    var requestCredentials = (
+      /** @type {const} */
+      ["omit", "same-origin", "include"]
+    );
+    var requestCache = (
+      /** @type {const} */
+      [
+        "default",
+        "no-store",
+        "reload",
+        "no-cache",
+        "force-cache",
+        "only-if-cached"
+      ]
+    );
+    var requestBodyHeader = (
+      /** @type {const} */
+      [
+        "content-encoding",
+        "content-language",
+        "content-location",
+        "content-type",
+        // See https://github.com/nodejs/undici/issues/2021
+        // 'Content-Length' is a forbidden header name, which is typically
+        // removed in the Headers implementation. However, undici doesn't
+        // filter out headers, so we add it here.
+        "content-length"
+      ]
+    );
+    var requestDuplex = (
+      /** @type {const} */
+      [
+        "half"
+      ]
+    );
+    var forbiddenMethods = (
+      /** @type {const} */
+      ["CONNECT", "TRACE", "TRACK"]
+    );
     var forbiddenMethodsSet = new Set(forbiddenMethods);
-    var subresource = [
-      "audio",
-      "audioworklet",
-      "font",
-      "image",
-      "manifest",
-      "paintworklet",
-      "script",
-      "style",
-      "track",
-      "video",
-      "xslt",
-      ""
-    ];
+    var subresource = (
+      /** @type {const} */
+      [
+        "audio",
+        "audioworklet",
+        "font",
+        "image",
+        "manifest",
+        "paintworklet",
+        "script",
+        "style",
+        "track",
+        "video",
+        "xslt",
+        ""
+      ]
+    );
     var subresourceSet = new Set(subresource);
     module2.exports = {
       subresource,
@@ -6964,9 +7246,9 @@ var require_constants4 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/global.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/global.js
 var require_global = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/global.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/global.js"(exports2, module2) {
     "use strict";
     var globalOrigin = Symbol.for("undici.globalOrigin.1");
     function getGlobalOrigin() {
@@ -7000,9 +7282,9 @@ var require_global = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/data-url.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/data-url.js
 var require_data_url = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/data-url.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/data-url.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var encoder = new TextEncoder();
@@ -7352,11 +7634,12 @@ var require_data_url = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/webidl.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/webidl.js
 var require_webidl = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/webidl.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/webidl.js"(exports2, module2) {
     "use strict";
     var { types, inspect } = require("node:util");
+    var { markAsUncloneable } = require("node:worker_threads");
     var { toUSVString } = require_util();
     var webidl = {};
     webidl.converters = {};
@@ -7431,6 +7714,8 @@ var require_webidl = __commonJS({
         }
       }
     };
+    webidl.util.markAsUncloneable = markAsUncloneable || (() => {
+    });
     webidl.util.ConvertToInt = function(V, bitLength, signedness, opts) {
       let upperBound;
       let lowerBound;
@@ -7768,9 +8053,9 @@ var require_webidl = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/util.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/util.js
 var require_util3 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/util.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/util.js"(exports2, module2) {
     "use strict";
     var { Transform } = require("node:stream");
     var zlib = require("node:zlib");
@@ -8003,7 +8288,21 @@ var require_util3 = __commonJS({
           return referrerOrigin;
         }
         case "strict-origin":
+        // eslint-disable-line
+        /**
+           * 1. If referrerURL is a potentially trustworthy URL and
+           * request’s current URL is not a potentially trustworthy URL,
+           * then return no referrer.
+           * 2. Return referrerOrigin
+          */
         case "no-referrer-when-downgrade":
+        // eslint-disable-line
+        /**
+         * 1. If referrerURL is a potentially trustworthy URL and
+         * request’s current URL is not a potentially trustworthy URL,
+         * then return no referrer.
+         * 2. Return referrerOrigin
+        */
         default:
           return isNonPotentiallyTrustWorthy ? "no-referrer" : referrerOrigin;
       }
@@ -8447,13 +8746,19 @@ var require_util3 = __commonJS({
       return contentRange;
     }
     var InflateStream = class extends Transform {
+      #zlibOptions;
+      /** @param {zlib.ZlibOptions} [zlibOptions] */
+      constructor(zlibOptions) {
+        super();
+        this.#zlibOptions = zlibOptions;
+      }
       _transform(chunk, encoding, callback) {
         if (!this._inflateStream) {
           if (chunk.length === 0) {
             callback();
             return;
           }
-          this._inflateStream = (chunk[0] & 15) === 8 ? zlib.createInflate() : zlib.createInflateRaw();
+          this._inflateStream = (chunk[0] & 15) === 8 ? zlib.createInflate(this.#zlibOptions) : zlib.createInflateRaw(this.#zlibOptions);
           this._inflateStream.on("data", this.push.bind(this));
           this._inflateStream.on("end", () => this.push(null));
           this._inflateStream.on("error", (err) => this.destroy(err));
@@ -8468,8 +8773,8 @@ var require_util3 = __commonJS({
         callback();
       }
     };
-    function createInflate() {
-      return new InflateStream();
+    function createInflate(zlibOptions) {
+      return new InflateStream(zlibOptions);
     }
     function extractMimeType(headers) {
       let charset = null;
@@ -8618,9 +8923,9 @@ var require_util3 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/symbols.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/symbols.js
 var require_symbols2 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/symbols.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/symbols.js"(exports2, module2) {
     "use strict";
     module2.exports = {
       kUrl: Symbol("url"),
@@ -8632,9 +8937,9 @@ var require_symbols2 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/file.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/file.js
 var require_file = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/file.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/file.js"(exports2, module2) {
     "use strict";
     var { Blob: Blob2, File } = require("node:buffer");
     var { kState } = require_symbols2();
@@ -8695,9 +9000,9 @@ var require_file = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/formdata.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/formdata.js
 var require_formdata = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/formdata.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/formdata.js"(exports2, module2) {
     "use strict";
     var { isBlobLike, iteratorMixin } = require_util3();
     var { kState } = require_symbols2();
@@ -8709,6 +9014,7 @@ var require_formdata = __commonJS({
     var File = globalThis.File ?? NativeFile;
     var FormData = class _FormData {
       constructor(form) {
+        webidl.util.markAsUncloneable(this);
         if (form !== void 0) {
           throw webidl.errors.conversionFailed({
             prefix: "FormData constructor",
@@ -8841,9 +9147,9 @@ var require_formdata = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/formdata-parser.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/formdata-parser.js
 var require_formdata_parser = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/formdata-parser.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/formdata-parser.js"(exports2, module2) {
     "use strict";
     var { isUSVString, bufferToLowerCasedHeaderName } = require_util();
     var { utf8DecodeBytes } = require_util3();
@@ -8887,9 +9193,16 @@ var require_formdata_parser = __commonJS({
       const boundary = Buffer.from(`--${boundaryString}`, "utf8");
       const entryList = [];
       const position = { position: 0 };
-      if (input[0] === 13 && input[1] === 10) {
+      while (input[position.position] === 13 && input[position.position + 1] === 10) {
         position.position += 2;
       }
+      let trailing = input.length;
+      while (input[trailing - 1] === 10 && input[trailing - 2] === 13) {
+        trailing -= 2;
+      }
+      if (trailing !== input.length) {
+        input = input.subarray(0, trailing);
+      }
       while (true) {
         if (input.subarray(position.position, position.position + boundary.length).equals(boundary)) {
           position.position += boundary.length;
@@ -9085,9 +9398,9 @@ var require_formdata_parser = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/body.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/body.js
 var require_body = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/web/fetch/body.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/web/fetch/body.js"(exports2, module2) {
     "use strict";
     var util = require_util();
     var {
@@ -9105,11 +9418,30 @@ var require_body = __commonJS({
     var { webidl } = require_webidl();
     var { Blob: Blob2 } = require("node:buffer");
     var assert5 = require("node:assert");
-    var { isErrored } = require_util();
+    var { isErrored, isDisturbed } = require("node:stream");
     var { isArrayBuffer } = require("node:util/types");
     var { serializeAMimeType } = require_data_url();
     var { multipartFormDataParser } = require_formdata_parser();
+    var random;
+    try {
+      const crypto = require("node:crypto");
+      random = (max) => crypto.randomInt(0, max);
+    } catch {
+      random = (max) => Math.floor(Math.random(max));
+    }
     var textEncoder = new TextEncoder();
+    function noop2() {
+    }
+    var hasFinalizationRegistry = globalThis.FinalizationRegistry && process.version.indexOf("v18") !== 0;
+    var streamRegistry;
+    if (hasFinalizationRegistry) {
+      streamRegistry = new FinalizationRegistry((weakRef) => {
+        const stream = weakRef.deref();
+        if (stream && !stream.locked && !isDisturbed(stream) && !isErrored(stream)) {
+          stream.cancel("Response object has been garbage collected").catch(noop2);
+        }
+      });
+    }
     function extractBody(object, keepalive = false) {
       let stream = null;
       if (object instanceof ReadableStream) {
@@ -9146,7 +9478,7 @@ var require_body = __commonJS({
       } else if (ArrayBuffer.isView(object)) {
         source = new Uint8Array(object.buffer.slice(object.byteOffset, object.byteOffset + object.byteLength));
       } else if (util.isFormDataLike(object)) {
-        const boundary = `----formdata-undici-0${`${Math.floor(Math.random() * 1e11)}`.padStart(11, "0")}`;
+        const boundary = `----formdata-undici-0${`${random(1e11)}`.padStart(11, "0")}`;
         const prefix = `--${boundary}\r
 Content-Disposition: form-data`;
         const escape = (str) => str.replace(/\n/g, "%0A").replace(/\r/g, "%0D").replace(/"/g, "%22");
@@ -9252,8 +9584,11 @@ Content-Type: ${value.type || "application/octet-stream"}\r
       }
       return extractBody(object, keepalive);
     }
-    function cloneBody(body) {
+    function cloneBody(instance, body) {
       const [out1, out2] = body.stream.tee();
+      if (hasFinalizationRegistry) {
+        streamRegistry.register(instance, new WeakRef(out1));
+      }
       body.stream = out1;
       return {
         stream: out2,
@@ -9332,7 +9667,7 @@ Content-Type: ${value.type || "application/octet-stream"}\r
     }
     async function consumeBody(object, convertBytesToJSValue, instance) {
       webidl.brandCheck(object, instance);
-      if (bodyUnusable(object[kState].body)) {
+      if (bodyUnusable(object)) {
         throw new TypeError("Body is unusable: Body has already been read");
       }
       throwIfAborted(object[kState]);
@@ -9352,7 +9687,8 @@ Content-Type: ${value.type || "application/octet-stream"}\r
       await fullyReadBody(object[kState].body, successSteps, errorSteps);
       return promise.promise;
     }
-    function bodyUnusable(body) {
+    function bodyUnusable(object) {
+      const body = object[kState].body;
       return body != null && (body.stream.locked || util.isDisturbed(body.stream));
     }
     function parseJSONFromBytes(bytes) {
@@ -9370,14 +9706,17 @@ Content-Type: ${value.type || "application/octet-stream"}\r
       extractBody,
       safelyExtractBody,
       cloneBody,
-      mixinBody
+      mixinBody,
+      streamRegistry,
+      hasFinalizationRegistry,
+      bodyUnusable
     };
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client-h1.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client-h1.js
 var require_client_h1 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client-h1.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client-h1.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var util = require_util();
@@ -9449,35 +9788,35 @@ var require_client_h1 = __commonJS({
             return 0;
           },
           wasm_on_status: (p, at, len) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             const start = at - currentBufferPtr + currentBufferRef.byteOffset;
             return currentParser.onStatus(new FastBuffer(currentBufferRef.buffer, start, len)) || 0;
           },
           wasm_on_message_begin: (p) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             return currentParser.onMessageBegin() || 0;
           },
           wasm_on_header_field: (p, at, len) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             const start = at - currentBufferPtr + currentBufferRef.byteOffset;
             return currentParser.onHeaderField(new FastBuffer(currentBufferRef.buffer, start, len)) || 0;
           },
           wasm_on_header_value: (p, at, len) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             const start = at - currentBufferPtr + currentBufferRef.byteOffset;
             return currentParser.onHeaderValue(new FastBuffer(currentBufferRef.buffer, start, len)) || 0;
           },
           wasm_on_headers_complete: (p, statusCode, upgrade, shouldKeepAlive) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             return currentParser.onHeadersComplete(statusCode, Boolean(upgrade), Boolean(shouldKeepAlive)) || 0;
           },
           wasm_on_body: (p, at, len) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             const start = at - currentBufferPtr + currentBufferRef.byteOffset;
             return currentParser.onBody(new FastBuffer(currentBufferRef.buffer, start, len)) || 0;
           },
           wasm_on_message_complete: (p) => {
-            assert5.strictEqual(currentParser.ptr, p);
+            assert5(currentParser.ptr === p);
             return currentParser.onMessageComplete() || 0;
           }
           /* eslint-enable camelcase */
@@ -9491,9 +9830,11 @@ var require_client_h1 = __commonJS({
     var currentBufferRef = null;
     var currentBufferSize = 0;
     var currentBufferPtr = null;
-    var TIMEOUT_HEADERS = 1;
-    var TIMEOUT_BODY = 2;
-    var TIMEOUT_IDLE = 3;
+    var USE_NATIVE_TIMER = 0;
+    var USE_FAST_TIMER = 1;
+    var TIMEOUT_HEADERS = 2 | USE_FAST_TIMER;
+    var TIMEOUT_BODY = 4 | USE_FAST_TIMER;
+    var TIMEOUT_KEEP_ALIVE = 8 | USE_NATIVE_TIMER;
     var Parser2 = class {
       constructor(client, socket, { exports: exports3 }) {
         assert5(Number.isFinite(client[kMaxHeadersSize]) && client[kMaxHeadersSize] > 0);
@@ -9519,24 +9860,27 @@ var require_client_h1 = __commonJS({
         this.connection = "";
         this.maxResponseSize = client[kMaxResponseSize];
       }
-      setTimeout(value, type) {
-        this.timeoutType = type;
-        if (value !== this.timeoutValue) {
-          timers.clearTimeout(this.timeout);
-          if (value) {
-            this.timeout = timers.setTimeout(onParserTimeout, value, this);
-            if (this.timeout.unref) {
+      setTimeout(delay, type) {
+        if (delay !== this.timeoutValue || type & USE_FAST_TIMER ^ this.timeoutType & USE_FAST_TIMER) {
+          if (this.timeout) {
+            timers.clearTimeout(this.timeout);
+            this.timeout = null;
+          }
+          if (delay) {
+            if (type & USE_FAST_TIMER) {
+              this.timeout = timers.setFastTimeout(onParserTimeout, delay, new WeakRef(this));
+            } else {
+              this.timeout = setTimeout(onParserTimeout, delay, new WeakRef(this));
               this.timeout.unref();
             }
-          } else {
-            this.timeout = null;
           }
-          this.timeoutValue = value;
+          this.timeoutValue = delay;
         } else if (this.timeout) {
           if (this.timeout.refresh) {
             this.timeout.refresh();
           }
         }
+        this.timeoutType = type;
       }
       resume() {
         if (this.socket.destroyed || !this.paused) {
@@ -9613,7 +9957,7 @@ var require_client_h1 = __commonJS({
         assert5(currentParser == null);
         this.llhttp.llhttp_free(this.ptr);
         this.ptr = null;
-        timers.clearTimeout(this.timeout);
+        this.timeout && timers.clearTimeout(this.timeout);
         this.timeout = null;
         this.timeoutValue = null;
         this.timeoutType = null;
@@ -9672,16 +10016,16 @@ var require_client_h1 = __commonJS({
       onUpgrade(head) {
         const { upgrade, client, socket, headers, statusCode } = this;
         assert5(upgrade);
-        const request = client[kQueue][client[kRunningIdx]];
-        assert5(request);
+        assert5(client[kSocket] === socket);
         assert5(!socket.destroyed);
-        assert5(socket === client[kSocket]);
         assert5(!this.paused);
+        assert5((headers.length & 1) === 0);
+        const request = client[kQueue][client[kRunningIdx]];
+        assert5(request);
         assert5(request.upgrade || request.method === "CONNECT");
         this.statusCode = null;
         this.statusText = "";
         this.shouldKeepAlive = null;
-        assert5(this.headers.length % 2 === 0);
         this.headers = [];
         this.headersSize = 0;
         socket.unshift(head);
@@ -9720,7 +10064,7 @@ var require_client_h1 = __commonJS({
           util.destroy(socket, new SocketError("bad upgrade", util.getSocketInfo(socket)));
           return -1;
         }
-        assert5.strictEqual(this.timeoutType, TIMEOUT_HEADERS);
+        assert5(this.timeoutType === TIMEOUT_HEADERS);
         this.statusCode = statusCode;
         this.shouldKeepAlive = shouldKeepAlive || // Override llhttp value which does not allow keepAlive for HEAD.
         request.method === "HEAD" && !socket[kReset] && this.connection.toLowerCase() === "keep-alive";
@@ -9742,7 +10086,7 @@ var require_client_h1 = __commonJS({
           this.upgrade = true;
           return 2;
         }
-        assert5(this.headers.length % 2 === 0);
+        assert5((this.headers.length & 1) === 0);
         this.headers = [];
         this.headersSize = 0;
         if (this.shouldKeepAlive && client[kPipelining]) {
@@ -9786,7 +10130,7 @@ var require_client_h1 = __commonJS({
         }
         const request = client[kQueue][client[kRunningIdx]];
         assert5(request);
-        assert5.strictEqual(this.timeoutType, TIMEOUT_BODY);
+        assert5(this.timeoutType === TIMEOUT_BODY);
         if (this.timeout) {
           if (this.timeout.refresh) {
             this.timeout.refresh();
@@ -9810,16 +10154,16 @@ var require_client_h1 = __commonJS({
         if (upgrade) {
           return;
         }
+        assert5(statusCode >= 100);
+        assert5((this.headers.length & 1) === 0);
         const request = client[kQueue][client[kRunningIdx]];
         assert5(request);
-        assert5(statusCode >= 100);
         this.statusCode = null;
         this.statusText = "";
         this.bytesRead = 0;
         this.contentLength = "";
         this.keepAlive = "";
         this.connection = "";
-        assert5(this.headers.length % 2 === 0);
         this.headers = [];
         this.headersSize = 0;
         if (statusCode < 200) {
@@ -9832,7 +10176,7 @@ var require_client_h1 = __commonJS({
         request.onComplete(headers);
         client[kQueue][client[kRunningIdx]++] = null;
         if (socket[kWriting]) {
-          assert5.strictEqual(client[kRunning], 0);
+          assert5(client[kRunning] === 0);
           util.destroy(socket, new InformationalError("reset"));
           return constants2.ERROR.PAUSED;
         } else if (!shouldKeepAlive) {
@@ -9849,17 +10193,17 @@ var require_client_h1 = __commonJS({
       }
     };
     function onParserTimeout(parser) {
-      const { socket, timeoutType, client } = parser;
+      const { socket, timeoutType, client, paused } = parser.deref();
       if (timeoutType === TIMEOUT_HEADERS) {
         if (!socket[kWriting] || socket.writableNeedDrain || client[kRunning] > 1) {
-          assert5(!parser.paused, "cannot be paused while waiting for headers");
+          assert5(!paused, "cannot be paused while waiting for headers");
           util.destroy(socket, new HeadersTimeoutError());
         }
       } else if (timeoutType === TIMEOUT_BODY) {
-        if (!parser.paused) {
+        if (!paused) {
           util.destroy(socket, new BodyTimeoutError());
         }
-      } else if (timeoutType === TIMEOUT_IDLE) {
+      } else if (timeoutType === TIMEOUT_KEEP_ALIVE) {
         assert5(client[kRunning] === 0 && client[kKeepAliveTimeoutValue]);
         util.destroy(socket, new InformationalError("socket idle timeout"));
       }
@@ -9876,8 +10220,8 @@ var require_client_h1 = __commonJS({
       socket[kBlocking] = false;
       socket[kParser] = new Parser2(client, socket, llhttpInstance);
       addListener(socket, "error", function(err) {
-        const parser = this[kParser];
         assert5(err.code !== "ERR_TLS_CERT_ALTNAME_INVALID");
+        const parser = this[kParser];
         if (err.code === "ECONNRESET" && parser.statusCode && !parser.shouldKeepAlive) {
           parser.onMessageComplete();
           return;
@@ -9984,8 +10328,8 @@ var require_client_h1 = __commonJS({
           socket[kNoRef] = false;
         }
         if (client[kSize] === 0) {
-          if (socket[kParser].timeoutType !== TIMEOUT_IDLE) {
-            socket[kParser].setTimeout(client[kKeepAliveTimeoutValue], TIMEOUT_IDLE);
+          if (socket[kParser].timeoutType !== TIMEOUT_KEEP_ALIVE) {
+            socket[kParser].setTimeout(client[kKeepAliveTimeoutValue], TIMEOUT_KEEP_ALIVE);
           }
         } else if (client[kRunning] > 0 && socket[kParser].statusCode < 200) {
           if (socket[kParser].timeoutType !== TIMEOUT_HEADERS) {
@@ -10002,7 +10346,7 @@ var require_client_h1 = __commonJS({
     function writeH1(client, request) {
       const { method, path: path16, host, upgrade, blocking, reset } = request;
       let { body, headers, contentLength } = request;
-      const expectsPayload = method === "PUT" || method === "POST" || method === "PATCH";
+      const expectsPayload = method === "PUT" || method === "POST" || method === "PATCH" || method === "QUERY" || method === "PROPFIND" || method === "PROPPATCH";
       if (util.isFormDataLike(body)) {
         if (!extractBody) {
           extractBody = require_body().extractBody;
@@ -10210,7 +10554,7 @@ upgrade: ${upgrade}\r
           socket.write(body);
           socket.uncork();
           request.onBodySent(body);
-          if (!expectsPayload) {
+          if (!expectsPayload && request.reset !== false) {
             socket[kReset] = true;
           }
         }
@@ -10235,7 +10579,7 @@ upgrade: ${upgrade}\r
         socket.uncork();
         request.onBodySent(buffer);
         request.onRequestSent();
-        if (!expectsPayload) {
+        if (!expectsPayload && request.reset !== false) {
           socket[kReset] = true;
         }
         client[kResume]();
@@ -10311,7 +10655,7 @@ upgrade: ${upgrade}\r
         }
         socket.cork();
         if (bytesWritten === 0) {
-          if (!expectsPayload) {
+          if (!expectsPayload && request.reset !== false) {
             socket[kReset] = true;
           }
           if (contentLength === null) {
@@ -10390,9 +10734,9 @@ ${len.toString(16)}\r
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client-h2.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client-h2.js
 var require_client_h2 = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client-h2.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client-h2.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var { pipeline } = require("node:stream");
@@ -10418,9 +10762,12 @@ var require_client_h2 = __commonJS({
       kOnError,
       kMaxConcurrentStreams,
       kHTTP2Session,
-      kResume
+      kResume,
+      kSize,
+      kHTTPContext
     } = require_symbols();
     var kOpenStreams = Symbol("open streams");
+    var extractBody;
     var h2ExperimentalWarned = false;
     var http2;
     try {
@@ -10515,9 +10862,10 @@ var require_client_h2 = __commonJS({
         version: "h2",
         defaultPipelining: Infinity,
         write(...args) {
-          writeH2(client, ...args);
+          return writeH2(client, ...args);
         },
         resume() {
+          resumeH2(client);
         },
         destroy(err, callback) {
           if (closed) {
@@ -10534,6 +10882,18 @@ var require_client_h2 = __commonJS({
         }
       };
     }
+    function resumeH2(client) {
+      const socket = client[kSocket];
+      if (socket?.destroyed === false) {
+        if (client[kSize] === 0 && client[kMaxConcurrentStreams] === 0) {
+          socket.unref();
+          client[kHTTP2Session].unref();
+        } else {
+          socket.ref();
+          client[kHTTP2Session].ref();
+        }
+      }
+    }
     function onHttp2SessionError(err) {
       assert5(err.code !== "ERR_TLS_CERT_ALTNAME_INVALID");
       this[kSocket][kError] = err;
@@ -10552,25 +10912,36 @@ var require_client_h2 = __commonJS({
       util.destroy(this[kSocket], err);
     }
     function onHTTP2GoAway(code2) {
-      const err = new RequestAbortedError(`HTTP/2: "GOAWAY" frame received with code ${code2}`);
-      this[kSocket][kError] = err;
-      this[kClient][kOnError](err);
-      this.unref();
+      const err = this[kError] || new SocketError(`HTTP/2: "GOAWAY" frame received with code ${code2}`, util.getSocketInfo(this));
+      const client = this[kClient];
+      client[kSocket] = null;
+      client[kHTTPContext] = null;
+      if (this[kHTTP2Session] != null) {
+        this[kHTTP2Session].destroy(err);
+        this[kHTTP2Session] = null;
+      }
       util.destroy(this[kSocket], err);
+      if (client[kRunningIdx] < client[kQueue].length) {
+        const request = client[kQueue][client[kRunningIdx]];
+        client[kQueue][client[kRunningIdx]++] = null;
+        util.errorRequest(client, request, err);
+        client[kPendingIdx] = client[kRunningIdx];
+      }
+      assert5(client[kRunning] === 0);
+      client.emit("disconnect", client[kUrl], [client], err);
+      client[kResume]();
     }
     function shouldSendContentLength(method) {
       return method !== "GET" && method !== "HEAD" && method !== "OPTIONS" && method !== "TRACE" && method !== "CONNECT";
     }
     function writeH2(client, request) {
       const session = client[kHTTP2Session];
-      const { body, method, path: path16, host, upgrade, expectContinue, signal, headers: reqHeaders } = request;
+      const { method, path: path16, host, upgrade, expectContinue, signal, headers: reqHeaders } = request;
+      let { body } = request;
       if (upgrade) {
         util.errorRequest(client, request, new Error("Upgrade not supported for H2"));
         return false;
       }
-      if (request.aborted) {
-        return false;
-      }
       const headers = {};
       for (let n = 0; n < reqHeaders.length; n += 2) {
         const key = reqHeaders[n + 0];
@@ -10601,22 +10972,29 @@ var require_client_h2 = __commonJS({
           util.destroy(stream, err);
         }
         util.destroy(body, err);
+        client[kQueue][client[kRunningIdx]++] = null;
+        client[kResume]();
       };
       try {
         request.onConnect(abort);
       } catch (err) {
         util.errorRequest(client, request, err);
       }
+      if (request.aborted) {
+        return false;
+      }
       if (method === "CONNECT") {
         session.ref();
         stream = session.request(headers, { endStream: false, signal });
         if (stream.id && !stream.pending) {
           request.onUpgrade(null, null, stream);
           ++session[kOpenStreams];
+          client[kQueue][client[kRunningIdx]++] = null;
         } else {
           stream.once("ready", () => {
             request.onUpgrade(null, null, stream);
             ++session[kOpenStreams];
+            client[kQueue][client[kRunningIdx]++] = null;
           });
         }
         stream.once("close", () => {
@@ -10632,6 +11010,13 @@ var require_client_h2 = __commonJS({
         body.read(0);
       }
       let contentLength = util.bodyLength(body);
+      if (util.isFormDataLike(body)) {
+        extractBody ??= require_body().extractBody;
+        const [bodyStream, contentType] = extractBody(body);
+        headers["content-type"] = contentType;
+        body = bodyStream.stream;
+        contentLength = bodyStream.length;
+      }
       if (contentLength == null) {
         contentLength = request.contentLength;
       }
@@ -10684,12 +11069,14 @@ var require_client_h2 = __commonJS({
       stream.once("end", () => {
         if (stream.state?.state == null || stream.state.state < 6) {
           request.onComplete([]);
-          return;
         }
         if (session[kOpenStreams] === 0) {
           session.unref();
         }
         abort(new InformationalError("HTTP/2: stream half-closed (remote)"));
+        client[kQueue][client[kRunningIdx]++] = null;
+        client[kPendingIdx] = client[kRunningIdx];
+        client[kResume]();
       });
       stream.once("close", () => {
         session[kOpenStreams] -= 1;
@@ -10888,9 +11275,9 @@ var require_client_h2 = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/handler/redirect-handler.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/handler/redirect-handler.js
 var require_redirect_handler = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/handler/redirect-handler.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/handler/redirect-handler.js"(exports2, module2) {
     "use strict";
     var util = require_util();
     var { kBodyUsed } = require_symbols();
@@ -11047,9 +11434,9 @@ var require_redirect_handler = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/interceptor/redirect-interceptor.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/interceptor/redirect-interceptor.js
 var require_redirect_interceptor = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/interceptor/redirect-interceptor.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/interceptor/redirect-interceptor.js"(exports2, module2) {
     "use strict";
     var RedirectHandler = require_redirect_handler();
     function createRedirectInterceptor({ maxRedirections: defaultMaxRedirections }) {
@@ -11069,9 +11456,9 @@ var require_redirect_interceptor = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client.js
 var require_client = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/client.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/client.js"(exports2, module2) {
     "use strict";
     var assert5 = require("node:assert");
     var net = require("node:net");
@@ -11132,6 +11519,8 @@ var require_client = __commonJS({
     var connectH2 = require_client_h2();
     var deprecatedInterceptorWarned = false;
     var kClosedResolve = Symbol("kClosedResolve");
+    var noop2 = () => {
+    };
     function getPipelining(client) {
       return client[kPipelining] ?? client[kHTTPContext]?.defaultPipelining ?? 1;
     }
@@ -11420,16 +11809,14 @@ var require_client = __commonJS({
           });
         });
         if (client.destroyed) {
-          util.destroy(socket.on("error", () => {
-          }), new ClientDestroyedError());
+          util.destroy(socket.on("error", noop2), new ClientDestroyedError());
           return;
         }
         assert5(socket);
         try {
           client[kHTTPContext] = socket.alpnProtocol === "h2" ? await connectH2(client, socket) : await connectH1(client, socket);
         } catch (err) {
-          socket.destroy().on("error", () => {
-          });
+          socket.destroy().on("error", noop2);
           throw err;
         }
         client[kConnecting] = false;
@@ -11569,9 +11956,9 @@ var require_client = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool.js
 var require_pool = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/pool.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/pool.js"(exports2, module2) {
     "use strict";
     var {
       PoolBase,
@@ -11652,9 +12039,9 @@ var require_pool = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/agent.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/agent.js
 var require_agent = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/agent.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/agent.js"(exports2, module2) {
     "use strict";
     var { InvalidArgumentError } = require_errors();
     var { kClients, kRunning, kClose, kDestroy, kDispatch, kInterceptors } = require_symbols();
@@ -11749,9 +12136,9 @@ var require_agent = __commonJS({
   }
 });
 
-// .yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/proxy-agent.js
+// .yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/proxy-agent.js
 var require_proxy_agent = __commonJS({
-  ".yarn/cache/undici-npm-6.19.2-a9aa1269bb-3b7b9238c0.zip/node_modules/undici/lib/dispatcher/proxy-agent.js"(exports2, module2) {
+  ".yarn/cache/undici-npm-6.21.1-0f7fc2c179-d604080e4f.zip/node_modules/undici/lib/dispatcher/proxy-agent.js"(exports2, module2) {
     "use strict";
     var { kProxy, kClose, kDestroy, kInterceptors } = require_symbols();
     var { URL: URL2 } = require("node:url");
@@ -11772,6 +12159,8 @@ var require_proxy_agent = __commonJS({
     function defaultFactory(origin, opts) {
       return new Pool(origin, opts);
     }
+    var noop2 = () => {
+    };
     var ProxyAgent2 = class extends DispatcherBase {
       constructor(opts) {
         super();
@@ -11821,8 +12210,7 @@ var require_proxy_agent = __commonJS({
                 servername: this[kProxyTls]?.servername || proxyHostname
               });
               if (statusCode !== 200) {
-                socket.on("error", () => {
-                }).destroy();
+                socket.on("error", noop2).destroy();
                 callback(new RequestAbortedError(`Proxy response (${statusCode}) !== 200 when HTTP Tunneling`));
               }
               if (opts2.protocol !== "https:") {
@@ -13174,10 +13562,10 @@ var init_esm2 = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/options.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/options.js
 var argmap, isSyncFile, isAsyncFile, isSyncNoFile, isAsyncNoFile, dealiasKey, dealias;
 var init_options = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/options.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/options.js"() {
     argmap = /* @__PURE__ */ new Map([
       ["C", "cwd"],
       ["f", "file"],
@@ -13227,10 +13615,10 @@ var init_options = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/make-command.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/make-command.js
 var makeCommand;
 var init_make_command = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/make-command.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/make-command.js"() {
     init_options();
     makeCommand = (syncFile, asyncFile, syncNoFile, asyncNoFile, validate) => {
       return Object.assign((opt_ = [], entries, cb) => {
@@ -14027,10 +14415,10 @@ var init_esm4 = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/large-numbers.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/large-numbers.js
 var encode, encodePositive, encodeNegative, parse, twos, pos, onesComp, twosComp;
 var init_large_numbers = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/large-numbers.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/large-numbers.js"() {
     encode = (num, buf) => {
       if (!Number.isSafeInteger(num)) {
         throw Error("cannot encode number outside of javascript safe integer range");
@@ -14113,10 +14501,10 @@ var init_large_numbers = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/types.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/types.js
 var isCode, name, code;
 var init_types = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/types.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/types.js"() {
     isCode = (c) => name.has(c);
     name = /* @__PURE__ */ new Map([
       ["0", "File"],
@@ -14161,10 +14549,10 @@ var init_types = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/header.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/header.js
 var import_node_path, Header, splitPrefix, decString, decDate, numToDate, decNumber, nanUndef, decSmallNumber, MAXNUM, encNumber, encSmallNumber, octalString, padOctal, encDate, NULLS, encString;
 var init_header = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/header.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/header.js"() {
     import_node_path = require("node:path");
     init_large_numbers();
     init_types();
@@ -14372,10 +14760,10 @@ var init_header = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/pax.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/pax.js
 var import_node_path2, Pax, merge, parseKV, parseKVLine;
 var init_pax = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/pax.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/pax.js"() {
     import_node_path2 = require("node:path");
     init_header();
     Pax = class _Pax {
@@ -14495,19 +14883,19 @@ var init_pax = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/normalize-windows-path.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/normalize-windows-path.js
 var platform, normalizeWindowsPath;
 var init_normalize_windows_path = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/normalize-windows-path.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/normalize-windows-path.js"() {
     platform = process.env.TESTING_TAR_FAKE_PLATFORM || process.platform;
     normalizeWindowsPath = platform !== "win32" ? (p) => p : (p) => p && p.replace(/\\/g, "/");
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/read-entry.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/read-entry.js
 var ReadEntry;
 var init_read_entry = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/read-entry.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/read-entry.js"() {
     init_esm();
     init_normalize_windows_path();
     ReadEntry = class extends Minipass {
@@ -14567,6 +14955,8 @@ var init_read_entry = __esm({
           case "OldExtendedHeader":
             this.meta = true;
             break;
+          // NOTE: gnutar and bsdtar treat unrecognized types as 'File'
+          // it may be worth doing the same, but with a warning.
           default:
             this.ignore = true;
         }
@@ -14626,10 +15016,10 @@ var init_read_entry = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/warn-method.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/warn-method.js
 var warnMethod;
 var init_warn_method = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/warn-method.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/warn-method.js"() {
     warnMethod = (self2, code2, message, data = {}) => {
       if (self2.file) {
         data.file = self2.file;
@@ -14654,10 +15044,10 @@ var init_warn_method = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/parse.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/parse.js
 var import_events3, maxMetaEntrySize, gzipHeader, STATE, WRITEENTRY, READENTRY, NEXTENTRY, PROCESSENTRY, EX, GEX, META, EMITMETA, BUFFER2, QUEUE, ENDED, EMITTEDEND, EMIT, UNZIP, CONSUMECHUNK, CONSUMECHUNKSUB, CONSUMEBODY, CONSUMEMETA, CONSUMEHEADER, CONSUMING, BUFFERCONCAT, MAYBEEND, WRITING, ABORTED2, DONE, SAW_VALID_ENTRY, SAW_NULL_BLOCK, SAW_EOF, CLOSESTREAM, noop, Parser;
 var init_parse = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/parse.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/parse.js"() {
     import_events3 = require("events");
     init_esm3();
     init_esm4();
@@ -14933,6 +15323,7 @@ var init_parse = __esm({
             ex.linkpath = this[META].replace(/\0.*/, "");
             break;
           }
+          /* c8 ignore start */
           default:
             throw new Error("unknown meta: " + entry.type);
         }
@@ -15087,6 +15478,7 @@ var init_parse = __esm({
             case "meta":
               position += this[CONSUMEMETA](chunk, position);
               break;
+            /* c8 ignore start */
             default:
               throw new Error("invalid state: " + this[STATE]);
           }
@@ -15137,10 +15529,10 @@ var init_parse = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/strip-trailing-slashes.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/strip-trailing-slashes.js
 var stripTrailingSlashes;
 var init_strip_trailing_slashes = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/strip-trailing-slashes.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/strip-trailing-slashes.js"() {
     stripTrailingSlashes = (str) => {
       let i = str.length - 1;
       let slashesStart = -1;
@@ -15153,7 +15545,7 @@ var init_strip_trailing_slashes = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/list.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/list.js
 var list_exports = {};
 __export(list_exports, {
   filesFilter: () => filesFilter,
@@ -15161,7 +15553,7 @@ __export(list_exports, {
 });
 var import_node_fs, import_path2, onReadEntryFunction, filesFilter, listFileSync, listFile, list;
 var init_list = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/list.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/list.js"() {
     init_esm2();
     import_node_fs = __toESM(require("node:fs"), 1);
     import_path2 = require("path");
@@ -15256,10 +15648,10 @@ var init_list = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/get-write-flag.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/get-write-flag.js
 var import_fs3, platform2, isWindows, O_CREAT, O_TRUNC, O_WRONLY, UV_FS_O_FILEMAP, fMapEnabled, fMapLimit, fMapFlag, getWriteFlag;
 var init_get_write_flag = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/get-write-flag.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/get-write-flag.js"() {
     import_fs3 = __toESM(require("fs"), 1);
     platform2 = process.env.__FAKE_PLATFORM__ || process.platform;
     isWindows = platform2 === "win32";
@@ -15622,10 +16014,10 @@ var init_mjs = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/cwd-error.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/cwd-error.js
 var CwdError;
 var init_cwd_error = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/cwd-error.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/cwd-error.js"() {
     CwdError = class extends Error {
       path;
       code;
@@ -15642,10 +16034,10 @@ var init_cwd_error = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/symlink-error.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/symlink-error.js
 var SymlinkError;
 var init_symlink_error = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/symlink-error.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/symlink-error.js"() {
     SymlinkError = class extends Error {
       path;
       symlink;
@@ -15663,10 +16055,10 @@ var init_symlink_error = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/mkdir.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/mkdir.js
 var import_fs6, import_node_path4, cGet, cSet, checkCwd, mkdir3, mkdir_, onmkdir, checkCwdSync, mkdirSync4;
 var init_mkdir = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/mkdir.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/mkdir.js"() {
     init_esm5();
     import_fs6 = __toESM(require("fs"), 1);
     init_mjs();
@@ -15841,10 +16233,10 @@ var init_mkdir = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/normalize-unicode.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/normalize-unicode.js
 var normalizeCache, hasOwnProperty, normalizeUnicode;
 var init_normalize_unicode = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/normalize-unicode.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/normalize-unicode.js"() {
     normalizeCache = /* @__PURE__ */ Object.create(null);
     ({ hasOwnProperty } = Object.prototype);
     normalizeUnicode = (s) => {
@@ -15856,10 +16248,10 @@ var init_normalize_unicode = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/strip-absolute-path.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/strip-absolute-path.js
 var import_node_path5, isAbsolute, parse4, stripAbsolutePath;
 var init_strip_absolute_path = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/strip-absolute-path.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/strip-absolute-path.js"() {
     import_node_path5 = require("node:path");
     ({ isAbsolute, parse: parse4 } = import_node_path5.win32);
     stripAbsolutePath = (path16) => {
@@ -15876,10 +16268,10 @@ var init_strip_absolute_path = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/winchars.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/winchars.js
 var raw, win, toWin, toRaw, encode2, decode;
 var init_winchars = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/winchars.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/winchars.js"() {
     raw = ["|", "<", ">", "?", ":"];
     win = raw.map((char) => String.fromCharCode(61440 + char.charCodeAt(0)));
     toWin = new Map(raw.map((char, i) => [char, win[i]]));
@@ -15889,10 +16281,10 @@ var init_winchars = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/path-reservations.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/path-reservations.js
 var import_node_path6, platform4, isWindows2, getDirs, PathReservations;
 var init_path_reservations = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/path-reservations.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/path-reservations.js"() {
     import_node_path6 = require("node:path");
     init_normalize_unicode();
     init_strip_trailing_slashes();
@@ -16029,10 +16421,10 @@ var init_path_reservations = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/unpack.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/unpack.js
 var import_node_assert, import_node_crypto, import_node_fs3, import_node_path7, ONENTRY, CHECKFS, CHECKFS2, PRUNECACHE, ISREUSABLE, MAKEFS, FILE, DIRECTORY, LINK, SYMLINK, HARDLINK, UNSUPPORTED, CHECKPATH, MKDIR, ONERROR, PENDING, PEND, UNPEND, ENDED2, MAYBECLOSE, SKIP, DOCHOWN, UID, GID, CHECKED_CWD, platform5, isWindows3, DEFAULT_MAX_DEPTH, unlinkFile, unlinkFileSync, uint32, cacheKeyNormalize, pruneCache, dropCache, Unpack, callSync, UnpackSync;
 var init_unpack = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/unpack.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/unpack.js"() {
     init_esm2();
     import_node_assert = __toESM(require("node:assert"), 1);
     import_node_crypto = require("node:crypto");
@@ -16285,6 +16677,7 @@ var init_unpack = __esm({
             if (entry.mode) {
               entry.mode = entry.mode | 448;
             }
+          // eslint-disable-next-line no-fallthrough
           case "File":
           case "OldFile":
           case "ContiguousFile":
@@ -16744,14 +17137,14 @@ var init_unpack = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/extract.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/extract.js
 var extract_exports = {};
 __export(extract_exports, {
   extract: () => extract
 });
 var import_node_fs4, extractFileSync, extractFile, extract;
 var init_extract = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/extract.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/extract.js"() {
     init_esm2();
     import_node_fs4 = __toESM(require("node:fs"), 1);
     init_list();
@@ -17072,6 +17465,22 @@ var require_v8_compile_cache = __commonJS({
   }
 });
 
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/satisfies.js
+var require_satisfies = __commonJS({
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/satisfies.js"(exports2, module2) {
+    var Range3 = require_range();
+    var satisfies = (version3, range, options) => {
+      try {
+        range = new Range3(range, options);
+      } catch (er) {
+        return false;
+      }
+      return range.test(version3);
+    };
+    module2.exports = satisfies;
+  }
+});
+
 // .yarn/cache/isexe-npm-3.1.1-9c0061eead-9ec2576540.zip/node_modules/isexe/dist/cjs/posix.js
 var require_posix = __commonJS({
   ".yarn/cache/isexe-npm-3.1.1-9c0061eead-9ec2576540.zip/node_modules/isexe/dist/cjs/posix.js"(exports2) {
@@ -17232,9 +17641,9 @@ var require_cjs = __commonJS({
   }
 });
 
-// .yarn/cache/which-npm-4.0.0-dd31cd4928-449fa5c44e.zip/node_modules/which/lib/index.js
+// .yarn/cache/which-npm-5.0.0-15aa39eb60-e556e4cd8b.zip/node_modules/which/lib/index.js
 var require_lib = __commonJS({
-  ".yarn/cache/which-npm-4.0.0-dd31cd4928-449fa5c44e.zip/node_modules/which/lib/index.js"(exports2, module2) {
+  ".yarn/cache/which-npm-5.0.0-15aa39eb60-e556e4cd8b.zip/node_modules/which/lib/index.js"(exports2, module2) {
     var { isexe, sync: isexeSync } = require_cjs();
     var { join: join3, delimiter, sep, posix } = require("path");
     var isWindows4 = process.platform === "win32";
@@ -18480,10 +18889,10 @@ ${nodePath ? "$env:NODE_PATH=$env_node_path\n" : ""}${prependPath ? "$env:PATH=$
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/mode-fix.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/mode-fix.js
 var modeFix;
 var init_mode_fix = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/mode-fix.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/mode-fix.js"() {
     modeFix = (mode, isDir, portable) => {
       mode &= 4095;
       if (portable) {
@@ -18505,10 +18914,10 @@ var init_mode_fix = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/write-entry.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/write-entry.js
 var import_fs14, import_path13, prefixPath, maxReadSize, PROCESS, FILE2, DIRECTORY2, SYMLINK2, HARDLINK2, HEADER, READ2, LSTAT, ONLSTAT, ONREAD, ONREADLINK, OPENFILE, ONOPENFILE, CLOSE, MODE, AWAITDRAIN, ONDRAIN, PREFIX, WriteEntry, WriteEntrySync, WriteEntryTar, getType;
 var init_write_entry = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/write-entry.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/write-entry.js"() {
     import_fs14 = __toESM(require("fs"), 1);
     init_esm();
     import_path13 = __toESM(require("path"), 1);
@@ -18577,7 +18986,7 @@ var init_write_entry = __esm({
       type;
       linkpath;
       stat;
-      /* c8 ignore start */
+      onWriteEntry;
       #hadError = false;
       constructor(p, opt_ = {}) {
         const opt = dealias(opt_);
@@ -18594,6 +19003,7 @@ var init_write_entry = __esm({
         this.noMtime = !!opt.noMtime;
         this.mtime = opt.mtime;
         this.prefix = opt.prefix ? normalizeWindowsPath(opt.prefix) : void 0;
+        this.onWriteEntry = opt.onWriteEntry;
         if (typeof opt.onwarn === "function") {
           this.on("warn", opt.onwarn);
         }
@@ -18662,6 +19072,7 @@ var init_write_entry = __esm({
             return this[DIRECTORY2]();
           case "SymbolicLink":
             return this[SYMLINK2]();
+          // unsupported types are ignored.
           default:
             return this.end();
         }
@@ -18679,6 +19090,7 @@ var init_write_entry = __esm({
         if (this.type === "Directory" && this.portable) {
           this.noMtime = true;
         }
+        this.onWriteEntry?.(this);
         this.header = new Header({
           path: this[PREFIX](this.path),
           // only apply the prefix to hard links.
@@ -18959,6 +19371,7 @@ var init_write_entry = __esm({
       ctime;
       linkpath;
       size;
+      onWriteEntry;
       warn(code2, message, data = {}) {
         return warnMethod(this, code2, message, data);
       }
@@ -18970,6 +19383,7 @@ var init_write_entry = __esm({
         this.strict = !!opt.strict;
         this.noPax = !!opt.noPax;
         this.noMtime = !!opt.noMtime;
+        this.onWriteEntry = opt.onWriteEntry;
         this.readEntry = readEntry;
         const { type } = readEntry;
         if (type === "Unsupported") {
@@ -19004,6 +19418,7 @@ var init_write_entry = __esm({
         }
         this.remain = readEntry.size;
         this.blockRemain = readEntry.startBlockSize;
+        this.onWriteEntry?.(this);
         this.header = new Header({
           path: this[PREFIX](this.path),
           linkpath: this.type === "Link" && this.linkpath !== void 0 ? this[PREFIX](this.linkpath) : this.linkpath,
@@ -19094,10 +19509,10 @@ var init_write_entry = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/pack.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/pack.js
 var import_fs15, import_path14, PackJob, EOF2, ONSTAT, ENDED3, QUEUE2, CURRENT, PROCESS2, PROCESSING, PROCESSJOB, JOBS, JOBDONE, ADDFSENTRY, ADDTARENTRY, STAT, READDIR, ONREADDIR, PIPE, ENTRY, ENTRYOPT, WRITEENTRYCLASS, WRITE, ONDRAIN2, Pack, PackSync;
 var init_pack = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/pack.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/pack.js"() {
     import_fs15 = __toESM(require("fs"), 1);
     init_write_entry();
     init_esm();
@@ -19235,16 +19650,24 @@ var init_pack = __esm({
         this.write(path16);
         return this;
       }
-      //@ts-ignore
-      end(path16) {
+      end(path16, encoding, cb) {
+        if (typeof path16 === "function") {
+          cb = path16;
+          path16 = void 0;
+        }
+        if (typeof encoding === "function") {
+          cb = encoding;
+          encoding = void 0;
+        }
         if (path16) {
           this.add(path16);
         }
         this[ENDED3] = true;
         this[PROCESS2]();
+        if (cb)
+          cb();
         return this;
       }
-      //@ts-ignore
       write(path16) {
         if (this[ENDED3]) {
           throw new Error("write after end");
@@ -19402,14 +19825,14 @@ var init_pack = __esm({
           statCache: this.statCache,
           noMtime: this.noMtime,
           mtime: this.mtime,
-          prefix: this.prefix
+          prefix: this.prefix,
+          onWriteEntry: this.onWriteEntry
         };
       }
       [ENTRY](job) {
         this[JOBS] += 1;
         try {
           const e = new this[WRITEENTRYCLASS](job.path, this[ENTRYOPT](job));
-          this.onWriteEntry?.(e);
           return e.on("end", () => this[JOBDONE](job)).on("error", (er) => this.emit("error", er));
         } catch (er) {
           this.emit("error", er);
@@ -19503,14 +19926,14 @@ var init_pack = __esm({
   }
 });
 
-// .yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/create.js
+// .yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/create.js
 var create_exports = {};
 __export(create_exports, {
   create: () => create
 });
 var import_node_path8, createFileSync, createFile, addFilesSync, addFilesAsync, createSync, createAsync, create;
 var init_create = __esm({
-  ".yarn/cache/tar-npm-7.4.0-2d244f1b3c-f4bab85fd1.zip/node_modules/tar/dist/esm/create.js"() {
+  ".yarn/cache/tar-npm-7.4.3-1dbbd1ffc3-d4679609bb.zip/node_modules/tar/dist/esm/create.js"() {
     init_esm2();
     import_node_path8 = __toESM(require("node:path"), 1);
     init_list();
@@ -19588,9 +20011,9 @@ var init_create = __esm({
   }
 });
 
-// .yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/major.js
+// .yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/major.js
 var require_major = __commonJS({
-  ".yarn/cache/semver-npm-7.6.3-57e82c14d5-88f33e148b.zip/node_modules/semver/functions/major.js"(exports2, module2) {
+  ".yarn/cache/semver-npm-7.7.1-4572475307-fd603a6fb9.zip/node_modules/semver/functions/major.js"(exports2, module2) {
     var SemVer3 = require_semver();
     var major = (a, loose) => new SemVer3(a, loose).major;
     module2.exports = major;
@@ -21260,21 +21683,21 @@ function String2(descriptor, ...args) {
 }
 
 // package.json
-var version = "0.31.0";
+var version = "0.32.0";
 
 // sources/Engine.ts
 var import_fs9 = __toESM(require("fs"));
 var import_path9 = __toESM(require("path"));
 var import_process3 = __toESM(require("process"));
 var import_rcompare = __toESM(require_rcompare());
-var import_valid2 = __toESM(require_valid());
-var import_valid3 = __toESM(require_valid2());
+var import_valid3 = __toESM(require_valid());
+var import_valid4 = __toESM(require_valid2());
 
 // config.json
 var config_default = {
   definitions: {
     npm: {
-      default: "11.0.0+sha1.7bba7c80740ef1f5b2c5d4cecc55e94912faa5e6",
+      default: "11.1.0+sha1.dba08f7d0f5301ebedaf968b4f74b2282f97a750",
       fetchLatestFrom: {
         type: "npm",
         package: "npm"
@@ -21311,7 +21734,7 @@ var config_default = {
       }
     },
     pnpm: {
-      default: "9.15.4+sha1.ffa0b5c573381e8035b354028ccff97c8e452047",
+      default: "10.5.2+sha1.ca68c0441df195b7e2992f1d1cb12fb731f82d78",
       fetchLatestFrom: {
         type: "npm",
         package: "pnpm"
@@ -21536,15 +21959,23 @@ async function fetchAsJson2(packageName, version3) {
   return fetchAsJson(`${npmRegistryUrl}/${packageName}${version3 ? `/${version3}` : ``}`, { headers });
 }
 function verifySignature({ signatures, integrity, packageName, version: version3 }) {
-  const { npm: keys } = process.env.COREPACK_INTEGRITY_KEYS ? JSON.parse(process.env.COREPACK_INTEGRITY_KEYS) : config_default.keys;
-  const key = keys.find(({ keyid }) => signatures.some((s) => s.keyid === keyid));
-  const signature = signatures.find(({ keyid }) => keyid === key?.keyid);
-  if (key == null || signature == null) throw new Error(`Cannot find matching keyid: ${JSON.stringify({ signatures, keys })}`);
+  if (!Array.isArray(signatures) || !signatures.length) throw new Error(`No compatible signature found in package metadata`);
+  const { npm: trustedKeys } = process.env.COREPACK_INTEGRITY_KEYS ? JSON.parse(process.env.COREPACK_INTEGRITY_KEYS) : config_default.keys;
+  let signature;
+  let key;
+  for (const k of trustedKeys) {
+    signature = signatures.find(({ keyid }) => keyid === k.keyid);
+    if (signature != null) {
+      key = k.key;
+      break;
+    }
+  }
+  if (signature?.sig == null) throw new UsageError(`The package was not signed by any trusted keys: ${JSON.stringify({ signatures, trustedKeys }, void 0, 2)}`);
   const verifier = (0, import_crypto.createVerify)(`SHA256`);
   verifier.end(`${packageName}@${version3}:${integrity}`);
   const valid = verifier.verify(
     `-----BEGIN PUBLIC KEY-----
-${key.key}
+${key}
 -----END PUBLIC KEY-----`,
     signature.sig,
     `base64`
@@ -21557,12 +21988,16 @@ async function fetchLatestStableVersion(packageName) {
   const metadata = await fetchAsJson2(packageName, `latest`);
   const { version: version3, dist: { integrity, signatures, shasum } } = metadata;
   if (!shouldSkipIntegrityCheck()) {
-    verifySignature({
-      packageName,
-      version: version3,
-      integrity,
-      signatures
-    });
+    try {
+      verifySignature({
+        packageName,
+        version: version3,
+        integrity,
+        signatures
+      });
+    } catch (cause) {
+      throw new Error(`Corepack cannot download the latest stable version of ${packageName}; you can disable signature verification by setting COREPACK_INTEGRITY_CHECK to 0 in your env, or instruct Corepack to use the latest stable release known by this version of Corepack by setting COREPACK_USE_LATEST to 0`, { cause });
+    }
   }
   return `${version3}+${integrity ? `sha512.${Buffer.from(integrity.slice(7), `base64`).toString(`hex`)}` : `sha1.${shasum}`}`;
 }
@@ -22028,7 +22463,10 @@ function satisfiesWithPrereleases(version3, range, loose = false) {
 // sources/specUtils.ts
 var import_fs8 = __toESM(require("fs"));
 var import_path8 = __toESM(require("path"));
+var import_satisfies = __toESM(require_satisfies());
 var import_valid = __toESM(require_valid());
+var import_valid2 = __toESM(require_valid2());
+var import_util = require("util");
 
 // sources/nodeUtils.ts
 var import_os2 = __toESM(require("os"));
@@ -22120,11 +22558,61 @@ function parseSpec(raw2, source, { enforceExactVersion = true } = {}) {
     range
   };
 }
+function warnOrThrow(errorMessage, onFail) {
+  switch (onFail) {
+    case `ignore`:
+      break;
+    case `error`:
+    case void 0:
+      throw new UsageError(errorMessage);
+    default:
+      console.warn(`! Corepack validation warning: ${errorMessage}`);
+  }
+}
+function parsePackageJSON(packageJSONContent) {
+  const { packageManager: pm } = packageJSONContent;
+  if (packageJSONContent.devEngines?.packageManager != null) {
+    const { packageManager } = packageJSONContent.devEngines;
+    if (typeof packageManager !== `object`) {
+      console.warn(`! Corepack only supports objects as valid value for devEngines.packageManager. The current value (${JSON.stringify(packageManager)}) will be ignored.`);
+      return pm;
+    }
+    if (Array.isArray(packageManager)) {
+      console.warn(`! Corepack does not currently support array values for devEngines.packageManager`);
+      return pm;
+    }
+    const { name: name2, version: version3, onFail } = packageManager;
+    if (typeof name2 !== `string` || name2.includes(`@`)) {
+      warnOrThrow(`The value of devEngines.packageManager.name ${JSON.stringify(name2)} is not a supported string value`, onFail);
+      return pm;
+    }
+    if (version3 != null && (typeof version3 !== `string` || !(0, import_valid2.default)(version3))) {
+      warnOrThrow(`The value of devEngines.packageManager.version ${JSON.stringify(version3)} is not a valid semver range`, onFail);
+      return pm;
+    }
+    log(`devEngines.packageManager defines that ${name2}@${version3} is the local package manager`);
+    if (pm) {
+      if (!pm.startsWith?.(`${name2}@`))
+        warnOrThrow(`"packageManager" field is set to ${JSON.stringify(pm)} which does not match the "devEngines.packageManager" field set to ${JSON.stringify(name2)}`, onFail);
+      else if (version3 != null && !(0, import_satisfies.default)(pm.slice(packageManager.name.length + 1), version3))
+        warnOrThrow(`"packageManager" field is set to ${JSON.stringify(pm)} which does not match the value defined in "devEngines.packageManager" for ${JSON.stringify(name2)} of ${JSON.stringify(version3)}`, onFail);
+      return pm;
+    }
+    return `${name2}@${version3 ?? `*`}`;
+  }
+  return pm;
+}
 async function setLocalPackageManager(cwd, info) {
   const lookup = await loadSpec(cwd);
+  const range = `range` in lookup && lookup.range;
+  if (range) {
+    if (info.locator.name !== range.name || !(0, import_satisfies.default)(info.locator.reference, range.range)) {
+      warnOrThrow(`The requested version of ${info.locator.name}@${info.locator.reference} does not match the devEngines specification (${range.name}@${range.range})`, range.onFail);
+    }
+  }
   const content = lookup.type !== `NoProject` ? await import_fs8.default.promises.readFile(lookup.target, `utf8`) : ``;
   const { data, indent } = readPackageJson(content);
-  const previousPackageManager = data.packageManager ?? `unknown`;
+  const previousPackageManager = data.packageManager ?? (range ? `${range.name}@${range.range}` : `unknown`);
   data.packageManager = `${info.locator.name}@${info.locator.reference}`;
   const newContent = normalizeLineEndings(content, `${JSON.stringify(data, null, indent)}
 `);
@@ -22158,17 +22646,53 @@ async function loadSpec(initialCwd) {
     }
     if (typeof data !== `object` || data === null)
       throw new UsageError(`Invalid package.json in ${import_path8.default.relative(initialCwd, manifestPath)}`);
-    selection = { data, manifestPath };
+    let localEnv;
+    const envFilePath2 = import_path8.default.resolve(currCwd, process.env.COREPACK_ENV_FILE ?? `.corepack.env`);
+    if (process.env.COREPACK_ENV_FILE == `0`) {
+      log(`Skipping env file as configured with COREPACK_ENV_FILE`);
+      localEnv = process.env;
+    } else if (typeof import_util.parseEnv !== `function`) {
+      log(`Skipping env file as it is not supported by the current version of Node.js`);
+      localEnv = process.env;
+    } else {
+      log(`Checking ${envFilePath2}`);
+      try {
+        localEnv = {
+          ...Object.fromEntries(Object.entries((0, import_util.parseEnv)(await import_fs8.default.promises.readFile(envFilePath2, `utf8`))).filter((e) => e[0].startsWith(`COREPACK_`))),
+          ...process.env
+        };
+        log(`Successfully loaded env file found at ${envFilePath2}`);
+      } catch (err) {
+        if (err?.code !== `ENOENT`)
+          throw err;
+        log(`No env file found at ${envFilePath2}`);
+        localEnv = process.env;
+      }
+    }
+    selection = { data, manifestPath, localEnv, envFilePath: envFilePath2 };
   }
   if (selection === null)
     return { type: `NoProject`, target: import_path8.default.join(initialCwd, `package.json`) };
-  const rawPmSpec = selection.data.packageManager;
+  let envFilePath;
+  if (selection.localEnv !== process.env) {
+    envFilePath = selection.envFilePath;
+    process.env = selection.localEnv;
+  }
+  const rawPmSpec = parsePackageJSON(selection.data);
   if (typeof rawPmSpec === `undefined`)
     return { type: `NoSpec`, target: selection.manifestPath };
+  log(`${selection.manifestPath} defines ${rawPmSpec} as local package manager`);
   return {
     type: `Found`,
     target: selection.manifestPath,
-    spec: parseSpec(rawPmSpec, import_path8.default.relative(initialCwd, selection.manifestPath))
+    envFilePath,
+    range: selection.data.devEngines?.packageManager?.version && {
+      name: selection.data.devEngines.packageManager.name,
+      range: selection.data.devEngines.packageManager.version,
+      onFail: selection.data.devEngines.packageManager.onFail
+    },
+    // Lazy-loading it so we do not throw errors on commands that do not need valid spec.
+    getSpec: () => parseSpec(rawPmSpec, import_path8.default.relative(initialCwd, selection.manifestPath))
   };
 }
 
@@ -22344,7 +22868,7 @@ var Engine = class {
    *
    * If the project doesn't include a specification file, we just assume that
    * whatever the user uses is exactly what they want to use. Since the version
-   * isn't explicited, we fallback on known good versions.
+   * isn't specified, we fallback on known good versions.
    *
    * Finally, if the project doesn't exist at all, we ask the user whether they
    * want to create one in the current project. If they do, we initialize a new
@@ -22353,17 +22877,25 @@ var Engine = class {
    */
   async findProjectSpec(initialCwd, locator, { transparent = false } = {}) {
     const fallbackDescriptor = { name: locator.name, range: `${locator.reference}` };
-    if (import_process3.default.env.COREPACK_ENABLE_PROJECT_SPEC === `0`)
+    if (import_process3.default.env.COREPACK_ENABLE_PROJECT_SPEC === `0`) {
+      if (typeof locator.reference === `function`)
+        fallbackDescriptor.range = await locator.reference();
       return fallbackDescriptor;
+    }
     if (import_process3.default.env.COREPACK_ENABLE_STRICT === `0`)
       transparent = true;
     while (true) {
       const result = await loadSpec(initialCwd);
       switch (result.type) {
-        case `NoProject`:
+        case `NoProject`: {
+          if (typeof locator.reference === `function`)
+            fallbackDescriptor.range = await locator.reference();
           log(`Falling back to ${fallbackDescriptor.name}@${fallbackDescriptor.range} as no project manifest were found`);
           return fallbackDescriptor;
+        }
         case `NoSpec`: {
+          if (typeof locator.reference === `function`)
+            fallbackDescriptor.range = await locator.reference();
           if (import_process3.default.env.COREPACK_ENABLE_AUTO_PIN !== `0`) {
             const resolved = await this.resolveDescriptor(fallbackDescriptor, { allowTags: true });
             if (resolved === null)
@@ -22378,16 +22910,19 @@ var Engine = class {
           return fallbackDescriptor;
         }
         case `Found`: {
-          if (result.spec.name !== locator.name) {
+          const spec = result.getSpec();
+          if (spec.name !== locator.name) {
             if (transparent) {
-              log(`Falling back to ${fallbackDescriptor.name}@${fallbackDescriptor.range} in a ${result.spec.name}@${result.spec.range} project`);
+              if (typeof locator.reference === `function`)
+                fallbackDescriptor.range = await locator.reference();
+              log(`Falling back to ${fallbackDescriptor.name}@${fallbackDescriptor.range} in a ${spec.name}@${spec.range} project`);
               return fallbackDescriptor;
             } else {
-              throw new UsageError(`This project is configured to use ${result.spec.name} because ${result.target} has a "packageManager" field`);
+              throw new UsageError(`This project is configured to use ${spec.name} because ${result.target} has a "packageManager" field`);
             }
           } else {
-            log(`Using ${result.spec.name}@${result.spec.range} as defined in project manifest ${result.target}`);
-            return result.spec;
+            log(`Using ${spec.name}@${spec.range} as defined in project manifest ${result.target}`);
+            return spec;
           }
         }
       }
@@ -22400,7 +22935,7 @@ var Engine = class {
     };
     let isTransparentCommand = false;
     if (packageManager != null) {
-      const defaultVersion = binaryVersion || await this.getDefaultVersion(packageManager);
+      const defaultVersion = binaryVersion || (() => this.getDefaultVersion(packageManager));
       const definition = this.config.definitions[packageManager];
       for (const transparentPath of definition.transparent.commands) {
         if (transparentPath[0] === binaryName && transparentPath.slice(1).every((segment, index) => segment === args[index])) {
@@ -22436,7 +22971,7 @@ var Engine = class {
     if (typeof definition === `undefined`)
       throw new UsageError(`This package manager (${descriptor.name}) isn't supported by this corepack build`);
     let finalDescriptor = descriptor;
-    if (!(0, import_valid2.default)(descriptor.range) && !(0, import_valid3.default)(descriptor.range)) {
+    if (!(0, import_valid3.default)(descriptor.range) && !(0, import_valid4.default)(descriptor.range)) {
       if (!allowTags)
         throw new UsageError(`Packages managers can't be referenced via tags in this context`);
       const ranges = Object.keys(definition.ranges);
@@ -22454,7 +22989,7 @@ var Engine = class {
     const cachedVersion = await findInstalledVersion(getInstallFolder(), finalDescriptor);
     if (cachedVersion !== null && useCache)
       return { name: finalDescriptor.name, reference: cachedVersion };
-    if ((0, import_valid2.default)(finalDescriptor.range))
+    if ((0, import_valid3.default)(finalDescriptor.range))
       return { name: finalDescriptor.name, reference: finalDescriptor.range };
     const versions = await Promise.all(Object.keys(definition.ranges).map(async (range) => {
       const packageManagerSpec = definition.ranges[range];
@@ -22640,11 +23175,11 @@ var BaseCommand = class extends Command {
       const lookup = await loadSpec(this.context.cwd);
       switch (lookup.type) {
         case `NoProject`:
-          throw new UsageError(`Couldn't find a project in the local directory - please explicit the package manager to pack, or run this command from a valid project`);
+          throw new UsageError(`Couldn't find a project in the local directory - please specify the package manager to pack, or run this command from a valid project`);
         case `NoSpec`:
-          throw new UsageError(`The local project doesn't feature a 'packageManager' field - please explicit the package manager to pack, or update the manifest to reference it`);
+          throw new UsageError(`The local project doesn't feature a 'packageManager' field nor a 'devEngines.packageManager' field - please specify the package manager to pack, or update the manifest to reference it`);
         default: {
-          return [lookup.spec];
+          return [lookup.range ?? lookup.getSpec()];
         }
       }
     }
@@ -22853,8 +23388,8 @@ var PackCommand = class extends BaseCommand {
 
 // sources/commands/Up.ts
 var import_major = __toESM(require_major());
-var import_valid4 = __toESM(require_valid());
-var import_valid5 = __toESM(require_valid2());
+var import_valid5 = __toESM(require_valid());
+var import_valid6 = __toESM(require_valid2());
 var UpCommand = class extends BaseCommand {
   static paths = [
     [`up`]
@@ -22880,7 +23415,7 @@ var UpCommand = class extends BaseCommand {
     const [descriptor] = await this.resolvePatternsToDescriptors({
       patterns: []
     });
-    if (!(0, import_valid4.default)(descriptor.range) && !(0, import_valid5.default)(descriptor.range))
+    if (!(0, import_valid5.default)(descriptor.range) && !(0, import_valid6.default)(descriptor.range))
       throw new UsageError(`The 'corepack up' command can only be used when your project's packageManager field is set to a semver version or semver range`);
     const resolved = await this.context.engine.resolveDescriptor(descriptor, { useCache: false });
     if (!resolved)
@@ -23007,11 +23542,11 @@ var PrepareCommand = class extends Command {
       const lookup = await loadSpec(this.context.cwd);
       switch (lookup.type) {
         case `NoProject`:
-          throw new UsageError(`Couldn't find a project in the local directory - please explicit the package manager to pack, or run this command from a valid project`);
+          throw new UsageError(`Couldn't find a project in the local directory - please specify the package manager to pack, or run this command from a valid project`);
         case `NoSpec`:
-          throw new UsageError(`The local project doesn't feature a 'packageManager' field - please explicit the package manager to pack, or update the manifest to reference it`);
+          throw new UsageError(`The local project doesn't feature a 'packageManager' field - please specify the package manager to pack, or update the manifest to reference it`);
         default: {
-          specs.push(lookup.spec);
+          specs.push(lookup.getSpec());
         }
       }
     }
@@ -23074,6 +23609,9 @@ function getPackageManagerRequestFromCli(parameter, engine) {
     binaryVersion: binaryVersion || null
   };
 }
+function isUsageError(error) {
+  return error?.name === `UsageError`;
+}
 async function runMain(argv) {
   const engine = new Engine();
   const [firstArg, ...restArgs] = argv;
@@ -23112,7 +23650,7 @@ async function runMain(argv) {
         args: restArgs
       });
     } catch (error) {
-      if (error?.name === `UsageError`) {
+      if (isUsageError(error)) {
         console.error(error.message);
         process.exit(1);
       }
diff --git a/deps/corepack/package.json b/deps/corepack/package.json
index 91b95f31d77b54..75f5725328a605 100644
--- a/deps/corepack/package.json
+++ b/deps/corepack/package.json
@@ -1,6 +1,6 @@
 {
   "name": "corepack",
-  "version": "0.31.0",
+  "version": "0.32.0",
   "homepage": "https://github.com/nodejs/corepack#readme",
   "bugs": {
     "url": "https://github.com/nodejs/corepack/issues"
@@ -16,7 +16,7 @@
     "./package.json": "./package.json"
   },
   "license": "MIT",
-  "packageManager": "yarn@4.3.1+sha224.934d21773e22af4b69a7032a2d3b4cb38c1f7c019624777cc9916b23",
+  "packageManager": "yarn@4.6.0+sha512.5383cc12567a95f1d668fbe762dfe0075c595b4bfff433be478dbbe24e05251a8e8c3eb992a986667c1d53b6c3a9c85b8398c35a960587fbd9fa3a0915406728",
   "devDependencies": {
     "@types/debug": "^4.1.5",
     "@types/node": "^20.4.6",
@@ -29,18 +29,18 @@
     "better-sqlite3": "^11.7.2",
     "clipanion": "patch:clipanion@npm%3A3.2.1#~/.yarn/patches/clipanion-npm-3.2.1-fc9187f56c.patch",
     "debug": "^4.1.1",
-    "esbuild": "^0.21.0",
+    "esbuild": "^0.25.0",
     "eslint": "^8.57.0",
     "proxy-from-env": "^1.1.0",
     "semver": "^7.6.3",
-    "supports-color": "^9.0.0",
+    "supports-color": "^10.0.0",
     "tar": "^7.4.0",
     "tsx": "^4.16.2",
-    "typescript": "^5.3.3",
+    "typescript": "^5.7.3",
     "undici": "^6.19.2",
     "v8-compile-cache": "^2.3.0",
-    "vitest": "^2.0.3",
-    "which": "^4.0.0"
+    "vitest": "^3.0.5",
+    "which": "^5.0.0"
   },
   "resolutions": {
     "undici-types": "6.x"
diff --git a/deps/icu-small/LICENSE b/deps/icu-small/LICENSE
index 180db98fcc66ca..0b9efcd9092f97 100644
--- a/deps/icu-small/LICENSE
+++ b/deps/icu-small/LICENSE
@@ -2,7 +2,7 @@ UNICODE LICENSE V3
 
 COPYRIGHT AND PERMISSION NOTICE
 
-Copyright © 2016-2024 Unicode, Inc.
+Copyright © 2016-2025 Unicode, Inc.
 
 NOTICE TO USER: Carefully read the following legal agreement. BY
 DOWNLOADING, INSTALLING, COPYING OR OTHERWISE USING DATA FILES, AND/OR
diff --git a/deps/icu-small/README-FULL-ICU.txt b/deps/icu-small/README-FULL-ICU.txt
index 46a34687d51b3e..ee3fdf50b5e59e 100644
--- a/deps/icu-small/README-FULL-ICU.txt
+++ b/deps/icu-small/README-FULL-ICU.txt
@@ -1,8 +1,8 @@
 ICU sources - auto generated by shrink-icu-src.py
 
 This directory contains the ICU subset used by --with-intl=full-icu
-It is a strict subset of ICU 76 source files with the following exception(s):
-* deps/icu-small/source/data/in/icudt76l.dat.bz2 : compressed data file
+It is a strict subset of ICU 77 source files with the following exception(s):
+* deps/icu-small/source/data/in/icudt77l.dat.bz2 : compressed data file
 
 
 To rebuild this directory, see ../../tools/icu/README.md
diff --git a/deps/icu-small/source/common/brkiter.cpp b/deps/icu-small/source/common/brkiter.cpp
index 4d945cc17e2bb6..44a13ee6a2acd1 100644
--- a/deps/icu-small/source/common/brkiter.cpp
+++ b/deps/icu-small/source/common/brkiter.cpp
@@ -59,7 +59,7 @@ BreakIterator::buildInstance(const Locale& loc, const char *type, UErrorCode &st
 {
     char fnbuff[256];
     char ext[4]={'\0'};
-    CharString actualLocale;
+    CharString actual;
     int32_t size;
     const char16_t* brkfname = nullptr;
     UResourceBundle brkRulesStack;
@@ -94,7 +94,7 @@ BreakIterator::buildInstance(const Locale& loc, const char *type, UErrorCode &st
 
         // Use the string if we found it
         if (U_SUCCESS(status) && brkfname) {
-            actualLocale.append(ures_getLocaleInternal(brkName, &status), -1, status);
+            actual.append(ures_getLocaleInternal(brkName, &status), -1, status);
 
             char16_t* extStart=u_strchr(brkfname, 0x002e);
             int len = 0;
@@ -123,10 +123,9 @@ BreakIterator::buildInstance(const Locale& loc, const char *type, UErrorCode &st
     if (U_SUCCESS(status) && result != nullptr) {
         U_LOCALE_BASED(locBased, *(BreakIterator*)result);
 
-        locBased.setLocaleIDs(ures_getLocaleByType(b, ULOC_VALID_LOCALE, &status), 
-                              actualLocale.data());
-        uprv_strncpy(result->requestLocale, loc.getName(), ULOC_FULLNAME_CAPACITY);
-        result->requestLocale[ULOC_FULLNAME_CAPACITY-1] = 0; // always terminate
+        locBased.setLocaleIDs(ures_getLocaleByType(b, ULOC_VALID_LOCALE, &status),
+                              actual.data(), status);
+        LocaleBased::setLocaleID(loc.getName(), result->requestLocale, status);
     }
 
     ures_close(b);
@@ -206,26 +205,32 @@ BreakIterator::getAvailableLocales(int32_t& count)
 
 BreakIterator::BreakIterator()
 {
-    *validLocale = *actualLocale = *requestLocale = 0;
 }
 
 BreakIterator::BreakIterator(const BreakIterator &other) : UObject(other) {
-    uprv_strncpy(actualLocale, other.actualLocale, sizeof(actualLocale));
-    uprv_strncpy(validLocale, other.validLocale, sizeof(validLocale));
-    uprv_strncpy(requestLocale, other.requestLocale, sizeof(requestLocale));
+    UErrorCode status = U_ZERO_ERROR;
+    U_LOCALE_BASED(locBased, *this);
+    locBased.setLocaleIDs(other.validLocale, other.actualLocale, status);
+    LocaleBased::setLocaleID(other.requestLocale, requestLocale, status);
+    U_ASSERT(U_SUCCESS(status));
 }
 
 BreakIterator &BreakIterator::operator =(const BreakIterator &other) {
     if (this != &other) {
-        uprv_strncpy(actualLocale, other.actualLocale, sizeof(actualLocale));
-        uprv_strncpy(validLocale, other.validLocale, sizeof(validLocale));
-        uprv_strncpy(requestLocale, other.requestLocale, sizeof(requestLocale));
+        UErrorCode status = U_ZERO_ERROR;
+        U_LOCALE_BASED(locBased, *this);
+        locBased.setLocaleIDs(other.validLocale, other.actualLocale, status);
+        LocaleBased::setLocaleID(other.requestLocale, requestLocale, status);
+        U_ASSERT(U_SUCCESS(status));
     }
     return *this;
 }
 
 BreakIterator::~BreakIterator()
 {
+    delete validLocale;
+    delete actualLocale;
+    delete requestLocale;
 }
 
 // ------------------------------------------
@@ -394,7 +399,7 @@ BreakIterator::createInstance(const Locale& loc, int32_t kind, UErrorCode& statu
         // revisit this in ICU 3.0 and clean it up/fix it/remove it.
         if (U_SUCCESS(status) && (result != nullptr) && *actualLoc.getName() != 0) {
             U_LOCALE_BASED(locBased, *result);
-            locBased.setLocaleIDs(actualLoc.getName(), actualLoc.getName());
+            locBased.setLocaleIDs(actualLoc.getName(), actualLoc.getName(), status);
         }
         return result;
     }
@@ -488,6 +493,7 @@ BreakIterator::makeInstance(const Locale& loc, int32_t kind, UErrorCode& status)
     }
 
     if (U_FAILURE(status)) {
+        delete result;
         return nullptr;
     }
 
@@ -496,20 +502,25 @@ BreakIterator::makeInstance(const Locale& loc, int32_t kind, UErrorCode& status)
 
 Locale
 BreakIterator::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
+    if (U_FAILURE(status)) {
+        return Locale::getRoot();
+    }
     if (type == ULOC_REQUESTED_LOCALE) {
-        return {requestLocale};
+        return requestLocale == nullptr ?
+            Locale::getRoot() : Locale(requestLocale->data());
     }
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocale(type, status);
+    return LocaleBased::getLocale(validLocale, actualLocale, type, status);
 }
 
 const char *
 BreakIterator::getLocaleID(ULocDataLocaleType type, UErrorCode& status) const {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
     if (type == ULOC_REQUESTED_LOCALE) {
-        return requestLocale;
+        return requestLocale == nullptr ?  "" : requestLocale->data();
     }
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocaleID(type, status);
+    return LocaleBased::getLocaleID(validLocale, actualLocale, type, status);
 }
 
 
@@ -536,8 +547,10 @@ int32_t BreakIterator::getRuleStatusVec(int32_t *fillInVec, int32_t capacity, UE
 }
 
 BreakIterator::BreakIterator (const Locale& valid, const Locale& actual) {
+  UErrorCode status = U_ZERO_ERROR;
   U_LOCALE_BASED(locBased, (*this));
-  locBased.setLocaleIDs(valid, actual);
+  locBased.setLocaleIDs(valid.getName(), actual.getName(), status);
+  U_ASSERT(U_SUCCESS(status));
 }
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/common/charstr.cpp b/deps/icu-small/source/common/charstr.cpp
index f76cc8a4dc90f3..dadc829b0b5605 100644
--- a/deps/icu-small/source/common/charstr.cpp
+++ b/deps/icu-small/source/common/charstr.cpp
@@ -70,6 +70,15 @@ CharString &CharString::copyFrom(const CharString &s, UErrorCode &errorCode) {
     return *this;
 }
 
+CharString &CharString::copyFrom(StringPiece s, UErrorCode &errorCode) {
+    if (U_FAILURE(errorCode)) {
+        return *this;
+    }
+    len = 0;
+    append(s, errorCode);
+    return *this;
+}
+
 int32_t CharString::lastIndexOf(char c) const {
     for(int32_t i=len; i>0;) {
         if(buffer[--i]==c) {
@@ -143,7 +152,7 @@ CharString &CharString::append(const char *s, int32_t sLength, UErrorCode &error
     return *this;
 }
 
-CharString &CharString::appendNumber(int32_t number, UErrorCode &status) {
+CharString &CharString::appendNumber(int64_t number, UErrorCode &status) {
     if (number < 0) {
         this->append('-', status);
         if (U_FAILURE(status)) {
diff --git a/deps/icu-small/source/common/charstr.h b/deps/icu-small/source/common/charstr.h
index 08283ca452ce2b..ea54ede735cd3d 100644
--- a/deps/icu-small/source/common/charstr.h
+++ b/deps/icu-small/source/common/charstr.h
@@ -74,6 +74,7 @@ class U_COMMON_API CharString : public UMemory {
      * use a UErrorCode where memory allocations might be needed.
      */
     CharString &copyFrom(const CharString &other, UErrorCode &errorCode);
+    CharString &copyFrom(StringPiece s, UErrorCode &errorCode);
 
     UBool isEmpty() const { return len==0; }
     int32_t length() const { return len; }
@@ -135,7 +136,7 @@ class U_COMMON_API CharString : public UMemory {
     }
     CharString &append(const char *s, int32_t sLength, UErrorCode &status);
 
-    CharString &appendNumber(int32_t number, UErrorCode &status);
+    CharString &appendNumber(int64_t number, UErrorCode &status);
 
     /**
      * Returns a writable buffer for appending and writes the buffer's capacity to
diff --git a/deps/icu-small/source/common/localefallback_data.h b/deps/icu-small/source/common/localefallback_data.h
index 0accf0324d7eb2..3b8ad8a3f398b8 100644
--- a/deps/icu-small/source/common/localefallback_data.h
+++ b/deps/icu-small/source/common/localefallback_data.h
@@ -11,11 +11,11 @@
 //======================================================================
 // Default script table
 const char scriptCodeChars[] =
-    "Aghb\0Ahom\0Arab\0Armi\0Armn\0Avst\0Bamu\0Bass\0Batk\0Beng\0Bopo\0"
-    "Brah\0Cakm\0Cans\0Cari\0Cham\0Cher\0Chrs\0Copt\0Cprt\0Cyrl\0Deva\0"
-    "Egyp\0Elym\0Ethi\0Geor\0Gong\0Gonm\0Goth\0Gran\0Grek\0Gujr\0Guru\0"
-    "Hang\0Hani\0Hans\0Hant\0Hebr\0Hluw\0Hmnp\0Ital\0Java\0Jpan\0Kali\0"
-    "Kana\0Kawi\0Khar\0Khmr\0Kits\0Knda\0Kore\0Lana\0Laoo\0Latf\0Latg\0"
+    "Aghb\0Ahom\0Arab\0Armi\0Armn\0Avst\0Bali\0Bamu\0Bass\0Batk\0Beng\0"
+    "Bopo\0Brah\0Cakm\0Cans\0Cari\0Cham\0Cher\0Chrs\0Copt\0Cprt\0Cyrl\0"
+    "Deva\0Egyp\0Elym\0Ethi\0Geor\0Gong\0Gonm\0Goth\0Gran\0Grek\0Gujr\0"
+    "Guru\0Hang\0Hani\0Hans\0Hant\0Hebr\0Hluw\0Hmnp\0Ital\0Java\0Jpan\0"
+    "Kali\0Kana\0Khar\0Khmr\0Kits\0Knda\0Kore\0Lana\0Laoo\0Latf\0Latg\0"
     "Lepc\0Lina\0Linb\0Lisu\0Lyci\0Lydi\0Mand\0Mani\0Marc\0Medf\0Merc\0"
     "Mlym\0Modi\0Mong\0Mroo\0Mtei\0Mymr\0Narb\0Newa\0Nkoo\0Nshu\0Ogam\0"
     "Olck\0Orkh\0Orya\0Osge\0Ougr\0Pauc\0Phli\0Phnx\0Plrd\0Prti\0Rjng\0"
@@ -48,70 +48,71 @@ const char dsLocaleIDChars[] =
     "gbz\0gdb\0gdo\0gdx\0gez\0ggg\0gha\0ghe\0gho\0ghr\0ght\0gig\0gin\0"
     "gjk\0gju\0gld\0glh\0glk\0gml\0gmv\0gmy\0goe\0gof\0goj\0gok\0gon\0"
     "got\0gra\0grc\0grt\0gru\0gu\0gvr\0gwc\0gwf\0gwt\0gyo\0gzi\0ha_CM\0"
-    "ha_SD\0hac\0hak\0har\0haz\0hbo\0hdy\0he\0hi\0hif\0hii\0hit\0hkh\0"
-    "hlb\0hlu\0hmd\0hmj\0hmq\0hnd\0hne\0hnj\0hno\0hoc\0hoh\0hoj\0how\0"
-    "hoy\0hpo\0hrt\0hrz\0hsn\0hss\0htx\0hut\0huy\0huz\0hy\0hyw\0ii\0"
-    "imy\0inh\0int\0ior\0iru\0isk\0itk\0itl\0iu\0iw\0ja\0jad\0jat\0"
-    "jbe\0jbn\0jct\0jda\0jdg\0jdt\0jee\0jge\0ji\0jje\0jkm\0jml\0jna\0"
-    "jnd\0jnl\0jns\0jog\0jpa\0jpr\0jrb\0jul\0jun\0juy\0jya\0jye\0ka\0"
-    "kaa\0kap\0kaw\0kbd\0kbg\0kbu\0kby\0kca\0kcy\0kdq\0kdt\0ket\0kev\0"
-    "kex\0key\0kfa\0kfb\0kfc\0kfd\0kfe\0kfg\0kfh\0kfi\0kfk\0kfm\0kfp\0"
-    "kfq\0kfr\0kfs\0kfu\0kfx\0kfy\0kgj\0kgy\0khb\0khf\0khg\0khn\0kho\0"
-    "kht\0khv\0khw\0kif\0kim\0kip\0kjg\0kjh\0kjl\0kjo\0kjp\0kjt\0kjz\0"
-    "kk\0kk_AF\0kk_CN\0kk_IR\0kk_MN\0kkf\0kkh\0kkt\0kle\0klj\0klr\0"
-    "km\0kmj\0kmz\0kn\0knn\0ko\0koi\0kok\0kpt\0kpy\0kqd\0kqy\0kra\0"
-    "krc\0krk\0krr\0kru\0krv\0ks\0ksu\0ksw\0ksz\0ktb\0kte\0ktl\0ktp\0"
-    "ku_LB\0kuf\0kum\0kv\0kva\0kvq\0kvt\0kvx\0kvy\0kxf\0kxk\0kxm\0"
-    "kxp\0ky\0ky_CN\0kyu\0kyv\0kyw\0lab\0lad\0lae\0lah\0lbe\0lbf\0"
-    "lbj\0lbm\0lbo\0lbr\0lcp\0lep\0lez\0lhm\0lhs\0lif\0lis\0lkh\0lki\0"
-    "lmh\0lmn\0lo\0loy\0lpo\0lrc\0lrk\0lrl\0lsa\0lsd\0lss\0ltc\0luk\0"
-    "luu\0luv\0luz\0lwl\0lwm\0lya\0lzh\0mag\0mai\0man_GN\0mby\0mde\0"
-    "mdf\0mdx\0mdy\0mfa\0mfi\0mga\0mgp\0mhj\0mid\0mjl\0mjq\0mjr\0mjt\0"
-    "mju\0mjv\0mjz\0mk\0mkb\0mke\0mki\0mkm\0ml\0mlf\0mn\0mn_CN\0mnc\0"
-    "mni\0mnj\0mns\0mnw\0mpz\0mr\0mra\0mrd\0mrj\0mro\0mrr\0ms_CC\0"
-    "mtm\0mtr\0mud\0muk\0mut\0muv\0muz\0mve\0mvf\0mvy\0mvz\0mwr\0mwt\0"
-    "mww\0my\0mym\0myv\0myz\0mzn\0nan\0nao\0ncd\0ncq\0ndf\0ne\0neg\0"
-    "neh\0nei\0new\0ngt\0nio\0nit\0niv\0nli\0nlm\0nlx\0nmm\0nnp\0nod\0"
-    "noe\0nog\0noi\0non\0nos\0npb\0nqo\0nrn\0nsd\0nsf\0nsk\0nst\0nsv\0"
-    "nty\0ntz\0nwc\0nwx\0nyl\0nyq\0nyw\0oaa\0oac\0oar\0oav\0obm\0obr\0"
-    "odk\0oht\0oj\0ojs\0okm\0oko\0okz\0ola\0ole\0omk\0omp\0omr\0omx\0"
-    "oon\0or\0ort\0oru\0orv\0os\0osa\0osc\0osi\0ota\0otb\0otk\0oty\0"
-    "oui\0pa\0pa_PK\0pal\0paq\0pbt\0pcb\0pce\0pcf\0pcg\0pch\0pci\0"
-    "pcj\0peg\0peo\0pgd\0pgg\0pgl\0pgn\0phd\0phk\0phl\0phn\0pho\0phr\0"
-    "pht\0phu\0phv\0phw\0pi\0pka\0pkr\0plk\0pll\0pmh\0pnt\0pra\0prc\0"
-    "prd\0prt\0prx\0ps\0psh\0psi\0pst\0psu\0pum\0pwo\0pwr\0pww\0pyx\0"
-    "qxq\0raa\0rab\0raf\0rah\0raj\0rav\0rbb\0rdb\0rei\0rhg\0rji\0rjs\0"
-    "rka\0rki\0rkt\0rmi\0rmt\0rmz\0rsk\0rtw\0ru\0rue\0rut\0rwr\0ryu\0"
-    "sa\0sah\0sam\0sat\0saz\0sbn\0sbu\0sck\0scl\0scp\0sct\0scu\0scx\0"
-    "sd\0sd_IN\0sdb\0sdf\0sdg\0sdh\0sdr\0sds\0sel\0sfm\0sga\0sgh\0"
-    "sgj\0sgr\0sgt\0sgw\0sgy\0shd\0shi\0shm\0shn\0shu\0shv\0si\0sia\0"
-    "sip\0siy\0siz\0sjd\0sjp\0sjt\0skb\0skj\0skr\0smh\0smp\0smu\0smy\0"
-    "soa\0sog\0soi\0sou\0spt\0spv\0sqo\0sqq\0sqt\0sr\0srb\0srh\0srx\0"
-    "srz\0ssh\0sss\0sts\0stv\0sty\0suz\0sva\0swb\0swi\0swv\0sxu\0syc\0"
-    "syl\0syn\0syr\0syw\0ta\0tab\0taj\0tbk\0tcn\0tco\0tcx\0tcy\0tda\0"
-    "tdb\0tdd\0tdg\0tdh\0te\0tes\0tg\0tg_PK\0tge\0tgf\0th\0the\0thf\0"
-    "thi\0thl\0thm\0thq\0thr\0ths\0ti\0tig\0tij\0tin\0tjl\0tjo\0tkb\0"
-    "tks\0tkt\0tmr\0tnv\0tov\0tpu\0tra\0trg\0trm\0trw\0tsd\0tsj\0tt\0"
-    "tth\0tto\0tts\0ttz\0tvn\0twm\0txg\0txo\0tyr\0tyv\0ude\0udg\0udi\0"
-    "udm\0ug\0ug_KZ\0ug_MN\0uga\0ugh\0ugo\0uk\0uki\0ulc\0unr\0unr_NP\0"
-    "unx\0ur\0urk\0ush\0uum\0uz_AF\0uz_CN\0uzs\0vaa\0vaf\0vah\0vai\0"
-    "vas\0vav\0vay\0vgr\0vjk\0vmd\0vmh\0wal\0wbk\0wbq\0wbr\0wle\0wlo\0"
-    "wme\0wne\0wni\0wsg\0wsv\0wtm\0wuu\0xag\0xal\0xan\0xas\0xco\0xcr\0"
-    "xdq\0xhe\0xhm\0xis\0xka\0xkc\0xkf\0xkj\0xkp\0xlc\0xld\0xly\0xmf\0"
-    "xmn\0xmr\0xna\0xnr\0xpg\0xpi\0xpm\0xpr\0xrm\0xrn\0xsa\0xsr\0xtq\0"
-    "xub\0xuj\0xve\0xvi\0xwo\0xzh\0yai\0ybh\0ybi\0ydg\0yea\0yej\0yeu\0"
-    "ygp\0yhd\0yi\0yig\0yih\0yiv\0ykg\0ykh\0yna\0ynk\0yoi\0yoy\0yrk\0"
-    "ysd\0ysn\0ysp\0ysr\0ysy\0yud\0yue\0yue_CN\0yug\0yux\0ywq\0ywu\0"
-    "zau\0zba\0zch\0zdj\0zeh\0zen\0zgb\0zgh\0zgm\0zgn\0zh\0zh_AU\0"
-    "zh_BN\0zh_GB\0zh_GF\0zh_HK\0zh_ID\0zh_MO\0zh_PA\0zh_PF\0zh_PH\0"
-    "zh_SR\0zh_TH\0zh_TW\0zh_US\0zh_VN\0zhd\0zhx\0zko\0zkt\0zkz\0zlj\0"
-    "zln\0zlq\0zqe\0zrg\0zrp\0zum\0zwa\0zyg\0zyn\0zzj\0";
+    "ha_SD\0hac\0hak\0hak_TW\0har\0haz\0hbo\0hdy\0he\0hi\0hif\0hii\0"
+    "hit\0hkh\0hlb\0hlu\0hmd\0hmj\0hmq\0hnd\0hne\0hnj\0hno\0hoc\0hoh\0"
+    "hoj\0how\0hoy\0hpo\0hrt\0hrz\0hsn\0hss\0htx\0hut\0huy\0huz\0hy\0"
+    "hyw\0ii\0imy\0inh\0int\0ior\0iru\0isk\0itk\0itl\0iu\0iw\0ja\0"
+    "jad\0jat\0jbe\0jbn\0jct\0jda\0jdg\0jdt\0jee\0jge\0ji\0jje\0jkm\0"
+    "jml\0jna\0jnd\0jnl\0jns\0jog\0jpa\0jpr\0jrb\0jul\0jun\0juy\0jya\0"
+    "jye\0ka\0kaa\0kap\0kaw\0kbd\0kbg\0kbu\0kby\0kca\0kcy\0kdq\0kdt\0"
+    "ket\0kev\0kex\0key\0kfa\0kfb\0kfc\0kfd\0kfe\0kfg\0kfh\0kfi\0kfk\0"
+    "kfm\0kfp\0kfq\0kfr\0kfs\0kfu\0kfx\0kfy\0kgj\0kgy\0khb\0khf\0khg\0"
+    "khn\0kho\0kht\0khv\0khw\0kif\0kim\0kip\0kjg\0kjh\0kjl\0kjo\0kjp\0"
+    "kjt\0kjz\0kk\0kk_AF\0kk_CN\0kk_IR\0kk_MN\0kkf\0kkh\0kkt\0kle\0"
+    "klj\0klr\0km\0kmj\0kmz\0kn\0knn\0ko\0koi\0kok\0kpt\0kpy\0kqd\0"
+    "kqy\0kra\0krc\0krk\0krr\0kru\0krv\0ks\0ksu\0ksw\0ksz\0ktb\0kte\0"
+    "ktl\0ktp\0ku_LB\0kuf\0kum\0kv\0kva\0kvq\0kvt\0kvx\0kvy\0kxf\0"
+    "kxk\0kxm\0kxp\0ky\0ky_CN\0kyu\0kyv\0kyw\0lab\0lad\0lae\0lah\0"
+    "lbe\0lbf\0lbj\0lbm\0lbo\0lbr\0lcp\0lep\0lez\0lhm\0lhs\0lif\0lis\0"
+    "lkh\0lki\0lmh\0lmn\0lo\0loy\0lpo\0lrc\0lrk\0lrl\0lsa\0lsd\0lss\0"
+    "ltc\0luk\0luu\0luv\0luz\0lwl\0lwm\0lya\0lzh\0lzz_GE\0mag\0mai\0"
+    "mby\0mde\0mdf\0mdx\0mdy\0mfa\0mfi\0mga\0mgp\0mhj\0mid\0mjl\0mjq\0"
+    "mjr\0mjt\0mju\0mjv\0mjz\0mk\0mkb\0mke\0mki\0mkm\0ml\0mlf\0mn\0"
+    "mn_CN\0mnc\0mni\0mnj\0mns\0mnw\0mpz\0mr\0mra\0mrd\0mrj\0mro\0"
+    "mrr\0ms_CC\0mtm\0mtr\0mud\0muk\0mut\0muv\0muz\0mve\0mvf\0mvy\0"
+    "mvz\0mwr\0mwt\0mww\0my\0mym\0myv\0myz\0mzn\0nan\0nan_TW\0nao\0"
+    "ncd\0ncq\0ndf\0ne\0neg\0neh\0nei\0new\0ngt\0nio\0nit\0niv\0nli\0"
+    "nlm\0nlx\0nmm\0nnp\0nod\0noe\0nog\0noi\0non\0nos\0npb\0nqo\0nrn\0"
+    "nsd\0nsf\0nsk\0nst\0nsv\0nty\0ntz\0nwc\0nwx\0nyl\0nyq\0nyw\0oaa\0"
+    "oac\0oar\0oav\0obm\0obr\0odk\0oht\0oj\0ojs\0okm\0oko\0okz\0ola\0"
+    "ole\0omk\0omp\0omr\0omx\0oon\0or\0ort\0oru\0orv\0os\0osa\0osc\0"
+    "osi\0ota\0otb\0otk\0oty\0oui\0pa\0pa_PK\0pal\0paq\0pbt\0pcb\0"
+    "pce\0pcf\0pcg\0pch\0pci\0pcj\0peg\0peo\0pgd\0pgg\0pgl\0pgn\0phd\0"
+    "phk\0phl\0phn\0pho\0phr\0pht\0phu\0phv\0phw\0pi\0pka\0pkr\0plk\0"
+    "pll\0pmh\0pnt\0pnt_RU\0pra\0prc\0prd\0prt\0prx\0ps\0psh\0psi\0"
+    "pst\0psu\0pum\0pwo\0pwr\0pww\0pyx\0qxq\0raa\0rab\0raf\0rah\0raj\0"
+    "rav\0rbb\0rdb\0rei\0rhg\0rji\0rjs\0rka\0rki\0rkt\0rmi\0rmt\0rmz\0"
+    "rsk\0rtw\0ru\0rue\0rut\0rwr\0ryu\0sa\0sah\0sam\0sat\0saz\0sbn\0"
+    "sbu\0sck\0scl\0scp\0sct\0scu\0scx\0sd\0sd_IN\0sdb\0sdf\0sdg\0"
+    "sdh\0sdr\0sds\0sel\0sfm\0sgh\0sgj\0sgr\0sgt\0sgw\0sgy\0shd\0shi\0"
+    "shm\0shn\0shu\0shv\0si\0sia\0sip\0siy\0siz\0sjd\0sjp\0sjt\0skb\0"
+    "skj\0skr\0smh\0smp\0smu\0smy\0soa\0sog\0soi\0sou\0spt\0spv\0sqo\0"
+    "sqq\0sqt\0sr\0srb\0srh\0srx\0srz\0ssh\0sss\0sts\0stv\0sty\0suz\0"
+    "sva\0swb\0swi\0swv\0sxu\0syc\0syl\0syn\0syr\0syw\0ta\0tab\0taj\0"
+    "tbk\0tcn\0tco\0tcx\0tcy\0tda\0tdb\0tdd\0tdg\0tdh\0te\0tes\0tg\0"
+    "tg_PK\0tge\0tgf\0th\0the\0thf\0thi\0thl\0thm\0thq\0thr\0ths\0"
+    "ti\0tig\0tij\0tin\0tjl\0tjo\0tkb\0tks\0tkt\0tmr\0tnv\0tov\0tpu\0"
+    "tra\0trg\0trm\0trw\0tsd\0tsj\0tt\0tth\0tto\0tts\0ttz\0tvn\0twm\0"
+    "txg\0txo\0tyr\0tyv\0ude\0udg\0udi\0udm\0ug\0ug_KZ\0ug_MN\0uga\0"
+    "ugh\0ugo\0uk\0uki\0ulc\0unr\0unr_NP\0unx\0ur\0urk\0ush\0uum\0"
+    "uz_AF\0uz_CN\0uzs\0vaa\0vaf\0vah\0vai\0vas\0vav\0vay\0vgr\0vjk\0"
+    "vmd\0vmh\0wal\0wbk\0wbq\0wbr\0wle\0wlo\0wme\0wne\0wni\0wsg\0wsv\0"
+    "wtm\0wuu\0xag\0xal\0xan\0xas\0xco\0xcr\0xdq\0xhe\0xhm\0xis\0xka\0"
+    "xkc\0xkf\0xkj\0xkp\0xlc\0xld\0xly\0xmf\0xmn\0xmr\0xna\0xnr\0xpg\0"
+    "xpi\0xpm\0xpr\0xrm\0xrn\0xsa\0xsr\0xtq\0xub\0xuj\0xve\0xvi\0xwo\0"
+    "xzh\0yai\0ybh\0ybi\0ydg\0yea\0yej\0yeu\0ygp\0yhd\0yi\0yig\0yih\0"
+    "yiv\0ykg\0ykh\0yna\0ynk\0yoi\0yoy\0yrk\0ysd\0ysn\0ysp\0ysr\0ysy\0"
+    "yud\0yue\0yue_CN\0yug\0yux\0ywq\0ywu\0zau\0zba\0zch\0zdj\0zeh\0"
+    "zen\0zgb\0zgh\0zgm\0zgn\0zh\0zh_AU\0zh_BN\0zh_GB\0zh_GF\0zh_HK\0"
+    "zh_ID\0zh_MO\0zh_PA\0zh_PF\0zh_PH\0zh_SR\0zh_TH\0zh_TW\0zh_US\0"
+    "zh_VN\0zhd\0zhx\0zko\0zkt\0zkz\0zlj\0zln\0zlq\0zqe\0zrg\0zrp\0"
+    "zum\0zwa\0zyg\0zyn\0zzj\0";
 
 const int32_t defaultScriptTable[] = {
     0, 330,  // aaf -> Mlym
     4, 10,  // aao -> Arab
-    8, 150,  // aat -> Grek
-    12, 100,  // ab -> Cyrl
+    8, 155,  // aat -> Grek
+    12, 105,  // ab -> Cyrl
     15, 10,  // abh -> Arab
     19, 435,  // abl -> Rjng
     23, 10,  // abv -> Arab
@@ -121,64 +122,64 @@ const int32_t defaultScriptTable[] = {
     39, 10,  // acx -> Arab
     43, 10,  // adf -> Arab
     47, 555,  // adx -> Tibt
-    51, 100,  // ady -> Cyrl
+    51, 105,  // ady -> Cyrl
     55, 25,  // ae -> Avst
     58, 10,  // aeb -> Arab
     62, 10,  // aec -> Arab
     66, 10,  // aee -> Arab
     70, 10,  // aeq -> Arab
     74, 10,  // afb -> Arab
-    78, 105,  // agi -> Deva
-    82, 120,  // agj -> Ethi
-    86, 100,  // agx -> Cyrl
-    90, 120,  // ahg -> Ethi
+    78, 110,  // agi -> Deva
+    82, 125,  // agj -> Ethi
+    86, 105,  // agx -> Cyrl
+    90, 125,  // ahg -> Ethi
     94, 5,  // aho -> Ahom
-    98, 105,  // ahr -> Deva
+    98, 110,  // ahr -> Deva
     102, 10,  // aib -> Arab
     106, 495,  // aii -> Syrc
-    110, 185,  // aij -> Hebr
-    114, 220,  // ain -> Kana
+    110, 190,  // aij -> Hebr
+    114, 225,  // ain -> Kana
     118, 355,  // aio -> Mymr
     122, 10,  // aiq -> Arab
     126, 590,  // akk -> Xsux
-    130, 100,  // akv -> Cyrl
+    130, 105,  // akv -> Cyrl
     134, 260,  // alk -> Laoo
     138, 330,  // all -> Mlym
-    142, 100,  // alr -> Cyrl
-    146, 100,  // alt -> Cyrl
-    150, 120,  // alw -> Ethi
-    154, 120,  // am -> Ethi
-    157, 210,  // ams -> Jpan
+    142, 105,  // alr -> Cyrl
+    146, 105,  // alt -> Cyrl
+    150, 125,  // alw -> Ethi
+    154, 125,  // am -> Ethi
+    157, 215,  // ams -> Jpan
     161, 495,  // amw -> Syrc
-    165, 100,  // ani -> Cyrl
-    169, 105,  // anp -> Deva
-    173, 105,  // anq -> Deva
-    177, 105,  // anr -> Deva
-    181, 120,  // anu -> Ethi
-    185, 45,  // aot -> Beng
+    165, 105,  // ani -> Cyrl
+    169, 110,  // anp -> Deva
+    173, 110,  // anq -> Deva
+    177, 110,  // anr -> Deva
+    181, 125,  // anu -> Ethi
+    185, 50,  // aot -> Beng
     189, 10,  // apc -> Arab
     193, 10,  // apd -> Arab
-    197, 105,  // aph -> Deva
-    201, 100,  // aqc -> Cyrl
+    197, 110,  // aph -> Deva
+    201, 105,  // aqc -> Cyrl
     205, 10,  // ar -> Arab
     208, 15,  // arc -> Armi
     212, 10,  // arq -> Arab
     216, 10,  // ars -> Arab
     220, 10,  // ary -> Arab
     224, 10,  // arz -> Arab
-    228, 45,  // as -> Beng
+    228, 50,  // as -> Beng
     231, 465,  // ase -> Sgnw
     235, 10,  // ask -> Arab
-    239, 105,  // asr -> Deva
+    239, 110,  // asr -> Deva
     243, 10,  // atn -> Arab
-    247, 100,  // atv -> Cyrl
+    247, 105,  // atv -> Cyrl
     251, 10,  // auj -> Arab
     255, 10,  // auz -> Arab
-    259, 100,  // av -> Cyrl
+    259, 105,  // av -> Cyrl
     262, 10,  // avd -> Arab
     266, 10,  // avl -> Arab
-    270, 105,  // awa -> Deva
-    274, 120,  // awn -> Ethi
+    270, 110,  // awa -> Deva
+    274, 125,  // awn -> Ethi
     278, 20,  // axm -> Armn
     282, 10,  // ayh -> Arab
     286, 10,  // ayl -> Arab
@@ -186,971 +187,973 @@ const int32_t defaultScriptTable[] = {
     294, 10,  // ayp -> Arab
     298, 10,  // az_IQ -> Arab
     304, 10,  // az_IR -> Arab
-    310, 100,  // az_RU -> Cyrl
+    310, 105,  // az_RU -> Cyrl
     316, 10,  // azb -> Arab
-    320, 100,  // ba -> Cyrl
+    320, 105,  // ba -> Cyrl
     323, 10,  // bal -> Arab
-    327, 105,  // bap -> Deva
-    331, 30,  // bax -> Bamu
-    335, 125,  // bbl -> Geor
-    339, 120,  // bcq -> Ethi
+    327, 110,  // bap -> Deva
+    331, 35,  // bax -> Bamu
+    335, 130,  // bbl -> Geor
+    339, 125,  // bcq -> Ethi
     343, 395,  // bdv -> Orya
     347, 10,  // bdz -> Arab
-    351, 100,  // be -> Cyrl
-    354, 105,  // bee -> Deva
+    351, 105,  // be -> Cyrl
+    354, 110,  // bee -> Deva
     358, 10,  // bej -> Arab
-    362, 105,  // bfb -> Deva
+    362, 110,  // bfb -> Deva
     366, 520,  // bfq -> Taml
     370, 10,  // bft -> Arab
     374, 555,  // bfu -> Tibt
     378, 395,  // bfw -> Orya
-    382, 105,  // bfy -> Deva
-    386, 105,  // bfz -> Deva
-    390, 100,  // bg -> Cyrl
-    393, 105,  // bgc -> Deva
-    397, 105,  // bgd -> Deva
+    382, 110,  // bfy -> Deva
+    386, 110,  // bfz -> Deva
+    390, 105,  // bg -> Cyrl
+    393, 110,  // bgc -> Deva
+    397, 110,  // bgd -> Deva
     401, 10,  // bgn -> Arab
     405, 10,  // bgp -> Arab
-    409, 105,  // bgq -> Deva
-    413, 105,  // bgw -> Deva
-    417, 150,  // bgx -> Grek
-    421, 105,  // bha -> Deva
-    425, 105,  // bhb -> Deva
-    429, 105,  // bhd -> Deva
+    409, 110,  // bgq -> Deva
+    413, 110,  // bgw -> Deva
+    417, 155,  // bgx -> Grek
+    421, 110,  // bha -> Deva
+    425, 110,  // bhb -> Deva
+    429, 110,  // bhd -> Deva
     433, 10,  // bhe -> Arab
-    437, 100,  // bhh -> Cyrl
-    441, 105,  // bhi -> Deva
-    445, 105,  // bhj -> Deva
+    437, 105,  // bhh -> Cyrl
+    441, 110,  // bhi -> Deva
+    445, 110,  // bhj -> Deva
     449, 10,  // bhm -> Arab
     453, 495,  // bhn -> Syrc
-    457, 105,  // bho -> Deva
-    461, 105,  // bht -> Deva
-    465, 105,  // bhu -> Deva
-    469, 105,  // biy -> Deva
+    457, 110,  // bho -> Deva
+    461, 110,  // bht -> Deva
+    465, 110,  // bhu -> Deva
+    469, 110,  // biy -> Deva
     473, 495,  // bjf -> Syrc
-    477, 105,  // bjj -> Deva
+    477, 110,  // bjj -> Deva
     481, 10,  // bjm -> Arab
     485, 555,  // bkk -> Tibt
     489, 355,  // blk -> Mymr
     493, 530,  // blt -> Tavt
-    497, 105,  // bmj -> Deva
-    501, 45,  // bn -> Beng
-    504, 105,  // bns -> Deva
+    497, 110,  // bmj -> Deva
+    501, 50,  // bn -> Beng
+    504, 110,  // bns -> Deva
     508, 555,  // bo -> Tibt
-    511, 100,  // bph -> Cyrl
-    515, 105,  // bpx -> Deva
-    519, 45,  // bpy -> Beng
+    511, 105,  // bph -> Cyrl
+    515, 110,  // bpx -> Deva
+    519, 50,  // bpy -> Beng
     523, 10,  // bqi -> Arab
-    527, 105,  // bra -> Deva
+    527, 110,  // bra -> Deva
     531, 235,  // brb -> Khmr
-    535, 105,  // brd -> Deva
+    535, 110,  // brd -> Deva
     539, 10,  // brh -> Arab
     543, 10,  // brk -> Arab
     547, 555,  // bro -> Tibt
     551, 260,  // brv -> Laoo
     555, 245,  // brw -> Knda
-    559, 105,  // brx -> Deva
+    559, 110,  // brx -> Deva
     563, 10,  // bsh -> Arab
     567, 10,  // bsk -> Arab
-    571, 35,  // bsq -> Bass
-    575, 120,  // bst -> Ethi
-    579, 40,  // btd -> Batk
-    583, 40,  // btm -> Batk
-    587, 105,  // btv -> Deva
-    591, 100,  // bua -> Cyrl
+    571, 40,  // bsq -> Bass
+    575, 125,  // bst -> Ethi
+    579, 45,  // btd -> Batk
+    583, 45,  // btm -> Batk
+    587, 110,  // btv -> Deva
+    591, 105,  // bua -> Cyrl
     595, 355,  // bwe -> Mymr
-    599, 100,  // bxm -> Cyrl
+    599, 105,  // bxm -> Cyrl
     603, 340,  // bxu -> Mong
-    607, 105,  // byh -> Deva
-    611, 120,  // byn -> Ethi
-    615, 105,  // byw -> Deva
+    607, 110,  // byh -> Deva
+    611, 125,  // byn -> Ethi
+    615, 110,  // byw -> Deva
     619, 550,  // bzi -> Thai
     623, 550,  // cbn -> Thai
-    627, 60,  // ccp -> Cakm
+    627, 65,  // ccp -> Cakm
     631, 535,  // cde -> Telu
-    635, 105,  // cdh -> Deva
-    639, 155,  // cdi -> Gujr
-    643, 105,  // cdj -> Deva
-    647, 105,  // cdm -> Deva
-    651, 175,  // cdo -> Hans
-    655, 45,  // cdz -> Beng
-    659, 100,  // ce -> Cyrl
+    635, 110,  // cdh -> Deva
+    639, 160,  // cdi -> Gujr
+    643, 110,  // cdj -> Deva
+    647, 110,  // cdm -> Deva
+    651, 180,  // cdo -> Hans
+    655, 50,  // cdz -> Beng
+    659, 105,  // ce -> Cyrl
     662, 555,  // cgk -> Tibt
     666, 10,  // chg -> Arab
-    670, 100,  // chm -> Cyrl
-    674, 80,  // chr -> Cher
-    678, 105,  // chx -> Deva
-    682, 105,  // cih -> Deva
+    670, 105,  // chm -> Cyrl
+    674, 85,  // chr -> Cher
+    678, 110,  // chx -> Deva
+    682, 110,  // cih -> Deva
     686, 10,  // cja -> Arab
-    690, 100,  // cji -> Cyrl
-    694, 75,  // cjm -> Cham
-    698, 175,  // cjy -> Hans
+    690, 105,  // cji -> Cyrl
+    694, 80,  // cjm -> Cham
+    698, 180,  // cjy -> Hans
     702, 10,  // ckb -> Arab
-    706, 100,  // ckt -> Cyrl
+    706, 105,  // ckt -> Cyrl
     710, 10,  // clh -> Arab
-    714, 100,  // clw -> Cyrl
+    714, 105,  // clw -> Cyrl
     718, 485,  // cmg -> Soyo
     722, 555,  // cna -> Tibt
-    726, 175,  // cnp -> Hans
+    726, 180,  // cnp -> Hans
     730, 550,  // cog -> Thai
-    734, 90,  // cop -> Copt
-    738, 150,  // cpg -> Grek
-    742, 65,  // cr -> Cans
-    745, 100,  // crh -> Cyrl
-    749, 65,  // crj -> Cans
-    753, 65,  // crk -> Cans
-    757, 65,  // crl -> Cans
-    761, 65,  // crm -> Cans
+    734, 95,  // cop -> Copt
+    738, 155,  // cpg -> Grek
+    742, 70,  // cr -> Cans
+    745, 105,  // crh -> Cyrl
+    749, 70,  // crj -> Cans
+    753, 70,  // crk -> Cans
+    757, 70,  // crl -> Cans
+    761, 70,  // crm -> Cans
     765, 355,  // csh -> Mymr
-    769, 175,  // csp -> Hans
-    773, 65,  // csw -> Cans
+    769, 180,  // csp -> Hans
+    773, 70,  // csw -> Cans
     777, 410,  // ctd -> Pauc
-    781, 45,  // ctg -> Beng
-    785, 105,  // ctn -> Deva
+    781, 50,  // ctg -> Beng
+    785, 110,  // ctn -> Deva
     789, 520,  // ctt -> Taml
     793, 520,  // cty -> Taml
-    797, 100,  // cu -> Cyrl
+    797, 105,  // cu -> Cyrl
     800, 255,  // cuu -> Lana
-    804, 100,  // cv -> Cyrl
-    807, 175,  // czh -> Hans
-    811, 185,  // czk -> Hebr
-    815, 105,  // daq -> Deva
-    819, 100,  // dar -> Cyrl
+    804, 105,  // cv -> Cyrl
+    807, 180,  // czh -> Hans
+    811, 190,  // czk -> Hebr
+    815, 110,  // daq -> Deva
+    819, 105,  // dar -> Cyrl
     823, 10,  // dcc -> Arab
-    827, 100,  // ddo -> Cyrl
+    827, 105,  // ddo -> Cyrl
     831, 10,  // def -> Arab
     835, 10,  // deh -> Arab
-    839, 45,  // der -> Beng
+    839, 50,  // der -> Beng
     843, 10,  // dgl -> Arab
-    847, 105,  // dhi -> Deva
-    851, 155,  // dhn -> Gujr
-    855, 105,  // dho -> Deva
-    859, 105,  // dhw -> Deva
+    847, 110,  // dhi -> Deva
+    851, 160,  // dhn -> Gujr
+    855, 110,  // dho -> Deva
+    859, 110,  // dhw -> Deva
     863, 555,  // dka -> Tibt
-    867, 100,  // dlg -> Cyrl
+    867, 105,  // dlg -> Cyrl
     871, 320,  // dmf -> Medf
     875, 10,  // dmk -> Arab
     879, 10,  // dml -> Arab
-    883, 100,  // dng -> Cyrl
+    883, 105,  // dng -> Cyrl
     887, 355,  // dnu -> Mymr
     891, 355,  // dnv -> Mymr
-    895, 105,  // doi -> Deva
-    899, 120,  // dox -> Ethi
+    895, 110,  // doi -> Deva
+    899, 125,  // dox -> Ethi
     903, 555,  // dre -> Tibt
-    907, 105,  // drq -> Deva
-    911, 120,  // drs -> Ethi
-    915, 105,  // dry -> Deva
+    907, 110,  // drq -> Deva
+    911, 125,  // drs -> Ethi
+    915, 110,  // dry -> Deva
     919, 395,  // dso -> Orya
-    923, 105,  // dty -> Deva
-    927, 155,  // dub -> Gujr
-    931, 105,  // duh -> Deva
-    935, 105,  // dus -> Deva
+    923, 110,  // dty -> Deva
+    927, 160,  // dub -> Gujr
+    931, 110,  // duh -> Deva
+    935, 110,  // dus -> Deva
     939, 545,  // dv -> Thaa
     942, 395,  // dwk -> Orya
-    946, 105,  // dwz -> Deva
+    946, 110,  // dwz -> Deva
     950, 555,  // dz -> Tibt
     953, 555,  // dzl -> Tibt
-    957, 150,  // ecr -> Grek
-    961, 95,  // ecy -> Cprt
-    965, 110,  // egy -> Egyp
-    969, 215,  // eky -> Kali
-    973, 150,  // el -> Grek
-    976, 105,  // emg -> Deva
-    980, 105,  // emu -> Deva
-    984, 100,  // enf -> Cyrl
-    988, 100,  // enh -> Cyrl
+    957, 155,  // ecr -> Grek
+    961, 100,  // ecy -> Cprt
+    965, 115,  // egy -> Egyp
+    969, 220,  // eky -> Kali
+    973, 155,  // el -> Grek
+    976, 110,  // emg -> Deva
+    980, 110,  // emu -> Deva
+    984, 105,  // enf -> Cyrl
+    988, 105,  // enh -> Cyrl
     992, 520,  // era -> Taml
-    996, 135,  // esg -> Gonm
+    996, 140,  // esg -> Gonm
     1000, 10,  // esh -> Arab
-    1004, 200,  // ett -> Ital
-    1008, 100,  // eve -> Cyrl
-    1012, 100,  // evn -> Cyrl
+    1004, 205,  // ett -> Ital
+    1008, 105,  // eve -> Cyrl
+    1012, 105,  // evn -> Cyrl
     1016, 10,  // fa -> Arab
     1019, 10,  // fay -> Arab
     1023, 10,  // faz -> Arab
     1027, 10,  // fia -> Arab
-    1031, 105,  // fmu -> Deva
+    1031, 110,  // fmu -> Deva
     1035, 10,  // fub -> Arab
-    1039, 175,  // gan -> Hans
+    1039, 180,  // gan -> Hans
     1043, 395,  // gaq -> Orya
-    1047, 155,  // gas -> Gujr
+    1047, 160,  // gas -> Gujr
     1051, 535,  // gau -> Telu
     1055, 395,  // gbj -> Orya
-    1059, 105,  // gbk -> Deva
-    1063, 155,  // gbl -> Gujr
-    1067, 105,  // gbm -> Deva
+    1059, 110,  // gbk -> Deva
+    1063, 160,  // gbl -> Gujr
+    1067, 110,  // gbm -> Deva
     1071, 10,  // gbz -> Arab
     1075, 395,  // gdb -> Orya
-    1079, 100,  // gdo -> Cyrl
-    1083, 105,  // gdx -> Deva
-    1087, 120,  // gez -> Ethi
+    1079, 105,  // gdo -> Cyrl
+    1083, 110,  // gdx -> Deva
+    1087, 125,  // gez -> Ethi
     1091, 10,  // ggg -> Arab
     1095, 10,  // gha -> Arab
-    1099, 105,  // ghe -> Deva
+    1099, 110,  // ghe -> Deva
     1103, 540,  // gho -> Tfng
     1107, 10,  // ghr -> Arab
     1111, 555,  // ght -> Tibt
     1115, 10,  // gig -> Arab
-    1119, 100,  // gin -> Cyrl
+    1119, 105,  // gin -> Cyrl
     1123, 10,  // gjk -> Arab
     1127, 10,  // gju -> Arab
-    1131, 100,  // gld -> Cyrl
+    1131, 105,  // gld -> Cyrl
     1135, 10,  // glh -> Arab
     1139, 10,  // glk -> Arab
     1143, 265,  // gml -> Latf
-    1147, 120,  // gmv -> Ethi
+    1147, 125,  // gmv -> Ethi
     1151, 285,  // gmy -> Linb
     1155, 555,  // goe -> Tibt
-    1159, 120,  // gof -> Ethi
-    1163, 105,  // goj -> Deva
-    1167, 105,  // gok -> Deva
-    1171, 105,  // gon -> Deva
-    1175, 140,  // got -> Goth
-    1179, 105,  // gra -> Deva
-    1183, 95,  // grc -> Cprt
-    1187, 45,  // grt -> Beng
-    1191, 120,  // gru -> Ethi
-    1195, 155,  // gu -> Gujr
-    1198, 105,  // gvr -> Deva
+    1159, 125,  // gof -> Ethi
+    1163, 110,  // goj -> Deva
+    1167, 110,  // gok -> Deva
+    1171, 110,  // gon -> Deva
+    1175, 145,  // got -> Goth
+    1179, 110,  // gra -> Deva
+    1183, 155,  // grc -> Grek
+    1187, 50,  // grt -> Beng
+    1191, 125,  // gru -> Ethi
+    1195, 160,  // gu -> Gujr
+    1198, 110,  // gvr -> Deva
     1202, 10,  // gwc -> Arab
     1206, 10,  // gwf -> Arab
     1210, 10,  // gwt -> Arab
-    1214, 105,  // gyo -> Deva
+    1214, 110,  // gyo -> Deva
     1218, 10,  // gzi -> Arab
     1222, 10,  // ha_CM -> Arab
     1228, 10,  // ha_SD -> Arab
     1234, 10,  // hac -> Arab
-    1238, 175,  // hak -> Hans
-    1242, 120,  // har -> Ethi
-    1246, 10,  // haz -> Arab
-    1250, 185,  // hbo -> Hebr
-    1254, 120,  // hdy -> Ethi
-    1258, 185,  // he -> Hebr
-    1261, 105,  // hi -> Deva
-    1264, 105,  // hif -> Deva
-    1268, 505,  // hii -> Takr
-    1272, 590,  // hit -> Xsux
-    1276, 10,  // hkh -> Arab
-    1280, 105,  // hlb -> Deva
-    1284, 190,  // hlu -> Hluw
-    1288, 425,  // hmd -> Plrd
-    1292, 50,  // hmj -> Bopo
-    1296, 50,  // hmq -> Bopo
-    1300, 10,  // hnd -> Arab
-    1304, 105,  // hne -> Deva
-    1308, 195,  // hnj -> Hmnp
-    1312, 10,  // hno -> Arab
-    1316, 105,  // hoc -> Deva
-    1320, 10,  // hoh -> Arab
-    1324, 105,  // hoj -> Deva
-    1328, 170,  // how -> Hani
-    1332, 105,  // hoy -> Deva
-    1336, 355,  // hpo -> Mymr
-    1340, 495,  // hrt -> Syrc
-    1344, 10,  // hrz -> Arab
-    1348, 175,  // hsn -> Hans
-    1352, 10,  // hss -> Arab
-    1356, 590,  // htx -> Xsux
-    1360, 105,  // hut -> Deva
-    1364, 185,  // huy -> Hebr
-    1368, 100,  // huz -> Cyrl
-    1372, 20,  // hy -> Armn
-    1375, 20,  // hyw -> Armn
-    1379, 595,  // ii -> Yiii
-    1382, 295,  // imy -> Lyci
-    1386, 100,  // inh -> Cyrl
-    1390, 355,  // int -> Mymr
-    1394, 120,  // ior -> Ethi
-    1398, 520,  // iru -> Taml
-    1402, 10,  // isk -> Arab
-    1406, 185,  // itk -> Hebr
-    1410, 100,  // itl -> Cyrl
-    1414, 65,  // iu -> Cans
-    1417, 185,  // iw -> Hebr
-    1420, 210,  // ja -> Jpan
-    1423, 10,  // jad -> Arab
-    1427, 10,  // jat -> Arab
-    1431, 185,  // jbe -> Hebr
-    1435, 10,  // jbn -> Arab
-    1439, 100,  // jct -> Cyrl
-    1443, 555,  // jda -> Tibt
-    1447, 10,  // jdg -> Arab
-    1451, 100,  // jdt -> Cyrl
-    1455, 105,  // jee -> Deva
-    1459, 125,  // jge -> Geor
-    1463, 185,  // ji -> Hebr
-    1466, 165,  // jje -> Hang
-    1470, 355,  // jkm -> Mymr
-    1474, 105,  // jml -> Deva
-    1478, 505,  // jna -> Takr
-    1482, 10,  // jnd -> Arab
-    1486, 105,  // jnl -> Deva
-    1490, 105,  // jns -> Deva
-    1494, 10,  // jog -> Arab
-    1498, 185,  // jpa -> Hebr
-    1502, 185,  // jpr -> Hebr
-    1506, 185,  // jrb -> Hebr
-    1510, 105,  // jul -> Deva
-    1514, 395,  // jun -> Orya
-    1518, 395,  // juy -> Orya
-    1522, 555,  // jya -> Tibt
-    1526, 185,  // jye -> Hebr
-    1530, 125,  // ka -> Geor
-    1533, 100,  // kaa -> Cyrl
-    1537, 100,  // kap -> Cyrl
-    1541, 225,  // kaw -> Kawi
-    1545, 100,  // kbd -> Cyrl
-    1549, 555,  // kbg -> Tibt
-    1553, 10,  // kbu -> Arab
-    1557, 10,  // kby -> Arab
-    1561, 100,  // kca -> Cyrl
-    1565, 10,  // kcy -> Arab
-    1569, 45,  // kdq -> Beng
-    1573, 550,  // kdt -> Thai
-    1577, 100,  // ket -> Cyrl
-    1581, 330,  // kev -> Mlym
-    1585, 105,  // kex -> Deva
-    1589, 535,  // key -> Telu
-    1593, 245,  // kfa -> Knda
-    1597, 105,  // kfb -> Deva
-    1601, 535,  // kfc -> Telu
-    1605, 245,  // kfd -> Knda
-    1609, 520,  // kfe -> Taml
-    1613, 245,  // kfg -> Knda
-    1617, 330,  // kfh -> Mlym
-    1621, 520,  // kfi -> Taml
-    1625, 105,  // kfk -> Deva
-    1629, 10,  // kfm -> Arab
-    1633, 105,  // kfp -> Deva
-    1637, 105,  // kfq -> Deva
-    1641, 105,  // kfr -> Deva
-    1645, 105,  // kfs -> Deva
-    1649, 105,  // kfu -> Deva
-    1653, 105,  // kfx -> Deva
-    1657, 105,  // kfy -> Deva
-    1661, 105,  // kgj -> Deva
-    1665, 105,  // kgy -> Deva
-    1669, 515,  // khb -> Talu
-    1673, 550,  // khf -> Thai
-    1677, 555,  // khg -> Tibt
-    1681, 105,  // khn -> Deva
-    1685, 55,  // kho -> Brah
-    1689, 355,  // kht -> Mymr
-    1693, 100,  // khv -> Cyrl
-    1697, 10,  // khw -> Arab
-    1701, 105,  // kif -> Deva
-    1705, 100,  // kim -> Cyrl
-    1709, 105,  // kip -> Deva
-    1713, 260,  // kjg -> Laoo
-    1717, 100,  // kjh -> Cyrl
-    1721, 105,  // kjl -> Deva
-    1725, 105,  // kjo -> Deva
-    1729, 355,  // kjp -> Mymr
-    1733, 550,  // kjt -> Thai
-    1737, 555,  // kjz -> Tibt
-    1741, 100,  // kk -> Cyrl
-    1744, 10,  // kk_AF -> Arab
-    1750, 10,  // kk_CN -> Arab
-    1756, 10,  // kk_IR -> Arab
-    1762, 10,  // kk_MN -> Arab
-    1768, 555,  // kkf -> Tibt
-    1772, 255,  // kkh -> Lana
-    1776, 105,  // kkt -> Deva
-    1780, 105,  // kle -> Deva
-    1784, 10,  // klj -> Arab
-    1788, 105,  // klr -> Deva
-    1792, 235,  // km -> Khmr
-    1795, 105,  // kmj -> Deva
-    1799, 10,  // kmz -> Arab
-    1803, 245,  // kn -> Knda
-    1806, 105,  // knn -> Deva
-    1810, 250,  // ko -> Kore
-    1813, 100,  // koi -> Cyrl
-    1817, 105,  // kok -> Deva
-    1821, 100,  // kpt -> Cyrl
-    1825, 100,  // kpy -> Cyrl
-    1829, 495,  // kqd -> Syrc
-    1833, 120,  // kqy -> Ethi
-    1837, 105,  // kra -> Deva
-    1841, 100,  // krc -> Cyrl
-    1845, 100,  // krk -> Cyrl
-    1849, 235,  // krr -> Khmr
-    1853, 105,  // kru -> Deva
-    1857, 235,  // krv -> Khmr
-    1861, 10,  // ks -> Arab
-    1864, 355,  // ksu -> Mymr
-    1868, 355,  // ksw -> Mymr
-    1872, 105,  // ksz -> Deva
-    1876, 120,  // ktb -> Ethi
-    1880, 105,  // kte -> Deva
-    1884, 10,  // ktl -> Arab
-    1888, 425,  // ktp -> Plrd
-    1892, 10,  // ku_LB -> Arab
-    1898, 260,  // kuf -> Laoo
-    1902, 100,  // kum -> Cyrl
-    1906, 100,  // kv -> Cyrl
-    1909, 100,  // kva -> Cyrl
-    1913, 355,  // kvq -> Mymr
-    1917, 355,  // kvt -> Mymr
-    1921, 10,  // kvx -> Arab
-    1925, 215,  // kvy -> Kali
-    1929, 355,  // kxf -> Mymr
-    1933, 355,  // kxk -> Mymr
-    1937, 550,  // kxm -> Thai
-    1941, 10,  // kxp -> Arab
-    1945, 100,  // ky -> Cyrl
-    1948, 10,  // ky_CN -> Arab
-    1954, 215,  // kyu -> Kali
-    1958, 105,  // kyv -> Deva
-    1962, 105,  // kyw -> Deva
-    1966, 280,  // lab -> Lina
-    1970, 185,  // lad -> Hebr
-    1974, 105,  // lae -> Deva
-    1978, 10,  // lah -> Arab
-    1982, 100,  // lbe -> Cyrl
-    1986, 105,  // lbf -> Deva
-    1990, 555,  // lbj -> Tibt
-    1994, 105,  // lbm -> Deva
-    1998, 260,  // lbo -> Laoo
-    2002, 105,  // lbr -> Deva
-    2006, 550,  // lcp -> Thai
-    2010, 275,  // lep -> Lepc
-    2014, 100,  // lez -> Cyrl
-    2018, 105,  // lhm -> Deva
-    2022, 495,  // lhs -> Syrc
-    2026, 105,  // lif -> Deva
-    2030, 290,  // lis -> Lisu
-    2034, 555,  // lkh -> Tibt
-    2038, 10,  // lki -> Arab
-    2042, 105,  // lmh -> Deva
-    2046, 535,  // lmn -> Telu
-    2050, 260,  // lo -> Laoo
-    2053, 105,  // loy -> Deva
-    2057, 425,  // lpo -> Plrd
-    2061, 10,  // lrc -> Arab
-    2065, 10,  // lrk -> Arab
-    2069, 10,  // lrl -> Arab
-    2073, 10,  // lsa -> Arab
-    2077, 185,  // lsd -> Hebr
-    2081, 10,  // lss -> Arab
-    2085, 180,  // ltc -> Hant
-    2089, 555,  // luk -> Tibt
-    2093, 105,  // luu -> Deva
-    2097, 10,  // luv -> Arab
-    2101, 10,  // luz -> Arab
-    2105, 550,  // lwl -> Thai
-    2109, 550,  // lwm -> Thai
-    2113, 555,  // lya -> Tibt
-    2117, 175,  // lzh -> Hans
-    2121, 105,  // mag -> Deva
-    2125, 105,  // mai -> Deva
-    2129, 370,  // man_GN -> Nkoo
-    2136, 10,  // mby -> Arab
-    2140, 10,  // mde -> Arab
-    2144, 100,  // mdf -> Cyrl
-    2148, 120,  // mdx -> Ethi
-    2152, 120,  // mdy -> Ethi
-    2156, 10,  // mfa -> Arab
-    2160, 10,  // mfi -> Arab
-    2164, 270,  // mga -> Latg
-    2168, 105,  // mgp -> Deva
-    2172, 10,  // mhj -> Arab
-    2176, 305,  // mid -> Mand
-    2180, 105,  // mjl -> Deva
-    2184, 330,  // mjq -> Mlym
-    2188, 330,  // mjr -> Mlym
-    2192, 105,  // mjt -> Deva
-    2196, 535,  // mju -> Telu
-    2200, 330,  // mjv -> Mlym
-    2204, 105,  // mjz -> Deva
-    2208, 100,  // mk -> Cyrl
-    2211, 105,  // mkb -> Deva
-    2215, 105,  // mke -> Deva
-    2219, 10,  // mki -> Arab
-    2223, 550,  // mkm -> Thai
-    2227, 330,  // ml -> Mlym
-    2230, 550,  // mlf -> Thai
-    2234, 100,  // mn -> Cyrl
-    2237, 340,  // mn_CN -> Mong
-    2243, 340,  // mnc -> Mong
-    2247, 45,  // mni -> Beng
-    2251, 10,  // mnj -> Arab
-    2255, 100,  // mns -> Cyrl
-    2259, 355,  // mnw -> Mymr
-    2263, 550,  // mpz -> Thai
-    2267, 105,  // mr -> Deva
-    2270, 550,  // mra -> Thai
-    2274, 105,  // mrd -> Deva
-    2278, 100,  // mrj -> Cyrl
-    2282, 345,  // mro -> Mroo
-    2286, 105,  // mrr -> Deva
-    2290, 10,  // ms_CC -> Arab
-    2296, 100,  // mtm -> Cyrl
-    2300, 105,  // mtr -> Deva
-    2304, 100,  // mud -> Cyrl
-    2308, 555,  // muk -> Tibt
-    2312, 105,  // mut -> Deva
-    2316, 520,  // muv -> Taml
-    2320, 120,  // muz -> Ethi
-    2324, 10,  // mve -> Arab
-    2328, 340,  // mvf -> Mong
-    2332, 10,  // mvy -> Arab
-    2336, 120,  // mvz -> Ethi
-    2340, 105,  // mwr -> Deva
-    2344, 355,  // mwt -> Mymr
-    2348, 195,  // mww -> Hmnp
-    2352, 355,  // my -> Mymr
-    2355, 120,  // mym -> Ethi
-    2359, 100,  // myv -> Cyrl
-    2363, 305,  // myz -> Mand
-    2367, 10,  // mzn -> Arab
-    2371, 175,  // nan -> Hans
-    2375, 105,  // nao -> Deva
-    2379, 105,  // ncd -> Deva
-    2383, 260,  // ncq -> Laoo
-    2387, 100,  // ndf -> Cyrl
-    2391, 105,  // ne -> Deva
-    2394, 100,  // neg -> Cyrl
-    2398, 555,  // neh -> Tibt
-    2402, 590,  // nei -> Xsux
-    2406, 105,  // new -> Deva
-    2410, 260,  // ngt -> Laoo
-    2414, 100,  // nio -> Cyrl
-    2418, 535,  // nit -> Telu
-    2422, 100,  // niv -> Cyrl
-    2426, 10,  // nli -> Arab
-    2430, 10,  // nlm -> Arab
-    2434, 105,  // nlx -> Deva
-    2438, 105,  // nmm -> Deva
-    2442, 580,  // nnp -> Wcho
-    2446, 255,  // nod -> Lana
-    2450, 105,  // noe -> Deva
-    2454, 100,  // nog -> Cyrl
-    2458, 105,  // noi -> Deva
-    2462, 445,  // non -> Runr
-    2466, 595,  // nos -> Yiii
-    2470, 555,  // npb -> Tibt
-    2474, 370,  // nqo -> Nkoo
-    2478, 445,  // nrn -> Runr
-    2482, 595,  // nsd -> Yiii
-    2486, 595,  // nsf -> Yiii
-    2490, 65,  // nsk -> Cans
-    2494, 560,  // nst -> Tnsa
-    2498, 595,  // nsv -> Yiii
-    2502, 595,  // nty -> Yiii
-    2506, 10,  // ntz -> Arab
-    2510, 365,  // nwc -> Newa
-    2514, 105,  // nwx -> Deva
-    2518, 550,  // nyl -> Thai
-    2522, 10,  // nyq -> Arab
-    2526, 550,  // nyw -> Thai
-    2530, 100,  // oaa -> Cyrl
-    2534, 100,  // oac -> Cyrl
-    2538, 495,  // oar -> Syrc
-    2542, 125,  // oav -> Geor
-    2546, 420,  // obm -> Phnx
-    2550, 355,  // obr -> Mymr
-    2554, 10,  // odk -> Arab
-    2558, 590,  // oht -> Xsux
-    2562, 65,  // oj -> Cans
-    2565, 65,  // ojs -> Cans
-    2569, 165,  // okm -> Hang
-    2573, 170,  // oko -> Hani
-    2577, 235,  // okz -> Khmr
-    2581, 105,  // ola -> Deva
-    2585, 555,  // ole -> Tibt
-    2589, 100,  // omk -> Cyrl
-    2593, 350,  // omp -> Mtei
-    2597, 335,  // omr -> Modi
-    2601, 355,  // omx -> Mymr
-    2605, 105,  // oon -> Deva
-    2609, 395,  // or -> Orya
-    2612, 535,  // ort -> Telu
-    2616, 10,  // oru -> Arab
-    2620, 100,  // orv -> Cyrl
-    2624, 100,  // os -> Cyrl
-    2627, 400,  // osa -> Osge
-    2631, 200,  // osc -> Ital
-    2635, 205,  // osi -> Java
-    2639, 10,  // ota -> Arab
-    2643, 555,  // otb -> Tibt
-    2647, 390,  // otk -> Orkh
-    2651, 145,  // oty -> Gran
-    2655, 405,  // oui -> Ougr
-    2659, 160,  // pa -> Guru
-    2662, 10,  // pa_PK -> Arab
-    2668, 415,  // pal -> Phli
-    2672, 100,  // paq -> Cyrl
-    2676, 10,  // pbt -> Arab
-    2680, 235,  // pcb -> Khmr
-    2684, 355,  // pce -> Mymr
-    2688, 330,  // pcf -> Mlym
-    2692, 330,  // pcg -> Mlym
-    2696, 105,  // pch -> Deva
-    2700, 105,  // pci -> Deva
-    2704, 535,  // pcj -> Telu
-    2708, 395,  // peg -> Orya
-    2712, 585,  // peo -> Xpeo
-    2716, 230,  // pgd -> Khar
-    2720, 105,  // pgg -> Deva
-    2724, 380,  // pgl -> Ogam
-    2728, 200,  // pgn -> Ital
-    2732, 105,  // phd -> Deva
-    2736, 355,  // phk -> Mymr
-    2740, 10,  // phl -> Arab
-    2744, 420,  // phn -> Phnx
-    2748, 260,  // pho -> Laoo
-    2752, 10,  // phr -> Arab
-    2756, 550,  // pht -> Thai
-    2760, 550,  // phu -> Thai
-    2764, 10,  // phv -> Arab
-    2768, 105,  // phw -> Deva
-    2772, 470,  // pi -> Sinh
-    2775, 55,  // pka -> Brah
-    2779, 330,  // pkr -> Mlym
-    2783, 10,  // plk -> Arab
-    2787, 355,  // pll -> Mymr
-    2791, 55,  // pmh -> Brah
-    2795, 150,  // pnt -> Grek
-    2799, 230,  // pra -> Khar
-    2803, 10,  // prc -> Arab
-    2807, 10,  // prd -> Arab
-    2811, 550,  // prt -> Thai
-    2815, 10,  // prx -> Arab
-    2819, 10,  // ps -> Arab
-    2822, 10,  // psh -> Arab
-    2826, 10,  // psi -> Arab
-    2830, 10,  // pst -> Arab
-    2834, 55,  // psu -> Brah
-    2838, 105,  // pum -> Deva
-    2842, 355,  // pwo -> Mymr
-    2846, 105,  // pwr -> Deva
-    2850, 550,  // pww -> Thai
-    2854, 355,  // pyx -> Mymr
-    2858, 10,  // qxq -> Arab
-    2862, 105,  // raa -> Deva
-    2866, 105,  // rab -> Deva
-    2870, 105,  // raf -> Deva
-    2874, 45,  // rah -> Beng
-    2878, 105,  // raj -> Deva
-    2882, 105,  // rav -> Deva
-    2886, 355,  // rbb -> Mymr
-    2890, 10,  // rdb -> Arab
-    2894, 395,  // rei -> Orya
-    2898, 440,  // rhg -> Rohg
-    2902, 105,  // rji -> Deva
-    2906, 105,  // rjs -> Deva
-    2910, 235,  // rka -> Khmr
-    2914, 355,  // rki -> Mymr
-    2918, 45,  // rkt -> Beng
-    2922, 20,  // rmi -> Armn
-    2926, 10,  // rmt -> Arab
-    2930, 355,  // rmz -> Mymr
-    2934, 100,  // rsk -> Cyrl
-    2938, 105,  // rtw -> Deva
-    2942, 100,  // ru -> Cyrl
-    2945, 100,  // rue -> Cyrl
-    2949, 100,  // rut -> Cyrl
-    2953, 105,  // rwr -> Deva
-    2957, 220,  // ryu -> Kana
-    2961, 105,  // sa -> Deva
-    2964, 100,  // sah -> Cyrl
-    2968, 450,  // sam -> Samr
-    2972, 385,  // sat -> Olck
-    2976, 460,  // saz -> Saur
-    2980, 10,  // sbn -> Arab
-    2984, 555,  // sbu -> Tibt
-    2988, 105,  // sck -> Deva
-    2992, 10,  // scl -> Arab
-    2996, 105,  // scp -> Deva
-    3000, 260,  // sct -> Laoo
-    3004, 505,  // scu -> Takr
-    3008, 150,  // scx -> Grek
-    3012, 10,  // sd -> Arab
-    3015, 105,  // sd_IN -> Deva
-    3021, 10,  // sdb -> Arab
-    3025, 10,  // sdf -> Arab
-    3029, 10,  // sdg -> Arab
-    3033, 10,  // sdh -> Arab
-    3037, 45,  // sdr -> Beng
-    3041, 10,  // sds -> Arab
-    3045, 100,  // sel -> Cyrl
-    3049, 425,  // sfm -> Plrd
-    3053, 380,  // sga -> Ogam
-    3057, 100,  // sgh -> Cyrl
-    3061, 105,  // sgj -> Deva
-    3065, 10,  // sgr -> Arab
-    3069, 555,  // sgt -> Tibt
-    3073, 120,  // sgw -> Ethi
-    3077, 10,  // sgy -> Arab
-    3081, 10,  // shd -> Arab
-    3085, 540,  // shi -> Tfng
-    3089, 10,  // shm -> Arab
-    3093, 355,  // shn -> Mymr
-    3097, 10,  // shu -> Arab
-    3101, 10,  // shv -> Arab
-    3105, 470,  // si -> Sinh
-    3108, 100,  // sia -> Cyrl
-    3112, 555,  // sip -> Tibt
-    3116, 10,  // siy -> Arab
-    3120, 10,  // siz -> Arab
-    3124, 100,  // sjd -> Cyrl
-    3128, 105,  // sjp -> Deva
-    3132, 100,  // sjt -> Cyrl
-    3136, 550,  // skb -> Thai
-    3140, 105,  // skj -> Deva
-    3144, 10,  // skr -> Arab
-    3148, 595,  // smh -> Yiii
-    3152, 450,  // smp -> Samr
-    3156, 235,  // smu -> Khmr
-    3160, 10,  // smy -> Arab
-    3164, 530,  // soa -> Tavt
-    3168, 475,  // sog -> Sogd
-    3172, 105,  // soi -> Deva
-    3176, 550,  // sou -> Thai
-    3180, 555,  // spt -> Tibt
-    3184, 395,  // spv -> Orya
-    3188, 10,  // sqo -> Arab
-    3192, 260,  // sqq -> Laoo
-    3196, 10,  // sqt -> Arab
-    3200, 100,  // sr -> Cyrl
-    3203, 480,  // srb -> Sora
-    3207, 10,  // srh -> Arab
-    3211, 105,  // srx -> Deva
-    3215, 10,  // srz -> Arab
-    3219, 10,  // ssh -> Arab
-    3223, 260,  // sss -> Laoo
-    3227, 10,  // sts -> Arab
-    3231, 120,  // stv -> Ethi
-    3235, 100,  // sty -> Cyrl
-    3239, 490,  // suz -> Sunu
-    3243, 125,  // sva -> Geor
-    3247, 10,  // swb -> Arab
-    3251, 170,  // swi -> Hani
-    3255, 105,  // swv -> Deva
-    3259, 445,  // sxu -> Runr
-    3263, 495,  // syc -> Syrc
-    3267, 45,  // syl -> Beng
-    3271, 495,  // syn -> Syrc
-    3275, 495,  // syr -> Syrc
-    3279, 105,  // syw -> Deva
-    3283, 520,  // ta -> Taml
-    3286, 100,  // tab -> Cyrl
-    3290, 105,  // taj -> Deva
-    3294, 500,  // tbk -> Tagb
-    3298, 555,  // tcn -> Tibt
-    3302, 355,  // tco -> Mymr
-    3306, 520,  // tcx -> Taml
-    3310, 245,  // tcy -> Knda
-    3314, 540,  // tda -> Tfng
-    3318, 105,  // tdb -> Deva
-    3322, 510,  // tdd -> Tale
-    3326, 105,  // tdg -> Deva
-    3330, 105,  // tdh -> Deva
-    3334, 535,  // te -> Telu
-    3337, 205,  // tes -> Java
-    3341, 100,  // tg -> Cyrl
-    3344, 10,  // tg_PK -> Arab
-    3350, 105,  // tge -> Deva
-    3354, 555,  // tgf -> Tibt
-    3358, 550,  // th -> Thai
-    3361, 105,  // the -> Deva
-    3365, 105,  // thf -> Deva
-    3369, 510,  // thi -> Tale
-    3373, 105,  // thl -> Deva
-    3377, 550,  // thm -> Thai
-    3381, 105,  // thq -> Deva
-    3385, 105,  // thr -> Deva
-    3389, 105,  // ths -> Deva
-    3393, 120,  // ti -> Ethi
-    3396, 120,  // tig -> Ethi
-    3400, 105,  // tij -> Deva
-    3404, 100,  // tin -> Cyrl
-    3408, 355,  // tjl -> Mymr
-    3412, 10,  // tjo -> Arab
-    3416, 105,  // tkb -> Deva
-    3420, 10,  // tks -> Arab
-    3424, 105,  // tkt -> Deva
-    3428, 495,  // tmr -> Syrc
-    3432, 60,  // tnv -> Cakm
-    3436, 10,  // tov -> Arab
-    3440, 235,  // tpu -> Khmr
-    3444, 10,  // tra -> Arab
-    3448, 185,  // trg -> Hebr
-    3452, 10,  // trm -> Arab
-    3456, 10,  // trw -> Arab
-    3460, 150,  // tsd -> Grek
-    3464, 555,  // tsj -> Tibt
-    3468, 100,  // tt -> Cyrl
-    3471, 260,  // tth -> Laoo
-    3475, 260,  // tto -> Laoo
-    3479, 550,  // tts -> Thai
-    3483, 105,  // ttz -> Deva
-    3487, 355,  // tvn -> Mymr
-    3491, 105,  // twm -> Deva
-    3495, 525,  // txg -> Tang
-    3499, 565,  // txo -> Toto
-    3503, 530,  // tyr -> Tavt
-    3507, 100,  // tyv -> Cyrl
-    3511, 100,  // ude -> Cyrl
-    3515, 330,  // udg -> Mlym
-    3519, 100,  // udi -> Cyrl
-    3523, 100,  // udm -> Cyrl
-    3527, 10,  // ug -> Arab
-    3530, 100,  // ug_KZ -> Cyrl
-    3536, 100,  // ug_MN -> Cyrl
-    3542, 570,  // uga -> Ugar
-    3546, 100,  // ugh -> Cyrl
-    3550, 550,  // ugo -> Thai
-    3554, 100,  // uk -> Cyrl
-    3557, 395,  // uki -> Orya
-    3561, 100,  // ulc -> Cyrl
-    3565, 45,  // unr -> Beng
-    3569, 105,  // unr_NP -> Deva
-    3576, 45,  // unx -> Beng
-    3580, 10,  // ur -> Arab
-    3583, 550,  // urk -> Thai
-    3587, 10,  // ush -> Arab
-    3591, 150,  // uum -> Grek
-    3595, 10,  // uz_AF -> Arab
-    3601, 100,  // uz_CN -> Cyrl
-    3607, 10,  // uzs -> Arab
-    3611, 520,  // vaa -> Taml
-    3615, 10,  // vaf -> Arab
-    3619, 105,  // vah -> Deva
-    3623, 575,  // vai -> Vaii
-    3627, 105,  // vas -> Deva
-    3631, 105,  // vav -> Deva
-    3635, 105,  // vay -> Deva
-    3639, 10,  // vgr -> Arab
-    3643, 105,  // vjk -> Deva
-    3647, 245,  // vmd -> Knda
-    3651, 10,  // vmh -> Arab
-    3655, 120,  // wal -> Ethi
-    3659, 10,  // wbk -> Arab
-    3663, 535,  // wbq -> Telu
-    3667, 105,  // wbr -> Deva
-    3671, 120,  // wle -> Ethi
-    3675, 10,  // wlo -> Arab
-    3679, 105,  // wme -> Deva
-    3683, 10,  // wne -> Arab
-    3687, 10,  // wni -> Arab
-    3691, 130,  // wsg -> Gong
-    3695, 10,  // wsv -> Arab
-    3699, 105,  // wtm -> Deva
-    3703, 175,  // wuu -> Hans
-    3707, 0,  // xag -> Aghb
-    3711, 100,  // xal -> Cyrl
-    3715, 120,  // xan -> Ethi
-    3719, 100,  // xas -> Cyrl
-    3723, 85,  // xco -> Chrs
-    3727, 70,  // xcr -> Cari
-    3731, 100,  // xdq -> Cyrl
-    3735, 10,  // xhe -> Arab
-    3739, 235,  // xhm -> Khmr
-    3743, 395,  // xis -> Orya
-    3747, 10,  // xka -> Arab
-    3751, 10,  // xkc -> Arab
-    3755, 555,  // xkf -> Tibt
-    3759, 10,  // xkj -> Arab
-    3763, 10,  // xkp -> Arab
-    3767, 295,  // xlc -> Lyci
-    3771, 300,  // xld -> Lydi
-    3775, 115,  // xly -> Elym
-    3779, 125,  // xmf -> Geor
-    3783, 310,  // xmn -> Mani
-    3787, 325,  // xmr -> Merc
-    3791, 360,  // xna -> Narb
-    3795, 105,  // xnr -> Deva
-    3799, 150,  // xpg -> Grek
-    3803, 380,  // xpi -> Ogam
-    3807, 100,  // xpm -> Cyrl
-    3811, 430,  // xpr -> Prti
-    3815, 100,  // xrm -> Cyrl
-    3819, 100,  // xrn -> Cyrl
-    3823, 455,  // xsa -> Sarb
-    3827, 105,  // xsr -> Deva
-    3831, 55,  // xtq -> Brah
-    3835, 520,  // xub -> Taml
-    3839, 520,  // xuj -> Taml
-    3843, 200,  // xve -> Ital
-    3847, 10,  // xvi -> Arab
-    3851, 100,  // xwo -> Cyrl
-    3855, 315,  // xzh -> Marc
-    3859, 100,  // yai -> Cyrl
-    3863, 105,  // ybh -> Deva
-    3867, 105,  // ybi -> Deva
-    3871, 10,  // ydg -> Arab
-    3875, 330,  // yea -> Mlym
-    3879, 150,  // yej -> Grek
-    3883, 535,  // yeu -> Telu
-    3887, 425,  // ygp -> Plrd
-    3891, 185,  // yhd -> Hebr
-    3895, 185,  // yi -> Hebr
-    3898, 595,  // yig -> Yiii
-    3902, 185,  // yih -> Hebr
-    3906, 595,  // yiv -> Yiii
-    3910, 100,  // ykg -> Cyrl
-    3914, 100,  // ykh -> Cyrl
-    3918, 425,  // yna -> Plrd
-    3922, 100,  // ynk -> Cyrl
-    3926, 210,  // yoi -> Jpan
-    3930, 550,  // yoy -> Thai
-    3934, 100,  // yrk -> Cyrl
-    3938, 595,  // ysd -> Yiii
-    3942, 595,  // ysn -> Yiii
-    3946, 595,  // ysp -> Yiii
-    3950, 100,  // ysr -> Cyrl
-    3954, 425,  // ysy -> Plrd
-    3958, 185,  // yud -> Hebr
-    3962, 180,  // yue -> Hant
-    3966, 175,  // yue_CN -> Hans
-    3973, 100,  // yug -> Cyrl
-    3977, 100,  // yux -> Cyrl
-    3981, 425,  // ywq -> Plrd
-    3985, 425,  // ywu -> Plrd
-    3989, 555,  // zau -> Tibt
-    3993, 10,  // zba -> Arab
-    3997, 170,  // zch -> Hani
-    4001, 10,  // zdj -> Arab
-    4005, 170,  // zeh -> Hani
-    4009, 540,  // zen -> Tfng
-    4013, 170,  // zgb -> Hani
-    4017, 540,  // zgh -> Tfng
-    4021, 170,  // zgm -> Hani
-    4025, 170,  // zgn -> Hani
-    4029, 175,  // zh -> Hans
-    4032, 180,  // zh_AU -> Hant
-    4038, 180,  // zh_BN -> Hant
-    4044, 180,  // zh_GB -> Hant
-    4050, 180,  // zh_GF -> Hant
-    4056, 180,  // zh_HK -> Hant
-    4062, 180,  // zh_ID -> Hant
-    4068, 180,  // zh_MO -> Hant
-    4074, 180,  // zh_PA -> Hant
-    4080, 180,  // zh_PF -> Hant
-    4086, 180,  // zh_PH -> Hant
-    4092, 180,  // zh_SR -> Hant
-    4098, 180,  // zh_TH -> Hant
-    4104, 180,  // zh_TW -> Hant
-    4110, 180,  // zh_US -> Hant
-    4116, 180,  // zh_VN -> Hant
-    4122, 170,  // zhd -> Hani
-    4126, 375,  // zhx -> Nshu
-    4130, 100,  // zko -> Cyrl
-    4134, 240,  // zkt -> Kits
-    4138, 100,  // zkz -> Cyrl
-    4142, 170,  // zlj -> Hani
-    4146, 170,  // zln -> Hani
-    4150, 170,  // zlq -> Hani
-    4154, 170,  // zqe -> Hani
-    4158, 395,  // zrg -> Orya
-    4162, 185,  // zrp -> Hebr
-    4166, 10,  // zum -> Arab
-    4170, 120,  // zwa -> Ethi
-    4174, 170,  // zyg -> Hani
-    4178, 170,  // zyn -> Hani
-    4182, 170,  // zzj -> Hani
+    1238, 180,  // hak -> Hans
+    1242, 185,  // hak_TW -> Hant
+    1249, 125,  // har -> Ethi
+    1253, 10,  // haz -> Arab
+    1257, 190,  // hbo -> Hebr
+    1261, 125,  // hdy -> Ethi
+    1265, 190,  // he -> Hebr
+    1268, 110,  // hi -> Deva
+    1271, 110,  // hif -> Deva
+    1275, 505,  // hii -> Takr
+    1279, 590,  // hit -> Xsux
+    1283, 10,  // hkh -> Arab
+    1287, 110,  // hlb -> Deva
+    1291, 195,  // hlu -> Hluw
+    1295, 425,  // hmd -> Plrd
+    1299, 55,  // hmj -> Bopo
+    1303, 55,  // hmq -> Bopo
+    1307, 10,  // hnd -> Arab
+    1311, 110,  // hne -> Deva
+    1315, 200,  // hnj -> Hmnp
+    1319, 10,  // hno -> Arab
+    1323, 110,  // hoc -> Deva
+    1327, 10,  // hoh -> Arab
+    1331, 110,  // hoj -> Deva
+    1335, 175,  // how -> Hani
+    1339, 110,  // hoy -> Deva
+    1343, 355,  // hpo -> Mymr
+    1347, 495,  // hrt -> Syrc
+    1351, 10,  // hrz -> Arab
+    1355, 180,  // hsn -> Hans
+    1359, 10,  // hss -> Arab
+    1363, 590,  // htx -> Xsux
+    1367, 110,  // hut -> Deva
+    1371, 190,  // huy -> Hebr
+    1375, 105,  // huz -> Cyrl
+    1379, 20,  // hy -> Armn
+    1382, 20,  // hyw -> Armn
+    1386, 595,  // ii -> Yiii
+    1389, 295,  // imy -> Lyci
+    1393, 105,  // inh -> Cyrl
+    1397, 355,  // int -> Mymr
+    1401, 125,  // ior -> Ethi
+    1405, 520,  // iru -> Taml
+    1409, 10,  // isk -> Arab
+    1413, 190,  // itk -> Hebr
+    1417, 105,  // itl -> Cyrl
+    1421, 70,  // iu -> Cans
+    1424, 190,  // iw -> Hebr
+    1427, 215,  // ja -> Jpan
+    1430, 10,  // jad -> Arab
+    1434, 10,  // jat -> Arab
+    1438, 190,  // jbe -> Hebr
+    1442, 10,  // jbn -> Arab
+    1446, 105,  // jct -> Cyrl
+    1450, 555,  // jda -> Tibt
+    1454, 10,  // jdg -> Arab
+    1458, 105,  // jdt -> Cyrl
+    1462, 110,  // jee -> Deva
+    1466, 130,  // jge -> Geor
+    1470, 190,  // ji -> Hebr
+    1473, 170,  // jje -> Hang
+    1477, 355,  // jkm -> Mymr
+    1481, 110,  // jml -> Deva
+    1485, 505,  // jna -> Takr
+    1489, 10,  // jnd -> Arab
+    1493, 110,  // jnl -> Deva
+    1497, 110,  // jns -> Deva
+    1501, 10,  // jog -> Arab
+    1505, 190,  // jpa -> Hebr
+    1509, 190,  // jpr -> Hebr
+    1513, 190,  // jrb -> Hebr
+    1517, 110,  // jul -> Deva
+    1521, 395,  // jun -> Orya
+    1525, 395,  // juy -> Orya
+    1529, 555,  // jya -> Tibt
+    1533, 190,  // jye -> Hebr
+    1537, 130,  // ka -> Geor
+    1540, 105,  // kaa -> Cyrl
+    1544, 105,  // kap -> Cyrl
+    1548, 30,  // kaw -> Bali
+    1552, 105,  // kbd -> Cyrl
+    1556, 555,  // kbg -> Tibt
+    1560, 10,  // kbu -> Arab
+    1564, 10,  // kby -> Arab
+    1568, 105,  // kca -> Cyrl
+    1572, 10,  // kcy -> Arab
+    1576, 50,  // kdq -> Beng
+    1580, 550,  // kdt -> Thai
+    1584, 105,  // ket -> Cyrl
+    1588, 330,  // kev -> Mlym
+    1592, 110,  // kex -> Deva
+    1596, 535,  // key -> Telu
+    1600, 245,  // kfa -> Knda
+    1604, 110,  // kfb -> Deva
+    1608, 535,  // kfc -> Telu
+    1612, 245,  // kfd -> Knda
+    1616, 520,  // kfe -> Taml
+    1620, 245,  // kfg -> Knda
+    1624, 330,  // kfh -> Mlym
+    1628, 520,  // kfi -> Taml
+    1632, 110,  // kfk -> Deva
+    1636, 10,  // kfm -> Arab
+    1640, 110,  // kfp -> Deva
+    1644, 110,  // kfq -> Deva
+    1648, 110,  // kfr -> Deva
+    1652, 110,  // kfs -> Deva
+    1656, 110,  // kfu -> Deva
+    1660, 110,  // kfx -> Deva
+    1664, 110,  // kfy -> Deva
+    1668, 110,  // kgj -> Deva
+    1672, 110,  // kgy -> Deva
+    1676, 515,  // khb -> Talu
+    1680, 550,  // khf -> Thai
+    1684, 555,  // khg -> Tibt
+    1688, 110,  // khn -> Deva
+    1692, 60,  // kho -> Brah
+    1696, 355,  // kht -> Mymr
+    1700, 105,  // khv -> Cyrl
+    1704, 10,  // khw -> Arab
+    1708, 110,  // kif -> Deva
+    1712, 105,  // kim -> Cyrl
+    1716, 110,  // kip -> Deva
+    1720, 260,  // kjg -> Laoo
+    1724, 105,  // kjh -> Cyrl
+    1728, 110,  // kjl -> Deva
+    1732, 110,  // kjo -> Deva
+    1736, 355,  // kjp -> Mymr
+    1740, 550,  // kjt -> Thai
+    1744, 555,  // kjz -> Tibt
+    1748, 105,  // kk -> Cyrl
+    1751, 10,  // kk_AF -> Arab
+    1757, 10,  // kk_CN -> Arab
+    1763, 10,  // kk_IR -> Arab
+    1769, 10,  // kk_MN -> Arab
+    1775, 555,  // kkf -> Tibt
+    1779, 255,  // kkh -> Lana
+    1783, 110,  // kkt -> Deva
+    1787, 110,  // kle -> Deva
+    1791, 10,  // klj -> Arab
+    1795, 110,  // klr -> Deva
+    1799, 235,  // km -> Khmr
+    1802, 110,  // kmj -> Deva
+    1806, 10,  // kmz -> Arab
+    1810, 245,  // kn -> Knda
+    1813, 110,  // knn -> Deva
+    1817, 250,  // ko -> Kore
+    1820, 105,  // koi -> Cyrl
+    1824, 110,  // kok -> Deva
+    1828, 105,  // kpt -> Cyrl
+    1832, 105,  // kpy -> Cyrl
+    1836, 495,  // kqd -> Syrc
+    1840, 125,  // kqy -> Ethi
+    1844, 110,  // kra -> Deva
+    1848, 105,  // krc -> Cyrl
+    1852, 105,  // krk -> Cyrl
+    1856, 235,  // krr -> Khmr
+    1860, 110,  // kru -> Deva
+    1864, 235,  // krv -> Khmr
+    1868, 10,  // ks -> Arab
+    1871, 355,  // ksu -> Mymr
+    1875, 355,  // ksw -> Mymr
+    1879, 110,  // ksz -> Deva
+    1883, 125,  // ktb -> Ethi
+    1887, 110,  // kte -> Deva
+    1891, 10,  // ktl -> Arab
+    1895, 425,  // ktp -> Plrd
+    1899, 10,  // ku_LB -> Arab
+    1905, 260,  // kuf -> Laoo
+    1909, 105,  // kum -> Cyrl
+    1913, 105,  // kv -> Cyrl
+    1916, 105,  // kva -> Cyrl
+    1920, 355,  // kvq -> Mymr
+    1924, 355,  // kvt -> Mymr
+    1928, 10,  // kvx -> Arab
+    1932, 220,  // kvy -> Kali
+    1936, 355,  // kxf -> Mymr
+    1940, 355,  // kxk -> Mymr
+    1944, 550,  // kxm -> Thai
+    1948, 10,  // kxp -> Arab
+    1952, 105,  // ky -> Cyrl
+    1955, 10,  // ky_CN -> Arab
+    1961, 220,  // kyu -> Kali
+    1965, 110,  // kyv -> Deva
+    1969, 110,  // kyw -> Deva
+    1973, 280,  // lab -> Lina
+    1977, 190,  // lad -> Hebr
+    1981, 110,  // lae -> Deva
+    1985, 10,  // lah -> Arab
+    1989, 105,  // lbe -> Cyrl
+    1993, 110,  // lbf -> Deva
+    1997, 555,  // lbj -> Tibt
+    2001, 110,  // lbm -> Deva
+    2005, 260,  // lbo -> Laoo
+    2009, 110,  // lbr -> Deva
+    2013, 550,  // lcp -> Thai
+    2017, 275,  // lep -> Lepc
+    2021, 105,  // lez -> Cyrl
+    2025, 110,  // lhm -> Deva
+    2029, 495,  // lhs -> Syrc
+    2033, 110,  // lif -> Deva
+    2037, 290,  // lis -> Lisu
+    2041, 555,  // lkh -> Tibt
+    2045, 10,  // lki -> Arab
+    2049, 110,  // lmh -> Deva
+    2053, 535,  // lmn -> Telu
+    2057, 260,  // lo -> Laoo
+    2060, 110,  // loy -> Deva
+    2064, 425,  // lpo -> Plrd
+    2068, 10,  // lrc -> Arab
+    2072, 10,  // lrk -> Arab
+    2076, 10,  // lrl -> Arab
+    2080, 10,  // lsa -> Arab
+    2084, 190,  // lsd -> Hebr
+    2088, 10,  // lss -> Arab
+    2092, 185,  // ltc -> Hant
+    2096, 555,  // luk -> Tibt
+    2100, 110,  // luu -> Deva
+    2104, 10,  // luv -> Arab
+    2108, 10,  // luz -> Arab
+    2112, 550,  // lwl -> Thai
+    2116, 550,  // lwm -> Thai
+    2120, 555,  // lya -> Tibt
+    2124, 180,  // lzh -> Hans
+    2128, 130,  // lzz_GE -> Geor
+    2135, 110,  // mag -> Deva
+    2139, 110,  // mai -> Deva
+    2143, 10,  // mby -> Arab
+    2147, 10,  // mde -> Arab
+    2151, 105,  // mdf -> Cyrl
+    2155, 125,  // mdx -> Ethi
+    2159, 125,  // mdy -> Ethi
+    2163, 10,  // mfa -> Arab
+    2167, 10,  // mfi -> Arab
+    2171, 270,  // mga -> Latg
+    2175, 110,  // mgp -> Deva
+    2179, 10,  // mhj -> Arab
+    2183, 305,  // mid -> Mand
+    2187, 110,  // mjl -> Deva
+    2191, 330,  // mjq -> Mlym
+    2195, 330,  // mjr -> Mlym
+    2199, 110,  // mjt -> Deva
+    2203, 535,  // mju -> Telu
+    2207, 330,  // mjv -> Mlym
+    2211, 110,  // mjz -> Deva
+    2215, 105,  // mk -> Cyrl
+    2218, 110,  // mkb -> Deva
+    2222, 110,  // mke -> Deva
+    2226, 10,  // mki -> Arab
+    2230, 550,  // mkm -> Thai
+    2234, 330,  // ml -> Mlym
+    2237, 550,  // mlf -> Thai
+    2241, 105,  // mn -> Cyrl
+    2244, 340,  // mn_CN -> Mong
+    2250, 340,  // mnc -> Mong
+    2254, 50,  // mni -> Beng
+    2258, 10,  // mnj -> Arab
+    2262, 105,  // mns -> Cyrl
+    2266, 355,  // mnw -> Mymr
+    2270, 550,  // mpz -> Thai
+    2274, 110,  // mr -> Deva
+    2277, 550,  // mra -> Thai
+    2281, 110,  // mrd -> Deva
+    2285, 105,  // mrj -> Cyrl
+    2289, 345,  // mro -> Mroo
+    2293, 110,  // mrr -> Deva
+    2297, 10,  // ms_CC -> Arab
+    2303, 105,  // mtm -> Cyrl
+    2307, 110,  // mtr -> Deva
+    2311, 105,  // mud -> Cyrl
+    2315, 555,  // muk -> Tibt
+    2319, 110,  // mut -> Deva
+    2323, 520,  // muv -> Taml
+    2327, 125,  // muz -> Ethi
+    2331, 10,  // mve -> Arab
+    2335, 340,  // mvf -> Mong
+    2339, 10,  // mvy -> Arab
+    2343, 125,  // mvz -> Ethi
+    2347, 110,  // mwr -> Deva
+    2351, 355,  // mwt -> Mymr
+    2355, 200,  // mww -> Hmnp
+    2359, 355,  // my -> Mymr
+    2362, 125,  // mym -> Ethi
+    2366, 105,  // myv -> Cyrl
+    2370, 305,  // myz -> Mand
+    2374, 10,  // mzn -> Arab
+    2378, 180,  // nan -> Hans
+    2382, 185,  // nan_TW -> Hant
+    2389, 110,  // nao -> Deva
+    2393, 110,  // ncd -> Deva
+    2397, 260,  // ncq -> Laoo
+    2401, 105,  // ndf -> Cyrl
+    2405, 110,  // ne -> Deva
+    2408, 105,  // neg -> Cyrl
+    2412, 555,  // neh -> Tibt
+    2416, 590,  // nei -> Xsux
+    2420, 110,  // new -> Deva
+    2424, 260,  // ngt -> Laoo
+    2428, 105,  // nio -> Cyrl
+    2432, 535,  // nit -> Telu
+    2436, 105,  // niv -> Cyrl
+    2440, 10,  // nli -> Arab
+    2444, 10,  // nlm -> Arab
+    2448, 110,  // nlx -> Deva
+    2452, 110,  // nmm -> Deva
+    2456, 580,  // nnp -> Wcho
+    2460, 255,  // nod -> Lana
+    2464, 110,  // noe -> Deva
+    2468, 105,  // nog -> Cyrl
+    2472, 110,  // noi -> Deva
+    2476, 445,  // non -> Runr
+    2480, 595,  // nos -> Yiii
+    2484, 555,  // npb -> Tibt
+    2488, 370,  // nqo -> Nkoo
+    2492, 445,  // nrn -> Runr
+    2496, 595,  // nsd -> Yiii
+    2500, 595,  // nsf -> Yiii
+    2504, 70,  // nsk -> Cans
+    2508, 560,  // nst -> Tnsa
+    2512, 595,  // nsv -> Yiii
+    2516, 595,  // nty -> Yiii
+    2520, 10,  // ntz -> Arab
+    2524, 365,  // nwc -> Newa
+    2528, 110,  // nwx -> Deva
+    2532, 550,  // nyl -> Thai
+    2536, 10,  // nyq -> Arab
+    2540, 550,  // nyw -> Thai
+    2544, 105,  // oaa -> Cyrl
+    2548, 105,  // oac -> Cyrl
+    2552, 495,  // oar -> Syrc
+    2556, 130,  // oav -> Geor
+    2560, 420,  // obm -> Phnx
+    2564, 355,  // obr -> Mymr
+    2568, 10,  // odk -> Arab
+    2572, 590,  // oht -> Xsux
+    2576, 70,  // oj -> Cans
+    2579, 70,  // ojs -> Cans
+    2583, 170,  // okm -> Hang
+    2587, 175,  // oko -> Hani
+    2591, 235,  // okz -> Khmr
+    2595, 110,  // ola -> Deva
+    2599, 555,  // ole -> Tibt
+    2603, 105,  // omk -> Cyrl
+    2607, 350,  // omp -> Mtei
+    2611, 335,  // omr -> Modi
+    2615, 355,  // omx -> Mymr
+    2619, 110,  // oon -> Deva
+    2623, 395,  // or -> Orya
+    2626, 535,  // ort -> Telu
+    2630, 10,  // oru -> Arab
+    2634, 105,  // orv -> Cyrl
+    2638, 105,  // os -> Cyrl
+    2641, 400,  // osa -> Osge
+    2645, 205,  // osc -> Ital
+    2649, 210,  // osi -> Java
+    2653, 10,  // ota -> Arab
+    2657, 555,  // otb -> Tibt
+    2661, 390,  // otk -> Orkh
+    2665, 150,  // oty -> Gran
+    2669, 405,  // oui -> Ougr
+    2673, 165,  // pa -> Guru
+    2676, 10,  // pa_PK -> Arab
+    2682, 415,  // pal -> Phli
+    2686, 105,  // paq -> Cyrl
+    2690, 10,  // pbt -> Arab
+    2694, 235,  // pcb -> Khmr
+    2698, 355,  // pce -> Mymr
+    2702, 330,  // pcf -> Mlym
+    2706, 330,  // pcg -> Mlym
+    2710, 110,  // pch -> Deva
+    2714, 110,  // pci -> Deva
+    2718, 535,  // pcj -> Telu
+    2722, 395,  // peg -> Orya
+    2726, 585,  // peo -> Xpeo
+    2730, 230,  // pgd -> Khar
+    2734, 110,  // pgg -> Deva
+    2738, 380,  // pgl -> Ogam
+    2742, 205,  // pgn -> Ital
+    2746, 110,  // phd -> Deva
+    2750, 355,  // phk -> Mymr
+    2754, 10,  // phl -> Arab
+    2758, 420,  // phn -> Phnx
+    2762, 260,  // pho -> Laoo
+    2766, 10,  // phr -> Arab
+    2770, 550,  // pht -> Thai
+    2774, 550,  // phu -> Thai
+    2778, 10,  // phv -> Arab
+    2782, 110,  // phw -> Deva
+    2786, 470,  // pi -> Sinh
+    2789, 60,  // pka -> Brah
+    2793, 330,  // pkr -> Mlym
+    2797, 10,  // plk -> Arab
+    2801, 355,  // pll -> Mymr
+    2805, 60,  // pmh -> Brah
+    2809, 155,  // pnt -> Grek
+    2813, 105,  // pnt_RU -> Cyrl
+    2820, 230,  // pra -> Khar
+    2824, 10,  // prc -> Arab
+    2828, 10,  // prd -> Arab
+    2832, 550,  // prt -> Thai
+    2836, 10,  // prx -> Arab
+    2840, 10,  // ps -> Arab
+    2843, 10,  // psh -> Arab
+    2847, 10,  // psi -> Arab
+    2851, 10,  // pst -> Arab
+    2855, 60,  // psu -> Brah
+    2859, 110,  // pum -> Deva
+    2863, 355,  // pwo -> Mymr
+    2867, 110,  // pwr -> Deva
+    2871, 550,  // pww -> Thai
+    2875, 355,  // pyx -> Mymr
+    2879, 10,  // qxq -> Arab
+    2883, 110,  // raa -> Deva
+    2887, 110,  // rab -> Deva
+    2891, 110,  // raf -> Deva
+    2895, 50,  // rah -> Beng
+    2899, 110,  // raj -> Deva
+    2903, 110,  // rav -> Deva
+    2907, 355,  // rbb -> Mymr
+    2911, 10,  // rdb -> Arab
+    2915, 395,  // rei -> Orya
+    2919, 440,  // rhg -> Rohg
+    2923, 110,  // rji -> Deva
+    2927, 110,  // rjs -> Deva
+    2931, 235,  // rka -> Khmr
+    2935, 355,  // rki -> Mymr
+    2939, 50,  // rkt -> Beng
+    2943, 20,  // rmi -> Armn
+    2947, 10,  // rmt -> Arab
+    2951, 355,  // rmz -> Mymr
+    2955, 105,  // rsk -> Cyrl
+    2959, 110,  // rtw -> Deva
+    2963, 105,  // ru -> Cyrl
+    2966, 105,  // rue -> Cyrl
+    2970, 105,  // rut -> Cyrl
+    2974, 110,  // rwr -> Deva
+    2978, 225,  // ryu -> Kana
+    2982, 110,  // sa -> Deva
+    2985, 105,  // sah -> Cyrl
+    2989, 450,  // sam -> Samr
+    2993, 385,  // sat -> Olck
+    2997, 460,  // saz -> Saur
+    3001, 10,  // sbn -> Arab
+    3005, 555,  // sbu -> Tibt
+    3009, 110,  // sck -> Deva
+    3013, 10,  // scl -> Arab
+    3017, 110,  // scp -> Deva
+    3021, 260,  // sct -> Laoo
+    3025, 505,  // scu -> Takr
+    3029, 155,  // scx -> Grek
+    3033, 10,  // sd -> Arab
+    3036, 110,  // sd_IN -> Deva
+    3042, 10,  // sdb -> Arab
+    3046, 10,  // sdf -> Arab
+    3050, 10,  // sdg -> Arab
+    3054, 10,  // sdh -> Arab
+    3058, 50,  // sdr -> Beng
+    3062, 10,  // sds -> Arab
+    3066, 105,  // sel -> Cyrl
+    3070, 425,  // sfm -> Plrd
+    3074, 105,  // sgh -> Cyrl
+    3078, 110,  // sgj -> Deva
+    3082, 10,  // sgr -> Arab
+    3086, 555,  // sgt -> Tibt
+    3090, 125,  // sgw -> Ethi
+    3094, 10,  // sgy -> Arab
+    3098, 10,  // shd -> Arab
+    3102, 540,  // shi -> Tfng
+    3106, 10,  // shm -> Arab
+    3110, 355,  // shn -> Mymr
+    3114, 10,  // shu -> Arab
+    3118, 10,  // shv -> Arab
+    3122, 470,  // si -> Sinh
+    3125, 105,  // sia -> Cyrl
+    3129, 555,  // sip -> Tibt
+    3133, 10,  // siy -> Arab
+    3137, 10,  // siz -> Arab
+    3141, 105,  // sjd -> Cyrl
+    3145, 110,  // sjp -> Deva
+    3149, 105,  // sjt -> Cyrl
+    3153, 550,  // skb -> Thai
+    3157, 110,  // skj -> Deva
+    3161, 10,  // skr -> Arab
+    3165, 595,  // smh -> Yiii
+    3169, 450,  // smp -> Samr
+    3173, 235,  // smu -> Khmr
+    3177, 10,  // smy -> Arab
+    3181, 530,  // soa -> Tavt
+    3185, 475,  // sog -> Sogd
+    3189, 110,  // soi -> Deva
+    3193, 550,  // sou -> Thai
+    3197, 555,  // spt -> Tibt
+    3201, 395,  // spv -> Orya
+    3205, 10,  // sqo -> Arab
+    3209, 260,  // sqq -> Laoo
+    3213, 10,  // sqt -> Arab
+    3217, 105,  // sr -> Cyrl
+    3220, 480,  // srb -> Sora
+    3224, 10,  // srh -> Arab
+    3228, 110,  // srx -> Deva
+    3232, 10,  // srz -> Arab
+    3236, 10,  // ssh -> Arab
+    3240, 260,  // sss -> Laoo
+    3244, 10,  // sts -> Arab
+    3248, 125,  // stv -> Ethi
+    3252, 105,  // sty -> Cyrl
+    3256, 490,  // suz -> Sunu
+    3260, 130,  // sva -> Geor
+    3264, 10,  // swb -> Arab
+    3268, 175,  // swi -> Hani
+    3272, 110,  // swv -> Deva
+    3276, 445,  // sxu -> Runr
+    3280, 495,  // syc -> Syrc
+    3284, 50,  // syl -> Beng
+    3288, 495,  // syn -> Syrc
+    3292, 495,  // syr -> Syrc
+    3296, 110,  // syw -> Deva
+    3300, 520,  // ta -> Taml
+    3303, 105,  // tab -> Cyrl
+    3307, 110,  // taj -> Deva
+    3311, 500,  // tbk -> Tagb
+    3315, 555,  // tcn -> Tibt
+    3319, 355,  // tco -> Mymr
+    3323, 520,  // tcx -> Taml
+    3327, 245,  // tcy -> Knda
+    3331, 540,  // tda -> Tfng
+    3335, 110,  // tdb -> Deva
+    3339, 510,  // tdd -> Tale
+    3343, 110,  // tdg -> Deva
+    3347, 110,  // tdh -> Deva
+    3351, 535,  // te -> Telu
+    3354, 210,  // tes -> Java
+    3358, 105,  // tg -> Cyrl
+    3361, 10,  // tg_PK -> Arab
+    3367, 110,  // tge -> Deva
+    3371, 555,  // tgf -> Tibt
+    3375, 550,  // th -> Thai
+    3378, 110,  // the -> Deva
+    3382, 110,  // thf -> Deva
+    3386, 510,  // thi -> Tale
+    3390, 110,  // thl -> Deva
+    3394, 550,  // thm -> Thai
+    3398, 110,  // thq -> Deva
+    3402, 110,  // thr -> Deva
+    3406, 110,  // ths -> Deva
+    3410, 125,  // ti -> Ethi
+    3413, 125,  // tig -> Ethi
+    3417, 110,  // tij -> Deva
+    3421, 105,  // tin -> Cyrl
+    3425, 355,  // tjl -> Mymr
+    3429, 10,  // tjo -> Arab
+    3433, 110,  // tkb -> Deva
+    3437, 10,  // tks -> Arab
+    3441, 110,  // tkt -> Deva
+    3445, 495,  // tmr -> Syrc
+    3449, 65,  // tnv -> Cakm
+    3453, 10,  // tov -> Arab
+    3457, 235,  // tpu -> Khmr
+    3461, 10,  // tra -> Arab
+    3465, 190,  // trg -> Hebr
+    3469, 10,  // trm -> Arab
+    3473, 10,  // trw -> Arab
+    3477, 155,  // tsd -> Grek
+    3481, 555,  // tsj -> Tibt
+    3485, 105,  // tt -> Cyrl
+    3488, 260,  // tth -> Laoo
+    3492, 260,  // tto -> Laoo
+    3496, 550,  // tts -> Thai
+    3500, 110,  // ttz -> Deva
+    3504, 355,  // tvn -> Mymr
+    3508, 110,  // twm -> Deva
+    3512, 525,  // txg -> Tang
+    3516, 565,  // txo -> Toto
+    3520, 530,  // tyr -> Tavt
+    3524, 105,  // tyv -> Cyrl
+    3528, 105,  // ude -> Cyrl
+    3532, 330,  // udg -> Mlym
+    3536, 105,  // udi -> Cyrl
+    3540, 105,  // udm -> Cyrl
+    3544, 10,  // ug -> Arab
+    3547, 105,  // ug_KZ -> Cyrl
+    3553, 105,  // ug_MN -> Cyrl
+    3559, 570,  // uga -> Ugar
+    3563, 105,  // ugh -> Cyrl
+    3567, 550,  // ugo -> Thai
+    3571, 105,  // uk -> Cyrl
+    3574, 395,  // uki -> Orya
+    3578, 105,  // ulc -> Cyrl
+    3582, 50,  // unr -> Beng
+    3586, 110,  // unr_NP -> Deva
+    3593, 50,  // unx -> Beng
+    3597, 10,  // ur -> Arab
+    3600, 550,  // urk -> Thai
+    3604, 10,  // ush -> Arab
+    3608, 155,  // uum -> Grek
+    3612, 10,  // uz_AF -> Arab
+    3618, 105,  // uz_CN -> Cyrl
+    3624, 10,  // uzs -> Arab
+    3628, 520,  // vaa -> Taml
+    3632, 10,  // vaf -> Arab
+    3636, 110,  // vah -> Deva
+    3640, 575,  // vai -> Vaii
+    3644, 110,  // vas -> Deva
+    3648, 110,  // vav -> Deva
+    3652, 110,  // vay -> Deva
+    3656, 10,  // vgr -> Arab
+    3660, 110,  // vjk -> Deva
+    3664, 245,  // vmd -> Knda
+    3668, 10,  // vmh -> Arab
+    3672, 125,  // wal -> Ethi
+    3676, 10,  // wbk -> Arab
+    3680, 535,  // wbq -> Telu
+    3684, 110,  // wbr -> Deva
+    3688, 125,  // wle -> Ethi
+    3692, 10,  // wlo -> Arab
+    3696, 110,  // wme -> Deva
+    3700, 10,  // wne -> Arab
+    3704, 10,  // wni -> Arab
+    3708, 135,  // wsg -> Gong
+    3712, 10,  // wsv -> Arab
+    3716, 110,  // wtm -> Deva
+    3720, 180,  // wuu -> Hans
+    3724, 0,  // xag -> Aghb
+    3728, 105,  // xal -> Cyrl
+    3732, 125,  // xan -> Ethi
+    3736, 105,  // xas -> Cyrl
+    3740, 90,  // xco -> Chrs
+    3744, 75,  // xcr -> Cari
+    3748, 105,  // xdq -> Cyrl
+    3752, 10,  // xhe -> Arab
+    3756, 235,  // xhm -> Khmr
+    3760, 395,  // xis -> Orya
+    3764, 10,  // xka -> Arab
+    3768, 10,  // xkc -> Arab
+    3772, 555,  // xkf -> Tibt
+    3776, 10,  // xkj -> Arab
+    3780, 10,  // xkp -> Arab
+    3784, 295,  // xlc -> Lyci
+    3788, 300,  // xld -> Lydi
+    3792, 120,  // xly -> Elym
+    3796, 130,  // xmf -> Geor
+    3800, 310,  // xmn -> Mani
+    3804, 325,  // xmr -> Merc
+    3808, 360,  // xna -> Narb
+    3812, 110,  // xnr -> Deva
+    3816, 155,  // xpg -> Grek
+    3820, 380,  // xpi -> Ogam
+    3824, 105,  // xpm -> Cyrl
+    3828, 430,  // xpr -> Prti
+    3832, 105,  // xrm -> Cyrl
+    3836, 105,  // xrn -> Cyrl
+    3840, 455,  // xsa -> Sarb
+    3844, 110,  // xsr -> Deva
+    3848, 60,  // xtq -> Brah
+    3852, 520,  // xub -> Taml
+    3856, 520,  // xuj -> Taml
+    3860, 205,  // xve -> Ital
+    3864, 10,  // xvi -> Arab
+    3868, 105,  // xwo -> Cyrl
+    3872, 315,  // xzh -> Marc
+    3876, 105,  // yai -> Cyrl
+    3880, 110,  // ybh -> Deva
+    3884, 110,  // ybi -> Deva
+    3888, 10,  // ydg -> Arab
+    3892, 330,  // yea -> Mlym
+    3896, 155,  // yej -> Grek
+    3900, 535,  // yeu -> Telu
+    3904, 425,  // ygp -> Plrd
+    3908, 190,  // yhd -> Hebr
+    3912, 190,  // yi -> Hebr
+    3915, 595,  // yig -> Yiii
+    3919, 190,  // yih -> Hebr
+    3923, 595,  // yiv -> Yiii
+    3927, 105,  // ykg -> Cyrl
+    3931, 105,  // ykh -> Cyrl
+    3935, 425,  // yna -> Plrd
+    3939, 105,  // ynk -> Cyrl
+    3943, 215,  // yoi -> Jpan
+    3947, 550,  // yoy -> Thai
+    3951, 105,  // yrk -> Cyrl
+    3955, 595,  // ysd -> Yiii
+    3959, 595,  // ysn -> Yiii
+    3963, 595,  // ysp -> Yiii
+    3967, 105,  // ysr -> Cyrl
+    3971, 425,  // ysy -> Plrd
+    3975, 190,  // yud -> Hebr
+    3979, 185,  // yue -> Hant
+    3983, 180,  // yue_CN -> Hans
+    3990, 105,  // yug -> Cyrl
+    3994, 105,  // yux -> Cyrl
+    3998, 425,  // ywq -> Plrd
+    4002, 425,  // ywu -> Plrd
+    4006, 555,  // zau -> Tibt
+    4010, 10,  // zba -> Arab
+    4014, 175,  // zch -> Hani
+    4018, 10,  // zdj -> Arab
+    4022, 175,  // zeh -> Hani
+    4026, 540,  // zen -> Tfng
+    4030, 175,  // zgb -> Hani
+    4034, 540,  // zgh -> Tfng
+    4038, 175,  // zgm -> Hani
+    4042, 175,  // zgn -> Hani
+    4046, 180,  // zh -> Hans
+    4049, 185,  // zh_AU -> Hant
+    4055, 185,  // zh_BN -> Hant
+    4061, 185,  // zh_GB -> Hant
+    4067, 185,  // zh_GF -> Hant
+    4073, 185,  // zh_HK -> Hant
+    4079, 185,  // zh_ID -> Hant
+    4085, 185,  // zh_MO -> Hant
+    4091, 185,  // zh_PA -> Hant
+    4097, 185,  // zh_PF -> Hant
+    4103, 185,  // zh_PH -> Hant
+    4109, 185,  // zh_SR -> Hant
+    4115, 185,  // zh_TH -> Hant
+    4121, 185,  // zh_TW -> Hant
+    4127, 185,  // zh_US -> Hant
+    4133, 185,  // zh_VN -> Hant
+    4139, 175,  // zhd -> Hani
+    4143, 375,  // zhx -> Nshu
+    4147, 105,  // zko -> Cyrl
+    4151, 240,  // zkt -> Kits
+    4155, 105,  // zkz -> Cyrl
+    4159, 175,  // zlj -> Hani
+    4163, 175,  // zln -> Hani
+    4167, 175,  // zlq -> Hani
+    4171, 175,  // zqe -> Hani
+    4175, 395,  // zrg -> Orya
+    4179, 190,  // zrp -> Hebr
+    4183, 10,  // zum -> Arab
+    4187, 125,  // zwa -> Ethi
+    4191, 175,  // zyg -> Hani
+    4195, 175,  // zyn -> Hani
+    4199, 175,  // zzj -> Hani
 };
 
 //======================================================================
@@ -1159,38 +1162,39 @@ const char parentLocaleChars[] =
     "az_Arab\0az_Cyrl\0bal_Latn\0blt_Latn\0bm_Nkoo\0bs_Cyrl\0byn_Latn\0"
     "cu_Glag\0dje_Arab\0dyo_Arab\0en_001\0en_150\0en_AG\0en_AI\0en_AT\0"
     "en_AU\0en_BB\0en_BE\0en_BM\0en_BS\0en_BW\0en_BZ\0en_CC\0en_CH\0"
-    "en_CK\0en_CM\0en_CX\0en_CY\0en_DE\0en_DG\0en_DK\0en_DM\0en_Dsrt\0"
-    "en_ER\0en_FI\0en_FJ\0en_FK\0en_FM\0en_GB\0en_GD\0en_GG\0en_GH\0"
-    "en_GI\0en_GM\0en_GY\0en_HK\0en_ID\0en_IE\0en_IL\0en_IM\0en_IN\0"
-    "en_IO\0en_JE\0en_JM\0en_KE\0en_KI\0en_KN\0en_KY\0en_LC\0en_LR\0"
-    "en_LS\0en_MG\0en_MO\0en_MS\0en_MT\0en_MU\0en_MV\0en_MW\0en_MY\0"
-    "en_NA\0en_NF\0en_NG\0en_NL\0en_NR\0en_NU\0en_NZ\0en_PG\0en_PK\0"
-    "en_PN\0en_PW\0en_RW\0en_SB\0en_SC\0en_SD\0en_SE\0en_SG\0en_SH\0"
-    "en_SI\0en_SL\0en_SS\0en_SX\0en_SZ\0en_Shaw\0en_TC\0en_TK\0en_TO\0"
-    "en_TT\0en_TV\0en_TZ\0en_UG\0en_VC\0en_VG\0en_VU\0en_WS\0en_ZA\0"
-    "en_ZM\0en_ZW\0es_419\0es_AR\0es_BO\0es_BR\0es_BZ\0es_CL\0es_CO\0"
-    "es_CR\0es_CU\0es_DO\0es_EC\0es_GT\0es_HN\0es_JP\0es_MX\0es_NI\0"
-    "es_PA\0es_PE\0es_PR\0es_PY\0es_SV\0es_US\0es_UY\0es_VE\0ff_Adlm\0"
-    "ff_Arab\0fr_HT\0ha_Arab\0hi_Latn\0ht\0iu_Latn\0kaa_Latn\0kk_Arab\0"
-    "kok_Latn\0ks_Deva\0ku_Arab\0kxv_Deva\0kxv_Orya\0kxv_Telu\0ky_Arab\0"
-    "ky_Latn\0ml_Arab\0mn_Mong\0mni_Mtei\0ms_Arab\0nb\0nn\0no\0no_NO\0"
-    "pa_Arab\0pt_AO\0pt_CH\0pt_CV\0pt_FR\0pt_GQ\0pt_GW\0pt_LU\0pt_MO\0"
-    "pt_MZ\0pt_PT\0pt_ST\0pt_TL\0root\0sat_Deva\0sd_Deva\0sd_Khoj\0"
-    "sd_Sind\0shi_Latn\0so_Arab\0sr_Latn\0sw_Arab\0tg_Arab\0ug_Cyrl\0"
-    "uz_Arab\0uz_Cyrl\0vai_Latn\0wo_Arab\0yo_Arab\0yue_Hans\0zh_Hant\0"
-    "zh_Hant_HK\0zh_Hant_MO\0";
+    "en_CK\0en_CM\0en_CX\0en_CY\0en_CZ\0en_DE\0en_DG\0en_DK\0en_DM\0"
+    "en_Dsrt\0en_ER\0en_ES\0en_FI\0en_FJ\0en_FK\0en_FM\0en_FR\0en_GB\0"
+    "en_GD\0en_GG\0en_GH\0en_GI\0en_GM\0en_GS\0en_GY\0en_HK\0en_HU\0"
+    "en_ID\0en_IE\0en_IL\0en_IM\0en_IN\0en_IO\0en_IT\0en_JE\0en_JM\0"
+    "en_KE\0en_KI\0en_KN\0en_KY\0en_LC\0en_LR\0en_LS\0en_MG\0en_MO\0"
+    "en_MS\0en_MT\0en_MU\0en_MV\0en_MW\0en_MY\0en_NA\0en_NF\0en_NG\0"
+    "en_NL\0en_NO\0en_NR\0en_NU\0en_NZ\0en_PG\0en_PK\0en_PL\0en_PN\0"
+    "en_PT\0en_PW\0en_RO\0en_RW\0en_SB\0en_SC\0en_SD\0en_SE\0en_SG\0"
+    "en_SH\0en_SI\0en_SK\0en_SL\0en_SS\0en_SX\0en_SZ\0en_Shaw\0en_TC\0"
+    "en_TK\0en_TO\0en_TT\0en_TV\0en_TZ\0en_UG\0en_VC\0en_VG\0en_VU\0"
+    "en_WS\0en_ZA\0en_ZM\0en_ZW\0es_419\0es_AR\0es_BO\0es_BR\0es_BZ\0"
+    "es_CL\0es_CO\0es_CR\0es_CU\0es_DO\0es_EC\0es_GT\0es_HN\0es_JP\0"
+    "es_MX\0es_NI\0es_PA\0es_PE\0es_PR\0es_PY\0es_SV\0es_US\0es_UY\0"
+    "es_VE\0ff_Adlm\0ff_Arab\0fr_HT\0ha_Arab\0hi_Latn\0ht\0iu_Latn\0"
+    "kaa_Latn\0kk_Arab\0kok_Latn\0ks_Deva\0ku_Arab\0kxv_Deva\0kxv_Orya\0"
+    "kxv_Telu\0ky_Arab\0ky_Latn\0ml_Arab\0mn_Mong\0mni_Mtei\0ms_Arab\0"
+    "nb\0nn\0no\0no_NO\0pa_Arab\0pt_AO\0pt_CH\0pt_CV\0pt_FR\0pt_GQ\0"
+    "pt_GW\0pt_LU\0pt_MO\0pt_MZ\0pt_PT\0pt_ST\0pt_TL\0root\0sat_Deva\0"
+    "sd_Deva\0sd_Khoj\0sd_Sind\0shi_Latn\0so_Arab\0sr_Latn\0sw_Arab\0"
+    "tg_Arab\0ug_Cyrl\0uz_Arab\0uz_Cyrl\0vai_Latn\0wo_Arab\0yo_Arab\0"
+    "yue_Hans\0zh_Hant\0zh_Hant_HK\0zh_Hant_MO\0";
 
 const int32_t parentLocaleTable[] = {
-    0, 1080,  // az_Arab -> root
-    8, 1080,  // az_Cyrl -> root
-    16, 1080,  // bal_Latn -> root
-    25, 1080,  // blt_Latn -> root
-    34, 1080,  // bm_Nkoo -> root
-    42, 1080,  // bs_Cyrl -> root
-    50, 1080,  // byn_Latn -> root
-    59, 1080,  // cu_Glag -> root
-    67, 1080,  // dje_Arab -> root
-    76, 1080,  // dyo_Arab -> root
+    0, 1146,  // az_Arab -> root
+    8, 1146,  // az_Cyrl -> root
+    16, 1146,  // bal_Latn -> root
+    25, 1146,  // blt_Latn -> root
+    34, 1146,  // bm_Nkoo -> root
+    42, 1146,  // bs_Cyrl -> root
+    50, 1146,  // byn_Latn -> root
+    59, 1146,  // cu_Glag -> root
+    67, 1146,  // dje_Arab -> root
+    76, 1146,  // dyo_Arab -> root
     92, 85,  // en_150 -> en_001
     99, 85,  // en_AG -> en_001
     105, 85,  // en_AI -> en_001
@@ -1208,161 +1212,172 @@ const int32_t parentLocaleTable[] = {
     177, 85,  // en_CM -> en_001
     183, 85,  // en_CX -> en_001
     189, 85,  // en_CY -> en_001
-    195, 92,  // en_DE -> en_150
-    201, 85,  // en_DG -> en_001
-    207, 92,  // en_DK -> en_150
-    213, 85,  // en_DM -> en_001
-    219, 1080,  // en_Dsrt -> root
-    227, 85,  // en_ER -> en_001
-    233, 92,  // en_FI -> en_150
-    239, 85,  // en_FJ -> en_001
-    245, 85,  // en_FK -> en_001
-    251, 85,  // en_FM -> en_001
-    257, 85,  // en_GB -> en_001
-    263, 85,  // en_GD -> en_001
-    269, 85,  // en_GG -> en_001
-    275, 85,  // en_GH -> en_001
-    281, 85,  // en_GI -> en_001
-    287, 85,  // en_GM -> en_001
-    293, 85,  // en_GY -> en_001
-    299, 85,  // en_HK -> en_001
-    305, 85,  // en_ID -> en_001
-    311, 85,  // en_IE -> en_001
-    317, 85,  // en_IL -> en_001
-    323, 85,  // en_IM -> en_001
-    329, 85,  // en_IN -> en_001
-    335, 85,  // en_IO -> en_001
-    341, 85,  // en_JE -> en_001
-    347, 85,  // en_JM -> en_001
-    353, 85,  // en_KE -> en_001
-    359, 85,  // en_KI -> en_001
-    365, 85,  // en_KN -> en_001
-    371, 85,  // en_KY -> en_001
-    377, 85,  // en_LC -> en_001
-    383, 85,  // en_LR -> en_001
-    389, 85,  // en_LS -> en_001
-    395, 85,  // en_MG -> en_001
-    401, 85,  // en_MO -> en_001
-    407, 85,  // en_MS -> en_001
-    413, 85,  // en_MT -> en_001
-    419, 85,  // en_MU -> en_001
-    425, 85,  // en_MV -> en_001
-    431, 85,  // en_MW -> en_001
-    437, 85,  // en_MY -> en_001
-    443, 85,  // en_NA -> en_001
-    449, 85,  // en_NF -> en_001
-    455, 85,  // en_NG -> en_001
-    461, 92,  // en_NL -> en_150
-    467, 85,  // en_NR -> en_001
-    473, 85,  // en_NU -> en_001
-    479, 85,  // en_NZ -> en_001
-    485, 85,  // en_PG -> en_001
-    491, 85,  // en_PK -> en_001
-    497, 85,  // en_PN -> en_001
-    503, 85,  // en_PW -> en_001
-    509, 85,  // en_RW -> en_001
-    515, 85,  // en_SB -> en_001
-    521, 85,  // en_SC -> en_001
-    527, 85,  // en_SD -> en_001
-    533, 92,  // en_SE -> en_150
-    539, 85,  // en_SG -> en_001
-    545, 85,  // en_SH -> en_001
-    551, 92,  // en_SI -> en_150
-    557, 85,  // en_SL -> en_001
-    563, 85,  // en_SS -> en_001
-    569, 85,  // en_SX -> en_001
-    575, 85,  // en_SZ -> en_001
-    581, 1080,  // en_Shaw -> root
-    589, 85,  // en_TC -> en_001
-    595, 85,  // en_TK -> en_001
-    601, 85,  // en_TO -> en_001
-    607, 85,  // en_TT -> en_001
-    613, 85,  // en_TV -> en_001
-    619, 85,  // en_TZ -> en_001
-    625, 85,  // en_UG -> en_001
-    631, 85,  // en_VC -> en_001
-    637, 85,  // en_VG -> en_001
-    643, 85,  // en_VU -> en_001
-    649, 85,  // en_WS -> en_001
-    655, 85,  // en_ZA -> en_001
-    661, 85,  // en_ZM -> en_001
-    667, 85,  // en_ZW -> en_001
-    680, 673,  // es_AR -> es_419
-    686, 673,  // es_BO -> es_419
-    692, 673,  // es_BR -> es_419
-    698, 673,  // es_BZ -> es_419
-    704, 673,  // es_CL -> es_419
-    710, 673,  // es_CO -> es_419
-    716, 673,  // es_CR -> es_419
-    722, 673,  // es_CU -> es_419
-    728, 673,  // es_DO -> es_419
-    734, 673,  // es_EC -> es_419
-    740, 673,  // es_GT -> es_419
-    746, 673,  // es_HN -> es_419
-    752, 673,  // es_JP -> es_419
-    758, 673,  // es_MX -> es_419
-    764, 673,  // es_NI -> es_419
-    770, 673,  // es_PA -> es_419
-    776, 673,  // es_PE -> es_419
-    782, 673,  // es_PR -> es_419
-    788, 673,  // es_PY -> es_419
-    794, 673,  // es_SV -> es_419
-    800, 673,  // es_US -> es_419
-    806, 673,  // es_UY -> es_419
-    812, 673,  // es_VE -> es_419
-    818, 1080,  // ff_Adlm -> root
-    826, 1080,  // ff_Arab -> root
-    840, 1080,  // ha_Arab -> root
-    848, 329,  // hi_Latn -> en_IN
-    856, 834,  // ht -> fr_HT
-    859, 1080,  // iu_Latn -> root
-    867, 1080,  // kaa_Latn -> root
-    876, 1080,  // kk_Arab -> root
-    884, 1080,  // kok_Latn -> root
-    893, 1080,  // ks_Deva -> root
-    901, 1080,  // ku_Arab -> root
-    909, 1080,  // kxv_Deva -> root
-    918, 1080,  // kxv_Orya -> root
-    927, 1080,  // kxv_Telu -> root
-    936, 1080,  // ky_Arab -> root
-    944, 1080,  // ky_Latn -> root
-    952, 1080,  // ml_Arab -> root
-    960, 1080,  // mn_Mong -> root
-    968, 1080,  // mni_Mtei -> root
-    977, 1080,  // ms_Arab -> root
-    985, 991,  // nb -> no
-    988, 991,  // nn -> no
-    994, 991,  // no_NO -> no
-    1000, 1080,  // pa_Arab -> root
-    1008, 1062,  // pt_AO -> pt_PT
-    1014, 1062,  // pt_CH -> pt_PT
-    1020, 1062,  // pt_CV -> pt_PT
-    1026, 1062,  // pt_FR -> pt_PT
-    1032, 1062,  // pt_GQ -> pt_PT
-    1038, 1062,  // pt_GW -> pt_PT
-    1044, 1062,  // pt_LU -> pt_PT
-    1050, 1062,  // pt_MO -> pt_PT
-    1056, 1062,  // pt_MZ -> pt_PT
-    1068, 1062,  // pt_ST -> pt_PT
-    1074, 1062,  // pt_TL -> pt_PT
-    1085, 1080,  // sat_Deva -> root
-    1094, 1080,  // sd_Deva -> root
-    1102, 1080,  // sd_Khoj -> root
-    1110, 1080,  // sd_Sind -> root
-    1118, 1080,  // shi_Latn -> root
-    1127, 1080,  // so_Arab -> root
-    1135, 1080,  // sr_Latn -> root
-    1143, 1080,  // sw_Arab -> root
-    1151, 1080,  // tg_Arab -> root
-    1159, 1080,  // ug_Cyrl -> root
-    1167, 1080,  // uz_Arab -> root
-    1175, 1080,  // uz_Cyrl -> root
-    1183, 1080,  // vai_Latn -> root
-    1192, 1080,  // wo_Arab -> root
-    1200, 1080,  // yo_Arab -> root
-    1208, 1080,  // yue_Hans -> root
-    1217, 1080,  // zh_Hant -> root
-    1236, 1225,  // zh_Hant_MO -> zh_Hant_HK
+    195, 92,  // en_CZ -> en_150
+    201, 92,  // en_DE -> en_150
+    207, 85,  // en_DG -> en_001
+    213, 92,  // en_DK -> en_150
+    219, 85,  // en_DM -> en_001
+    225, 1146,  // en_Dsrt -> root
+    233, 85,  // en_ER -> en_001
+    239, 92,  // en_ES -> en_150
+    245, 92,  // en_FI -> en_150
+    251, 85,  // en_FJ -> en_001
+    257, 85,  // en_FK -> en_001
+    263, 85,  // en_FM -> en_001
+    269, 92,  // en_FR -> en_150
+    275, 85,  // en_GB -> en_001
+    281, 85,  // en_GD -> en_001
+    287, 85,  // en_GG -> en_001
+    293, 85,  // en_GH -> en_001
+    299, 85,  // en_GI -> en_001
+    305, 85,  // en_GM -> en_001
+    311, 85,  // en_GS -> en_001
+    317, 85,  // en_GY -> en_001
+    323, 85,  // en_HK -> en_001
+    329, 92,  // en_HU -> en_150
+    335, 85,  // en_ID -> en_001
+    341, 85,  // en_IE -> en_001
+    347, 85,  // en_IL -> en_001
+    353, 85,  // en_IM -> en_001
+    359, 85,  // en_IN -> en_001
+    365, 85,  // en_IO -> en_001
+    371, 92,  // en_IT -> en_150
+    377, 85,  // en_JE -> en_001
+    383, 85,  // en_JM -> en_001
+    389, 85,  // en_KE -> en_001
+    395, 85,  // en_KI -> en_001
+    401, 85,  // en_KN -> en_001
+    407, 85,  // en_KY -> en_001
+    413, 85,  // en_LC -> en_001
+    419, 85,  // en_LR -> en_001
+    425, 85,  // en_LS -> en_001
+    431, 85,  // en_MG -> en_001
+    437, 85,  // en_MO -> en_001
+    443, 85,  // en_MS -> en_001
+    449, 85,  // en_MT -> en_001
+    455, 85,  // en_MU -> en_001
+    461, 85,  // en_MV -> en_001
+    467, 85,  // en_MW -> en_001
+    473, 85,  // en_MY -> en_001
+    479, 85,  // en_NA -> en_001
+    485, 85,  // en_NF -> en_001
+    491, 85,  // en_NG -> en_001
+    497, 92,  // en_NL -> en_150
+    503, 92,  // en_NO -> en_150
+    509, 85,  // en_NR -> en_001
+    515, 85,  // en_NU -> en_001
+    521, 85,  // en_NZ -> en_001
+    527, 85,  // en_PG -> en_001
+    533, 85,  // en_PK -> en_001
+    539, 92,  // en_PL -> en_150
+    545, 85,  // en_PN -> en_001
+    551, 92,  // en_PT -> en_150
+    557, 85,  // en_PW -> en_001
+    563, 92,  // en_RO -> en_150
+    569, 85,  // en_RW -> en_001
+    575, 85,  // en_SB -> en_001
+    581, 85,  // en_SC -> en_001
+    587, 85,  // en_SD -> en_001
+    593, 92,  // en_SE -> en_150
+    599, 85,  // en_SG -> en_001
+    605, 85,  // en_SH -> en_001
+    611, 92,  // en_SI -> en_150
+    617, 92,  // en_SK -> en_150
+    623, 85,  // en_SL -> en_001
+    629, 85,  // en_SS -> en_001
+    635, 85,  // en_SX -> en_001
+    641, 85,  // en_SZ -> en_001
+    647, 1146,  // en_Shaw -> root
+    655, 85,  // en_TC -> en_001
+    661, 85,  // en_TK -> en_001
+    667, 85,  // en_TO -> en_001
+    673, 85,  // en_TT -> en_001
+    679, 85,  // en_TV -> en_001
+    685, 85,  // en_TZ -> en_001
+    691, 85,  // en_UG -> en_001
+    697, 85,  // en_VC -> en_001
+    703, 85,  // en_VG -> en_001
+    709, 85,  // en_VU -> en_001
+    715, 85,  // en_WS -> en_001
+    721, 85,  // en_ZA -> en_001
+    727, 85,  // en_ZM -> en_001
+    733, 85,  // en_ZW -> en_001
+    746, 739,  // es_AR -> es_419
+    752, 739,  // es_BO -> es_419
+    758, 739,  // es_BR -> es_419
+    764, 739,  // es_BZ -> es_419
+    770, 739,  // es_CL -> es_419
+    776, 739,  // es_CO -> es_419
+    782, 739,  // es_CR -> es_419
+    788, 739,  // es_CU -> es_419
+    794, 739,  // es_DO -> es_419
+    800, 739,  // es_EC -> es_419
+    806, 739,  // es_GT -> es_419
+    812, 739,  // es_HN -> es_419
+    818, 739,  // es_JP -> es_419
+    824, 739,  // es_MX -> es_419
+    830, 739,  // es_NI -> es_419
+    836, 739,  // es_PA -> es_419
+    842, 739,  // es_PE -> es_419
+    848, 739,  // es_PR -> es_419
+    854, 739,  // es_PY -> es_419
+    860, 739,  // es_SV -> es_419
+    866, 739,  // es_US -> es_419
+    872, 739,  // es_UY -> es_419
+    878, 739,  // es_VE -> es_419
+    884, 1146,  // ff_Adlm -> root
+    892, 1146,  // ff_Arab -> root
+    906, 1146,  // ha_Arab -> root
+    914, 359,  // hi_Latn -> en_IN
+    922, 900,  // ht -> fr_HT
+    925, 1146,  // iu_Latn -> root
+    933, 1146,  // kaa_Latn -> root
+    942, 1146,  // kk_Arab -> root
+    950, 1146,  // kok_Latn -> root
+    959, 1146,  // ks_Deva -> root
+    967, 1146,  // ku_Arab -> root
+    975, 1146,  // kxv_Deva -> root
+    984, 1146,  // kxv_Orya -> root
+    993, 1146,  // kxv_Telu -> root
+    1002, 1146,  // ky_Arab -> root
+    1010, 1146,  // ky_Latn -> root
+    1018, 1146,  // ml_Arab -> root
+    1026, 1146,  // mn_Mong -> root
+    1034, 1146,  // mni_Mtei -> root
+    1043, 1146,  // ms_Arab -> root
+    1051, 1057,  // nb -> no
+    1054, 1057,  // nn -> no
+    1060, 1057,  // no_NO -> no
+    1066, 1146,  // pa_Arab -> root
+    1074, 1128,  // pt_AO -> pt_PT
+    1080, 1128,  // pt_CH -> pt_PT
+    1086, 1128,  // pt_CV -> pt_PT
+    1092, 1128,  // pt_FR -> pt_PT
+    1098, 1128,  // pt_GQ -> pt_PT
+    1104, 1128,  // pt_GW -> pt_PT
+    1110, 1128,  // pt_LU -> pt_PT
+    1116, 1128,  // pt_MO -> pt_PT
+    1122, 1128,  // pt_MZ -> pt_PT
+    1134, 1128,  // pt_ST -> pt_PT
+    1140, 1128,  // pt_TL -> pt_PT
+    1151, 1146,  // sat_Deva -> root
+    1160, 1146,  // sd_Deva -> root
+    1168, 1146,  // sd_Khoj -> root
+    1176, 1146,  // sd_Sind -> root
+    1184, 1146,  // shi_Latn -> root
+    1193, 1146,  // so_Arab -> root
+    1201, 1146,  // sr_Latn -> root
+    1209, 1146,  // sw_Arab -> root
+    1217, 1146,  // tg_Arab -> root
+    1225, 1146,  // ug_Cyrl -> root
+    1233, 1146,  // uz_Arab -> root
+    1241, 1146,  // uz_Cyrl -> root
+    1249, 1146,  // vai_Latn -> root
+    1258, 1146,  // wo_Arab -> root
+    1266, 1146,  // yo_Arab -> root
+    1274, 1146,  // yue_Hans -> root
+    1283, 1146,  // zh_Hant -> root
+    1302, 1291,  // zh_Hant_MO -> zh_Hant_HK
 };
 
 
diff --git a/deps/icu-small/source/common/locbased.cpp b/deps/icu-small/source/common/locbased.cpp
index 832bc3e88b142c..6f35e72210fef6 100644
--- a/deps/icu-small/source/common/locbased.cpp
+++ b/deps/icu-small/source/common/locbased.cpp
@@ -12,44 +12,84 @@
 */
 #include "locbased.h"
 #include "cstring.h"
+#include "charstr.h"
 
 U_NAMESPACE_BEGIN
 
-Locale LocaleBased::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
-    const char* id = getLocaleID(type, status);
+Locale LocaleBased::getLocale(const CharString* valid, const CharString* actual,
+                              ULocDataLocaleType type, UErrorCode& status) {
+    const char* id = getLocaleID(valid, actual, type, status);
     return Locale(id != nullptr ? id : "");
 }
 
-const char* LocaleBased::getLocaleID(ULocDataLocaleType type, UErrorCode& status) const {
+const char* LocaleBased::getLocaleID(const CharString* valid, const CharString* actual,
+                                     ULocDataLocaleType type, UErrorCode& status) {
     if (U_FAILURE(status)) {
         return nullptr;
     }
 
     switch(type) {
     case ULOC_VALID_LOCALE:
-        return valid;
+        return valid == nullptr ? "" : valid->data();
     case ULOC_ACTUAL_LOCALE:
-        return actual;
+        return actual == nullptr ? "" : actual->data();
     default:
         status = U_ILLEGAL_ARGUMENT_ERROR;
         return nullptr;
     }
 }
 
-void LocaleBased::setLocaleIDs(const char* validID, const char* actualID) {
-    if (validID != nullptr) {
-      uprv_strncpy(valid, validID, ULOC_FULLNAME_CAPACITY);
-      valid[ULOC_FULLNAME_CAPACITY-1] = 0; // always terminate
+void LocaleBased::setLocaleIDs(const CharString* validID, const CharString* actualID, UErrorCode& status) {
+    setValidLocaleID(validID, status);
+    setActualLocaleID(actualID,status);
+}
+void LocaleBased::setLocaleIDs(const char* validID, const char* actualID, UErrorCode& status) {
+    setValidLocaleID(validID, status);
+    setActualLocaleID(actualID,status);
+}
+
+void LocaleBased::setLocaleID(const char* id, CharString*& dest, UErrorCode& status) {
+    if (U_FAILURE(status)) { return; }
+    if (id == nullptr || *id == 0) {
+        delete dest;
+        dest = nullptr;
+    } else {
+        if (dest == nullptr) {
+            dest = new CharString(id, status);
+            if (dest == nullptr) {
+                status = U_MEMORY_ALLOCATION_ERROR;
+                return;
+            }
+        } else {
+            dest->copyFrom(id, status);
+        }
     }
-    if (actualID != nullptr) {
-      uprv_strncpy(actual, actualID, ULOC_FULLNAME_CAPACITY);
-      actual[ULOC_FULLNAME_CAPACITY-1] = 0; // always terminate
+}
+
+void LocaleBased::setLocaleID(const CharString* id, CharString*& dest, UErrorCode& status) {
+    if (U_FAILURE(status)) { return; }
+    if (id == nullptr || id->isEmpty()) {
+        delete dest;
+        dest = nullptr;
+    } else {
+        if (dest == nullptr) {
+            dest = new CharString(*id, status);
+            if (dest == nullptr) {
+                status = U_MEMORY_ALLOCATION_ERROR;
+                return;
+            }
+        } else {
+            dest->copyFrom(*id, status);
+        }
     }
 }
 
-void LocaleBased::setLocaleIDs(const Locale& validID, const Locale& actualID) {
-  uprv_strcpy(valid, validID.getName());
-  uprv_strcpy(actual, actualID.getName());
+bool LocaleBased::equalIDs(const CharString* left, const CharString* right) {
+    // true if both are nullptr
+    if (left == nullptr && right == nullptr) return true;
+    // false if only one is nullptr
+    if (left == nullptr || right == nullptr) return false;
+    return *left == *right;
 }
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/common/locbased.h b/deps/icu-small/source/common/locbased.h
index 2d260b527873d3..9441eb823107e3 100644
--- a/deps/icu-small/source/common/locbased.h
+++ b/deps/icu-small/source/common/locbased.h
@@ -19,13 +19,14 @@
 /**
  * Macro to declare a locale LocaleBased wrapper object for the given
  * object, which must have two members named `validLocale' and
- * `actualLocale' of size ULOC_FULLNAME_CAPACITY
+ * `actualLocale' of which are pointers to the internal icu::CharString.
  */
 #define U_LOCALE_BASED(varname, objname) \
   LocaleBased varname((objname).validLocale, (objname).actualLocale)
 
 U_NAMESPACE_BEGIN
 
+class CharString;
 /**
  * A utility class that unifies the implementation of getLocale() by
  * various ICU services.  This class is likely to be removed in the
@@ -41,33 +42,35 @@ class U_COMMON_API LocaleBased : public UMemory {
      * Construct a LocaleBased wrapper around the two pointers.  These
      * will be aliased for the lifetime of this object.
      */
-    inline LocaleBased(char* validAlias, char* actualAlias);
-
-    /**
-     * Construct a LocaleBased wrapper around the two const pointers.
-     * These will be aliased for the lifetime of this object.
-     */
-    inline LocaleBased(const char* validAlias, const char* actualAlias);
+    inline LocaleBased(CharString*& validAlias, CharString*& actualAlias);
 
     /**
      * Return locale meta-data for the service object wrapped by this
      * object.  Either the valid or the actual locale may be
      * retrieved.
+     * @param valid The valid locale.
+     * @param actual The actual locale.
      * @param type either ULOC_VALID_LOCALE or ULOC_ACTUAL_LOCALE
      * @param status input-output error code
      * @return the indicated locale
      */
-    Locale getLocale(ULocDataLocaleType type, UErrorCode& status) const;
+    static Locale getLocale(
+        const CharString* valid, const CharString* actual,
+        ULocDataLocaleType type, UErrorCode& status);
 
     /**
      * Return the locale ID for the service object wrapped by this
      * object.  Either the valid or the actual locale may be
      * retrieved.
+     * @param valid The valid locale.
+     * @param actual The actual locale.
      * @param type either ULOC_VALID_LOCALE or ULOC_ACTUAL_LOCALE
      * @param status input-output error code
      * @return the indicated locale ID
      */
-    const char* getLocaleID(ULocDataLocaleType type, UErrorCode& status) const;
+    static const char* getLocaleID(
+        const CharString* valid, const CharString* actual,
+        ULocDataLocaleType type, UErrorCode& status);
 
     /**
      * Set the locale meta-data for the service object wrapped by this
@@ -75,31 +78,40 @@ class U_COMMON_API LocaleBased : public UMemory {
      * @param valid the ID of the valid locale
      * @param actual the ID of the actual locale
      */
-    void setLocaleIDs(const char* valid, const char* actual);
+    void setLocaleIDs(const char* valid, const char* actual, UErrorCode& status);
+    void setLocaleIDs(const CharString* valid, const CharString* actual, UErrorCode& status);
 
-    /**
-     * Set the locale meta-data for the service object wrapped by this
-     * object.
-     * @param valid the ID of the valid locale
-     * @param actual the ID of the actual locale
-     */
-    void setLocaleIDs(const Locale& valid, const Locale& actual);
+    static void setLocaleID(const char* id, CharString*& dest, UErrorCode& status);
+    static void setLocaleID(const CharString* id, CharString*& dest, UErrorCode& status);
+
+    static bool equalIDs(const CharString* left, const CharString* right);
 
  private:
 
-    char* valid;
-    
-    char* actual;
+    void setValidLocaleID(const CharString* id, UErrorCode& status);
+    void setActualLocaleID(const CharString* id, UErrorCode& status);
+    void setValidLocaleID(const char* id, UErrorCode& status);
+    void setActualLocaleID(const char* id, UErrorCode& status);
+
+    CharString*& valid;
+    CharString*& actual;
 };
 
-inline LocaleBased::LocaleBased(char* validAlias, char* actualAlias) :
+inline LocaleBased::LocaleBased(CharString*& validAlias, CharString*& actualAlias) :
     valid(validAlias), actual(actualAlias) {
 }
 
-inline LocaleBased::LocaleBased(const char* validAlias,
-                                const char* actualAlias) :
-    // ugh: cast away const
-    valid(const_cast<char*>(validAlias)), actual(const_cast<char*>(actualAlias)) {
+inline void LocaleBased::setValidLocaleID(const CharString* id, UErrorCode& status) {
+    setLocaleID(id, valid, status);
+}
+inline void LocaleBased::setActualLocaleID(const CharString* id, UErrorCode& status) {
+    setLocaleID(id, actual, status);
+}
+inline void LocaleBased::setValidLocaleID(const char* id, UErrorCode& status) {
+    setLocaleID(id, valid, status);
+}
+inline void LocaleBased::setActualLocaleID(const char* id, UErrorCode& status) {
+    setLocaleID(id, actual, status);
 }
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/common/locdispnames.cpp b/deps/icu-small/source/common/locdispnames.cpp
index ddf7687a2bf07b..d3521e879b60c8 100644
--- a/deps/icu-small/source/common/locdispnames.cpp
+++ b/deps/icu-small/source/common/locdispnames.cpp
@@ -19,6 +19,8 @@
 *   that then do not depend on resource bundle code and display name data.
 */
 
+#include <string_view>
+
 #include "unicode/utypes.h"
 #include "unicode/brkiter.h"
 #include "unicode/locid.h"
@@ -359,7 +361,7 @@ _getStringOrCopyKey(const char *path, const char *locale,
     return u_terminateUChars(dest, destCapacity, length, &errorCode);
 }
 
-using UDisplayNameGetter = icu::CharString(const char*, UErrorCode&);
+using UDisplayNameGetter = icu::CharString(std::string_view, UErrorCode&);
 
 int32_t
 _getDisplayNameForComponent(const char *locale,
@@ -377,6 +379,10 @@ _getDisplayNameForComponent(const char *locale,
         return 0;
     }
 
+    if (locale == nullptr) {
+        locale = uloc_getDefault();
+    }
+
     localStatus = U_ZERO_ERROR;
     icu::CharString localeBuffer = (*getter)(locale, localStatus);
     if (U_FAILURE(localStatus)) {
diff --git a/deps/icu-small/source/common/locid.cpp b/deps/icu-small/source/common/locid.cpp
index 4a73f559205232..e7e86079ae9169 100644
--- a/deps/icu-small/source/common/locid.cpp
+++ b/deps/icu-small/source/common/locid.cpp
@@ -1828,8 +1828,13 @@ ulocimp_isCanonicalizedLocaleForTest(const char* localeName)
 
 U_NAMESPACE_BEGIN
 
-/*This function initializes a Locale from a C locale ID*/
 Locale& Locale::init(const char* localeID, UBool canonicalize)
+{
+    return localeID == nullptr ? *this = getDefault() : init(StringPiece{localeID}, canonicalize);
+}
+
+/*This function initializes a Locale from a C locale ID*/
+Locale& Locale::init(StringPiece localeID, UBool canonicalize)
 {
     fIsBogus = false;
     /* Free our current storage */
@@ -1854,19 +1859,28 @@ Locale& Locale::init(const char* localeID, UBool canonicalize)
         int32_t length;
         UErrorCode err;
 
-        if(localeID == nullptr) {
-            // not an error, just set the default locale
-            return *this = getDefault();
-        }
-
         /* preset all fields to empty */
         language[0] = script[0] = country[0] = 0;
 
+        const auto parse = [canonicalize](std::string_view localeID,
+                                          char* name,
+                                          int32_t nameCapacity,
+                                          UErrorCode& status) {
+            return ByteSinkUtil::viaByteSinkToTerminatedChars(
+                name, nameCapacity,
+                [&](ByteSink& sink, UErrorCode& status) {
+                    if (canonicalize) {
+                        ulocimp_canonicalize(localeID, sink, status);
+                    } else {
+                        ulocimp_getName(localeID, sink, status);
+                    }
+                },
+                status);
+        };
+
         // "canonicalize" the locale ID to ICU/Java format
         err = U_ZERO_ERROR;
-        length = canonicalize ?
-            uloc_canonicalize(localeID, fullName, sizeof(fullNameBuffer), &err) :
-            uloc_getName(localeID, fullName, sizeof(fullNameBuffer), &err);
+        length = parse(localeID, fullName, sizeof fullNameBuffer, err);
 
         if (err == U_BUFFER_OVERFLOW_ERROR || length >= static_cast<int32_t>(sizeof(fullNameBuffer))) {
             U_ASSERT(baseName == nullptr);
@@ -1877,9 +1891,7 @@ Locale& Locale::init(const char* localeID, UBool canonicalize)
             }
             fullName = newFullName;
             err = U_ZERO_ERROR;
-            length = canonicalize ?
-                uloc_canonicalize(localeID, fullName, length+1, &err) :
-                uloc_getName(localeID, fullName, length+1, &err);
+            length = parse(localeID, fullName, length + 1, err);
         }
         if(U_FAILURE(err) || err == U_STRING_NOT_TERMINATED_WARNING) {
             /* should never occur */
@@ -2200,6 +2212,13 @@ Locale::createFromName (const char *name)
     }
 }
 
+Locale U_EXPORT2
+Locale::createFromName(StringPiece name) {
+    Locale loc("");
+    loc.init(name, false);
+    return loc;
+}
+
 Locale U_EXPORT2
 Locale::createCanonical(const char* name) {
     Locale loc("");
diff --git a/deps/icu-small/source/common/loclikely.cpp b/deps/icu-small/source/common/loclikely.cpp
index ccbcbfa7a5d7a1..f87fd8dd61ca13 100644
--- a/deps/icu-small/source/common/loclikely.cpp
+++ b/deps/icu-small/source/common/loclikely.cpp
@@ -300,6 +300,9 @@ ulocimp_addLikelySubtags(const char* localeID,
                          icu::ByteSink& sink,
                          UErrorCode& status) {
     if (U_FAILURE(status)) { return; }
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     icu::CharString localeBuffer = ulocimp_canonicalize(localeID, status);
     _uloc_addLikelySubtags(localeBuffer.data(), sink, status);
 }
@@ -334,6 +337,9 @@ ulocimp_minimizeSubtags(const char* localeID,
                         bool favorScript,
                         UErrorCode& status) {
     if (U_FAILURE(status)) { return; }
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     icu::CharString localeBuffer = ulocimp_canonicalize(localeID, status);
     _uloc_minimizeSubtags(localeBuffer.data(), sink, favorScript, status);
 }
@@ -349,7 +355,9 @@ uloc_isRightToLeft(const char *locale) {
     UErrorCode errorCode = U_ZERO_ERROR;
     icu::CharString lang;
     icu::CharString script;
-    ulocimp_getSubtags(locale, &lang, &script, nullptr, nullptr, nullptr, errorCode);
+    ulocimp_getSubtags(
+        locale == nullptr ? uloc_getDefault() : locale,
+        &lang, &script, nullptr, nullptr, nullptr, errorCode);
     if (U_FAILURE(errorCode) || script.isEmpty()) {
         // Fastpath: We know the likely scripts and their writing direction
         // for some common languages.
@@ -369,7 +377,7 @@ uloc_isRightToLeft(const char *locale) {
         if (U_FAILURE(errorCode)) {
             return false;
         }
-        ulocimp_getSubtags(likely.data(), nullptr, &script, nullptr, nullptr, nullptr, errorCode);
+        ulocimp_getSubtags(likely.toStringPiece(), nullptr, &script, nullptr, nullptr, nullptr, errorCode);
         if (U_FAILURE(errorCode) || script.isEmpty()) {
             return false;
         }
@@ -430,7 +438,7 @@ ulocimp_getRegionForSupplementalData(const char *localeID, bool inferRegion,
     icu::CharString rgBuf = GetRegionFromKey(localeID, "rg", status);
     if (U_SUCCESS(status) && rgBuf.isEmpty()) {
         // No valid rg keyword value, try for unicode_region_subtag
-        rgBuf = ulocimp_getRegion(localeID, status);
+        rgBuf = ulocimp_getRegion(localeID == nullptr ? uloc_getDefault() : localeID, status);
         if (U_SUCCESS(status) && rgBuf.isEmpty() && inferRegion) {
             // Second check for sd keyword value
             rgBuf = GetRegionFromKey(localeID, "sd", status);
@@ -439,7 +447,7 @@ ulocimp_getRegionForSupplementalData(const char *localeID, bool inferRegion,
                 UErrorCode rgStatus = U_ZERO_ERROR;
                 icu::CharString locBuf = ulocimp_addLikelySubtags(localeID, rgStatus);
                 if (U_SUCCESS(rgStatus)) {
-                    rgBuf = ulocimp_getRegion(locBuf.data(), status);
+                    rgBuf = ulocimp_getRegion(locBuf.toStringPiece(), status);
                 }
             }
         }
diff --git a/deps/icu-small/source/common/loclikelysubtags.cpp b/deps/icu-small/source/common/loclikelysubtags.cpp
index 7c6131197d894b..7245a779816ce7 100644
--- a/deps/icu-small/source/common/loclikelysubtags.cpp
+++ b/deps/icu-small/source/common/loclikelysubtags.cpp
@@ -527,7 +527,7 @@ LSR LikelySubtags::makeMaximizedLsrFrom(const Locale &locale,
         return {};
     }
     const char *name = locale.getName();
-    if (uprv_isAtSign(name[0]) && name[1] == 'x' && name[2] == '=') {  // name.startsWith("@x=")
+    if (!returnInputIfUnmatch && uprv_isAtSign(name[0]) && name[1] == 'x' && name[2] == '=') {  // name.startsWith("@x=")
         // Private use language tag x-subtag-subtag... which CLDR changes to
         // und-x-subtag-subtag...
         return LSR(name, "", "", LSR::EXPLICIT_LSR);
diff --git a/deps/icu-small/source/common/locresdata.cpp b/deps/icu-small/source/common/locresdata.cpp
index 725e6609159896..ba7163fa2dbac7 100644
--- a/deps/icu-small/source/common/locresdata.cpp
+++ b/deps/icu-small/source/common/locresdata.cpp
@@ -161,6 +161,9 @@ _uloc_getOrientationHelper(const char* localeId,
 
     if (U_FAILURE(status)) { return result; }
 
+    if (localeId == nullptr) {
+        localeId = uloc_getDefault();
+    }
     icu::CharString localeBuffer = ulocimp_canonicalize(localeId, status);
 
     if (U_FAILURE(status)) { return result; }
diff --git a/deps/icu-small/source/common/punycode.cpp b/deps/icu-small/source/common/punycode.cpp
index aa02298c5e6d07..1868a07a856a78 100644
--- a/deps/icu-small/source/common/punycode.cpp
+++ b/deps/icu-small/source/common/punycode.cpp
@@ -193,7 +193,7 @@ u_strToPunycode(const char16_t *src, int32_t srcLength,
         return 0;
     }
 
-    if(src==nullptr || srcLength<-1 || (dest==nullptr && destCapacity!=0)) {
+    if(src==nullptr || srcLength<-1 || destCapacity<0 || (dest==nullptr && destCapacity!=0)) {
         *pErrorCode=U_ILLEGAL_ARGUMENT_ERROR;
         return 0;
     }
diff --git a/deps/icu-small/source/common/putil.cpp b/deps/icu-small/source/common/putil.cpp
index 4cf07797ba350e..ea15fdff0b0c67 100644
--- a/deps/icu-small/source/common/putil.cpp
+++ b/deps/icu-small/source/common/putil.cpp
@@ -76,7 +76,7 @@
 #include <float.h>
 
 #ifndef U_COMMON_IMPLEMENTATION
-#error U_COMMON_IMPLEMENTATION not set - must be set for all ICU source files in common/ - see https://unicode-org.github.io/icu/userguide/howtouseicu
+#error U_COMMON_IMPLEMENTATION not set - must be set for all ICU source files in common/ - see https://unicode-org.github.io/icu/userguide/icu/howtouseicu.html
 #endif
 
 
diff --git a/deps/icu-small/source/common/rbbinode.cpp b/deps/icu-small/source/common/rbbinode.cpp
index 71407b9e684eea..849ee7180a22f1 100644
--- a/deps/icu-small/source/common/rbbinode.cpp
+++ b/deps/icu-small/source/common/rbbinode.cpp
@@ -47,7 +47,10 @@ static int  gLastSerial = 0;
 //    Constructor.   Just set the fields to reasonable default values.
 //
 //-------------------------------------------------------------------------
-RBBINode::RBBINode(NodeType t) : UMemory() {
+RBBINode::RBBINode(NodeType t, UErrorCode& status) : UMemory() {
+    if (U_FAILURE(status)) {
+        return;
+    }
 #ifdef RBBI_DEBUG
     fSerialNum    = ++gLastSerial;
 #endif
@@ -65,10 +68,13 @@ RBBINode::RBBINode(NodeType t) : UMemory() {
     fVal          = 0;
     fPrecedence   = precZero;
 
-    UErrorCode     status = U_ZERO_ERROR;
-    fFirstPosSet  = new UVector(status);  // TODO - get a real status from somewhere
+    fFirstPosSet  = new UVector(status);
     fLastPosSet   = new UVector(status);
     fFollowPos    = new UVector(status);
+    if (U_SUCCESS(status) &&
+        (fFirstPosSet == nullptr || fLastPosSet == nullptr || fFollowPos == nullptr)) {
+        status =  U_MEMORY_ALLOCATION_ERROR;
+    }
     if      (t==opCat)    {fPrecedence = precOpCat;}
     else if (t==opOr)     {fPrecedence = precOpOr;}
     else if (t==opStart)  {fPrecedence = precStart;}
@@ -77,7 +83,10 @@ RBBINode::RBBINode(NodeType t) : UMemory() {
 }
 
 
-RBBINode::RBBINode(const RBBINode &other) : UMemory(other) {
+RBBINode::RBBINode(const RBBINode &other, UErrorCode& status) : UMemory(other) {
+    if (U_FAILURE(status)) {
+        return;
+    }
 #ifdef RBBI_DEBUG
     fSerialNum   = ++gLastSerial;
 #endif
@@ -94,10 +103,13 @@ RBBINode::RBBINode(const RBBINode &other) : UMemory(other) {
     fVal         = other.fVal;
     fRuleRoot    = false;
     fChainIn     = other.fChainIn;
-    UErrorCode     status = U_ZERO_ERROR;
     fFirstPosSet = new UVector(status);   // TODO - get a real status from somewhere
     fLastPosSet  = new UVector(status);
     fFollowPos   = new UVector(status);
+    if (U_SUCCESS(status) &&
+        (fFirstPosSet == nullptr || fLastPosSet == nullptr || fFollowPos == nullptr)) {
+        status =  U_MEMORY_ALLOCATION_ERROR;
+    }
 }
 
 
@@ -193,27 +205,54 @@ void RBBINode::NRDeleteNode(RBBINode *node) {
 //                  references in preparation for generating the DFA tables.
 //
 //-------------------------------------------------------------------------
-RBBINode *RBBINode::cloneTree() {
+constexpr int kRecursiveDepthLimit = 3500;
+RBBINode *RBBINode::cloneTree(UErrorCode &status, int depth) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    // If the depth of the stack is too deep, we return U_INPUT_TOO_LONG_ERROR
+    // to avoid stack overflow crash.
+    if (depth > kRecursiveDepthLimit) {
+        status = U_INPUT_TOO_LONG_ERROR;
+        return nullptr;
+    }
     RBBINode    *n;
 
     if (fType == RBBINode::varRef) {
         // If the current node is a variable reference, skip over it
         //   and clone the definition of the variable instead.
-        n = fLeftChild->cloneTree();
+        n = fLeftChild->cloneTree(status, depth+1);
+        if (U_FAILURE(status)) {
+            return nullptr;
+        }
     } else if (fType == RBBINode::uset) {
         n = this;
     } else {
-        n = new RBBINode(*this);
+        n = new RBBINode(*this, status);
+        if (U_FAILURE(status)) {
+            delete n;
+            return nullptr;
+        }
         // Check for null pointer.
-        if (n != nullptr) {
-            if (fLeftChild != nullptr) {
-                n->fLeftChild          = fLeftChild->cloneTree();
-                n->fLeftChild->fParent = n;
+        if (n == nullptr) {
+            status =  U_MEMORY_ALLOCATION_ERROR;
+            return nullptr;
+        }
+        if (fLeftChild != nullptr) {
+            n->fLeftChild          = fLeftChild->cloneTree(status, depth+1);
+            if (U_FAILURE(status)) {
+                delete n;
+                return nullptr;
             }
-            if (fRightChild != nullptr) {
-                n->fRightChild          = fRightChild->cloneTree();
-                n->fRightChild->fParent = n;
+            n->fLeftChild->fParent = n;
+        }
+        if (fRightChild != nullptr) {
+            n->fRightChild          = fRightChild->cloneTree(status, depth+1);
+            if (U_FAILURE(status)) {
+                delete n;
+                return nullptr;
             }
+            n->fRightChild->fParent = n;
         }
     }
     return n;
@@ -239,7 +278,6 @@ RBBINode *RBBINode::cloneTree() {
 //                      nested references are handled by cloneTree(), not here.
 //
 //-------------------------------------------------------------------------
-constexpr int kRecursiveDepthLimit = 3500;
 RBBINode *RBBINode::flattenVariables(UErrorCode& status, int depth) {
     if (U_FAILURE(status)) {
         return this;
@@ -251,21 +289,34 @@ RBBINode *RBBINode::flattenVariables(UErrorCode& status, int depth) {
         return this;
     }
     if (fType == varRef) {
-        RBBINode *retNode  = fLeftChild->cloneTree();
-        if (retNode != nullptr) {
-            retNode->fRuleRoot = this->fRuleRoot;
-            retNode->fChainIn  = this->fChainIn;
+        RBBINode *retNode  = fLeftChild->cloneTree(status, depth+1);
+        if (U_FAILURE(status)) {
+            return this;
         }
+        retNode->fRuleRoot = this->fRuleRoot;
+        retNode->fChainIn  = this->fChainIn;
         delete this;   // TODO: undefined behavior. Fix.
         return retNode;
     }
 
     if (fLeftChild != nullptr) {
         fLeftChild = fLeftChild->flattenVariables(status, depth+1);
+        if (fLeftChild == nullptr) {
+            status = U_MEMORY_ALLOCATION_ERROR;
+        }
+        if (U_FAILURE(status)) {
+            return this;
+        }
         fLeftChild->fParent  = this;
     }
     if (fRightChild != nullptr) {
         fRightChild = fRightChild->flattenVariables(status, depth+1);
+        if (fRightChild == nullptr) {
+            status = U_MEMORY_ALLOCATION_ERROR;
+        }
+        if (U_FAILURE(status)) {
+            return this;
+        }
         fRightChild->fParent = this;
     }
     return this;
@@ -280,7 +331,16 @@ RBBINode *RBBINode::flattenVariables(UErrorCode& status, int depth) {
 //                 the left child of the uset node.
 //
 //-------------------------------------------------------------------------
-void RBBINode::flattenSets() {
+void RBBINode::flattenSets(UErrorCode &status, int depth) {
+    if (U_FAILURE(status)) {
+        return;
+    }
+    // If the depth of the stack is too deep, we return U_INPUT_TOO_LONG_ERROR
+    // to avoid stack overflow crash.
+    if (depth > kRecursiveDepthLimit) {
+        status = U_INPUT_TOO_LONG_ERROR;
+        return;
+    }
     U_ASSERT(fType != setRef);
 
     if (fLeftChild != nullptr) {
@@ -288,11 +348,15 @@ void RBBINode::flattenSets() {
             RBBINode *setRefNode = fLeftChild;
             RBBINode *usetNode   = setRefNode->fLeftChild;
             RBBINode *replTree   = usetNode->fLeftChild;
-            fLeftChild           = replTree->cloneTree();
+            fLeftChild           = replTree->cloneTree(status, depth+1);
+            if (U_FAILURE(status)) {
+                delete setRefNode;
+                return;
+            }
             fLeftChild->fParent  = this;
             delete setRefNode;
         } else {
-            fLeftChild->flattenSets();
+            fLeftChild->flattenSets(status, depth+1);
         }
     }
 
@@ -301,11 +365,15 @@ void RBBINode::flattenSets() {
             RBBINode *setRefNode = fRightChild;
             RBBINode *usetNode   = setRefNode->fLeftChild;
             RBBINode *replTree   = usetNode->fLeftChild;
-            fRightChild           = replTree->cloneTree();
+            fRightChild           = replTree->cloneTree(status, depth+1);
+            if (U_FAILURE(status)) {
+                delete setRefNode;
+                return;
+            }
             fRightChild->fParent  = this;
             delete setRefNode;
         } else {
-            fRightChild->flattenSets();
+            fRightChild->flattenSets(status, depth+1);
         }
     }
 }
diff --git a/deps/icu-small/source/common/rbbinode.h b/deps/icu-small/source/common/rbbinode.h
index 497a31b8d098b3..8fbc9d1b588e05 100644
--- a/deps/icu-small/source/common/rbbinode.h
+++ b/deps/icu-small/source/common/rbbinode.h
@@ -91,14 +91,14 @@ class RBBINode : public UMemory {
         UVector       *fFollowPos;
 
 
-        RBBINode(NodeType t);
-        RBBINode(const RBBINode &other);
+        RBBINode(NodeType t, UErrorCode& status);
+        RBBINode(const RBBINode &other, UErrorCode& status);
         ~RBBINode();
         static void  NRDeleteNode(RBBINode *node);
         
-        RBBINode    *cloneTree();
+        RBBINode    *cloneTree(UErrorCode &status, int depth=0);
         RBBINode    *flattenVariables(UErrorCode &status, int depth=0);
-        void         flattenSets();
+        void         flattenSets(UErrorCode &status, int depth=0);
         void         findNodes(UVector *dest, RBBINode::NodeType kind, UErrorCode &status);
 
 #ifdef RBBI_DEBUG
diff --git a/deps/icu-small/source/common/rbbiscan.cpp b/deps/icu-small/source/common/rbbiscan.cpp
index cf2d63cd807b0f..77fc3bcd9b7e29 100644
--- a/deps/icu-small/source/common/rbbiscan.cpp
+++ b/deps/icu-small/source/common/rbbiscan.cpp
@@ -767,15 +767,24 @@ void RBBIRuleScanner::findSetFor(const UnicodeString &s, RBBINode *node, Unicode
             c = s.char32At(0);
             setToAdopt = new UnicodeSet(c, c);
         }
+        if (setToAdopt == nullptr) {
+            error(U_MEMORY_ALLOCATION_ERROR);
+            return;
+        }
     }
 
     //
     // Make a new uset node to refer to this UnicodeSet
     // This new uset node becomes the child of the caller's setReference node.
     //
-    RBBINode *usetNode    = new RBBINode(RBBINode::uset);
+    UErrorCode localStatus = U_ZERO_ERROR;
+    RBBINode *usetNode    = new RBBINode(RBBINode::uset, localStatus);
     if (usetNode == nullptr) {
-        error(U_MEMORY_ALLOCATION_ERROR);
+        localStatus = U_MEMORY_ALLOCATION_ERROR;
+    }
+    if (U_FAILURE(localStatus)) {
+        delete usetNode;
+        error(localStatus);
         delete setToAdopt;
         return;
     }
@@ -1191,7 +1200,7 @@ RBBINode  *RBBIRuleScanner::pushNewNode(RBBINode::NodeType  t) {
         return nullptr;
     }
     fNodeStackPtr++;
-    fNodeStack[fNodeStackPtr] = new RBBINode(t);
+    fNodeStack[fNodeStackPtr] = new RBBINode(t, *fRB->fStatus);
     if (fNodeStack[fNodeStackPtr] == nullptr) {
         *fRB->fStatus = U_MEMORY_ALLOCATION_ERROR;
     }
diff --git a/deps/icu-small/source/common/rbbisetb.cpp b/deps/icu-small/source/common/rbbisetb.cpp
index 6c22cf470f8b2f..df94fc8bc4fb86 100644
--- a/deps/icu-small/source/common/rbbisetb.cpp
+++ b/deps/icu-small/source/common/rbbisetb.cpp
@@ -375,7 +375,11 @@ void  RBBISetBuilder::addValToSets(UVector *sets, uint32_t val) {
 }
 
 void  RBBISetBuilder::addValToSet(RBBINode *usetNode, uint32_t val) {
-    RBBINode *leafNode = new RBBINode(RBBINode::leafChar);
+    RBBINode *leafNode = new RBBINode(RBBINode::leafChar, *fStatus);
+    if (U_FAILURE(*fStatus)) {
+        delete leafNode;
+        return;
+    }
     if (leafNode == nullptr) {
         *fStatus = U_MEMORY_ALLOCATION_ERROR;
         return;
@@ -388,9 +392,13 @@ void  RBBISetBuilder::addValToSet(RBBINode *usetNode, uint32_t val) {
         // There are already input symbols present for this set.
         // Set up an OR node, with the previous stuff as the left child
         //   and the new value as the right child.
-        RBBINode *orNode = new RBBINode(RBBINode::opOr);
+        RBBINode *orNode = new RBBINode(RBBINode::opOr, *fStatus);
         if (orNode == nullptr) {
             *fStatus = U_MEMORY_ALLOCATION_ERROR;
+        }
+        if (U_FAILURE(*fStatus)) {
+            delete orNode;
+            delete leafNode;
             return;
         }
         orNode->fLeftChild  = usetNode->fLeftChild;
diff --git a/deps/icu-small/source/common/rbbitblb.cpp b/deps/icu-small/source/common/rbbitblb.cpp
index 4d95137601efb1..b89909807c2023 100644
--- a/deps/icu-small/source/common/rbbitblb.cpp
+++ b/deps/icu-small/source/common/rbbitblb.cpp
@@ -99,13 +99,22 @@ void  RBBITableBuilder::buildForwardTable() {
     //   {bof} fake character.
     // 
     if (fRB->fSetBuilder->sawBOF()) {
-        RBBINode *bofTop    = new RBBINode(RBBINode::opCat);
-        RBBINode *bofLeaf   = new RBBINode(RBBINode::leafChar);
-        // Delete and exit if memory allocation failed.
-        if (bofTop == nullptr || bofLeaf == nullptr) {
+        RBBINode *bofTop    = new RBBINode(RBBINode::opCat, *fStatus);
+        if (bofTop == nullptr) {
             *fStatus = U_MEMORY_ALLOCATION_ERROR;
+        }
+        if (U_FAILURE(*fStatus)) {
             delete bofTop;
+            return;
+        }
+        RBBINode *bofLeaf   = new RBBINode(RBBINode::leafChar, *fStatus);
+        // Delete and exit if memory allocation failed.
+        if (bofLeaf == nullptr) {
+            *fStatus = U_MEMORY_ALLOCATION_ERROR;
+        }
+        if (U_FAILURE(*fStatus)) {
             delete bofLeaf;
+            delete bofTop;
             return;
         }
         bofTop->fLeftChild  = bofLeaf;
@@ -120,18 +129,23 @@ void  RBBITableBuilder::buildForwardTable() {
     //   Appears as a cat-node, left child being the original tree,
     //   right child being the end marker.
     //
-    RBBINode *cn = new RBBINode(RBBINode::opCat);
+    RBBINode *cn = new RBBINode(RBBINode::opCat, *fStatus);
     // Exit if memory allocation failed.
     if (cn == nullptr) {
         *fStatus = U_MEMORY_ALLOCATION_ERROR;
+    }
+    if (U_FAILURE(*fStatus)) {
+        delete cn;
         return;
     }
     cn->fLeftChild = fTree;
     fTree->fParent = cn;
-    RBBINode *endMarkerNode = cn->fRightChild = new RBBINode(RBBINode::endMark);
+    RBBINode *endMarkerNode = cn->fRightChild = new RBBINode(RBBINode::endMark, *fStatus);
     // Delete and exit if memory allocation failed.
     if (cn->fRightChild == nullptr) {
         *fStatus = U_MEMORY_ALLOCATION_ERROR;
+    }
+    if (U_FAILURE(*fStatus)) {
         delete cn;
         return;
     }
@@ -142,7 +156,7 @@ void  RBBITableBuilder::buildForwardTable() {
     //  Replace all references to UnicodeSets with the tree for the equivalent
     //      expression.
     //
-    fTree->flattenSets();
+    fTree->flattenSets(*fStatus, 0);
 #ifdef RBBI_DEBUG
     if (fRB->fDebugEnv && uprv_strstr(fRB->fDebugEnv, "stree")) {
         RBBIDebugPuts("\nParse tree after flattening Unicode Set references.");
diff --git a/deps/icu-small/source/common/resbund.cpp b/deps/icu-small/source/common/resbund.cpp
index 41337cdc155daf..4c14dae133eefc 100644
--- a/deps/icu-small/source/common/resbund.cpp
+++ b/deps/icu-small/source/common/resbund.cpp
@@ -388,7 +388,7 @@ const Locale &ResourceBundle::getLocale() const {
     return ncThis->fLocale != nullptr ? *ncThis->fLocale : Locale::getDefault();
 }
 
-const Locale ResourceBundle::getLocale(ULocDataLocaleType type, UErrorCode &status) const
+Locale ResourceBundle::getLocale(ULocDataLocaleType type, UErrorCode &status) const
 {
   return ures_getLocaleByType(fResource, type, &status);
 }
diff --git a/deps/icu-small/source/common/ucnvmbcs.cpp b/deps/icu-small/source/common/ucnvmbcs.cpp
index f5507043bf35be..d65c284746163b 100644
--- a/deps/icu-small/source/common/ucnvmbcs.cpp
+++ b/deps/icu-small/source/common/ucnvmbcs.cpp
@@ -3146,11 +3146,8 @@ ucnv_MBCSGetNextUChar(UConverterToUnicodeArgs *pArgs,
     if(c<0) {
         if(U_SUCCESS(*pErrorCode) && source==sourceLimit && lastSource<source) {
             /* incomplete character byte sequence */
-            uint8_t *bytes=cnv->toUBytes;
             cnv->toULength = static_cast<int8_t>(source - lastSource);
-            do {
-                *bytes++=*lastSource++;
-            } while(lastSource<source);
+            uprv_memcpy(cnv->toUBytes, lastSource, cnv->toULength);
             *pErrorCode=U_TRUNCATED_CHAR_FOUND;
         } else if(U_FAILURE(*pErrorCode)) {
             /* callback(illegal) */
diff --git a/deps/icu-small/source/common/ucurr.cpp b/deps/icu-small/source/common/ucurr.cpp
index b74a80a676afc5..cccf1130ae8e63 100644
--- a/deps/icu-small/source/common/ucurr.cpp
+++ b/deps/icu-small/source/common/ucurr.cpp
@@ -372,12 +372,8 @@ struct CReg : public icu::UMemory {
     CReg(const char16_t* _iso, const char* _id)
         : next(nullptr)
     {
-        int32_t len = static_cast<int32_t>(uprv_strlen(_id));
-        if (len > static_cast<int32_t>(sizeof(id) - 1)) {
-            len = (sizeof(id)-1);
-        }
-        uprv_strncpy(id, _id, len);
-        id[len] = 0;
+        uprv_strncpy(id, _id, sizeof(id)-1);
+        id[sizeof(id)-1] = 0;
         u_memcpy(iso, _iso, ISO_CURRENCY_CODE_LENGTH);
         iso[ISO_CURRENCY_CODE_LENGTH] = 0;
     }
@@ -682,6 +678,9 @@ ucurr_getName(const char16_t* currency,
     // this function.
     UErrorCode ec2 = U_ZERO_ERROR;
 
+    if (locale == nullptr) {
+        locale = uloc_getDefault();
+    }
     CharString loc = ulocimp_getName(locale, ec2);
     if (U_FAILURE(ec2)) {
         *ec = U_ILLEGAL_ARGUMENT_ERROR;
@@ -780,6 +779,9 @@ ucurr_getPluralName(const char16_t* currency,
     // this function.
     UErrorCode ec2 = U_ZERO_ERROR;
 
+    if (locale == nullptr) {
+        locale = uloc_getDefault();
+    }
     CharString loc = ulocimp_getName(locale, ec2);
     if (U_FAILURE(ec2)) {
         *ec = U_ILLEGAL_ARGUMENT_ERROR;
@@ -973,6 +975,9 @@ collectCurrencyNames(const char* locale,
     // Look up the Currencies resource for the given locale.
     UErrorCode ec2 = U_ZERO_ERROR;
 
+    if (locale == nullptr) {
+        locale = uloc_getDefault();
+    }
     CharString loc = ulocimp_getName(locale, ec2);
     if (U_FAILURE(ec2)) {
         ec = U_ILLEGAL_ARGUMENT_ERROR;
diff --git a/deps/icu-small/source/common/uloc.cpp b/deps/icu-small/source/common/uloc.cpp
index 51887c97c3e1e3..bea4827a04984c 100644
--- a/deps/icu-small/source/common/uloc.cpp
+++ b/deps/icu-small/source/common/uloc.cpp
@@ -482,8 +482,8 @@ constexpr CanonicalizationMap CANONICALIZE_MAP[] = {
 
 /* ### BCP47 Conversion *******************************************/
 /* Gets the size of the shortest subtag in the given localeID. */
-int32_t getShortestSubtagLength(const char *localeID) {
-    int32_t localeIDLength = static_cast<int32_t>(uprv_strlen(localeID));
+int32_t getShortestSubtagLength(std::string_view localeID) {
+    int32_t localeIDLength = static_cast<int32_t>(localeID.length());
     int32_t length = localeIDLength;
     int32_t tmpLength = 0;
     int32_t i;
@@ -507,8 +507,8 @@ int32_t getShortestSubtagLength(const char *localeID) {
     return length;
 }
 /* Test if the locale id has BCP47 u extension and does not have '@' */
-inline bool _hasBCP47Extension(const char *id) {
-    return id != nullptr && uprv_strstr(id, "@") == nullptr && getShortestSubtagLength(id) == 1;
+inline bool _hasBCP47Extension(std::string_view id) {
+    return id.find('@') == std::string_view::npos && getShortestSubtagLength(id) == 1;
 }
 
 /* ### Keywords **************************************************/
@@ -523,10 +523,9 @@ inline bool UPRV_OK_VALUE_PUNCTUATION(char c) { return c == '_' || c == '-' || c
 #define ULOC_MAX_NO_KEYWORDS 25
 
 U_CAPI const char * U_EXPORT2
-locale_getKeywordsStart(const char *localeID) {
-    const char *result = nullptr;
-    if((result = uprv_strchr(localeID, '@')) != nullptr) {
-        return result;
+locale_getKeywordsStart(std::string_view localeID) {
+    if (size_t pos = localeID.find('@'); pos != std::string_view::npos) {
+        return localeID.data() + pos;
     }
 #if (U_CHARSET_FAMILY == U_EBCDIC_FAMILY)
     else {
@@ -536,8 +535,8 @@ locale_getKeywordsStart(const char *localeID) {
         static const uint8_t ebcdicSigns[] = { 0x7C, 0x44, 0x66, 0x80, 0xAC, 0xAE, 0xAF, 0xB5, 0xEC, 0xEF, 0x00 };
         const uint8_t *charToFind = ebcdicSigns;
         while(*charToFind) {
-            if((result = uprv_strchr(localeID, *charToFind)) != nullptr) {
-                return result;
+            if (size_t pos = localeID.find(*charToFind); pos != std::string_view::npos) {
+                return localeID.data() + pos;
             }
             charToFind++;
         }
@@ -590,7 +589,7 @@ compareKeywordStructs(const void * /*context*/, const void *left, const void *ri
 }  // namespace
 
 U_EXPORT CharString
-ulocimp_getKeywords(const char* localeID,
+ulocimp_getKeywords(std::string_view localeID,
                     char prev,
                     bool valuesToo,
                     UErrorCode& status)
@@ -607,7 +606,7 @@ ulocimp_getKeywords(const char* localeID,
 }
 
 U_EXPORT void
-ulocimp_getKeywords(const char* localeID,
+ulocimp_getKeywords(std::string_view localeID,
                     char prev,
                     ByteSink& sink,
                     bool valuesToo,
@@ -619,9 +618,8 @@ ulocimp_getKeywords(const char* localeID,
 
     int32_t maxKeywords = ULOC_MAX_NO_KEYWORDS;
     int32_t numKeywords = 0;
-    const char* pos = localeID;
-    const char* equalSign = nullptr;
-    const char* semicolon = nullptr;
+    size_t equalSign = std::string_view::npos;
+    size_t semicolon = std::string_view::npos;
     int32_t i = 0, j, n;
 
     if(prev == '@') { /* start of keyword definition */
@@ -629,74 +627,72 @@ ulocimp_getKeywords(const char* localeID,
         do {
             bool duplicate = false;
             /* skip leading spaces */
-            while(*pos == ' ') {
-                pos++;
+            while (localeID.front() == ' ') {
+                localeID.remove_prefix(1);
             }
-            if (!*pos) { /* handle trailing "; " */
+            if (localeID.empty()) { /* handle trailing "; " */
                 break;
             }
             if(numKeywords == maxKeywords) {
                 status = U_INTERNAL_PROGRAM_ERROR;
                 return;
             }
-            equalSign = uprv_strchr(pos, '=');
-            semicolon = uprv_strchr(pos, ';');
+            equalSign = localeID.find('=');
+            semicolon = localeID.find(';');
             /* lack of '=' [foo@currency] is illegal */
             /* ';' before '=' [foo@currency;collation=pinyin] is illegal */
-            if(!equalSign || (semicolon && semicolon<equalSign)) {
+            if (equalSign == std::string_view::npos ||
+                (semicolon != std::string_view::npos && semicolon < equalSign)) {
+                status = U_INVALID_FORMAT_ERROR;
+                return;
+            }
+            /* zero-length keyword is an error. */
+            if (equalSign == 0) {
                 status = U_INVALID_FORMAT_ERROR;
                 return;
             }
             /* need to normalize both keyword and keyword name */
-            if(equalSign - pos >= ULOC_KEYWORD_BUFFER_LEN) {
+            if (equalSign >= ULOC_KEYWORD_BUFFER_LEN) {
                 /* keyword name too long for internal buffer */
                 status = U_INTERNAL_PROGRAM_ERROR;
                 return;
             }
-            for(i = 0, n = 0; i < equalSign - pos; ++i) {
-                if (pos[i] != ' ') {
-                    keywordList[numKeywords].keyword[n++] = uprv_tolower(pos[i]);
+            for (i = 0, n = 0; static_cast<size_t>(i) < equalSign; ++i) {
+                if (localeID[i] != ' ') {
+                    keywordList[numKeywords].keyword[n++] = uprv_tolower(localeID[i]);
                 }
             }
 
-            /* zero-length keyword is an error. */
-            if (n == 0) {
-                status = U_INVALID_FORMAT_ERROR;
-                return;
-            }
-
             keywordList[numKeywords].keyword[n] = 0;
             keywordList[numKeywords].keywordLen = n;
             /* now grab the value part. First we skip the '=' */
             equalSign++;
             /* then we leading spaces */
-            while(*equalSign == ' ') {
+            while (equalSign < localeID.length() && localeID[equalSign] == ' ') {
                 equalSign++;
             }
 
             /* Premature end or zero-length value */
-            if (!*equalSign || equalSign == semicolon) {
+            if (equalSign == localeID.length() || equalSign == semicolon) {
                 status = U_INVALID_FORMAT_ERROR;
                 return;
             }
 
-            keywordList[numKeywords].valueStart = equalSign;
+            keywordList[numKeywords].valueStart = localeID.data() + equalSign;
 
-            pos = semicolon;
-            i = 0;
-            if(pos) {
-                while(*(pos - i - 1) == ' ') {
-                    i++;
-                }
-                keywordList[numKeywords].valueLen = static_cast<int32_t>(pos - equalSign - i);
-                pos++;
+            std::string_view value = localeID;
+            if (semicolon != std::string_view::npos) {
+                value.remove_suffix(value.length() - semicolon);
+                localeID.remove_prefix(semicolon + 1);
             } else {
-                i = static_cast<int32_t>(uprv_strlen(equalSign));
-                while(i && equalSign[i-1] == ' ') {
-                    i--;
-                }
-                keywordList[numKeywords].valueLen = i;
+                localeID = {};
             }
+            value.remove_prefix(equalSign);
+            if (size_t last = value.find_last_not_of(' '); last != std::string_view::npos) {
+                value.remove_suffix(value.length() - last - 1);
+            }
+            keywordList[numKeywords].valueLen = static_cast<int32_t>(value.length());
+
             /* If this is a duplicate keyword, then ignore it */
             for (j=0; j<numKeywords; ++j) {
                 if (uprv_strcmp(keywordList[j].keyword, keywordList[numKeywords].keyword) == 0) {
@@ -707,7 +703,7 @@ ulocimp_getKeywords(const char* localeID,
             if (!duplicate) {
                 ++numKeywords;
             }
-        } while(pos);
+        } while (!localeID.empty());
 
         /* now we have a list of keywords */
         /* we need to sort it */
@@ -784,7 +780,7 @@ ulocimp_getKeywordValue(const char* localeID,
       return;
     }
 
-    if (_hasBCP47Extension(localeID)) {
+    if (localeID != nullptr && _hasBCP47Extension(localeID)) {
         tempBuffer = ulocimp_forLanguageTag(localeID, -1, nullptr, status);
         tmpLocaleID = U_SUCCESS(status) && !tempBuffer.isEmpty() ? tempBuffer.data() : localeID;
     } else {
@@ -889,7 +885,8 @@ uloc_setKeywordValue(const char* keywordName,
         return 0;
     }
 
-    char* keywords = const_cast<char*>(locale_getKeywordsStart(buffer));
+    char* keywords = const_cast<char*>(
+        locale_getKeywordsStart({buffer, static_cast<std::string_view::size_type>(bufLen)}));
     int32_t baseLen = keywords == nullptr ? bufLen : keywords - buffer;
     // Remove -1 from the capacity so that this function can guarantee NUL termination.
     CheckedArrayByteSink sink(keywords == nullptr ? buffer + bufLen : keywords,
@@ -921,7 +918,7 @@ ulocimp_setKeywordValue(std::string_view keywordName,
 {
     if (U_FAILURE(status)) { return; }
     std::string_view keywords;
-    if (const char* start = locale_getKeywordsStart(localeID.data()); start != nullptr) {
+    if (const char* start = locale_getKeywordsStart(localeID.toStringPiece()); start != nullptr) {
         // This is safe because CharString::truncate() doesn't actually erase any
         // data, but simply sets the position for where new data will be written.
         int32_t size = start - localeID.data();
@@ -1138,15 +1135,18 @@ inline bool _isPrefixLetter(char a) { return a == 'x' || a == 'X' || a == 'i' ||
 
 /*returns true if one of the special prefixes is here (s=string)
   'x-' or 'i-' */
-inline bool _isIDPrefix(const char *s) { return _isPrefixLetter(s[0]) && _isIDSeparator(s[1]); }
+inline bool _isIDPrefix(std::string_view s) {
+    return s.size() >= 2 && _isPrefixLetter(s[0]) && _isIDSeparator(s[1]);
+}
 
 /* Dot terminates it because of POSIX form  where dot precedes the codepage
  * except for variant
  */
-inline bool _isTerminator(char a) { return a == 0 || a == '.' || a == '@'; }
+inline bool _isTerminator(char a) { return a == '.' || a == '@'; }
 
-inline bool _isBCP47Extension(const char* p) {
-    return p[0] == '-' &&
+inline bool _isBCP47Extension(std::string_view p) {
+    return p.size() >= 3 &&
+           p[0] == '-' &&
            (p[1] == 't' || p[1] == 'T' ||
             p[1] == 'u' || p[1] == 'U' ||
             p[1] == 'x' || p[1] == 'X') &&
@@ -1202,49 +1202,44 @@ namespace {
  * TODO try to use this in Locale
  */
 
-void
-_getLanguage(const char* localeID,
-             ByteSink* sink,
-             const char** pEnd,
-             UErrorCode& status) {
-    U_ASSERT(pEnd != nullptr);
-    *pEnd = localeID;
-
-    if (uprv_stricmp(localeID, "root") == 0) {
-        localeID += 4;
-    } else if (uprv_strnicmp(localeID, "und", 3) == 0 &&
-               (localeID[3] == '\0' ||
+size_t _getLanguage(std::string_view localeID, ByteSink* sink, UErrorCode& status) {
+    size_t skip = 0;
+    if (localeID.size() == 4 && uprv_strnicmp(localeID.data(), "root", 4) == 0) {
+        skip = 4;
+        localeID.remove_prefix(skip);
+    } else if (localeID.size() >= 3 && uprv_strnicmp(localeID.data(), "und", 3) == 0 &&
+               (localeID.size() == 3 ||
                 localeID[3] == '-' ||
                 localeID[3] == '_' ||
                 localeID[3] == '@')) {
-        localeID += 3;
+        skip = 3;
+        localeID.remove_prefix(skip);
     }
 
     constexpr int32_t MAXLEN = ULOC_LANG_CAPACITY - 1;  // Minus NUL.
 
     /* if it starts with i- or x- then copy that prefix */
-    int32_t len = _isIDPrefix(localeID) ? 2 : 0;
-    while (!_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len])) {
+    size_t len = _isIDPrefix(localeID) ? 2 : 0;
+    while (len < localeID.size() && !_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len])) {
         if (len == MAXLEN) {
             status = U_ILLEGAL_ARGUMENT_ERROR;
-            return;
+            return 0;
         }
         len++;
     }
 
-    *pEnd = localeID + len;
-    if (sink == nullptr || len == 0) { return; }
+    if (sink == nullptr || len == 0) { return skip + len; }
 
-    int32_t minCapacity = uprv_max(len, 4);  // Minimum 3 letters plus NUL.
+    int32_t minCapacity = uprv_max(static_cast<int32_t>(len), 4);  // Minimum 3 letters plus NUL.
     char scratch[MAXLEN];
     int32_t capacity = 0;
     char* buffer = sink->GetAppendBuffer(
             minCapacity, minCapacity, scratch, UPRV_LENGTHOF(scratch), &capacity);
 
-    for (int32_t i = 0; i < len; ++i) {
+    for (size_t i = 0; i < len; ++i) {
         buffer[i] = uprv_tolower(localeID[i]);
     }
-    if (_isIDSeparator(localeID[1])) {
+    if (localeID.size() >= 2 && _isIDSeparator(localeID[1])) {
         buffer[1] = '-';
     }
 
@@ -1256,32 +1251,26 @@ _getLanguage(const char* localeID,
         if (offset.has_value()) {
             const char* const alias = LANGUAGES[*offset];
             sink->Append(alias, static_cast<int32_t>(uprv_strlen(alias)));
-            return;
+            return skip + len;
         }
     }
 
-    sink->Append(buffer, len);
+    sink->Append(buffer, static_cast<int32_t>(len));
+    return skip + len;
 }
 
-void
-_getScript(const char* localeID,
-           ByteSink* sink,
-           const char** pEnd) {
-    U_ASSERT(pEnd != nullptr);
-    *pEnd = localeID;
-
+size_t _getScript(std::string_view localeID, ByteSink* sink) {
     constexpr int32_t LENGTH = 4;
 
-    int32_t len = 0;
-    while (!_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len]) &&
+    size_t len = 0;
+    while (len < localeID.size() && !_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len]) &&
             uprv_isASCIILetter(localeID[len])) {
-        if (len == LENGTH) { return; }
+        if (len == LENGTH) { return 0; }
         len++;
     }
-    if (len != LENGTH) { return; }
+    if (len != LENGTH) { return 0; }
 
-    *pEnd = localeID + LENGTH;
-    if (sink == nullptr) { return; }
+    if (sink == nullptr) { return len; }
 
     char scratch[LENGTH];
     int32_t capacity = 0;
@@ -1294,27 +1283,21 @@ _getScript(const char* localeID,
     }
 
     sink->Append(buffer, LENGTH);
+    return len;
 }
 
-void
-_getRegion(const char* localeID,
-           ByteSink* sink,
-           const char** pEnd) {
-    U_ASSERT(pEnd != nullptr);
-    *pEnd = localeID;
-
+size_t _getRegion(std::string_view localeID, ByteSink* sink) {
     constexpr int32_t MINLEN = 2;
     constexpr int32_t MAXLEN = ULOC_COUNTRY_CAPACITY - 1;  // Minus NUL.
 
-    int32_t len = 0;
-    while (!_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len])) {
-        if (len == MAXLEN) { return; }
+    size_t len = 0;
+    while (len < localeID.size() && !_isTerminator(localeID[len]) && !_isIDSeparator(localeID[len])) {
+        if (len == MAXLEN) { return 0; }
         len++;
     }
-    if (len < MINLEN) { return; }
+    if (len < MINLEN) { return 0; }
 
-    *pEnd = localeID + len;
-    if (sink == nullptr) { return; }
+    if (sink == nullptr) { return len; }
 
     char scratch[ULOC_COUNTRY_CAPACITY];
     int32_t capacity = 0;
@@ -1325,7 +1308,7 @@ _getRegion(const char* localeID,
             UPRV_LENGTHOF(scratch),
             &capacity);
 
-    for (int32_t i = 0; i < len; ++i) {
+    for (size_t i = 0; i < len; ++i) {
         buffer[i] = uprv_toupper(localeID[i]);
     }
 
@@ -1337,26 +1320,25 @@ _getRegion(const char* localeID,
         if (offset.has_value()) {
             const char* const alias = COUNTRIES[*offset];
             sink->Append(alias, static_cast<int32_t>(uprv_strlen(alias)));
-            return;
+            return len;
         }
     }
 
-    sink->Append(buffer, len);
+    sink->Append(buffer, static_cast<int32_t>(len));
+    return len;
 }
 
 /**
  * @param needSeparator if true, then add leading '_' if any variants
  * are added to 'variant'
  */
-void
-_getVariant(const char* localeID,
+size_t
+_getVariant(std::string_view localeID,
             char prev,
             ByteSink* sink,
-            const char** pEnd,
             bool needSeparator,
             UErrorCode& status) {
-    if (U_FAILURE(status)) return;
-    if (pEnd != nullptr) { *pEnd = localeID; }
+    if (U_FAILURE(status) || localeID.empty()) return 0;
 
     // Reasonable upper limit for variants
     // There are no strict limitation of the syntax of variant in the legacy
@@ -1369,63 +1351,82 @@ _getVariant(const char* localeID,
     constexpr int32_t MAX_VARIANTS_LENGTH = 179;
 
     /* get one or more variant tags and separate them with '_' */
-    int32_t index = 0;
+    size_t index = 0;
     if (_isIDSeparator(prev)) {
         /* get a variant string after a '-' or '_' */
-        for (index=0; !_isTerminator(localeID[index]); index++) {
-            if (index >= MAX_VARIANTS_LENGTH) { // same as length > MAX_VARIANTS_LENGTH
+        for (std::string_view sub = localeID;;) {
+            size_t next = sub.find_first_of(".@_-");
+            // For historical reasons, a trailing separator is included in the variant.
+            bool finished = next == std::string_view::npos || next + 1 == sub.length();
+            size_t limit = finished ? sub.length() : next;
+            index += limit;
+            if (index > MAX_VARIANTS_LENGTH) {
                 status = U_ILLEGAL_ARGUMENT_ERROR;
-                return;
+                return 0;
             }
-            if (needSeparator) {
-                if (sink != nullptr) {
+
+            if (sink != nullptr) {
+                if (needSeparator) {
                     sink->Append("_", 1);
+                } else {
+                    needSeparator = true;
                 }
-                needSeparator = false;
-            }
-            if (sink != nullptr) {
-                char c = uprv_toupper(localeID[index]);
-                if (c == '-') c = '_';
-                sink->Append(&c, 1);
+
+                int32_t length = static_cast<int32_t>(limit);
+                int32_t minCapacity = uprv_min(length, MAX_VARIANTS_LENGTH);
+                char scratch[MAX_VARIANTS_LENGTH];
+                int32_t capacity = 0;
+                char* buffer = sink->GetAppendBuffer(
+                        minCapacity, minCapacity, scratch, UPRV_LENGTHOF(scratch), &capacity);
+
+                for (size_t i = 0; i < limit; ++i) {
+                    buffer[i] = uprv_toupper(sub[i]);
+                }
+                sink->Append(buffer, length);
             }
+
+            if (finished) { return index; }
+            sub.remove_prefix(next);
+            if (_isTerminator(sub.front()) || _isBCP47Extension(sub)) { return index; }
+            sub.remove_prefix(1);
+            index++;
         }
-        if (pEnd != nullptr) { *pEnd = localeID+index; }
     }
 
+    size_t skip = 0;
     /* if there is no variant tag after a '-' or '_' then look for '@' */
-    if (index == 0) {
-        if (prev=='@') {
-            /* keep localeID */
-        } else if((localeID=locale_getKeywordsStart(localeID))!=nullptr) {
-            ++localeID; /* point after the '@' */
-        } else {
-            return;
+    if (prev == '@') {
+        /* keep localeID */
+    } else if (const char* p = locale_getKeywordsStart(localeID); p != nullptr) {
+        skip = 1 + p - localeID.data(); /* point after the '@' */
+        localeID.remove_prefix(skip);
+    } else {
+        return 0;
+    }
+    for (; index < localeID.size() && !_isTerminator(localeID[index]); index++) {
+        if (index >= MAX_VARIANTS_LENGTH) { // same as length > MAX_VARIANTS_LENGTH
+            status = U_ILLEGAL_ARGUMENT_ERROR;
+            return 0;
         }
-        for(; !_isTerminator(localeID[index]); index++) {
-            if (index >= MAX_VARIANTS_LENGTH) { // same as length > MAX_VARIANTS_LENGTH
-                status = U_ILLEGAL_ARGUMENT_ERROR;
-                return;
-            }
-            if (needSeparator) {
-                if (sink != nullptr) {
-                    sink->Append("_", 1);
-                }
-                needSeparator = false;
-            }
+        if (needSeparator) {
             if (sink != nullptr) {
-                char c = uprv_toupper(localeID[index]);
-                if (c == '-' || c == ',') c = '_';
-                sink->Append(&c, 1);
+                sink->Append("_", 1);
             }
+            needSeparator = false;
+        }
+        if (sink != nullptr) {
+            char c = uprv_toupper(localeID[index]);
+            if (c == '-' || c == ',') c = '_';
+            sink->Append(&c, 1);
         }
-        if (pEnd != nullptr) { *pEnd = localeID + index; }
     }
+    return skip + index;
 }
 
 }  // namespace
 
 U_EXPORT CharString
-ulocimp_getLanguage(const char* localeID, UErrorCode& status) {
+ulocimp_getLanguage(std::string_view localeID, UErrorCode& status) {
     return ByteSinkUtil::viaByteSinkToCharString(
         [&](ByteSink& sink, UErrorCode& status) {
             ulocimp_getSubtags(
@@ -1441,7 +1442,7 @@ ulocimp_getLanguage(const char* localeID, UErrorCode& status) {
 }
 
 U_EXPORT CharString
-ulocimp_getScript(const char* localeID, UErrorCode& status) {
+ulocimp_getScript(std::string_view localeID, UErrorCode& status) {
     return ByteSinkUtil::viaByteSinkToCharString(
         [&](ByteSink& sink, UErrorCode& status) {
             ulocimp_getSubtags(
@@ -1457,7 +1458,7 @@ ulocimp_getScript(const char* localeID, UErrorCode& status) {
 }
 
 U_EXPORT CharString
-ulocimp_getRegion(const char* localeID, UErrorCode& status) {
+ulocimp_getRegion(std::string_view localeID, UErrorCode& status) {
     return ByteSinkUtil::viaByteSinkToCharString(
         [&](ByteSink& sink, UErrorCode& status) {
             ulocimp_getSubtags(
@@ -1473,7 +1474,7 @@ ulocimp_getRegion(const char* localeID, UErrorCode& status) {
 }
 
 U_EXPORT CharString
-ulocimp_getVariant(const char* localeID, UErrorCode& status) {
+ulocimp_getVariant(std::string_view localeID, UErrorCode& status) {
     return ByteSinkUtil::viaByteSinkToCharString(
         [&](ByteSink& sink, UErrorCode& status) {
             ulocimp_getSubtags(
@@ -1490,7 +1491,7 @@ ulocimp_getVariant(const char* localeID, UErrorCode& status) {
 
 U_EXPORT void
 ulocimp_getSubtags(
-        const char* localeID,
+        std::string_view localeID,
         CharString* language,
         CharString* script,
         CharString* region,
@@ -1521,7 +1522,7 @@ ulocimp_getSubtags(
 
 U_EXPORT void
 ulocimp_getSubtags(
-        const char* localeID,
+        std::string_view localeID,
         ByteSink* language,
         ByteSink* script,
         ByteSink* region,
@@ -1531,7 +1532,7 @@ ulocimp_getSubtags(
     if (U_FAILURE(status)) { return; }
 
     if (pEnd != nullptr) {
-        *pEnd = localeID;
+        *pEnd = localeID.data();
     } else if (language == nullptr &&
                script == nullptr &&
                region == nullptr &&
@@ -1539,62 +1540,94 @@ ulocimp_getSubtags(
         return;
     }
 
+    if (localeID.empty()) { return; }
+
     bool hasRegion = false;
 
-    if (localeID == nullptr) {
-        localeID = uloc_getDefault();
+    {
+        size_t len = _getLanguage(localeID, language, status);
+        if (U_FAILURE(status)) { return; }
+        if (len > 0) {
+            localeID.remove_prefix(len);
+        }
     }
 
-    _getLanguage(localeID, language, &localeID, status);
-    if (U_FAILURE(status)) { return; }
-    U_ASSERT(localeID != nullptr);
-
     if (pEnd != nullptr) {
-        *pEnd = localeID;
+        *pEnd = localeID.data();
     } else if (script == nullptr &&
                region == nullptr &&
                variant == nullptr) {
         return;
     }
 
-    if (_isIDSeparator(*localeID)) {
-        const char* begin = localeID + 1;
-        const char* end = nullptr;
-        _getScript(begin, script, &end);
-        U_ASSERT(end != nullptr);
-        if (end != begin) {
-            localeID = end;
-            if (pEnd != nullptr) { *pEnd = localeID; }
+    if (localeID.empty()) { return; }
+
+    if (_isIDSeparator(localeID.front())) {
+        std::string_view sub = localeID;
+        sub.remove_prefix(1);
+        size_t len = _getScript(sub, script);
+        if (len > 0) {
+            localeID.remove_prefix(len + 1);
+            if (pEnd != nullptr) { *pEnd = localeID.data(); }
         }
     }
 
-    if (region == nullptr && variant == nullptr && pEnd == nullptr) { return; }
+    if ((region == nullptr && variant == nullptr && pEnd == nullptr) || localeID.empty()) { return; }
 
-    if (_isIDSeparator(*localeID)) {
-        const char* begin = localeID + 1;
-        const char* end = nullptr;
-        _getRegion(begin, region, &end);
-        U_ASSERT(end != nullptr);
-        if (end != begin) {
+    if (_isIDSeparator(localeID.front())) {
+        std::string_view sub = localeID;
+        sub.remove_prefix(1);
+        size_t len = _getRegion(sub, region);
+        if (len > 0) {
             hasRegion = true;
-            localeID = end;
-            if (pEnd != nullptr) { *pEnd = localeID; }
+            localeID.remove_prefix(len + 1);
+            if (pEnd != nullptr) { *pEnd = localeID.data(); }
         }
     }
 
-    if (variant == nullptr && pEnd == nullptr) { return; }
+    if ((variant == nullptr && pEnd == nullptr) || localeID.empty()) { return; }
 
-    if (_isIDSeparator(*localeID) && !_isBCP47Extension(localeID)) {
+    bool hasVariant = false;
+
+    if (_isIDSeparator(localeID.front()) && !_isBCP47Extension(localeID)) {
+        std::string_view sub = localeID;
         /* If there was no country ID, skip a possible extra IDSeparator */
-        if (!hasRegion && _isIDSeparator(localeID[1])) {
-            localeID++;
-        }
-        const char* begin = localeID + 1;
-        const char* end = nullptr;
-        _getVariant(begin, *localeID, variant, &end, false, status);
+        size_t skip = !hasRegion && localeID.size() > 1 && _isIDSeparator(localeID[1]) ? 2 : 1;
+        sub.remove_prefix(skip);
+        size_t len = _getVariant(sub, localeID[0], variant, false, status);
         if (U_FAILURE(status)) { return; }
-        U_ASSERT(end != nullptr);
-        if (end != begin && pEnd != nullptr) { *pEnd = end; }
+        if (len > 0) {
+            hasVariant = true;
+            localeID.remove_prefix(skip + len);
+            if (pEnd != nullptr) { *pEnd = localeID.data(); }
+        }
+    }
+
+    if ((variant == nullptr && pEnd == nullptr) || localeID.empty()) { return; }
+
+    if (_isBCP47Extension(localeID)) {
+        localeID.remove_prefix(2);
+        constexpr char vaposix[] = "-va-posix";
+        constexpr size_t length = sizeof vaposix - 1;
+        for (size_t next;; localeID.remove_prefix(next)) {
+            next = localeID.find('-', 1);
+            if (next == std::string_view::npos) { break; }
+            next = localeID.find('-', next + 1);
+            bool finished = next == std::string_view::npos;
+            std::string_view sub = localeID;
+            if (!finished) { sub.remove_suffix(sub.length() - next); }
+
+            if (sub.length() == length && uprv_strnicmp(sub.data(), vaposix, length) == 0) {
+                if (variant != nullptr) {
+                    if (hasVariant) { variant->Append("_", 1); }
+                    constexpr char posix[] = "POSIX";
+                    variant->Append(posix, sizeof posix - 1);
+                }
+                if (pEnd != nullptr) { *pEnd = localeID.data() + length; }
+            }
+
+            if (finished) { break; }
+        }
     }
 }
 
@@ -1700,7 +1733,7 @@ uloc_openKeywords(const char* localeID,
     CharString tempBuffer;
     const char* tmpLocaleID;
 
-    if (_hasBCP47Extension(localeID)) {
+    if (localeID != nullptr && _hasBCP47Extension(localeID)) {
         tempBuffer = ulocimp_forLanguageTag(localeID, -1, nullptr, *status);
         tmpLocaleID = U_SUCCESS(*status) && !tempBuffer.isEmpty() ? tempBuffer.data() : localeID;
     } else {
@@ -1753,7 +1786,7 @@ constexpr int32_t I_DEFAULT_LENGTH = UPRV_LENGTHOF(i_default);
  * This is the code underlying uloc_getName and uloc_canonicalize.
  */
 void
-_canonicalize(const char* localeID,
+_canonicalize(std::string_view localeID,
               ByteSink& sink,
               uint32_t options,
               UErrorCode& err) {
@@ -1764,33 +1797,30 @@ _canonicalize(const char* localeID,
     int32_t j, fieldCount=0;
     CharString tempBuffer;  // if localeID has a BCP47 extension, tmpLocaleID points to this
     CharString localeIDWithHyphens;  // if localeID has a BPC47 extension and have _, tmpLocaleID points to this
-    const char* origLocaleID;
-    const char* tmpLocaleID;
-    const char* keywordAssign = nullptr;
-    const char* separatorIndicator = nullptr;
+    std::string_view origLocaleID;
+    std::string_view tmpLocaleID;
+    size_t keywordAssign = std::string_view::npos;
+    size_t separatorIndicator = std::string_view::npos;
 
     if (_hasBCP47Extension(localeID)) {
-        const char* localeIDPtr = localeID;
+        std::string_view localeIDPtr = localeID;
 
         // convert all underbars to hyphens, unless the "BCP47 extension" comes at the beginning of the string
-        if (uprv_strchr(localeID, '_') != nullptr && localeID[1] != '-' && localeID[1] != '_') {
-            localeIDWithHyphens.append(localeID, -1, err);
+        if (localeID.size() >= 2 && localeID.find('_') != std::string_view::npos && localeID[1] != '-' && localeID[1] != '_') {
+            localeIDWithHyphens.append(localeID, err);
             if (U_SUCCESS(err)) {
                 for (char* p = localeIDWithHyphens.data(); *p != '\0'; ++p) {
                     if (*p == '_') {
                         *p = '-';
                     }
                 }
-                localeIDPtr = localeIDWithHyphens.data();
+                localeIDPtr = localeIDWithHyphens.toStringPiece();
             }
         }
 
-        tempBuffer = ulocimp_forLanguageTag(localeIDPtr, -1, nullptr, err);
-        tmpLocaleID = U_SUCCESS(err) && !tempBuffer.isEmpty() ? tempBuffer.data() : localeIDPtr;
+        tempBuffer = ulocimp_forLanguageTag(localeIDPtr.data(), static_cast<int32_t>(localeIDPtr.size()), nullptr, err);
+        tmpLocaleID = U_SUCCESS(err) && !tempBuffer.isEmpty() ? static_cast<std::string_view>(tempBuffer.toStringPiece()) : localeIDPtr;
     } else {
-        if (localeID==nullptr) {
-           localeID=uloc_getDefault();
-        }
         tmpLocaleID=localeID;
     }
 
@@ -1801,20 +1831,25 @@ _canonicalize(const char* localeID,
     CharString script;
     CharString country;
     CharString variant;
+    const char* end = nullptr;
     ulocimp_getSubtags(
             tmpLocaleID,
             &tag,
             &script,
             &country,
             &variant,
-            &tmpLocaleID,
+            &end,
             err);
     if (U_FAILURE(err)) {
         return;
     }
+    U_ASSERT(end != nullptr);
+    if (end > tmpLocaleID.data()) {
+        tmpLocaleID.remove_prefix(end - tmpLocaleID.data());
+    }
 
-    if (tag.length() == I_DEFAULT_LENGTH &&
-            uprv_strncmp(origLocaleID, i_default, I_DEFAULT_LENGTH) == 0) {
+    if (tag.length() == I_DEFAULT_LENGTH && origLocaleID.length() >= I_DEFAULT_LENGTH &&
+            uprv_strncmp(origLocaleID.data(), i_default, I_DEFAULT_LENGTH) == 0) {
         tag.clear();
         tag.append(uloc_getDefault(), err);
     } else {
@@ -1839,15 +1874,14 @@ _canonicalize(const char* localeID,
     }
 
     /* Copy POSIX-style charset specifier, if any [mr.utf8] */
-    if (!OPTION_SET(options, _ULOC_CANONICALIZE) && *tmpLocaleID == '.') {
+    if (!OPTION_SET(options, _ULOC_CANONICALIZE) && !tmpLocaleID.empty() && tmpLocaleID.front() == '.') {
         tag.append('.', err);
-        ++tmpLocaleID;
-        const char *atPos = nullptr;
+        tmpLocaleID.remove_prefix(1);
         size_t length;
-        if((atPos = uprv_strchr(tmpLocaleID, '@')) != nullptr) {
-            length = atPos - tmpLocaleID;
+        if (size_t atPos = tmpLocaleID.find('@'); atPos != std::string_view::npos) {
+            length = atPos;
         } else {
-            length = uprv_strlen(tmpLocaleID);
+            length = tmpLocaleID.length();
         }
         // The longest charset name we found in IANA charset registry
         // https://www.iana.org/assignments/character-sets/ is
@@ -1859,33 +1893,34 @@ _canonicalize(const char* localeID,
            err = U_ILLEGAL_ARGUMENT_ERROR; /* malformed keyword name */
            return;
         }
-        tag.append(tmpLocaleID, static_cast<int32_t>(length), err);
-        tmpLocaleID += length;
+        if (length > 0) {
+            tag.append(tmpLocaleID.data(), static_cast<int32_t>(length), err);
+            tmpLocaleID.remove_prefix(length);
+        }
     }
 
     /* Scan ahead to next '@' and determine if it is followed by '=' and/or ';'
-       After this, tmpLocaleID either points to '@' or is nullptr */
-    if ((tmpLocaleID=locale_getKeywordsStart(tmpLocaleID))!=nullptr) {
-        keywordAssign = uprv_strchr(tmpLocaleID, '=');
-        separatorIndicator = uprv_strchr(tmpLocaleID, ';');
+       After this, tmpLocaleID either starts at '@' or is empty. */
+    if (const char* start = locale_getKeywordsStart(tmpLocaleID); start != nullptr) {
+        if (start > tmpLocaleID.data()) {
+            tmpLocaleID.remove_prefix(start - tmpLocaleID.data());
+        }
+        keywordAssign = tmpLocaleID.find('=');
+        separatorIndicator = tmpLocaleID.find(';');
+    } else {
+        tmpLocaleID = {};
     }
 
     /* Copy POSIX-style variant, if any [mr@FOO] */
     if (!OPTION_SET(options, _ULOC_CANONICALIZE) &&
-        tmpLocaleID != nullptr && keywordAssign == nullptr) {
-        for (;;) {
-            char c = *tmpLocaleID;
-            if (c == 0) {
-                break;
-            }
-            tag.append(c, err);
-            ++tmpLocaleID;
-        }
+        !tmpLocaleID.empty() && keywordAssign == std::string_view::npos) {
+        tag.append(tmpLocaleID, err);
+        tmpLocaleID = {};
     }
 
     if (OPTION_SET(options, _ULOC_CANONICALIZE)) {
         /* Handle @FOO variant if @ is present and not followed by = */
-        if (tmpLocaleID!=nullptr && keywordAssign==nullptr) {
+        if (!tmpLocaleID.empty() && keywordAssign == std::string_view::npos) {
             /* Add missing '_' if needed */
             if (fieldCount < 2 || (fieldCount < 3 && !script.isEmpty())) {
                 do {
@@ -1895,7 +1930,9 @@ _canonicalize(const char* localeID,
             }
 
             CharStringByteSink s(&tag);
-            _getVariant(tmpLocaleID+1, '@', &s, nullptr, !variant.isEmpty(), err);
+            std::string_view sub = tmpLocaleID;
+            sub.remove_prefix(1);
+            _getVariant(sub, '@', &s, !variant.isEmpty(), err);
             if (U_FAILURE(err)) { return; }
         }
 
@@ -1903,7 +1940,7 @@ _canonicalize(const char* localeID,
         for (j=0; j<UPRV_LENGTHOF(CANONICALIZE_MAP); j++) {
             StringPiece id(CANONICALIZE_MAP[j].id);
             if (tag == id) {
-                if (id.empty() && tmpLocaleID != nullptr) {
+                if (id.empty() && !tmpLocaleID.empty()) {
                     break; /* Don't remap "" if keywords present */
                 }
                 tag.clear();
@@ -1916,11 +1953,12 @@ _canonicalize(const char* localeID,
     sink.Append(tag.data(), tag.length());
 
     if (!OPTION_SET(options, _ULOC_STRIP_KEYWORDS)) {
-        if (tmpLocaleID!=nullptr && keywordAssign!=nullptr &&
-            (!separatorIndicator || separatorIndicator > keywordAssign)) {
+        if (!tmpLocaleID.empty() && keywordAssign != std::string_view::npos &&
+            (separatorIndicator == std::string_view::npos || separatorIndicator > keywordAssign)) {
             sink.Append("@", 1);
             ++fieldCount;
-            ulocimp_getKeywords(tmpLocaleID+1, '@', sink, true, err);
+            tmpLocaleID.remove_prefix(1);
+            ulocimp_getKeywords(tmpLocaleID, '@', sink, true, err);
         }
     }
 }
@@ -1989,6 +2027,10 @@ uloc_getLanguage(const char*    localeID,
          int32_t languageCapacity,
          UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
+
     /* uloc_getLanguage will return a 2 character iso-639 code if one exists. *CWB*/
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         language, languageCapacity,
@@ -2011,6 +2053,10 @@ uloc_getScript(const char*    localeID,
          int32_t scriptCapacity,
          UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
+
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         script, scriptCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2032,6 +2078,10 @@ uloc_getCountry(const char* localeID,
             int32_t countryCapacity,
             UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
+
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         country, countryCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2053,6 +2103,10 @@ uloc_getVariant(const char* localeID,
                 int32_t variantCapacity,
                 UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
+
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         variant, variantCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2074,6 +2128,9 @@ uloc_getName(const char* localeID,
              int32_t nameCapacity,
              UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         name, nameCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2083,7 +2140,7 @@ uloc_getName(const char* localeID,
 }
 
 U_EXPORT CharString
-ulocimp_getName(const char* localeID,
+ulocimp_getName(std::string_view localeID,
                 UErrorCode& err)
 {
     return ByteSinkUtil::viaByteSinkToCharString(
@@ -2094,7 +2151,7 @@ ulocimp_getName(const char* localeID,
 }
 
 U_EXPORT void
-ulocimp_getName(const char* localeID,
+ulocimp_getName(std::string_view localeID,
                 ByteSink& sink,
                 UErrorCode& err)
 {
@@ -2107,6 +2164,9 @@ uloc_getBaseName(const char* localeID,
                  int32_t nameCapacity,
                  UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         name, nameCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2116,7 +2176,7 @@ uloc_getBaseName(const char* localeID,
 }
 
 U_EXPORT CharString
-ulocimp_getBaseName(const char* localeID,
+ulocimp_getBaseName(std::string_view localeID,
                     UErrorCode& err)
 {
     return ByteSinkUtil::viaByteSinkToCharString(
@@ -2127,7 +2187,7 @@ ulocimp_getBaseName(const char* localeID,
 }
 
 U_EXPORT void
-ulocimp_getBaseName(const char* localeID,
+ulocimp_getBaseName(std::string_view localeID,
                     ByteSink& sink,
                     UErrorCode& err)
 {
@@ -2140,6 +2200,9 @@ uloc_canonicalize(const char* localeID,
                   int32_t nameCapacity,
                   UErrorCode* err)
 {
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     return ByteSinkUtil::viaByteSinkToTerminatedChars(
         name, nameCapacity,
         [&](ByteSink& sink, UErrorCode& status) {
@@ -2149,7 +2212,7 @@ uloc_canonicalize(const char* localeID,
 }
 
 U_EXPORT CharString
-ulocimp_canonicalize(const char* localeID,
+ulocimp_canonicalize(std::string_view localeID,
                      UErrorCode& err)
 {
     return ByteSinkUtil::viaByteSinkToCharString(
@@ -2160,7 +2223,7 @@ ulocimp_canonicalize(const char* localeID,
 }
 
 U_EXPORT void
-ulocimp_canonicalize(const char* localeID,
+ulocimp_canonicalize(std::string_view localeID,
                      ByteSink& sink,
                      UErrorCode& err)
 {
diff --git a/deps/icu-small/source/common/uloc_tag.cpp b/deps/icu-small/source/common/uloc_tag.cpp
index 7b3b1e73a37cf4..b2e9946f48a7e5 100644
--- a/deps/icu-small/source/common/uloc_tag.cpp
+++ b/deps/icu-small/source/common/uloc_tag.cpp
@@ -1043,7 +1043,7 @@ _initializeULanguageTag(ULanguageTag* langtag) {
 }
 
 void
-_appendLanguageToLanguageTag(const char* localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
+_appendLanguageToLanguageTag(std::string_view localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
     UErrorCode tmpStatus = U_ZERO_ERROR;
 
     if (U_FAILURE(status)) {
@@ -1088,7 +1088,7 @@ _appendLanguageToLanguageTag(const char* localeID, icu::ByteSink& sink, bool str
 }
 
 void
-_appendScriptToLanguageTag(const char* localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
+_appendScriptToLanguageTag(std::string_view localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
     UErrorCode tmpStatus = U_ZERO_ERROR;
 
     if (U_FAILURE(status)) {
@@ -1118,7 +1118,7 @@ _appendScriptToLanguageTag(const char* localeID, icu::ByteSink& sink, bool stric
 }
 
 void
-_appendRegionToLanguageTag(const char* localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
+_appendRegionToLanguageTag(std::string_view localeID, icu::ByteSink& sink, bool strict, UErrorCode& status) {
     UErrorCode tmpStatus = U_ZERO_ERROR;
 
     if (U_FAILURE(status)) {
@@ -1169,7 +1169,7 @@ void _sortVariants(VariantListEntry* first) {
 }
 
 void
-_appendVariantsToLanguageTag(const char* localeID, icu::ByteSink& sink, bool strict, bool& hadPosix, UErrorCode& status) {
+_appendVariantsToLanguageTag(std::string_view localeID, icu::ByteSink& sink, bool strict, bool& hadPosix, UErrorCode& status) {
     if (U_FAILURE(status)) { return; }
 
     UErrorCode tmpStatus = U_ZERO_ERROR;
@@ -1872,7 +1872,7 @@ _appendKeywords(ULanguageTag* langtag, icu::ByteSink& sink, UErrorCode& status)
 }
 
 void
-_appendPrivateuseToLanguageTag(const char* localeID, icu::ByteSink& sink, bool strict, bool /*hadPosix*/, UErrorCode& status) {
+_appendPrivateuseToLanguageTag(std::string_view localeID, icu::ByteSink& sink, bool strict, bool /*hadPosix*/, UErrorCode& status) {
     if (U_FAILURE(status)) { return; }
 
     UErrorCode tmpStatus = U_ZERO_ERROR;
@@ -2596,6 +2596,9 @@ ulocimp_toLanguageTag(const char* localeID,
     bool hadPosix = false;
     const char* pKeywordStart;
 
+    if (localeID == nullptr) {
+        localeID = uloc_getDefault();
+    }
     /* Note: uloc_canonicalize returns "en_US_POSIX" for input locale ID "".  See #6835 */
     icu::CharString canonical = ulocimp_canonicalize(localeID, tmpStatus);
     if (U_FAILURE(tmpStatus)) {
@@ -2604,7 +2607,7 @@ ulocimp_toLanguageTag(const char* localeID,
     }
 
     /* For handling special case - private use only tag */
-    pKeywordStart = locale_getKeywordsStart(canonical.data());
+    pKeywordStart = locale_getKeywordsStart(canonical.toStringPiece());
     if (pKeywordStart == canonical.data()) {
         int kwdCnt = 0;
         bool done = false;
@@ -2642,12 +2645,12 @@ ulocimp_toLanguageTag(const char* localeID,
         }
     }
 
-    _appendLanguageToLanguageTag(canonical.data(), sink, strict, status);
-    _appendScriptToLanguageTag(canonical.data(), sink, strict, status);
-    _appendRegionToLanguageTag(canonical.data(), sink, strict, status);
-    _appendVariantsToLanguageTag(canonical.data(), sink, strict, hadPosix, status);
+    _appendLanguageToLanguageTag(canonical.toStringPiece(), sink, strict, status);
+    _appendScriptToLanguageTag(canonical.toStringPiece(), sink, strict, status);
+    _appendRegionToLanguageTag(canonical.toStringPiece(), sink, strict, status);
+    _appendVariantsToLanguageTag(canonical.toStringPiece(), sink, strict, hadPosix, status);
     _appendKeywordsToLanguageTag(canonical.data(), sink, strict, hadPosix, status);
-    _appendPrivateuseToLanguageTag(canonical.data(), sink, strict, hadPosix, status);
+    _appendPrivateuseToLanguageTag(canonical.toStringPiece(), sink, strict, hadPosix, status);
 }
 
 
diff --git a/deps/icu-small/source/common/ulocale.cpp b/deps/icu-small/source/common/ulocale.cpp
index f2f81bc97109bd..33814713dc1bfd 100644
--- a/deps/icu-small/source/common/ulocale.cpp
+++ b/deps/icu-small/source/common/ulocale.cpp
@@ -10,7 +10,6 @@
 #include "unicode/locid.h"
 
 #include "bytesinkutil.h"
-#include "charstr.h"
 #include "cmemory.h"
 
 U_NAMESPACE_USE
@@ -24,9 +23,7 @@ ulocale_openForLocaleID(const char* localeID, int32_t length, UErrorCode* err) {
     if (length < 0) {
         return EXTERNAL(icu::Locale::createFromName(localeID).clone());
     }
-    CharString str(localeID, length, *err);  // Make a NUL terminated copy.
-    if (U_FAILURE(*err)) { return nullptr; }
-    return EXTERNAL(icu::Locale::createFromName(str.data()).clone());
+    return EXTERNAL(icu::Locale::createFromName(StringPiece{localeID, length}).clone());
 }
 
 ULocale*
diff --git a/deps/icu-small/source/common/ulocimp.h b/deps/icu-small/source/common/ulocimp.h
index 1887e2a849ab0e..7f09748c8ac1c8 100644
--- a/deps/icu-small/source/common/ulocimp.h
+++ b/deps/icu-small/source/common/ulocimp.h
@@ -68,42 +68,42 @@ U_EXPORT std::optional<std::string_view>
 ulocimp_toLegacyTypeWithFallback(std::string_view keyword, std::string_view value);
 
 U_EXPORT icu::CharString
-ulocimp_getKeywords(const char* localeID,
+ulocimp_getKeywords(std::string_view localeID,
                     char prev,
                     bool valuesToo,
                     UErrorCode& status);
 
 U_EXPORT void
-ulocimp_getKeywords(const char* localeID,
+ulocimp_getKeywords(std::string_view localeID,
                     char prev,
                     icu::ByteSink& sink,
                     bool valuesToo,
                     UErrorCode& status);
 
 U_EXPORT icu::CharString
-ulocimp_getName(const char* localeID,
+ulocimp_getName(std::string_view localeID,
                 UErrorCode& err);
 
 U_EXPORT void
-ulocimp_getName(const char* localeID,
+ulocimp_getName(std::string_view localeID,
                 icu::ByteSink& sink,
                 UErrorCode& err);
 
 U_EXPORT icu::CharString
-ulocimp_getBaseName(const char* localeID,
+ulocimp_getBaseName(std::string_view localeID,
                     UErrorCode& err);
 
 U_EXPORT void
-ulocimp_getBaseName(const char* localeID,
+ulocimp_getBaseName(std::string_view localeID,
                     icu::ByteSink& sink,
                     UErrorCode& err);
 
 U_EXPORT icu::CharString
-ulocimp_canonicalize(const char* localeID,
+ulocimp_canonicalize(std::string_view localeID,
                      UErrorCode& err);
 
 U_EXPORT void
-ulocimp_canonicalize(const char* localeID,
+ulocimp_canonicalize(std::string_view localeID,
                      icu::ByteSink& sink,
                      UErrorCode& err);
 
@@ -119,16 +119,16 @@ ulocimp_getKeywordValue(const char* localeID,
                         UErrorCode& status);
 
 U_EXPORT icu::CharString
-ulocimp_getLanguage(const char* localeID, UErrorCode& status);
+ulocimp_getLanguage(std::string_view localeID, UErrorCode& status);
 
 U_EXPORT icu::CharString
-ulocimp_getScript(const char* localeID, UErrorCode& status);
+ulocimp_getScript(std::string_view localeID, UErrorCode& status);
 
 U_EXPORT icu::CharString
-ulocimp_getRegion(const char* localeID, UErrorCode& status);
+ulocimp_getRegion(std::string_view localeID, UErrorCode& status);
 
 U_EXPORT icu::CharString
-ulocimp_getVariant(const char* localeID, UErrorCode& status);
+ulocimp_getVariant(std::string_view localeID, UErrorCode& status);
 
 U_EXPORT void
 ulocimp_setKeywordValue(std::string_view keywordName,
@@ -145,7 +145,7 @@ ulocimp_setKeywordValue(std::string_view keywords,
 
 U_EXPORT void
 ulocimp_getSubtags(
-        const char* localeID,
+        std::string_view localeID,
         icu::CharString* language,
         icu::CharString* script,
         icu::CharString* region,
@@ -155,7 +155,7 @@ ulocimp_getSubtags(
 
 U_EXPORT void
 ulocimp_getSubtags(
-        const char* localeID,
+        std::string_view localeID,
         icu::ByteSink* language,
         icu::ByteSink* script,
         icu::ByteSink* region,
@@ -165,7 +165,7 @@ ulocimp_getSubtags(
 
 inline void
 ulocimp_getSubtags(
-        const char* localeID,
+        std::string_view localeID,
         std::nullptr_t,
         std::nullptr_t,
         std::nullptr_t,
@@ -364,7 +364,7 @@ ulocimp_minimizeSubtags(const char* localeID,
                         UErrorCode& err);
 
 U_CAPI const char * U_EXPORT2
-locale_getKeywordsStart(const char *localeID);
+locale_getKeywordsStart(std::string_view localeID);
 
 bool
 ultag_isExtensionSubtags(const char* s, int32_t len);
diff --git a/deps/icu-small/source/common/umapfile.cpp b/deps/icu-small/source/common/umapfile.cpp
index b58ac37f4d4593..3ba0251df9c52a 100644
--- a/deps/icu-small/source/common/umapfile.cpp
+++ b/deps/icu-small/source/common/umapfile.cpp
@@ -237,8 +237,13 @@ typedef HANDLE MemoryMap;
         pData->map = (char *)data + length;
         pData->pHeader=(const DataHeader *)data;
         pData->mapAddr = data;
-#if U_PLATFORM == U_PF_IPHONE
+#if U_PLATFORM == U_PF_IPHONE || U_PLATFORM == U_PF_ANDROID
+    // Apparently supported from Android 23 and higher:
+    //   https://github.com/ggml-org/llama.cpp/pull/3631
+    // Checking for the flag itself is safer than checking for __ANDROID_API__.
+#   ifdef POSIX_MADV_RANDOM
         posix_madvise(data, length, POSIX_MADV_RANDOM);
+#   endif
 #endif
         return true;
     }
diff --git a/deps/icu-small/source/common/unicode/brkiter.h b/deps/icu-small/source/common/unicode/brkiter.h
index 30c59c4a94ace1..d953925bd72a0c 100644
--- a/deps/icu-small/source/common/unicode/brkiter.h
+++ b/deps/icu-small/source/common/unicode/brkiter.h
@@ -58,6 +58,8 @@ U_NAMESPACE_END
 
 U_NAMESPACE_BEGIN
 
+class CharString;
+
 /**
  * The BreakIterator class implements methods for finding the location
  * of boundaries in text. BreakIterator is an abstract base class.
@@ -646,9 +648,9 @@ class U_COMMON_API BreakIterator : public UObject {
 private:
 
     /** @internal (private) */
-    char actualLocale[ULOC_FULLNAME_CAPACITY];
-    char validLocale[ULOC_FULLNAME_CAPACITY];
-    char requestLocale[ULOC_FULLNAME_CAPACITY];
+    CharString* actualLocale = nullptr;
+    CharString* validLocale = nullptr;
+    CharString* requestLocale = nullptr;
 };
 
 #ifndef U_HIDE_DEPRECATED_API
diff --git a/deps/icu-small/source/common/unicode/char16ptr.h b/deps/icu-small/source/common/unicode/char16ptr.h
index daf35cd43ba2d6..049de9efee810f 100644
--- a/deps/icu-small/source/common/unicode/char16ptr.h
+++ b/deps/icu-small/source/common/unicode/char16ptr.h
@@ -9,10 +9,13 @@
 
 #include "unicode/utypes.h"
 
-#if U_SHOW_CPLUSPLUS_API
+#if U_SHOW_CPLUSPLUS_API || U_SHOW_CPLUSPLUS_HEADER_API
 
 #include <cstddef>
 #include <string_view>
+#include <type_traits>
+
+#endif
 
 /**
  * \file
@@ -21,8 +24,6 @@
  *        Also conversion functions from char16_t * to UChar * and OldUChar *.
  */
 
-U_NAMESPACE_BEGIN
-
 /**
  * \def U_ALIASING_BARRIER
  * Barrier for pointer anti-aliasing optimizations even across function boundaries.
@@ -36,6 +37,11 @@ U_NAMESPACE_BEGIN
 #   define U_ALIASING_BARRIER(ptr)
 #endif
 
+// ICU DLL-exported
+#if U_SHOW_CPLUSPLUS_API
+
+U_NAMESPACE_BEGIN
+
 /**
  * char16_t * wrapper with implicit conversion from distinct but bit-compatible pointer types.
  * @stable ICU 59
@@ -251,6 +257,60 @@ const char16_t *ConstChar16Ptr::get() const { return u_.cp; }
 #endif
 /// \endcond
 
+U_NAMESPACE_END
+
+#endif  // U_SHOW_CPLUSPLUS_API
+
+// Usable in header-only definitions
+#if U_SHOW_CPLUSPLUS_API || U_SHOW_CPLUSPLUS_HEADER_API
+
+namespace U_ICU_NAMESPACE_OR_INTERNAL {
+
+#ifndef U_FORCE_HIDE_INTERNAL_API
+/** @internal */
+template<typename T, typename = std::enable_if_t<std::is_same_v<T, UChar>>>
+inline const char16_t *uprv_char16PtrFromUChar(const T *p) {
+    if constexpr (std::is_same_v<UChar, char16_t>) {
+        return p;
+    } else {
+#if U_SHOW_CPLUSPLUS_API
+        return ConstChar16Ptr(p).get();
+#else
+#ifdef U_ALIASING_BARRIER
+        U_ALIASING_BARRIER(p);
+#endif
+        return reinterpret_cast<const char16_t *>(p);
+#endif
+    }
+}
+#if !U_CHAR16_IS_TYPEDEF && (!defined(_LIBCPP_VERSION) || _LIBCPP_VERSION < 180000)
+/** @internal */
+inline const char16_t *uprv_char16PtrFromUint16(const uint16_t *p) {
+#if U_SHOW_CPLUSPLUS_API
+    return ConstChar16Ptr(p).get();
+#else
+#ifdef U_ALIASING_BARRIER
+    U_ALIASING_BARRIER(p);
+#endif
+    return reinterpret_cast<const char16_t *>(p);
+#endif
+}
+#endif
+#if U_SIZEOF_WCHAR_T==2
+/** @internal */
+inline const char16_t *uprv_char16PtrFromWchar(const wchar_t *p) {
+#if U_SHOW_CPLUSPLUS_API
+    return ConstChar16Ptr(p).get();
+#else
+#ifdef U_ALIASING_BARRIER
+    U_ALIASING_BARRIER(p);
+#endif
+    return reinterpret_cast<const char16_t *>(p);
+#endif
+}
+#endif
+#endif
+
 /**
  * Converts from const char16_t * to const UChar *.
  * Includes an aliasing barrier if available.
@@ -307,6 +367,15 @@ inline OldUChar *toOldUCharPtr(char16_t *p) {
     return reinterpret_cast<OldUChar *>(p);
 }
 
+}  // U_ICU_NAMESPACE_OR_INTERNAL
+
+#endif  // U_SHOW_CPLUSPLUS_API || U_SHOW_CPLUSPLUS_HEADER_API
+
+// ICU DLL-exported
+#if U_SHOW_CPLUSPLUS_API
+
+U_NAMESPACE_BEGIN
+
 #ifndef U_FORCE_HIDE_INTERNAL_API
 /**
  * Is T convertible to a std::u16string_view or some other 16-bit string view?
@@ -379,6 +448,6 @@ inline std::u16string_view toU16StringViewNullable(const T& text) {
 
 U_NAMESPACE_END
 
-#endif /* U_SHOW_CPLUSPLUS_API */
+#endif  // U_SHOW_CPLUSPLUS_API
 
 #endif  // __CHAR16PTR_H__
diff --git a/deps/icu-small/source/common/unicode/locid.h b/deps/icu-small/source/common/unicode/locid.h
index e1afd598cf9bc9..a394cd9347dbde 100644
--- a/deps/icu-small/source/common/unicode/locid.h
+++ b/deps/icu-small/source/common/unicode/locid.h
@@ -449,6 +449,11 @@ class U_COMMON_API Locale : public UObject {
      */
     static Locale U_EXPORT2 createFromName(const char *name);
 
+#ifndef U_HIDE_INTERNAL_API
+    /** @internal */
+    static Locale U_EXPORT2 createFromName(StringPiece name);
+#endif  /* U_HIDE_INTERNAL_API */
+
     /**
      * Creates a locale from the given string after canonicalizing
      * the string according to CLDR by calling uloc_canonicalize().
@@ -1133,7 +1138,9 @@ class U_COMMON_API Locale : public UObject {
      * @param cLocaleID The new locale name.
      * @param canonicalize whether to call uloc_canonicalize on cLocaleID
      */
-    Locale& init(const char* cLocaleID, UBool canonicalize);
+    Locale& init(const char* localeID, UBool canonicalize);
+    /** @internal */
+    Locale& init(StringPiece localeID, UBool canonicalize);
 
     /*
      * Internal constructor to allow construction of a locale object with
diff --git a/deps/icu-small/source/common/unicode/resbund.h b/deps/icu-small/source/common/unicode/resbund.h
index 3965371729d814..03ff6faee239a7 100644
--- a/deps/icu-small/source/common/unicode/resbund.h
+++ b/deps/icu-small/source/common/unicode/resbund.h
@@ -450,7 +450,7 @@ class U_COMMON_API ResourceBundle : public UObject {
      * @return a Locale object
      * @stable ICU 2.8
      */
-    const Locale
+    Locale
       getLocale(ULocDataLocaleType type, UErrorCode &status) const;
 #ifndef U_HIDE_INTERNAL_API
     /**
diff --git a/deps/icu-small/source/common/unicode/uchar.h b/deps/icu-small/source/common/unicode/uchar.h
index 0daa7dd2141cd0..82ec63ab524e05 100644
--- a/deps/icu-small/source/common/unicode/uchar.h
+++ b/deps/icu-small/source/common/unicode/uchar.h
@@ -675,14 +675,14 @@ typedef enum UProperty {
      * @stable ICU 63
      */
     UCHAR_VERTICAL_ORIENTATION=0x1018,
-#ifndef U_HIDE_DRAFT_API
     /**
      * Enumerated property Identifier_Status.
      * Used for UTS #39 General Security Profile for Identifiers
      * (https://www.unicode.org/reports/tr39/#General_Security_Profile).
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UCHAR_IDENTIFIER_STATUS=0x1019,
+#ifndef U_HIDE_DRAFT_API
     /**
      * Enumerated property Indic_Conjunct_Break.
      * Used in the grapheme cluster break algorithm in UAX #29.
@@ -796,7 +796,6 @@ typedef enum UProperty {
     UCHAR_SCRIPT_EXTENSIONS=0x7000,
     /** First constant for Unicode properties with unusual value types. @stable ICU 4.6 */
     UCHAR_OTHER_PROPERTY_START=UCHAR_SCRIPT_EXTENSIONS,
-#ifndef U_HIDE_DRAFT_API
     /**
      * Miscellaneous property Identifier_Type.
      * Used for UTS #39 General Security Profile for Identifiers
@@ -808,10 +807,9 @@ typedef enum UProperty {
      *
      * @see u_hasIDType
      * @see u_getIDTypes
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UCHAR_IDENTIFIER_TYPE=0x7001,
-#endif  // U_HIDE_DRAFT_API
 #ifndef U_HIDE_DEPRECATED_API
     /**
      * One more than the last constant for Unicode properties with unusual value types.
@@ -2791,13 +2789,12 @@ typedef enum UVerticalOrientation {
     U_VO_UPRIGHT,
 } UVerticalOrientation;
 
-#ifndef U_HIDE_DRAFT_API
 /**
  * Identifier Status constants.
  * See https://www.unicode.org/reports/tr39/#Identifier_Status_and_Type.
  *
  * @see UCHAR_IDENTIFIER_STATUS
- * @draft ICU 75
+ * @stable ICU 75
  */
 typedef enum UIdentifierStatus {
     /*
@@ -2806,9 +2803,9 @@ typedef enum UIdentifierStatus {
      *     U_ID_STATUS_<Unicode Identifier_Status value name>
      */
 
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_STATUS_RESTRICTED,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_STATUS_ALLOWED,
 } UIdentifierStatus;
 
@@ -2817,7 +2814,7 @@ typedef enum UIdentifierStatus {
  * See https://www.unicode.org/reports/tr39/#Identifier_Status_and_Type.
  *
  * @see UCHAR_IDENTIFIER_TYPE
- * @draft ICU 75
+ * @stable ICU 75
  */
 typedef enum UIdentifierType {
     /*
@@ -2826,32 +2823,31 @@ typedef enum UIdentifierType {
      *     U_ID_TYPE_<Unicode Identifier_Type value name>
      */
 
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_NOT_CHARACTER,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_DEPRECATED,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_DEFAULT_IGNORABLE,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_NOT_NFKC,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_NOT_XID,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_EXCLUSION,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_OBSOLETE,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_TECHNICAL,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_UNCOMMON_USE,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_LIMITED_USE,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_INCLUSION,
-    /** @draft ICU 75 */
+    /** @stable ICU 75 */
     U_ID_TYPE_RECOMMENDED,
 } UIdentifierType;
-#endif  // U_HIDE_DRAFT_API
 
 /**
  * Check a binary Unicode property for a code point.
@@ -4057,7 +4053,6 @@ u_isIDStart(UChar32 c);
 U_CAPI UBool U_EXPORT2
 u_isIDPart(UChar32 c);
 
-#ifndef U_HIDE_DRAFT_API
 /**
  * Does the set of Identifier_Type values code point c contain the given type?
  *
@@ -4069,7 +4064,7 @@ u_isIDPart(UChar32 c);
  * @param c code point
  * @param type Identifier_Type to check
  * @return true if type is in Identifier_Type(c)
- * @draft ICU 75
+ * @stable ICU 75
  */
 U_CAPI bool U_EXPORT2
 u_hasIDType(UChar32 c, UIdentifierType type);
@@ -4104,11 +4099,10 @@ u_hasIDType(UChar32 c, UIdentifierType type);
  *                   function chaining. (See User Guide for details.)
  * @return number of values in c's Identifier_Type,
  *         written to types unless U_BUFFER_OVERFLOW_ERROR indicates insufficient capacity
- * @draft ICU 75
+ * @stable ICU 75
  */
 U_CAPI int32_t U_EXPORT2
 u_getIDTypes(UChar32 c, UIdentifierType *types, int32_t capacity, UErrorCode *pErrorCode);
-#endif  // U_HIDE_DRAFT_API
 
 /**
  * Determines if the specified character should be regarded
diff --git a/deps/icu-small/source/common/unicode/uniset.h b/deps/icu-small/source/common/unicode/uniset.h
index d070fd631a22d9..6b1ac9ba26202d 100644
--- a/deps/icu-small/source/common/unicode/uniset.h
+++ b/deps/icu-small/source/common/unicode/uniset.h
@@ -1173,10 +1173,12 @@ class U_COMMON_API UnicodeSet final : public UnicodeFilter {
     inline U_HEADER_NESTED_NAMESPACE::USetStrings strings() const {
         return U_HEADER_NESTED_NAMESPACE::USetStrings(toUSet());
     }
+#endif  // U_HIDE_DRAFT_API
 
+#ifndef U_HIDE_DRAFT_API
     /**
      * Returns a C++ iterator for iterating over all of the elements of this set.
-     * Convenient all-in one iteration, but creates a UnicodeString for each
+     * Convenient all-in one iteration, but creates a std::u16string for each
      * code point or string.
      * (Similar to how Java UnicodeSet *is an* Iterable&lt;String&gt;.)
      *
@@ -1185,13 +1187,14 @@ class U_COMMON_API UnicodeSet final : public UnicodeFilter {
      * \code
      * UnicodeSet set(u"[abcçカ🚴{}{abc}{de}]", errorCode);
      * for (auto el : set) {
+     *     UnicodeString us(el);
      *     std::string u8;
-     *     printf("set.string length %ld \"%s\"\n", (long)el.length(), el.toUTF8String(u8).c_str());
+     *     printf("set.element length %ld \"%s\"\n", (long)us.length(), us.toUTF8String(u8).c_str());
      * }
      * \endcode
      *
      * @return an all-elements iterator.
-     * @draft ICU 76
+     * @draft ICU 77
      * @see end
      * @see codePoints
      * @see ranges
@@ -1203,7 +1206,7 @@ class U_COMMON_API UnicodeSet final : public UnicodeFilter {
 
     /**
      * @return an exclusive-end sentinel for iterating over all of the elements of this set.
-     * @draft ICU 76
+     * @draft ICU 77
      * @see begin
      * @see codePoints
      * @see ranges
diff --git a/deps/icu-small/source/common/unicode/uset.h b/deps/icu-small/source/common/unicode/uset.h
index c8f9b5592df2ea..c5e7f23901b6b7 100644
--- a/deps/icu-small/source/common/unicode/uset.h
+++ b/deps/icu-small/source/common/unicode/uset.h
@@ -32,12 +32,13 @@
 #include "unicode/utypes.h"
 #include "unicode/uchar.h"
 
-#if U_SHOW_CPLUSPLUS_API
+#if U_SHOW_CPLUSPLUS_API || U_SHOW_CPLUSPLUS_HEADER_API
+#include <string>
 #include <string_view>
 #include "unicode/char16ptr.h"
 #include "unicode/localpointer.h"
-#include "unicode/unistr.h"
-#endif   // U_SHOW_CPLUSPLUS_API
+#include "unicode/utf16.h"
+#endif
 
 #ifndef USET_DEFINED
 
@@ -1392,8 +1393,8 @@ class USetCodePointIterator {
 private:
     friend class USetCodePoints;
 
-    USetCodePointIterator(const USet *uset, int32_t rangeIndex, int32_t rangeCount)
-            : uset(uset), rangeIndex(rangeIndex), rangeCount(rangeCount),
+    USetCodePointIterator(const USet *pUset, int32_t nRangeIndex, int32_t nRangeCount)
+            : uset(pUset), rangeIndex(nRangeIndex), rangeCount(nRangeCount),
                 c(U_SENTINEL), end(U_SENTINEL) {
         // Fetch the first range.
         operator++();
@@ -1429,7 +1430,7 @@ class USetCodePoints {
      * Constructs a C++ "range" object over the code points of the USet.
      * @draft ICU 76
      */
-    USetCodePoints(const USet *uset) : uset(uset), rangeCount(uset_getRangeCount(uset)) {}
+    USetCodePoints(const USet *pUset) : uset(pUset), rangeCount(uset_getRangeCount(pUset)) {}
 
     /** @draft ICU 76 */
     USetCodePoints(const USetCodePoints &other) = default;
@@ -1460,7 +1461,7 @@ struct CodePointRange {
     /** @draft ICU 76 */
     struct iterator {
         /** @draft ICU 76 */
-        iterator(UChar32 c) : c(c) {}
+        iterator(UChar32 aC) : c(aC) {}
 
         /** @draft ICU 76 */
         bool operator==(const iterator &other) const { return c == other.c; }
@@ -1573,8 +1574,8 @@ class USetRangeIterator {
 private:
     friend class USetRanges;
 
-    USetRangeIterator(const USet *uset, int32_t rangeIndex, int32_t rangeCount)
-            : uset(uset), rangeIndex(rangeIndex), rangeCount(rangeCount) {}
+    USetRangeIterator(const USet *pUset, int32_t nRangeIndex, int32_t nRangeCount)
+            : uset(pUset), rangeIndex(nRangeIndex), rangeCount(nRangeCount) {}
 
     const USet *uset;
     int32_t rangeIndex;
@@ -1610,7 +1611,7 @@ class USetRanges {
      * Constructs a C++ "range" object over the code point ranges of the USet.
      * @draft ICU 76
      */
-    USetRanges(const USet *uset) : uset(uset), rangeCount(uset_getRangeCount(uset)) {}
+    USetRanges(const USet *pUset) : uset(pUset), rangeCount(uset_getRangeCount(pUset)) {}
 
     /** @draft ICU 76 */
     USetRanges(const USetRanges &other) = default;
@@ -1657,7 +1658,7 @@ class USetStringIterator {
             int32_t length;
             const UChar *uchars = uset_getString(uset, index, &length);
             // assert uchars != nullptr;
-            return {ConstChar16Ptr(uchars), static_cast<uint32_t>(length)};
+            return {uprv_char16PtrFromUChar(uchars), static_cast<size_t>(length)};
         }
         return {};
     }
@@ -1684,8 +1685,8 @@ class USetStringIterator {
 private:
     friend class USetStrings;
 
-    USetStringIterator(const USet *uset, int32_t index, int32_t count)
-            : uset(uset), index(index), count(count) {}
+    USetStringIterator(const USet *pUset, int32_t nIndex, int32_t nCount)
+            : uset(pUset), index(nIndex), count(nCount) {}
 
     const USet *uset;
     int32_t index;
@@ -1699,9 +1700,11 @@ class USetStringIterator {
  * using U_HEADER_NESTED_NAMESPACE::USetStrings;
  * LocalUSetPointer uset(uset_openPattern(u"[abcçカ🚴{}{abc}{de}]", -1, &errorCode));
  * for (auto s : USetStrings(uset.getAlias())) {
- *     UnicodeString us(s);
- *     std::string u8;
- *     printf("uset.string length %ld \"%s\"\n", (long)s.length(), us.toUTF8String(u8).c_str());
+ *     int32_t len32 = s.length();
+ *     char utf8[200];
+ *     u_strToUTF8WithSub(utf8, int32_t{sizeof(utf8) - 1}, nullptr,
+ *                        s.data(), len32, 0xFFFD, nullptr, errorCode);
+ *     printf("uset.string length %ld \"%s\"\n", long{len32}, utf8);
  * }
  * \endcode
  *
@@ -1718,7 +1721,7 @@ class USetStrings {
      * Constructs a C++ "range" object over the strings of the USet.
      * @draft ICU 76
      */
-    USetStrings(const USet *uset) : uset(uset), count(uset_getStringCount(uset)) {}
+    USetStrings(const USet *pUset) : uset(pUset), count(uset_getStringCount(pUset)) {}
 
     /** @draft ICU 76 */
     USetStrings(const USetStrings &other) = default;
@@ -1737,17 +1740,19 @@ class USetStrings {
     const USet *uset;
     int32_t count;
 };
+#endif  // U_HIDE_DRAFT_API
 
+#ifndef U_HIDE_DRAFT_API
 /**
  * Iterator returned by USetElements.
- * @draft ICU 76
+ * @draft ICU 77
  */
 class USetElementIterator {
 public:
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     USetElementIterator(const USetElementIterator &other) = default;
 
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     bool operator==(const USetElementIterator &other) const {
         // No need to compare rangeCount & end given private constructor
         // and assuming we don't compare iterators across the set being modified.
@@ -1756,26 +1761,28 @@ class USetElementIterator {
         return uset == other.uset && c == other.c && index == other.index;
     }
 
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     bool operator!=(const USetElementIterator &other) const { return !operator==(other); }
 
-    /** @draft ICU 76 */
-    UnicodeString operator*() const {
+    /** @draft ICU 77 */
+    std::u16string operator*() const {
         if (c >= 0) {
-            return UnicodeString(c);
+            return c <= 0xffff ?
+                std::u16string({static_cast<char16_t>(c)}) :
+                std::u16string({U16_LEAD(c), U16_TRAIL(c)});
         } else if (index < totalCount) {
             int32_t length;
             const UChar *uchars = uset_getString(uset, index - rangeCount, &length);
             // assert uchars != nullptr;
-            return UnicodeString(uchars, length);
+            return {uprv_char16PtrFromUChar(uchars), static_cast<size_t>(length)};
         } else {
-            return UnicodeString();
+            return {};
         }
     }
 
     /**
      * Pre-increment.
-     * @draft ICU 76
+     * @draft ICU 77
      */
     USetElementIterator &operator++() {
         if (c < end) {
@@ -1800,7 +1807,7 @@ class USetElementIterator {
 
     /**
      * Post-increment.
-     * @draft ICU 76
+     * @draft ICU 77
      */
     USetElementIterator operator++(int) {
         USetElementIterator result(*this);
@@ -1811,8 +1818,8 @@ class USetElementIterator {
 private:
     friend class USetElements;
 
-    USetElementIterator(const USet *uset, int32_t index, int32_t rangeCount, int32_t totalCount)
-            : uset(uset), index(index), rangeCount(rangeCount), totalCount(totalCount),
+    USetElementIterator(const USet *pUset, int32_t nIndex, int32_t nRangeCount, int32_t nTotalCount)
+            : uset(pUset), index(nIndex), rangeCount(nRangeCount), totalCount(nTotalCount),
                 c(U_SENTINEL), end(U_SENTINEL) {
         if (index < rangeCount) {
             // Fetch the first range.
@@ -1840,7 +1847,7 @@ class USetElementIterator {
 
 /**
  * A C++ "range" for iterating over all of the elements of a USet.
- * Convenient all-in one iteration, but creates a UnicodeString for each
+ * Convenient all-in one iteration, but creates a std::u16string for each
  * code point or string.
  *
  * Code points are returned first, then empty and multi-character strings.
@@ -1849,15 +1856,18 @@ class USetElementIterator {
  * using U_HEADER_NESTED_NAMESPACE::USetElements;
  * LocalUSetPointer uset(uset_openPattern(u"[abcçカ🚴{}{abc}{de}]", -1, &errorCode));
  * for (auto el : USetElements(uset.getAlias())) {
- *     std::string u8;
- *     printf("uset.string length %ld \"%s\"\n", (long)el.length(), el.toUTF8String(u8).c_str());
+ *     int32_t len32 = el.length();
+ *     char utf8[200];
+ *     u_strToUTF8WithSub(utf8, int32_t{sizeof(utf8) - 1}, nullptr,
+ *                        el.data(), len32, 0xFFFD, nullptr, errorCode);
+ *     printf("uset.element length %ld \"%s\"\n", long{len32}, utf8);
  * }
  * \endcode
  *
  * C++ UnicodeSet has member functions for iteration, including begin() and end().
  *
  * @return an all-elements iterator.
- * @draft ICU 76
+ * @draft ICU 77
  * @see USetCodePoints
  * @see USetRanges
  * @see USetStrings
@@ -1866,21 +1876,21 @@ class USetElements {
 public:
     /**
      * Constructs a C++ "range" object over all of the elements of the USet.
-     * @draft ICU 76
+     * @draft ICU 77
      */
-    USetElements(const USet *uset)
-        : uset(uset), rangeCount(uset_getRangeCount(uset)),
-            stringCount(uset_getStringCount(uset)) {}
+    USetElements(const USet *pUset)
+        : uset(pUset), rangeCount(uset_getRangeCount(pUset)),
+            stringCount(uset_getStringCount(pUset)) {}
 
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     USetElements(const USetElements &other) = default;
 
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     USetElementIterator begin() const {
         return USetElementIterator(uset, 0, rangeCount, rangeCount + stringCount);
     }
 
-    /** @draft ICU 76 */
+    /** @draft ICU 77 */
     USetElementIterator end() const {
         return USetElementIterator(uset, rangeCount + stringCount, rangeCount, rangeCount + stringCount);
     }
diff --git a/deps/icu-small/source/common/unicode/utf8.h b/deps/icu-small/source/common/unicode/utf8.h
index 5a07435fcf6096..96ad46161aa1e8 100644
--- a/deps/icu-small/source/common/unicode/utf8.h
+++ b/deps/icu-small/source/common/unicode/utf8.h
@@ -124,7 +124,7 @@
  * @internal
  */
 U_CAPI UChar32 U_EXPORT2
-utf8_nextCharSafeBody(const uint8_t *s, int32_t *pi, int32_t length, UChar32 c, UBool strict);
+utf8_nextCharSafeBody(const uint8_t *s, int32_t *pi, int32_t length, UChar32 c, int8_t strict);
 
 /**
  * Function for handling "append code point" with error-checking.
@@ -148,7 +148,7 @@ utf8_appendCharSafeBody(uint8_t *s, int32_t i, int32_t length, UChar32 c, UBool
  * @internal
  */
 U_CAPI UChar32 U_EXPORT2
-utf8_prevCharSafeBody(const uint8_t *s, int32_t start, int32_t *pi, UChar32 c, UBool strict);
+utf8_prevCharSafeBody(const uint8_t *s, int32_t start, int32_t *pi, UChar32 c, int8_t strict);
 
 /**
  * Function for handling "skip backward one code point" with error-checking.
diff --git a/deps/icu-small/source/common/unicode/utypes.h b/deps/icu-small/source/common/unicode/utypes.h
index 0151ebd4701576..ecdee51643166e 100644
--- a/deps/icu-small/source/common/unicode/utypes.h
+++ b/deps/icu-small/source/common/unicode/utypes.h
@@ -598,12 +598,13 @@ typedef enum UErrorCode {
     U_MF_DUPLICATE_DECLARATION_ERROR, /**< The same variable is declared in more than one .local or .input declaration. @internal ICU 75 technology preview @deprecated This API is for technology preview only. */
     U_MF_OPERAND_MISMATCH_ERROR,     /**< An operand provided to a function does not have the required form for that function @internal ICU 75 technology preview @deprecated This API is for technology preview only. */
     U_MF_DUPLICATE_VARIANT_ERROR, /**< A message includes a variant with the same key list as another variant. @internal ICU 76 technology preview @deprecated This API is for technology preview only. */
+    U_MF_BAD_OPTION,             /**< An option value provided to a function does not have the required form for that option. @internal ICU 77 technology preview @deprecated This API is for technology preview only. */
 #ifndef U_HIDE_DEPRECATED_API
     /**
      * One more than the highest normal formatting API error code.
      * @deprecated ICU 58 The numeric value may change over time, see ICU ticket #12420.
      */
-    U_FMT_PARSE_ERROR_LIMIT = 0x10120,
+    U_FMT_PARSE_ERROR_LIMIT = 0x10121,
 #endif  // U_HIDE_DEPRECATED_API
 
     /*
diff --git a/deps/icu-small/source/common/unicode/uvernum.h b/deps/icu-small/source/common/unicode/uvernum.h
index a3cb882623be86..847c49f4cfb8cd 100644
--- a/deps/icu-small/source/common/unicode/uvernum.h
+++ b/deps/icu-small/source/common/unicode/uvernum.h
@@ -53,7 +53,7 @@
  *  This value will change in the subsequent releases of ICU
  *  @stable ICU 2.4
  */
-#define U_ICU_VERSION_MAJOR_NUM 76
+#define U_ICU_VERSION_MAJOR_NUM 77
 
 /** The current ICU minor version as an integer.
  *  This value will change in the subsequent releases of ICU
@@ -79,7 +79,7 @@
  *  This value will change in the subsequent releases of ICU
  *  @stable ICU 2.6
  */
-#define U_ICU_VERSION_SUFFIX _76
+#define U_ICU_VERSION_SUFFIX _77
 
 /**
  * \def U_DEF2_ICU_ENTRY_POINT_RENAME
@@ -132,7 +132,7 @@
  *  This value will change in the subsequent releases of ICU
  *  @stable ICU 2.4
  */
-#define U_ICU_VERSION "76.1"
+#define U_ICU_VERSION "77.1"
 
 /**
  * The current ICU library major version number as a string, for library name suffixes.
@@ -145,13 +145,13 @@
  *
  * @stable ICU 2.6
  */
-#define U_ICU_VERSION_SHORT "76"
+#define U_ICU_VERSION_SHORT "77"
 
 #ifndef U_HIDE_INTERNAL_API
 /** Data version in ICU4C.
  * @internal ICU 4.4 Internal Use Only
  **/
-#define U_ICU_DATA_VERSION "76.1"
+#define U_ICU_DATA_VERSION "77.1"
 #endif  /* U_HIDE_INTERNAL_API */
 
 /*===========================================================================
diff --git a/deps/icu-small/source/common/unicode/uversion.h b/deps/icu-small/source/common/unicode/uversion.h
index 25d73a3aeb5449..a29bf21efda597 100644
--- a/deps/icu-small/source/common/unicode/uversion.h
+++ b/deps/icu-small/source/common/unicode/uversion.h
@@ -125,7 +125,7 @@ typedef uint8_t UVersionInfo[U_MAX_VERSION_LENGTH];
         U_NAMESPACE_USE
 #   endif
 
-#ifndef U_HIDE_DRAFT_API
+#ifndef U_FORCE_HIDE_DRAFT_API
 /**
  * \def U_HEADER_NESTED_NAMESPACE
  * Nested namespace used inside U_ICU_NAMESPACE for header-only APIs.
@@ -150,22 +150,37 @@ typedef uint8_t UVersionInfo[U_MAX_VERSION_LENGTH];
  * @draft ICU 76
  */
 
+/**
+ * \def U_ICU_NAMESPACE_OR_INTERNAL
+ * Namespace used for header-only APIs that used to be regular C++ APIs.
+ * Different when used inside ICU to prevent public use of internal instantiations.
+ * Similar to U_HEADER_ONLY_NAMESPACE, but the public definition is the same as U_ICU_NAMESPACE.
+ * "U_ICU_NAMESPACE" or "U_ICU_NAMESPACE::internal".
+ *
+ * @draft ICU 77
+ */
+
 // The first test is the same as for defining U_EXPORT for Windows.
 #if defined(_MSC_VER) || (UPRV_HAS_DECLSPEC_ATTRIBUTE(__dllexport__) && \
                           UPRV_HAS_DECLSPEC_ATTRIBUTE(__dllimport__))
 #   define U_HEADER_NESTED_NAMESPACE header
+#   define U_ICU_NAMESPACE_OR_INTERNAL U_ICU_NAMESPACE
 #elif defined(U_COMBINED_IMPLEMENTATION) || defined(U_COMMON_IMPLEMENTATION) || \
         defined(U_I18N_IMPLEMENTATION) || defined(U_IO_IMPLEMENTATION) || \
         defined(U_LAYOUTEX_IMPLEMENTATION) || defined(U_TOOLUTIL_IMPLEMENTATION)
 #   define U_HEADER_NESTED_NAMESPACE internal
+#   define U_ICU_NAMESPACE_OR_INTERNAL U_ICU_NAMESPACE::internal
+    namespace U_ICU_NAMESPACE_OR_INTERNAL {}
+    using namespace U_ICU_NAMESPACE_OR_INTERNAL;
 #else
 #   define U_HEADER_NESTED_NAMESPACE header
+#   define U_ICU_NAMESPACE_OR_INTERNAL U_ICU_NAMESPACE
 #endif
 
 #define U_HEADER_ONLY_NAMESPACE U_ICU_NAMESPACE::U_HEADER_NESTED_NAMESPACE
 
 namespace U_HEADER_ONLY_NAMESPACE {}
-#endif  // U_HIDE_DRAFT_API
+#endif  // U_FORCE_HIDE_DRAFT_API
 
 #endif /* __cplusplus */
 
diff --git a/deps/icu-small/source/common/unistr.cpp b/deps/icu-small/source/common/unistr.cpp
index a720245772e637..4e29bad1d3b971 100644
--- a/deps/icu-small/source/common/unistr.cpp
+++ b/deps/icu-small/source/common/unistr.cpp
@@ -1945,6 +1945,13 @@ UnicodeString::cloneArrayIfNeeded(int32_t newCapacity,
       growCapacity = newCapacity;
     } else if(newCapacity <= US_STACKBUF_SIZE && growCapacity > US_STACKBUF_SIZE) {
       growCapacity = US_STACKBUF_SIZE;
+    } else if(newCapacity > growCapacity) {
+      setToBogus();
+      return false;  // bad inputs
+    }
+    if(growCapacity > kMaxCapacity) {
+      setToBogus();
+      return false;
     }
 
     // save old values
diff --git a/deps/icu-small/source/common/uresbund.cpp b/deps/icu-small/source/common/uresbund.cpp
index 3a09cbf3bcaac6..afda2770fd3fc2 100644
--- a/deps/icu-small/source/common/uresbund.cpp
+++ b/deps/icu-small/source/common/uresbund.cpp
@@ -2716,6 +2716,9 @@ ures_openWithType(UResourceBundle *r, const char* path, const char* localeID,
 
     UResourceDataEntry *entry;
     if(openType != URES_OPEN_DIRECT) {
+        if (localeID == nullptr) {
+            localeID = uloc_getDefault();
+        }
         /* first "canonicalize" the locale ID */
         CharString canonLocaleID = ulocimp_getBaseName(localeID, *status);
         if(U_FAILURE(*status)) {
@@ -3080,6 +3083,9 @@ ures_getFunctionalEquivalent(char *result, int32_t resultCapacity,
             kwVal.clear();
         }
     }
+    if (locid == nullptr) {
+        locid = uloc_getDefault();
+    }
     CharString base = ulocimp_getBaseName(locid, subStatus);
 #if defined(URES_TREE_DEBUG)
     fprintf(stderr, "getFunctionalEquivalent: \"%s\" [%s=%s] in %s - %s\n", 
@@ -3244,7 +3250,7 @@ ures_getFunctionalEquivalent(char *result, int32_t resultCapacity,
             const char *validLoc = ures_getLocaleByType(res, ULOC_VALID_LOCALE, &subStatus);
             if (U_SUCCESS(subStatus) && validLoc != nullptr && validLoc[0] != 0 && uprv_strcmp(validLoc, "root") != 0) {
                 CharString validLang = ulocimp_getLanguage(validLoc, subStatus);
-                CharString parentLang = ulocimp_getLanguage(parent.data(), subStatus);
+                CharString parentLang = ulocimp_getLanguage(parent.toStringPiece(), subStatus);
                 if (U_SUCCESS(subStatus) && validLang != parentLang) {
                     // validLoc is not root and has a different language than parent, use it instead
                     found.clear().append(validLoc, subStatus);
diff --git a/deps/icu-small/source/common/uscript.cpp b/deps/icu-small/source/common/uscript.cpp
index c48a28fd14345a..ce40d354958fdd 100644
--- a/deps/icu-small/source/common/uscript.cpp
+++ b/deps/icu-small/source/common/uscript.cpp
@@ -59,6 +59,9 @@ getCodesFromLocale(const char *locale,
     if (U_FAILURE(*err)) { return 0; }
     icu::CharString lang;
     icu::CharString script;
+    if (locale == nullptr) {
+        locale = uloc_getDefault();
+    }
     ulocimp_getSubtags(locale, &lang, &script, nullptr, nullptr, nullptr, *err);
     if (U_FAILURE(*err)) { return 0; }
     // Multi-script languages, equivalent to the LocaleScript data
diff --git a/deps/icu-small/source/common/ushape.cpp b/deps/icu-small/source/common/ushape.cpp
index 00125635cb2672..b7946dc3ce03f2 100644
--- a/deps/icu-small/source/common/ushape.cpp
+++ b/deps/icu-small/source/common/ushape.cpp
@@ -28,6 +28,7 @@
 #include "ubidi_props.h"
 #include "uassert.h"
 
+#include <limits>
 /*
  * This implementation is designed for 16-bit Unicode strings.
  * The main assumption is that the Arabic characters and their
@@ -747,6 +748,10 @@ handleGeneratedSpaces(char16_t *dest, int32_t sourceLength,
         }
     }
 
+    if (static_cast<size_t>(sourceLength) + 1 > std::numeric_limits<size_t>::max() / U_SIZEOF_UCHAR) {
+        *pErrorCode = U_INDEX_OUTOFBOUNDS_ERROR;
+        return 0;
+    }
     tempbuffer = static_cast<char16_t*>(uprv_malloc((sourceLength + 1) * U_SIZEOF_UCHAR));
     /* Test for nullptr */
     if(tempbuffer == nullptr) {
diff --git a/deps/icu-small/source/common/usprep.cpp b/deps/icu-small/source/common/usprep.cpp
index 477b8f2309db53..048b423645f825 100644
--- a/deps/icu-small/source/common/usprep.cpp
+++ b/deps/icu-small/source/common/usprep.cpp
@@ -126,7 +126,7 @@ compareEntries(const UHashTok p1, const UHashTok p2) {
     name2.pointer = b2->name;
     path1.pointer = b1->path;
     path2.pointer = b2->path;
-    return uhash_compareChars(name1, name2) & uhash_compareChars(path1, path2);
+    return uhash_compareChars(name1, name2) && uhash_compareChars(path1, path2);
 }
 
 static void 
diff --git a/deps/icu-small/source/common/utf_impl.cpp b/deps/icu-small/source/common/utf_impl.cpp
index 827a82daf403a3..7da10c9b2d36d5 100644
--- a/deps/icu-small/source/common/utf_impl.cpp
+++ b/deps/icu-small/source/common/utf_impl.cpp
@@ -124,11 +124,9 @@ errorValue(int32_t count, int8_t strict) {
  * >0  Obsolete "strict" behavior of UTF8_NEXT_CHAR_SAFE(..., true):
  *     Same as the obsolete "safe" behavior, but non-characters are also treated
  *     like illegal sequences.
- *
- * Note that a UBool is the same as an int8_t.
  */
 U_CAPI UChar32 U_EXPORT2
-utf8_nextCharSafeBody(const uint8_t *s, int32_t *pi, int32_t length, UChar32 c, UBool strict) {
+utf8_nextCharSafeBody(const uint8_t *s, int32_t *pi, int32_t length, UChar32 c, int8_t strict) {
     // *pi is one after byte c.
     int32_t i=*pi;
     // length can be negative for NUL-terminated strings: Read and validate one byte at a time.
@@ -233,7 +231,7 @@ utf8_appendCharSafeBody(uint8_t *s, int32_t i, int32_t length, UChar32 c, UBool
 }
 
 U_CAPI UChar32 U_EXPORT2
-utf8_prevCharSafeBody(const uint8_t *s, int32_t start, int32_t *pi, UChar32 c, UBool strict) {
+utf8_prevCharSafeBody(const uint8_t *s, int32_t start, int32_t *pi, UChar32 c, int8_t strict) {
     // *pi is the index of byte c.
     int32_t i=*pi;
     if(U8_IS_TRAIL(c) && i>start) {
diff --git a/deps/icu-small/source/common/utypes.cpp b/deps/icu-small/source/common/utypes.cpp
index 4602314147f19e..4d4c1f81b5e6b8 100644
--- a/deps/icu-small/source/common/utypes.cpp
+++ b/deps/icu-small/source/common/utypes.cpp
@@ -140,7 +140,8 @@ _uFmtErrorName[U_FMT_PARSE_ERROR_LIMIT - U_FMT_PARSE_ERROR_START] = {
     "U_MF_MISSING_SELECTOR_ANNOTATION_ERROR",
     "U_MF_DUPLICATE_DECLARATION_ERROR",
     "U_MF_OPERAND_MISMATCH_ERROR",
-    "U_MF_DUPLICATE_VARIANT_ERROR"
+    "U_MF_DUPLICATE_VARIANT_ERROR",
+    "U_MF_BAD_OPTION"
 };
 
 static const char * const
diff --git a/deps/icu-small/source/data/in/icudt76l.dat.bz2 b/deps/icu-small/source/data/in/icudt77l.dat.bz2
similarity index 57%
rename from deps/icu-small/source/data/in/icudt76l.dat.bz2
rename to deps/icu-small/source/data/in/icudt77l.dat.bz2
index 8843d4561126d3..1f298a4c2affe7 100644
Binary files a/deps/icu-small/source/data/in/icudt76l.dat.bz2 and b/deps/icu-small/source/data/in/icudt77l.dat.bz2 differ
diff --git a/deps/icu-small/source/i18n/basictz.cpp b/deps/icu-small/source/i18n/basictz.cpp
index a2c1ec7fb9142d..610a31ad5dc559 100644
--- a/deps/icu-small/source/i18n/basictz.cpp
+++ b/deps/icu-small/source/i18n/basictz.cpp
@@ -160,12 +160,13 @@ BasicTimeZone::getSimpleRulesNear(UDate date, InitialTimeZoneRule*& initial,
               || (tr.getFrom()->getDSTSavings() != 0 && tr.getTo()->getDSTSavings() == 0))
             && (date + MILLIS_PER_YEAR > nextTransitionTime)) {
  
-            int32_t year, month, dom, dow, doy, mid;
+            int32_t year, mid;
+            int8_t month, dom, dow;
             UDate d;
 
             // Get local wall time for the next transition time
             Grego::timeToFields(nextTransitionTime + initialRaw + initialDst,
-                year, month, dom, dow, doy, mid, status);
+                year, month, dom, dow, mid, status);
             if (U_FAILURE(status)) return;
             int32_t weekInMonth = Grego::dayOfWeekInMonth(year, month, dom);
             // Create DOW rule
@@ -193,7 +194,7 @@ BasicTimeZone::getSimpleRulesNear(UDate date, InitialTimeZoneRule*& initial,
 
                         // Get local wall time for the next transition time
                         Grego::timeToFields(tr.getTime() + tr.getFrom()->getRawOffset() + tr.getFrom()->getDSTSavings(),
-                            year, month, dom, dow, doy, mid, status);
+                            year, month, dom, dow, mid, status);
                         if (U_FAILURE(status)) return;
                         weekInMonth = Grego::dayOfWeekInMonth(year, month, dom);
                         // Generate another DOW rule
@@ -225,7 +226,7 @@ BasicTimeZone::getSimpleRulesNear(UDate date, InitialTimeZoneRule*& initial,
 
                         // Generate another DOW rule
                         Grego::timeToFields(tr.getTime() + tr.getFrom()->getRawOffset() + tr.getFrom()->getDSTSavings(),
-                            year, month, dom, dow, doy, mid, status);
+                            year, month, dom, dow, mid, status);
                         if (U_FAILURE(status)) return;
                         weekInMonth = Grego::dayOfWeekInMonth(year, month, dom);
                         dtr = new DateTimeRule(month, weekInMonth, dow, mid, DateTimeRule::WALL_TIME);
@@ -486,8 +487,7 @@ BasicTimeZone::getTimeZoneRulesAfter(UDate start, InitialTimeZoneRule*& initial,
                 }
             } else {
                 // Calculate the transition year
-                int32_t year, month, dom, dow, doy, mid;
-                Grego::timeToFields(tzt.getTime(), year, month, dom, dow, doy, mid, status);
+                int32_t year = Grego::timeToYear(tzt.getTime(), status);
                 if (U_FAILURE(status)) {
                     return;
                 }
diff --git a/deps/icu-small/source/i18n/buddhcal.cpp b/deps/icu-small/source/i18n/buddhcal.cpp
index 7723ade105d2c5..c99b97e26713c8 100644
--- a/deps/icu-small/source/i18n/buddhcal.cpp
+++ b/deps/icu-small/source/i18n/buddhcal.cpp
@@ -36,7 +36,6 @@ static const int32_t kGregorianEpoch = 1970;    // used as the default value of
 BuddhistCalendar::BuddhistCalendar(const Locale& aLocale, UErrorCode& success)
 :   GregorianCalendar(aLocale, success)
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 BuddhistCalendar::~BuddhistCalendar()
@@ -48,12 +47,6 @@ BuddhistCalendar::BuddhistCalendar(const BuddhistCalendar& source)
 {
 }
 
-BuddhistCalendar& BuddhistCalendar::operator= ( const BuddhistCalendar& right)
-{
-    GregorianCalendar::operator=(right);
-    return *this;
-}
-
 BuddhistCalendar* BuddhistCalendar::clone() const
 {
     return new BuddhistCalendar(*this);
diff --git a/deps/icu-small/source/i18n/buddhcal.h b/deps/icu-small/source/i18n/buddhcal.h
index 1fa8395b35b216..196b21311fdaa9 100644
--- a/deps/icu-small/source/i18n/buddhcal.h
+++ b/deps/icu-small/source/i18n/buddhcal.h
@@ -82,13 +82,6 @@ class BuddhistCalendar : public GregorianCalendar {
      */
     BuddhistCalendar(const BuddhistCalendar& source);
 
-    /**
-     * Default assignment operator
-     * @param right    the object to be copied.
-     * @internal
-     */
-    BuddhistCalendar& operator=(const BuddhistCalendar& right);
-
     /**
      * Create and return a polymorphic copy of this calendar.
      * @return    return a polymorphic copy of this calendar.
diff --git a/deps/icu-small/source/i18n/calendar.cpp b/deps/icu-small/source/i18n/calendar.cpp
index 96247174f70d6b..cac237d2b67d8f 100644
--- a/deps/icu-small/source/i18n/calendar.cpp
+++ b/deps/icu-small/source/i18n/calendar.cpp
@@ -156,7 +156,7 @@ U_CFUNC void ucal_dump(UCalendar* cal) {
 #endif
 
 /* Max value for stamp allowable before recalculation */
-#define STAMP_MAX 10000
+#define STAMP_MAX 127
 
 static const char * const gCalTypes[] = {
     "gregorian",
@@ -700,15 +700,10 @@ fIsTimeSet(false),
 fAreFieldsSet(false),
 fAreAllFieldsSet(false),
 fAreFieldsVirtuallySet(false),
-fNextStamp(static_cast<int32_t>(kMinimumUserStamp)),
-fTime(0),
 fLenient(true),
-fZone(nullptr),
 fRepeatedWallTime(UCAL_WALLTIME_LAST),
 fSkippedWallTime(UCAL_WALLTIME_LAST)
 {
-    validLocale[0] = 0;
-    actualLocale[0] = 0;
     clear();
     if (U_FAILURE(success)) {
         return;
@@ -722,26 +717,21 @@ fSkippedWallTime(UCAL_WALLTIME_LAST)
 
 // -------------------------------------
 
-Calendar::Calendar(TimeZone* zone, const Locale& aLocale, UErrorCode& success)
+Calendar::Calendar(TimeZone* adoptZone, const Locale& aLocale, UErrorCode& success)
 :   UObject(),
 fIsTimeSet(false),
 fAreFieldsSet(false),
 fAreAllFieldsSet(false),
 fAreFieldsVirtuallySet(false),
-fNextStamp(static_cast<int32_t>(kMinimumUserStamp)),
-fTime(0),
 fLenient(true),
-fZone(nullptr),
 fRepeatedWallTime(UCAL_WALLTIME_LAST),
 fSkippedWallTime(UCAL_WALLTIME_LAST)
 {
-    validLocale[0] = 0;
-    actualLocale[0] = 0;
+    LocalPointer<TimeZone> zone(adoptZone, success);
     if (U_FAILURE(success)) {
-        delete zone;
         return;
     }
-    if (zone == nullptr) {
+    if (zone.isNull()) {
 #if defined (U_DEBUG_CAL)
         fprintf(stderr, "%s:%d: ILLEGAL ARG because timezone cannot be 0\n",
             __FILE__, __LINE__);
@@ -751,7 +741,7 @@ fSkippedWallTime(UCAL_WALLTIME_LAST)
     }
 
     clear();
-    fZone = zone;
+    fZone = zone.orphan();
     setWeekData(aLocale, nullptr, success);
 }
 
@@ -763,15 +753,10 @@ fIsTimeSet(false),
 fAreFieldsSet(false),
 fAreAllFieldsSet(false),
 fAreFieldsVirtuallySet(false),
-fNextStamp(static_cast<int32_t>(kMinimumUserStamp)),
-fTime(0),
 fLenient(true),
-fZone(nullptr),
 fRepeatedWallTime(UCAL_WALLTIME_LAST),
 fSkippedWallTime(UCAL_WALLTIME_LAST)
 {
-    validLocale[0] = 0;
-    actualLocale[0] = 0;
     if (U_FAILURE(success)) {
         return;
     }
@@ -779,6 +764,7 @@ fSkippedWallTime(UCAL_WALLTIME_LAST)
     fZone = zone.clone();
     if (fZone == nullptr) {
         success = U_MEMORY_ALLOCATION_ERROR;
+        return;
     }
     setWeekData(aLocale, nullptr, success);
 }
@@ -788,6 +774,8 @@ fSkippedWallTime(UCAL_WALLTIME_LAST)
 Calendar::~Calendar()
 {
     delete fZone;
+    delete actualLocale;
+    delete validLocale;
 }
 
 // -------------------------------------
@@ -795,7 +783,6 @@ Calendar::~Calendar()
 Calendar::Calendar(const Calendar &source)
 :   UObject(source)
 {
-    fZone = nullptr;
     *this = source;
 }
 
@@ -806,7 +793,6 @@ Calendar::operator=(const Calendar &right)
 {
     if (this != &right) {
         uprv_arrayCopy(right.fFields, fFields, UCAL_FIELD_COUNT);
-        uprv_arrayCopy(right.fIsSet, fIsSet, UCAL_FIELD_COUNT);
         uprv_arrayCopy(right.fStamp, fStamp, UCAL_FIELD_COUNT);
         fTime                    = right.fTime;
         fIsTimeSet               = right.fIsTimeSet;
@@ -828,10 +814,10 @@ Calendar::operator=(const Calendar &right)
         fWeekendCease            = right.fWeekendCease;
         fWeekendCeaseMillis      = right.fWeekendCeaseMillis;
         fNextStamp               = right.fNextStamp;
-        uprv_strncpy(validLocale, right.validLocale, sizeof(validLocale));
-        uprv_strncpy(actualLocale, right.actualLocale, sizeof(actualLocale));
-        validLocale[sizeof(validLocale)-1] = 0;
-        actualLocale[sizeof(validLocale)-1] = 0;
+        UErrorCode status = U_ZERO_ERROR;
+        U_LOCALE_BASED(locBased, *this);
+        locBased.setLocaleIDs(right.validLocale, right.actualLocale, status);
+        U_ASSERT(U_SUCCESS(status));
     }
 
     return *this;
@@ -1167,13 +1153,9 @@ Calendar::setTimeInMillis( double millis, UErrorCode& status ) {
     fAreFieldsSet = fAreAllFieldsSet = false;
     fIsTimeSet = fAreFieldsVirtuallySet = true;
 
-    for (int32_t i=0; i<UCAL_FIELD_COUNT; ++i) {
-        fFields[i]     = 0;
-        fStamp[i]     = kUnset;
-        fIsSet[i]     = false;
-    }
-
-
+    uprv_memset(fFields, 0, sizeof(fFields));
+    uprv_memset(fStamp, kUnset, sizeof(fStamp));
+    fNextStamp = kMinimumUserStamp;
 }
 
 // -------------------------------------
@@ -1208,12 +1190,11 @@ Calendar::set(UCalendarDateFields field, int32_t value)
         computeFields(ec);
     }
     fFields[field]     = value;
-    /* Ensure that the fNextStamp value doesn't go pass max value for int32_t */
+    /* Ensure that the fNextStamp value doesn't go pass max value for int8_t */
     if (fNextStamp == STAMP_MAX) {
         recalculateStamp();
     }
     fStamp[field]     = fNextStamp++;
-    fIsSet[field]     = true; // Remove later
     fIsTimeSet = fAreFieldsSet = fAreFieldsVirtuallySet = false;
 }
 
@@ -1270,11 +1251,9 @@ void Calendar::setRelatedYear(int32_t year)
 void
 Calendar::clear()
 {
-    for (int32_t i=0; i<UCAL_FIELD_COUNT; ++i) {
-        fFields[i]     = 0; // Must do this; other code depends on it
-        fStamp[i]     = kUnset;
-        fIsSet[i]     = false; // Remove later
-    }
+    uprv_memset(fFields, 0, sizeof(fFields));
+    uprv_memset(fStamp, kUnset, sizeof(fStamp));
+    fNextStamp = kMinimumUserStamp;
     fIsTimeSet = fAreFieldsSet = fAreAllFieldsSet = fAreFieldsVirtuallySet = false;
     // fTime is not 'cleared' - may be used if no fields are set.
 }
@@ -1296,12 +1275,10 @@ Calendar::clear(UCalendarDateFields field)
     if (field == UCAL_MONTH) {
         fFields[UCAL_ORDINAL_MONTH]         = 0;
         fStamp[UCAL_ORDINAL_MONTH]         = kUnset;
-        fIsSet[UCAL_ORDINAL_MONTH]         = false; // Remove later
     }
     if (field == UCAL_ORDINAL_MONTH) {
         fFields[UCAL_MONTH]         = 0;
         fStamp[UCAL_MONTH]         = kUnset;
-        fIsSet[UCAL_MONTH]         = false; // Remove later
     }
     fIsTimeSet = fAreFieldsSet = fAreAllFieldsSet = fAreFieldsVirtuallySet = false;
 }
@@ -1434,10 +1411,8 @@ void Calendar::computeFields(UErrorCode &ec)
     for (int32_t i=0; i<UCAL_FIELD_COUNT; ++i) {
         if ((mask & 1) == 0) {
             fStamp[i] = kInternallySet;
-            fIsSet[i] = true; // Remove later
         } else {
             fStamp[i] = kUnset;
-            fIsSet[i] = false; // Remove later
         }
         mask >>= 1;
     }
@@ -1467,7 +1442,7 @@ void Calendar::computeFields(UErrorCode &ec)
     //__FILE__, __LINE__, fFields[UCAL_JULIAN_DAY], localMillis);
 #endif
 
-    computeGregorianAndDOWFields(fFields[UCAL_JULIAN_DAY], ec);
+    computeGregorianFields(fFields[UCAL_JULIAN_DAY], ec);
 
     // Call framework method to have subclass compute its fields.
     // These must include, at a minimum, MONTH, DAY_OF_MONTH,
@@ -1538,32 +1513,6 @@ uint8_t Calendar::julianDayToDayOfWeek(int32_t julian)
     return result;
 }
 
-/**
-* Compute the Gregorian calendar year, month, and day of month from
-* the given Julian day.  These values are not stored in fields, but in
-* member variables gregorianXxx.  Also compute the DAY_OF_WEEK and
-* DOW_LOCAL fields.
-*/
-void Calendar::computeGregorianAndDOWFields(int32_t julianDay, UErrorCode &ec)
-{
-    computeGregorianFields(julianDay, ec);
-    if (U_FAILURE(ec)) {
-        return;
-    }
-
-    // Compute day of week: JD 0 = Monday
-    int32_t dow = julianDayToDayOfWeek(julianDay);
-    internalSet(UCAL_DAY_OF_WEEK,dow);
-
-    // Calculate 1-based localized day of week
-    int32_t dowLocal = dow - getFirstDayOfWeek() + 1;
-    if (dowLocal < 1) {
-        dowLocal += 7;
-    }
-    internalSet(UCAL_DOW_LOCAL,dowLocal);
-    fFields[UCAL_DOW_LOCAL] = dowLocal;
-}
-
 /**
 * Compute the Gregorian calendar year, month, and day of month from the
 * Julian day.  These values are not stored in fields, but in member
@@ -1575,14 +1524,13 @@ void Calendar::computeGregorianFields(int32_t julianDay, UErrorCode& ec) {
     if (U_FAILURE(ec)) {
         return;
     }
-    int32_t gregorianDayOfWeekUnused;
     if (uprv_add32_overflow(
             julianDay, -kEpochStartAsJulianDay, &julianDay)) {
         ec = U_ILLEGAL_ARGUMENT_ERROR;
         return;
     }
     Grego::dayToFields(julianDay, fGregorianYear, fGregorianMonth,
-                       fGregorianDayOfMonth, gregorianDayOfWeekUnused,
+                       fGregorianDayOfMonth,
                        fGregorianDayOfYear, ec);
 }
 
@@ -1610,8 +1558,19 @@ void Calendar::computeWeekFields(UErrorCode &ec) {
     if(U_FAILURE(ec)) {
         return;
     }
+
+    // Compute day of week: JD 0 = Monday
+    int32_t dayOfWeek = julianDayToDayOfWeek(fFields[UCAL_JULIAN_DAY]);
+    internalSet(UCAL_DAY_OF_WEEK, dayOfWeek);
+    int32_t firstDayOfWeek = getFirstDayOfWeek();
+    // Calculate 1-based localized day of week
+    int32_t dowLocal = dayOfWeek - firstDayOfWeek + 1;
+    if (dowLocal < 1) {
+        dowLocal += 7;
+    }
+    internalSet(UCAL_DOW_LOCAL,dowLocal);
+
     int32_t eyear = fFields[UCAL_EXTENDED_YEAR];
-    int32_t dayOfWeek = fFields[UCAL_DAY_OF_WEEK];
     int32_t dayOfYear = fFields[UCAL_DAY_OF_YEAR];
 
     // WEEK_OF_YEAR start
@@ -1624,10 +1583,11 @@ void Calendar::computeWeekFields(UErrorCode &ec) {
     // first week of the next year.  ASSUME that the year length is less than
     // 7000 days.
     int32_t yearOfWeekOfYear = eyear;
-    int32_t relDow = (dayOfWeek + 7 - getFirstDayOfWeek()) % 7; // 0..6
-    int32_t relDowJan1 = (dayOfWeek - dayOfYear + 7001 - getFirstDayOfWeek()) % 7; // 0..6
+    int32_t relDow = (dayOfWeek + 7 - firstDayOfWeek) % 7; // 0..6
+    int32_t relDowJan1 = (dayOfWeek - dayOfYear + 7001 - firstDayOfWeek) % 7; // 0..6
     int32_t woy = (dayOfYear - 1 + relDowJan1) / 7; // 0..53
-    if ((7 - relDowJan1) >= getMinimalDaysInFirstWeek()) {
+    int32_t minimalDaysInFirstWeek = getMinimalDaysInFirstWeek();
+    if ((7 - relDowJan1) >= minimalDaysInFirstWeek) {
         ++woy;
     }
 
@@ -1639,11 +1599,13 @@ void Calendar::computeWeekFields(UErrorCode &ec) {
         // to handle the case in which we are the first week of the
         // next year.
 
-        int32_t prevDoy = dayOfYear + handleGetYearLength(eyear - 1);
+        int32_t prevDoy = dayOfYear + handleGetYearLength(eyear - 1, ec);
+        if(U_FAILURE(ec)) return;
         woy = weekNumber(prevDoy, dayOfWeek);
         yearOfWeekOfYear--;
     } else {
-        int32_t lastDoy = handleGetYearLength(eyear);
+        int32_t lastDoy = handleGetYearLength(eyear, ec);
+        if(U_FAILURE(ec)) return;
         // Fast check: For it to be week 1 of the next year, the DOY
         // must be on or after L-5, where L is yearLength(), then it
         // cannot possibly be week 1 of the next year:
@@ -1655,7 +1617,7 @@ void Calendar::computeWeekFields(UErrorCode &ec) {
             if (lastRelDow < 0) {
                 lastRelDow += 7;
             }
-            if (((6 - lastRelDow) >= getMinimalDaysInFirstWeek()) &&
+            if (((6 - lastRelDow) >= minimalDaysInFirstWeek) &&
                 ((dayOfYear + 7 - relDow) > lastDoy)) {
                     woy = 1;
                     yearOfWeekOfYear++;
@@ -2946,7 +2908,7 @@ void Calendar::validateField(UCalendarDateFields field, UErrorCode &status) {
         if (U_FAILURE(status)) {
            return;
         }
-        validateField(field, 1, handleGetYearLength(y), status);
+        validateField(field, 1, handleGetYearLength(y, status), status);
         break;
     case UCAL_DAY_OF_WEEK_IN_MONTH:
         if (internalGet(field) == 0) {
@@ -3607,9 +3569,19 @@ int32_t Calendar::handleComputeJulianDay(UCalendarDateFields bestField, UErrorCo
                     fprintf(stderr, "%s:%d - y=%d, y-1=%d doy%d, njd%d (C.F. %d)\n",
                         __FILE__, __LINE__, year, year-1, testDate, julianDay+testDate, nextJulianDay);
 #endif
-                    if(julianDay+testDate > nextJulianDay) { // is it past Dec 31?  (nextJulianDay is day BEFORE year+1's  Jan 1)
+                    if (uprv_add32_overflow(julianDay, testDate, &testDate)) {
+                        status = U_ILLEGAL_ARGUMENT_ERROR;
+                        return 0;
+                    }
+
+                    if(testDate > nextJulianDay) { // is it past Dec 31?  (nextJulianDay is day BEFORE year+1's  Jan 1)
                         // Fire up the calculating engines.. retry YWOY = (year-1)
-                        julianDay = handleComputeMonthStart(year-1, 0, false, status); // jd before Jan 1 of previous year
+                        int32_t prevYear;
+                        if (uprv_add32_overflow(year, -1, &prevYear)) {
+                            status = U_ILLEGAL_ARGUMENT_ERROR;
+                            return 0;
+                        }
+                        julianDay = handleComputeMonthStart(prevYear, 0, false, status); // jd before Jan 1 of previous year
                         if (U_FAILURE(status)) {
                             return 0;
                         }
@@ -3834,16 +3806,20 @@ int32_t Calendar::handleGetExtendedYearFromWeekFields(int32_t yearWoy, int32_t w
 
 int32_t Calendar::handleGetMonthLength(int32_t extendedYear, int32_t month, UErrorCode& status) const
 {
-    return handleComputeMonthStart(extendedYear, month+1, true, status) -
+    int32_t nextMonth;
+    if (uprv_add32_overflow(month, 1, &nextMonth)) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return 0;
+    }
+    return handleComputeMonthStart(extendedYear, nextMonth, true, status) -
         handleComputeMonthStart(extendedYear, month, true, status);
 }
 
-int32_t Calendar::handleGetYearLength(int32_t eyear) const
+int32_t Calendar::handleGetYearLength(int32_t eyear, UErrorCode& status) const
 {
-    UErrorCode status = U_ZERO_ERROR;
     int32_t result = handleComputeMonthStart(eyear+1, 0, false, status) -
         handleComputeMonthStart(eyear, 0, false, status);
-    U_ASSERT(U_SUCCESS(status));
+    if (U_FAILURE(status)) return 0;
     return result;
 }
 
@@ -3882,7 +3858,7 @@ Calendar::getActualMaximum(UCalendarDateFields field, UErrorCode& status) const
             }
             cal->setLenient(true);
             cal->prepareGetActual(field,false,status);
-            result = handleGetYearLength(cal->get(UCAL_EXTENDED_YEAR, status));
+            result = handleGetYearLength(cal->get(UCAL_EXTENDED_YEAR, status), status);
             delete cal;
         }
         break;
@@ -4141,7 +4117,7 @@ Calendar::setWeekData(const Locale& desiredLocale, const char *type, UErrorCode&
     if (U_SUCCESS(status)) {
         U_LOCALE_BASED(locBased,*this);
         locBased.setLocaleIDs(ures_getLocaleByType(monthNames.getAlias(), ULOC_VALID_LOCALE, &status),
-                              ures_getLocaleByType(monthNames.getAlias(), ULOC_ACTUAL_LOCALE, &status));
+                              ures_getLocaleByType(monthNames.getAlias(), ULOC_ACTUAL_LOCALE, &status), status);
     } else {
         status = U_USING_FALLBACK_WARNING;
         return;
@@ -4229,14 +4205,12 @@ Calendar::updateTime(UErrorCode& status)
 
 Locale
 Calendar::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocale(type, status);
+    return LocaleBased::getLocale(validLocale, actualLocale, type, status);
 }
 
 const char *
 Calendar::getLocaleID(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocaleID(type, status);
+    return LocaleBased::getLocaleID(validLocale, actualLocale, type, status);
 }
 
 void
@@ -4245,7 +4219,7 @@ Calendar::recalculateStamp() {
     int32_t currentValue;
     int32_t j, i;
 
-    fNextStamp = 1;
+    fNextStamp = kInternallySet;
 
     for (j = 0; j < UCAL_FIELD_COUNT; j++) {
         currentValue = STAMP_MAX;
diff --git a/deps/icu-small/source/i18n/cecal.cpp b/deps/icu-small/source/i18n/cecal.cpp
index 7771c32efb30c8..33d32adab731f5 100644
--- a/deps/icu-small/source/i18n/cecal.cpp
+++ b/deps/icu-small/source/i18n/cecal.cpp
@@ -53,7 +53,6 @@ static const int32_t LIMITS[UCAL_FIELD_COUNT][4] = {
 CECalendar::CECalendar(const Locale& aLocale, UErrorCode& success)
 :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success)
 {
-    setTimeInMillis(getNow(), success);
 }
 
 CECalendar::CECalendar (const CECalendar& other) 
@@ -65,13 +64,6 @@ CECalendar::~CECalendar()
 {
 }
 
-CECalendar&
-CECalendar::operator=(const CECalendar& right)
-{
-    Calendar::operator=(right);
-    return *this;
-}
-
 //-------------------------------------------------------------------------
 // Calendar framework
 //-------------------------------------------------------------------------
diff --git a/deps/icu-small/source/i18n/cecal.h b/deps/icu-small/source/i18n/cecal.h
index 9c3332f3b84a82..e6514e18f857cc 100644
--- a/deps/icu-small/source/i18n/cecal.h
+++ b/deps/icu-small/source/i18n/cecal.h
@@ -82,13 +82,6 @@ class U_I18N_API CECalendar : public Calendar {
      */
     virtual ~CECalendar();
 
-    /**
-     * Default assignment operator
-     * @param right    Calendar object to be copied
-     * @internal
-     */
-    CECalendar& operator=(const CECalendar& right);
-
 protected:
     //-------------------------------------------------------------------------
     // Calendar framework
diff --git a/deps/icu-small/source/i18n/chnsecal.cpp b/deps/icu-small/source/i18n/chnsecal.cpp
index 050994fcbaf67c..afb16d3e25a83b 100644
--- a/deps/icu-small/source/i18n/chnsecal.cpp
+++ b/deps/icu-small/source/i18n/chnsecal.cpp
@@ -130,7 +130,6 @@ ChineseCalendar::ChineseCalendar(const Locale& aLocale, UErrorCode& success)
 :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success),
     hasLeapMonthBetweenWinterSolstices(false)
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 ChineseCalendar::ChineseCalendar(const ChineseCalendar& other) : Calendar(other) {
@@ -219,7 +218,9 @@ int32_t ChineseCalendar::handleGetExtendedYear(UErrorCode& status) {
     }
 
     int32_t year;
-    if (newestStamp(UCAL_ERA, UCAL_YEAR, kUnset) <= fStamp[UCAL_EXTENDED_YEAR]) {
+    // if UCAL_EXTENDED_YEAR is not older than UCAL_ERA nor UCAL_YEAR
+    if (newerField(UCAL_EXTENDED_YEAR, newerField(UCAL_ERA, UCAL_YEAR)) ==
+        UCAL_EXTENDED_YEAR) {
         year = internalGet(UCAL_EXTENDED_YEAR, 1); // Default to year 1
     } else {
         // adjust to the instance specific epoch
@@ -252,11 +253,16 @@ int32_t ChineseCalendar::handleGetExtendedYear(UErrorCode& status) {
  * @stable ICU 2.8
  */
 int32_t ChineseCalendar::handleGetMonthLength(int32_t extendedYear, int32_t month, UErrorCode& status) const {
+    bool isLeapMonth = internalGet(UCAL_IS_LEAP_MONTH) == 1;
+    return handleGetMonthLengthWithLeap(extendedYear, month, isLeapMonth, status);
+}
+
+int32_t ChineseCalendar::handleGetMonthLengthWithLeap(int32_t extendedYear, int32_t month, bool leap, UErrorCode& status) const {
     const Setting setting = getSetting(status);
     if (U_FAILURE(status)) {
         return 0;
     }
-    int32_t thisStart = handleComputeMonthStart(extendedYear, month, true, status);
+    int32_t thisStart = handleComputeMonthStartWithLeap(extendedYear, month, leap, status);
     if (U_FAILURE(status)) {
         return 0;
     }
@@ -332,18 +338,24 @@ struct MonthInfo computeMonthInfo(
  * @stable ICU 2.8
  */
 int64_t ChineseCalendar::handleComputeMonthStart(int32_t eyear, int32_t month, UBool useMonth, UErrorCode& status) const {
+    bool isLeapMonth = false;
+    if (useMonth) {
+        isLeapMonth = internalGet(UCAL_IS_LEAP_MONTH) != 0;
+    }
+    return handleComputeMonthStartWithLeap(eyear, month, isLeapMonth, status);
+}
+
+int64_t ChineseCalendar::handleComputeMonthStartWithLeap(int32_t eyear, int32_t month, bool isLeapMonth, UErrorCode& status) const {
     if (U_FAILURE(status)) {
        return 0;
     }
     // If the month is out of range, adjust it into range, and
     // modify the extended year value accordingly.
     if (month < 0 || month > 11) {
-        double m = month;
-        if (uprv_add32_overflow(eyear, ClockMath::floorDivide(m, 12.0, &m), &eyear)) {
+        if (uprv_add32_overflow(eyear, ClockMath::floorDivide(month, 12, &month), &eyear)) {
             status = U_ILLEGAL_ARGUMENT_ERROR;
             return 0;
         }
-        month = static_cast<int32_t>(m);
     }
 
     const Setting setting = getSetting(status);
@@ -362,19 +374,9 @@ int64_t ChineseCalendar::handleComputeMonthStart(int32_t eyear, int32_t month, U
        return 0;
     }
 
-    // Ignore IS_LEAP_MONTH field if useMonth is false
-    bool isLeapMonth = false;
-    if (useMonth) {
-        isLeapMonth = internalGet(UCAL_IS_LEAP_MONTH) != 0;
-    }
-
-    int32_t unusedMonth;
-    int32_t unusedDayOfWeek;
-    int32_t unusedDayOfMonth;
-    int32_t unusedDayOfYear;
-    Grego::dayToFields(newMoon, gyear, unusedMonth, unusedDayOfWeek, unusedDayOfMonth, unusedDayOfYear, status);
+    int32_t newMonthYear = Grego::dayToYear(newMoon, status);
 
-    struct MonthInfo monthInfo = computeMonthInfo(setting, gyear, newMoon, status);
+    struct MonthInfo monthInfo = computeMonthInfo(setting, newMonthYear, newMoon, status);
     if (U_FAILURE(status)) {
        return 0;
     }
@@ -794,6 +796,9 @@ struct MonthInfo computeMonthInfo(
         solsticeBefore = solsticeAfter;
         solsticeAfter = winterSolstice(setting, gyear + 1, status);
     }
+    if (!(solsticeBefore <= days && days < solsticeAfter)) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+    }
     if (U_FAILURE(status)) {
         return output;
     }
@@ -1043,7 +1048,12 @@ void ChineseCalendar::offsetMonth(int32_t newMoon, int32_t dayOfMonth, int32_t d
     }
 
     // Find the target dayOfMonth
-    int32_t jd = newMoon + kEpochStartAsJulianDay - 1 + dayOfMonth;
+    int32_t jd;
+    if (uprv_add32_overflow(newMoon, kEpochStartAsJulianDay - 1, &jd) ||
+        uprv_add32_overflow(jd, dayOfMonth, &jd)) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return;
+    }
 
     // Pin the dayOfMonth.  In this calendar all months are 29 or 30 days
     // so pinning just means handling dayOfMonth 30.
@@ -1182,6 +1192,27 @@ ChineseCalendar::Setting ChineseCalendar::getSetting(UErrorCode&) const {
   };
 }
 
+int32_t
+ChineseCalendar::getActualMaximum(UCalendarDateFields field, UErrorCode& status) const
+{
+    if (U_FAILURE(status)) {
+       return 0;
+    }
+    if (field == UCAL_DATE) {
+        LocalPointer<ChineseCalendar> cal(clone(), status);
+        if(U_FAILURE(status)) {
+            return 0;
+        }
+        cal->setLenient(true);
+        cal->prepareGetActual(field,false,status);
+        int32_t year = cal->get(UCAL_EXTENDED_YEAR, status);
+        int32_t month = cal->get(UCAL_MONTH, status);
+        bool leap = cal->get(UCAL_IS_LEAP_MONTH, status) != 0;
+        return handleGetMonthLengthWithLeap(year, month, leap, status);
+    }
+    return Calendar::getActualMaximum(field, status);
+}
+
 U_NAMESPACE_END
 
 #endif
diff --git a/deps/icu-small/source/i18n/chnsecal.h b/deps/icu-small/source/i18n/chnsecal.h
index 41bd3557fcbff1..410a5a0222cf60 100644
--- a/deps/icu-small/source/i18n/chnsecal.h
+++ b/deps/icu-small/source/i18n/chnsecal.h
@@ -194,6 +194,10 @@ class U_I18N_API ChineseCalendar : public Calendar {
   virtual void handleComputeFields(int32_t julianDay, UErrorCode &status) override;
   virtual const UFieldResolutionTable* getFieldResolutionTable() const override;
 
+ private:
+  int32_t handleGetMonthLengthWithLeap(int32_t extendedYear, int32_t month, bool isLeap, UErrorCode& status) const;
+  int64_t handleComputeMonthStartWithLeap(int32_t eyear, int32_t month, bool isLeap, UErrorCode& status) const;
+
  public:
   virtual void add(UCalendarDateFields field, int32_t amount, UErrorCode &status) override;
   virtual void add(EDateFields field, int32_t amount, UErrorCode &status) override;
@@ -254,6 +258,8 @@ class U_I18N_API ChineseCalendar : public Calendar {
    */
   virtual const char * getType() const override;
 
+  virtual int32_t getActualMaximum(UCalendarDateFields field, UErrorCode& status) const override;
+
   struct Setting {
       int32_t epochYear;
       const TimeZone* zoneAstroCalc;
diff --git a/deps/icu-small/source/i18n/collationruleparser.cpp b/deps/icu-small/source/i18n/collationruleparser.cpp
index b20d2c9428c48b..f5608cde7dc643 100644
--- a/deps/icu-small/source/i18n/collationruleparser.cpp
+++ b/deps/icu-small/source/i18n/collationruleparser.cpp
@@ -613,18 +613,24 @@ CollationRuleParser::parseSetting(UErrorCode &errorCode) {
                 return;
             }
             // localeID minus all keywords
-            char baseID[ULOC_FULLNAME_CAPACITY];
-            int32_t length = uloc_getBaseName(localeID.data(), baseID, ULOC_FULLNAME_CAPACITY, &errorCode);
-            if(U_FAILURE(errorCode) || length >= ULOC_KEYWORDS_CAPACITY) {
+            CharString baseID = ulocimp_getBaseName(localeID.toStringPiece(), errorCode);
+            if (U_FAILURE(errorCode)) {
                 errorCode = U_ZERO_ERROR;
                 setParseError("expected language tag in [import langTag]", errorCode);
                 return;
             }
-            if(length == 0) {
-                uprv_strcpy(baseID, "root");
-            } else if(*baseID == '_') {
-                uprv_memmove(baseID + 3, baseID, length + 1);
-                uprv_memcpy(baseID, "und", 3);
+            if (baseID.isEmpty()) {
+                baseID.copyFrom("root", errorCode);
+            } else if (baseID[0] == '_') {
+                // CharString doesn't have any insert() method, only append().
+                constexpr char und[] = "und";
+                constexpr int32_t length = sizeof und - 1;
+                int32_t dummy;
+                char* tail = baseID.getAppendBuffer(length, length, dummy, errorCode);
+                char* head = baseID.data();
+                uprv_memmove(head + length, head, baseID.length());
+                uprv_memcpy(head, und, length);
+                baseID.append(tail, length, errorCode);
             }
             // @collation=type, or length=0 if not specified
             CharString collationType = ulocimp_getKeywordValue(localeID.data(), "collation", errorCode);
@@ -637,7 +643,7 @@ CollationRuleParser::parseSetting(UErrorCode &errorCode) {
                 setParseError("[import langTag] is not supported", errorCode);
             } else {
                 UnicodeString importedRules;
-                importer->getRules(baseID,
+                importer->getRules(baseID.data(),
                                    !collationType.isEmpty() ? collationType.data() : "standard",
                                    importedRules, errorReason, errorCode);
                 if(U_FAILURE(errorCode)) {
diff --git a/deps/icu-small/source/i18n/datefmt.cpp b/deps/icu-small/source/i18n/datefmt.cpp
index 655cfbd1239daa..de9efc7ca8f971 100644
--- a/deps/icu-small/source/i18n/datefmt.cpp
+++ b/deps/icu-small/source/i18n/datefmt.cpp
@@ -40,6 +40,7 @@
 #if defined( U_DEBUG_CALSVC ) || defined (U_DEBUG_CAL)
 #include <stdio.h>
 #endif
+#include <typeinfo>
 
 // *****************************************************************************
 // class DateFormat
@@ -279,9 +280,8 @@ UnicodeString&
 DateFormat::format(UDate date, UnicodeString& appendTo, FieldPosition& fieldPosition) const {
     if (fCalendar != nullptr) {
         UErrorCode ec = U_ZERO_ERROR;
-        const auto* calType = fCalendar->getType();
         // Avoid a heap allocation and corresponding free for the common case
-        if (uprv_strcmp(calType, "gregorian") == 0) {
+        if (typeid(*fCalendar) == typeid(GregorianCalendar)) {
             GregorianCalendar cal(*static_cast<GregorianCalendar*>(fCalendar));
             cal.setTime(date, ec);
             if (U_SUCCESS(ec)) {
@@ -309,9 +309,8 @@ DateFormat::format(UDate date, UnicodeString& appendTo, FieldPositionIterator* p
                    UErrorCode& status) const {
     if (fCalendar != nullptr) {
         UErrorCode ec = U_ZERO_ERROR;
-        const auto* calType = fCalendar->getType();
         // Avoid a heap allocation and corresponding free for the common case
-        if (uprv_strcmp(calType, "gregorian") == 0) {
+        if (typeid(*fCalendar) == typeid(GregorianCalendar)) {
             GregorianCalendar cal(*static_cast<GregorianCalendar*>(fCalendar));
             cal.setTime(date, ec);
             if (U_SUCCESS(ec)) {
diff --git a/deps/icu-small/source/i18n/dcfmtsym.cpp b/deps/icu-small/source/i18n/dcfmtsym.cpp
index b4c90e6765a7df..b85f3ad134a289 100644
--- a/deps/icu-small/source/i18n/dcfmtsym.cpp
+++ b/deps/icu-small/source/i18n/dcfmtsym.cpp
@@ -118,7 +118,6 @@ DecimalFormatSymbols::DecimalFormatSymbols(const Locale& loc, const NumberingSys
 
 DecimalFormatSymbols::DecimalFormatSymbols()
         : UObject(), locale(Locale::getRoot()) {
-    *validLocale = *actualLocale = 0;
     initialize();
 }
 
@@ -136,6 +135,8 @@ DecimalFormatSymbols::createWithLastResortData(UErrorCode& status) {
 
 DecimalFormatSymbols::~DecimalFormatSymbols()
 {
+    delete actualLocale;
+    delete validLocale;
 }
 
 // -------------------------------------
@@ -163,8 +164,12 @@ DecimalFormatSymbols::operator=(const DecimalFormatSymbols& rhs)
             currencySpcAfterSym[i].fastCopyFrom(rhs.currencySpcAfterSym[i]);
         }
         locale = rhs.locale;
-        uprv_strcpy(validLocale, rhs.validLocale);
-        uprv_strcpy(actualLocale, rhs.actualLocale);
+
+        UErrorCode status = U_ZERO_ERROR;
+        U_LOCALE_BASED(locBased, *this);
+        locBased.setLocaleIDs(rhs.validLocale, rhs.actualLocale, status);
+        U_ASSERT(U_SUCCESS(status));
+
         fIsCustomCurrencySymbol = rhs.fIsCustomCurrencySymbol; 
         fIsCustomIntlCurrencySymbol = rhs.fIsCustomIntlCurrencySymbol; 
         fCodePointZero = rhs.fCodePointZero;
@@ -203,8 +208,8 @@ DecimalFormatSymbols::operator==(const DecimalFormatSymbols& that) const
     }
     // No need to check fCodePointZero since it is based on fSymbols
     return locale == that.locale &&
-        uprv_strcmp(validLocale, that.validLocale) == 0 &&
-        uprv_strcmp(actualLocale, that.actualLocale) == 0;
+        LocaleBased::equalIDs(actualLocale, that.actualLocale) &&
+        LocaleBased::equalIDs(validLocale, that.validLocale);
 }
 
 // -------------------------------------
@@ -353,7 +358,6 @@ DecimalFormatSymbols::initialize(const Locale& loc, UErrorCode& status,
     UBool useLastResortData, const NumberingSystem* ns)
 {
     if (U_FAILURE(status)) { return; }
-    *validLocale = *actualLocale = 0;
 
     // First initialize all the symbols to the fallbacks for anything we can't find
     initialize();
@@ -409,7 +413,8 @@ DecimalFormatSymbols::initialize(const Locale& loc, UErrorCode& status,
             ULOC_VALID_LOCALE, &status),
         ures_getLocaleByType(
             numberElementsRes.getAlias(),
-            ULOC_ACTUAL_LOCALE, &status));
+            ULOC_ACTUAL_LOCALE, &status),
+        status);
 
     // Now load the rest of the data from the data sink.
     // Start with loading this nsName if it is not Latin.
@@ -568,8 +573,7 @@ void DecimalFormatSymbols::setCurrency(const char16_t* currency, UErrorCode& sta
 
 Locale
 DecimalFormatSymbols::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocale(type, status);
+    return LocaleBased::getLocale(validLocale, actualLocale, type, status);
 }
 
 const UnicodeString&
diff --git a/deps/icu-small/source/i18n/dtfmtsym.cpp b/deps/icu-small/source/i18n/dtfmtsym.cpp
index 23cea3eba20ae4..339db48dba8565 100644
--- a/deps/icu-small/source/i18n/dtfmtsym.cpp
+++ b/deps/icu-small/source/i18n/dtfmtsym.cpp
@@ -402,9 +402,8 @@ void
 DateFormatSymbols::copyData(const DateFormatSymbols& other) {
     UErrorCode status = U_ZERO_ERROR;
     U_LOCALE_BASED(locBased, *this);
-    locBased.setLocaleIDs(
-        other.getLocale(ULOC_VALID_LOCALE, status),
-        other.getLocale(ULOC_ACTUAL_LOCALE, status));
+    locBased.setLocaleIDs(other.validLocale, other.actualLocale, status);
+    U_ASSERT(U_SUCCESS(status));
     assignArray(fEras, fErasCount, other.fEras, other.fErasCount);
     assignArray(fEraNames, fEraNamesCount, other.fEraNames, other.fEraNamesCount);
     assignArray(fNarrowEras, fNarrowErasCount, other.fNarrowEras, other.fNarrowErasCount);
@@ -497,6 +496,8 @@ DateFormatSymbols& DateFormatSymbols::operator=(const DateFormatSymbols& other)
 DateFormatSymbols::~DateFormatSymbols()
 {
     dispose();
+    delete actualLocale;
+    delete validLocale;
 }
 
 void DateFormatSymbols::dispose()
@@ -536,6 +537,10 @@ void DateFormatSymbols::dispose()
     delete[] fStandaloneWideDayPeriods;
     delete[] fStandaloneNarrowDayPeriods;
 
+    delete actualLocale;
+    actualLocale = nullptr;
+    delete validLocale;
+    validLocale = nullptr;
     disposeZoneStrings();
 }
 
@@ -2302,7 +2307,7 @@ DateFormatSymbols::initializeData(const Locale& locale, const char *type, UError
     // of it that we need except for the time-zone and localized-pattern data, which
     // are stored in a separate file
     locBased.setLocaleIDs(ures_getLocaleByType(cb.getAlias(), ULOC_VALID_LOCALE, &status),
-                          ures_getLocaleByType(cb.getAlias(), ULOC_ACTUAL_LOCALE, &status));
+                          ures_getLocaleByType(cb.getAlias(), ULOC_ACTUAL_LOCALE, &status), status);
 
     // Load eras
     initField(&fEras, fErasCount, calendarSink, buildResourcePath(path, gErasTag, gNamesAbbrTag, status), status);
@@ -2528,8 +2533,7 @@ DateFormatSymbols::initializeData(const Locale& locale, const char *type, UError
 
 Locale
 DateFormatSymbols::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocale(type, status);
+    return LocaleBased::getLocale(validLocale, actualLocale, type, status);
 }
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/i18n/erarules.cpp b/deps/icu-small/source/i18n/erarules.cpp
index 8ab6f00ae04b68..d23d7d516c7202 100644
--- a/deps/icu-small/source/i18n/erarules.cpp
+++ b/deps/icu-small/source/i18n/erarules.cpp
@@ -305,8 +305,9 @@ void EraRules::initCurrentEra() {
         localMillis += (rawOffset + dstOffset);
     }
 
-    int year, month0, dom, dow, doy, mid;
-    Grego::timeToFields(localMillis, year, month0, dom, dow, doy, mid, ec);
+    int32_t year, mid;
+    int8_t  month0, dom;
+    Grego::timeToFields(localMillis, year, month0, dom, mid, ec);
     if (U_FAILURE(ec)) return;
     int currentEncodedDate = encodeDate(year, month0 + 1 /* changes to 1-base */, dom);
     int eraIdx = numEras - 1;
diff --git a/deps/icu-small/source/i18n/format.cpp b/deps/icu-small/source/i18n/format.cpp
index 10856a4acba286..b4aec1d9811547 100644
--- a/deps/icu-small/source/i18n/format.cpp
+++ b/deps/icu-small/source/i18n/format.cpp
@@ -24,6 +24,7 @@
 #include "utypeinfo.h"  // for 'typeid' to work
 
 #include "unicode/utypes.h"
+#include "charstr.h"
 
 #ifndef U_I18N_IMPLEMENTATION
 #error U_I18N_IMPLEMENTATION not set - must be set for all ICU source files in i18n/ - see https://unicode-org.github.io/icu/userguide/howtouseicu
@@ -72,13 +73,14 @@ FieldPosition::clone() const {
 Format::Format()
     : UObject()
 {
-    *validLocale = *actualLocale = 0;
 }
 
 // -------------------------------------
 
 Format::~Format()
 {
+    delete actualLocale;
+    delete validLocale;
 }
 
 // -------------------------------------
@@ -97,8 +99,10 @@ Format&
 Format::operator=(const Format& that)
 {
     if (this != &that) {
-        uprv_strcpy(validLocale, that.validLocale);
-        uprv_strcpy(actualLocale, that.actualLocale);
+        UErrorCode status = U_ZERO_ERROR;
+        U_LOCALE_BASED(locBased, *this);
+        locBased.setLocaleIDs(that.validLocale, that.actualLocale, status);
+        U_ASSERT(U_SUCCESS(status));
     }
     return *this;
 }
@@ -196,20 +200,20 @@ void Format::syntaxError(const UnicodeString& pattern,
 
 Locale
 Format::getLocale(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocale(type, status);
+    return LocaleBased::getLocale(validLocale, actualLocale, type, status);
 }
 
 const char *
 Format::getLocaleID(ULocDataLocaleType type, UErrorCode& status) const {
-    U_LOCALE_BASED(locBased, *this);
-    return locBased.getLocaleID(type, status);
+    return LocaleBased::getLocaleID(validLocale,actualLocale, type, status);
 }
 
 void
 Format::setLocaleIDs(const char* valid, const char* actual) {
     U_LOCALE_BASED(locBased, *this);
-    locBased.setLocaleIDs(valid, actual);
+    UErrorCode status = U_ZERO_ERROR;
+    locBased.setLocaleIDs(valid, actual, status);
+    U_ASSERT(U_SUCCESS(status));
 }
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/i18n/formattedvalue.cpp b/deps/icu-small/source/i18n/formattedvalue.cpp
index aacd6ac70e090f..f2bfdda6e465f2 100644
--- a/deps/icu-small/source/i18n/formattedvalue.cpp
+++ b/deps/icu-small/source/i18n/formattedvalue.cpp
@@ -193,6 +193,11 @@ ucfpos_close(UConstrainedFieldPosition* ptr) {
 }
 
 
+// -Wreturn-local-addr first found in https://gcc.gnu.org/onlinedocs/gcc-4.8.5/gcc/Warning-Options.html#Warning-Options
+#if U_GCC_MAJOR_MINOR >= 409
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wreturn-local-addr"
+#endif
 U_CAPI const char16_t* U_EXPORT2
 ufmtval_getString(
         const UFormattedValue* ufmtval,
@@ -213,6 +218,9 @@ ufmtval_getString(
     // defined to return memory owned by the ufmtval argument.
     return readOnlyAlias.getBuffer();
 }
+#if U_GCC_MAJOR_MINOR >= 409
+#pragma GCC diagnostic pop
+#endif
 
 
 U_CAPI UBool U_EXPORT2
diff --git a/deps/icu-small/source/i18n/gregocal.cpp b/deps/icu-small/source/i18n/gregocal.cpp
index 23366c7ab7a333..8a4bb15c16d60c 100644
--- a/deps/icu-small/source/i18n/gregocal.cpp
+++ b/deps/icu-small/source/i18n/gregocal.cpp
@@ -147,6 +147,7 @@ UOBJECT_DEFINE_RTTI_IMPLEMENTATION(GregorianCalendar)
 // in Java, -12219292800000L
 //const UDate GregorianCalendar::kPapalCutover = -12219292800000L;
 static const uint32_t kCutoverJulianDay = 2299161;
+static const int32_t kDefaultCutoverYear = 1582;
 static const UDate kPapalCutover = (2299161.0 - kEpochStartAsJulianDay) * U_MILLIS_PER_DAY;
 //static const UDate kPapalCutoverJulian = (2299161.0 - kEpochStartAsJulianDay);
 
@@ -155,7 +156,7 @@ static const UDate kPapalCutover = (2299161.0 - kEpochStartAsJulianDay) * U_MILL
 GregorianCalendar::GregorianCalendar(UErrorCode& status)
 :   Calendar(status),
 fGregorianCutover(kPapalCutover),
-fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
+fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(kDefaultCutoverYear),
 fIsGregorian(true), fInvertGregorian(false)
 {
     setTimeInMillis(getNow(), status);
@@ -164,34 +165,22 @@ fIsGregorian(true), fInvertGregorian(false)
 // -------------------------------------
 
 GregorianCalendar::GregorianCalendar(TimeZone* zone, UErrorCode& status)
-:   Calendar(zone, Locale::getDefault(), status),
-fGregorianCutover(kPapalCutover),
-fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
-fIsGregorian(true), fInvertGregorian(false)
+:   GregorianCalendar(zone, Locale::getDefault(), status)
 {
-    setTimeInMillis(getNow(), status);
 }
 
 // -------------------------------------
 
 GregorianCalendar::GregorianCalendar(const TimeZone& zone, UErrorCode& status)
-:   Calendar(zone, Locale::getDefault(), status),
-fGregorianCutover(kPapalCutover),
-fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
-fIsGregorian(true), fInvertGregorian(false)
+:   GregorianCalendar(zone, Locale::getDefault(), status)
 {
-    setTimeInMillis(getNow(), status);
 }
 
 // -------------------------------------
 
 GregorianCalendar::GregorianCalendar(const Locale& aLocale, UErrorCode& status)
-:   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, status),
-fGregorianCutover(kPapalCutover),
-fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
-fIsGregorian(true), fInvertGregorian(false)
+:   GregorianCalendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, status)
 {
-    setTimeInMillis(getNow(), status);
 }
 
 // -------------------------------------
@@ -200,7 +189,7 @@ GregorianCalendar::GregorianCalendar(TimeZone* zone, const Locale& aLocale,
                                      UErrorCode& status)
                                      :   Calendar(zone, aLocale, status),
                                      fGregorianCutover(kPapalCutover),
-                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
+                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(kDefaultCutoverYear),
                                      fIsGregorian(true), fInvertGregorian(false)
 {
     setTimeInMillis(getNow(), status);
@@ -212,7 +201,7 @@ GregorianCalendar::GregorianCalendar(const TimeZone& zone, const Locale& aLocale
                                      UErrorCode& status)
                                      :   Calendar(zone, aLocale, status),
                                      fGregorianCutover(kPapalCutover),
-                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
+                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(kDefaultCutoverYear),
                                      fIsGregorian(true), fInvertGregorian(false)
 {
     setTimeInMillis(getNow(), status);
@@ -224,7 +213,7 @@ GregorianCalendar::GregorianCalendar(int32_t year, int32_t month, int32_t date,
                                      UErrorCode& status)
                                      :   Calendar(TimeZone::createDefault(), Locale::getDefault(), status),
                                      fGregorianCutover(kPapalCutover),
-                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
+                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(kDefaultCutoverYear),
                                      fIsGregorian(true), fInvertGregorian(false)
 {
     set(UCAL_ERA, AD);
@@ -237,15 +226,8 @@ GregorianCalendar::GregorianCalendar(int32_t year, int32_t month, int32_t date,
 
 GregorianCalendar::GregorianCalendar(int32_t year, int32_t month, int32_t date,
                                      int32_t hour, int32_t minute, UErrorCode& status)
-                                     :   Calendar(TimeZone::createDefault(), Locale::getDefault(), status),
-                                     fGregorianCutover(kPapalCutover),
-                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
-                                     fIsGregorian(true), fInvertGregorian(false)
+                                     :   GregorianCalendar(year, month, date, status)
 {
-    set(UCAL_ERA, AD);
-    set(UCAL_YEAR, year);
-    set(UCAL_MONTH, month);
-    set(UCAL_DATE, date);
     set(UCAL_HOUR_OF_DAY, hour);
     set(UCAL_MINUTE, minute);
 }
@@ -255,17 +237,8 @@ GregorianCalendar::GregorianCalendar(int32_t year, int32_t month, int32_t date,
 GregorianCalendar::GregorianCalendar(int32_t year, int32_t month, int32_t date,
                                      int32_t hour, int32_t minute, int32_t second,
                                      UErrorCode& status)
-                                     :   Calendar(TimeZone::createDefault(), Locale::getDefault(), status),
-                                     fGregorianCutover(kPapalCutover),
-                                     fCutoverJulianDay(kCutoverJulianDay), fNormalizedGregorianCutover(fGregorianCutover), fGregorianCutoverYear(1582),
-                                     fIsGregorian(true), fInvertGregorian(false)
+                                     :   GregorianCalendar(year, month, date, hour, minute, status)
 {
-    set(UCAL_ERA, AD);
-    set(UCAL_YEAR, year);
-    set(UCAL_MONTH, month);
-    set(UCAL_DATE, date);
-    set(UCAL_HOUR_OF_DAY, hour);
-    set(UCAL_MINUTE, minute);
     set(UCAL_SECOND, second);
 }
 
@@ -601,7 +574,8 @@ int32_t GregorianCalendar::handleGetMonthLength(int32_t extendedYear, int32_t mo
     return isLeapYear(extendedYear) ? kLeapMonthLength[month] : kMonthLength[month];
 }
 
-int32_t GregorianCalendar::handleGetYearLength(int32_t eyear) const {
+int32_t GregorianCalendar::handleGetYearLength(int32_t eyear, UErrorCode& status) const {
+    if (U_FAILURE(status)) return 0;
     return isLeapYear(eyear) ? 366 : 365;
 }
 
@@ -871,13 +845,14 @@ GregorianCalendar::roll(UCalendarDateFields field, int32_t amount, UErrorCode& s
         }
         if (month == UCAL_JANUARY) {
             if (woy >= 52) {
-                isoDoy += handleGetYearLength(isoYear);
+                isoDoy += handleGetYearLength(isoYear, status);
             }
         } else {
             if (woy == 1) {
-                isoDoy -= handleGetYearLength(isoYear - 1);
+                isoDoy -= handleGetYearLength(isoYear - 1, status);
             }
         }
+        if (U_FAILURE(status)) return;
         if (uprv_add32_overflow(woy, amount, &woy)) {
             status = U_ILLEGAL_ARGUMENT_ERROR;
             return;
@@ -890,7 +865,8 @@ GregorianCalendar::roll(UCalendarDateFields field, int32_t amount, UErrorCode& s
             // days at the end of the year are going to fall into
             // week 1 of the next year, we drop the last week by
             // subtracting 7 from the last day of the year.
-            int32_t lastDoy = handleGetYearLength(isoYear);
+            int32_t lastDoy = handleGetYearLength(isoYear, status);
+            if (U_FAILURE(status)) return;
             int32_t lastRelDow = (lastDoy - isoDoy + internalGet(UCAL_DAY_OF_WEEK) -
                 getFirstDayOfWeek()) % 7;
             if (lastRelDow < 0) lastRelDow += 7;
@@ -1186,14 +1162,10 @@ int32_t GregorianCalendar::handleGetExtendedYear(UErrorCode& status) {
     int32_t year = kEpochYear;
 
     // year field to use
-    int32_t yearField = UCAL_EXTENDED_YEAR;
-
     // There are three separate fields which could be used to
     // derive the proper year.  Use the one most recently set.
-    if (fStamp[yearField] < fStamp[UCAL_YEAR])
-        yearField = UCAL_YEAR;
-    if (fStamp[yearField] < fStamp[UCAL_YEAR_WOY])
-        yearField = UCAL_YEAR_WOY;
+    UCalendarDateFields yearField = newerField(
+        newerField(UCAL_EXTENDED_YEAR, UCAL_YEAR), UCAL_YEAR_WOY);
 
     // based on the "best" year field, get the year
     switch(yearField) {
diff --git a/deps/icu-small/source/i18n/gregoimp.cpp b/deps/icu-small/source/i18n/gregoimp.cpp
index d5c8437a9b80f0..03bf9d2c9fdfec 100644
--- a/deps/icu-small/source/i18n/gregoimp.cpp
+++ b/deps/icu-small/source/i18n/gregoimp.cpp
@@ -117,57 +117,110 @@ int64_t Grego::fieldsToDay(int32_t year, int32_t month, int32_t dom) {
     return julian - JULIAN_1970_CE; // JD => epoch day
 }
 
-void Grego::dayToFields(int32_t day, int32_t& year, int32_t& month,
-                        int32_t& dom, int32_t& dow, int32_t& doy, UErrorCode& status) {
-
+void Grego::dayToFields(int32_t day, int32_t& year, int8_t& month,
+                        int8_t& dom, int8_t& dow, int16_t& doy, UErrorCode& status) {
+    year = dayToYear(day, doy, status); // one-based doy
     if (U_FAILURE(status)) return;
+
     // Convert from 1970 CE epoch to 1 CE epoch (Gregorian calendar)
     if (uprv_add32_overflow(day, JULIAN_1970_CE - JULIAN_1_CE, &day)) {
         status = U_ILLEGAL_ARGUMENT_ERROR;
         return;
     }
 
-    // Convert from the day number to the multiple radix
-    // representation.  We use 400-year, 100-year, and 4-year cycles.
-    // For example, the 4-year cycle has 4 years + 1 leap day; giving
-    // 1461 == 365*4 + 1 days.
-    int32_t n400 = ClockMath::floorDivide(day, 146097, &doy); // 400-year cycle length
-    int32_t n100 = ClockMath::floorDivide(doy, 36524, &doy); // 100-year cycle length
-    int32_t n4   = ClockMath::floorDivide(doy, 1461, &doy); // 4-year cycle length
-    int32_t n1   = ClockMath::floorDivide(doy, 365, &doy);
-    year = 400*n400 + 100*n100 + 4*n4 + n1;
-    if (n100 == 4 || n1 == 4) {
-        doy = 365; // Dec 31 at end of 4- or 400-year cycle
-    } else {
-        ++year;
-    }
-    
-    UBool isLeap = isLeapYear(year);
-    
     // Gregorian day zero is a Monday.
     dow = (day + 1) % 7;
     dow += (dow < 0) ? (UCAL_SUNDAY + 7) : UCAL_SUNDAY;
 
     // Common Julian/Gregorian calculation
     int32_t correction = 0;
+    bool isLeap = isLeapYear(year);
     int32_t march1 = isLeap ? 60 : 59; // zero-based DOY for March 1
-    if (doy >= march1) {
+    if (doy > march1) {
         correction = isLeap ? 1 : 2;
     }
-    month = (12 * (doy + correction) + 6) / 367; // zero-based month
-    dom = doy - DAYS_BEFORE[month + (isLeap ? 12 : 0)] + 1; // one-based DOM
+    month = (12 * (doy - 1 + correction) + 6) / 367; // zero-based month
+    dom = doy - DAYS_BEFORE[month + (isLeap ? 12 : 0)]; // one-based DOM
+}
+
+int32_t Grego::dayToYear(int32_t day, UErrorCode& status) {
+    int16_t unusedDOY;
+    return dayToYear(day, unusedDOY, status);
+}
+
+int32_t Grego::dayToYear(int32_t day, int16_t& doy, UErrorCode& status) {
+    if (U_FAILURE(status)) return 0;
+    // Convert from 1970 CE epoch to 1 CE epoch (Gregorian calendar)
+    if (uprv_add32_overflow(day, JULIAN_1970_CE - JULIAN_1_CE, &day)) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return 0;
+    }
+
+    // Convert from the day number to the multiple radix
+    // representation.  We use 400-year, 100-year, and 4-year cycles.
+    // For example, the 4-year cycle has 4 years + 1 leap day; giving
+    // 1461 == 365*4 + 1 days.
+    int32_t doy32;
+    int32_t n400 = ClockMath::floorDivide(day, 146097, &doy32); // 400-year cycle length
+    int32_t n100 = ClockMath::floorDivide(doy32, 36524, &doy32); // 100-year cycle length
+    int32_t n4   = ClockMath::floorDivide(doy32, 1461, &doy32); // 4-year cycle length
+    int32_t n1   = ClockMath::floorDivide(doy32, 365, &doy32);
+    int32_t year = 400*n400 + 100*n100 + 4*n4 + n1;
+    if (n100 == 4 || n1 == 4) {
+        doy = 365; // Dec 31 at end of 4- or 400-year cycle
+    } else {
+        doy = doy32;
+        ++year;
+    }
     doy++; // one-based doy
+    return year;
+}
+
+void Grego::dayToFields(int32_t day, int32_t& year, int8_t& month,
+                        int8_t& dom, int8_t& dow, UErrorCode& status) {
+    int16_t unusedDOY;
+    dayToFields(day, year, month, dom, dow, unusedDOY, status);
+}
+
+void Grego::dayToFields(int32_t day, int32_t& year, int8_t& month,
+                        int8_t& dom, int16_t& doy, UErrorCode& status) {
+    int8_t unusedDOW;
+    dayToFields(day, year, month, dom, unusedDOW, doy, status);
 }
 
-void Grego::timeToFields(UDate time, int32_t& year, int32_t& month,
-                        int32_t& dom, int32_t& dow, int32_t& doy, int32_t& mid, UErrorCode& status) {
+void Grego::timeToFields(UDate time, int32_t& year, int8_t& month,
+                        int8_t& dom, int32_t& mid, UErrorCode& status) {
+    int8_t unusedDOW;
+    timeToFields(time, year, month, dom, unusedDOW, mid, status);
+}
+
+void Grego::timeToFields(UDate time, int32_t& year, int8_t& month,
+                        int8_t& dom, int8_t& dow, int32_t& mid, UErrorCode& status) {
+    int16_t unusedDOY;
+    timeToFields(time, year, month, dom, dow, unusedDOY, mid, status);
+}
+
+void Grego::timeToFields(UDate time, int32_t& year, int8_t& month,
+                        int8_t& dom, int8_t& dow, int16_t& doy, int32_t& mid, UErrorCode& status) {
     if (U_FAILURE(status)) return;
-    double millisInDay;
-    double day = ClockMath::floorDivide(static_cast<double>(time), static_cast<double>(U_MILLIS_PER_DAY), &millisInDay);
-    mid = static_cast<int32_t>(millisInDay);
+    double day = ClockMath::floorDivide(time, U_MILLIS_PER_DAY, &mid);
+    if (day > INT32_MAX || day < INT32_MIN) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return;
+    }
     dayToFields(day, year, month, dom, dow, doy, status);
 }
 
+int32_t Grego::timeToYear(UDate time, UErrorCode& status) {
+    if (U_FAILURE(status)) return 0;
+    double day = ClockMath::floorDivide(time, double(U_MILLIS_PER_DAY));
+    if (day > INT32_MAX || day < INT32_MIN) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return 0;
+    }
+    return Grego::dayToYear(day, status);
+}
+
 int32_t Grego::dayOfWeek(int32_t day) {
     int32_t dow;
     ClockMath::floorDivide(day + int{UCAL_THURSDAY}, 7, &dow);
diff --git a/deps/icu-small/source/i18n/gregoimp.h b/deps/icu-small/source/i18n/gregoimp.h
index e069fb60de7909..39881c0eefbe6c 100644
--- a/deps/icu-small/source/i18n/gregoimp.h
+++ b/deps/icu-small/source/i18n/gregoimp.h
@@ -210,8 +210,21 @@ class Grego {
      * @param doy output parameter to receive day-of-year (1-based)
      * @param status error code.
      */
-    static void dayToFields(int32_t day, int32_t& year, int32_t& month,
-                            int32_t& dom, int32_t& dow, int32_t& doy, UErrorCode& status);
+    static void dayToFields(int32_t day, int32_t& year, int8_t& month,
+                            int8_t& dom, int8_t& dow, int16_t& doy, UErrorCode& status);
+
+    /**
+     * Convert a 1970-epoch day number to proleptic Gregorian year,
+     * month, day-of-month, and day-of-week.
+     * @param day 1970-epoch day
+     * @param year output parameter to receive year
+     * @param month output parameter to receive month (0-based, 0==Jan)
+     * @param dom output parameter to receive day-of-month (1-based)
+     * @param doy output parameter to receive day-of-year (1-based)
+     * @param status error code.
+     */
+    static void dayToFields(int32_t day, int32_t& year, int8_t& month,
+                            int8_t& dom, int16_t& doy, UErrorCode& status);
 
     /**
      * Convert a 1970-epoch day number to proleptic Gregorian year,
@@ -223,8 +236,24 @@ class Grego {
      * @param dow output parameter to receive day-of-week (1-based, 1==Sun)
      * @param status error code.
      */
-    static inline void dayToFields(int32_t day, int32_t& year, int32_t& month,
-                                   int32_t& dom, int32_t& dow, UErrorCode& status);
+    static void dayToFields(int32_t day, int32_t& year, int8_t& month,
+                            int8_t& dom, int8_t& dow, UErrorCode& status);
+
+    /**
+     * Convert a 1970-epoch day number to proleptic Gregorian year.
+     * @param day 1970-epoch day
+     * @param status error code.
+     * @return year.
+     */
+    static int32_t dayToYear(int32_t day, UErrorCode& status);
+    /**
+     * Convert a 1970-epoch day number to proleptic Gregorian year.
+     * @param day 1970-epoch day
+     * @param doy output parameter to receive day-of-year (1-based)
+     * @param status error code.
+     * @return year.
+     */
+    static int32_t dayToYear(int32_t day, int16_t& doy, UErrorCode& status);
 
     /**
      * Convert a 1970-epoch milliseconds to proleptic Gregorian year,
@@ -238,8 +267,43 @@ class Grego {
      * @param mid output parameter to receive millis-in-day
      * @param status error code.
      */
-    static void timeToFields(UDate time, int32_t& year, int32_t& month,
-                            int32_t& dom, int32_t& dow, int32_t& doy, int32_t& mid, UErrorCode& status);
+    static void timeToFields(UDate time, int32_t& year, int8_t& month,
+                            int8_t& dom, int8_t& dow, int16_t& doy, int32_t& mid, UErrorCode& status);
+
+    /**
+     * Convert a 1970-epoch milliseconds to proleptic Gregorian year,
+     * month, day-of-month, and day-of-week, day of year and millis-in-day.
+     * @param time 1970-epoch milliseconds
+     * @param year output parameter to receive year
+     * @param month output parameter to receive month (0-based, 0==Jan)
+     * @param dom output parameter to receive day-of-month (1-based)
+     * @param dow output parameter to receive day-of-week (1-based, 1==Sun)
+     * @param mid output parameter to receive millis-in-day
+     * @param status error code.
+     */
+    static void timeToFields(UDate time, int32_t& year, int8_t& month,
+                            int8_t& dom, int8_t& dow, int32_t& mid, UErrorCode& status);
+
+    /**
+     * Convert a 1970-epoch milliseconds to proleptic Gregorian year,
+     * month, day-of-month, and day-of-week, day of year and millis-in-day.
+     * @param time 1970-epoch milliseconds
+     * @param year output parameter to receive year
+     * @param month output parameter to receive month (0-based, 0==Jan)
+     * @param dom output parameter to receive day-of-month (1-based)
+     * @param mid output parameter to receive millis-in-day
+     * @param status error code.
+     */
+    static void timeToFields(UDate time, int32_t& year, int8_t& month,
+                            int8_t& dom, int32_t& mid, UErrorCode& status);
+
+    /**
+     * Convert a 1970-epoch milliseconds to proleptic Gregorian year.
+     * @param time 1970-epoch milliseconds
+     * @param status error code.
+     * @return year.
+     */
+    static int32_t timeToYear(UDate time, UErrorCode& status);
 
     /**
      * Return the day of week on the 1970-epoch day
@@ -305,12 +369,6 @@ Grego::previousMonthLength(int y, int m) {
   return (m > 0) ? monthLength(y, m-1) : 31;
 }
 
-inline void Grego::dayToFields(int32_t day, int32_t& year, int32_t& month,
-                               int32_t& dom, int32_t& dow, UErrorCode& status) {
-  int32_t doy_unused;
-  dayToFields(day,year,month,dom,dow,doy_unused, status);
-}
-
 inline double Grego::julianDayToMillis(int32_t julian)
 {
   return (static_cast<double>(julian) - kEpochStartAsJulianDay) * kOneDay;
diff --git a/deps/icu-small/source/i18n/hebrwcal.cpp b/deps/icu-small/source/i18n/hebrwcal.cpp
index ef70a48f2353e3..70dfe9b7c735e5 100644
--- a/deps/icu-small/source/i18n/hebrwcal.cpp
+++ b/deps/icu-small/source/i18n/hebrwcal.cpp
@@ -164,7 +164,6 @@ HebrewCalendar::HebrewCalendar(const Locale& aLocale, UErrorCode& success)
 :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success)
 
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 
@@ -591,13 +590,8 @@ int32_t HebrewCalendar::handleGetMonthLength(int32_t extendedYear, int32_t month
 * Returns the number of days in the given Hebrew year
 * @internal
 */
-int32_t HebrewCalendar::handleGetYearLength(int32_t eyear) const {
-    UErrorCode status = U_ZERO_ERROR;
-    int32_t len = daysInYear(eyear, status);
-    if (U_FAILURE(status)) {
-        return 12;
-    }
-    return len;
+int32_t HebrewCalendar::handleGetYearLength(int32_t eyear, UErrorCode& status) const {
+    return daysInYear(eyear, status);
 }
 
 void HebrewCalendar::validateField(UCalendarDateFields field, UErrorCode &status) {
diff --git a/deps/icu-small/source/i18n/hebrwcal.h b/deps/icu-small/source/i18n/hebrwcal.h
index 5fb10993d307b6..33db1b7860ab48 100644
--- a/deps/icu-small/source/i18n/hebrwcal.h
+++ b/deps/icu-small/source/i18n/hebrwcal.h
@@ -326,9 +326,9 @@ class U_I18N_API HebrewCalendar : public Calendar {
      * calendar system.  Subclasses should override this method if they can
      * provide a more correct or more efficient implementation than the
      * default implementation in Calendar.
-     * @stable ICU 2.0
+     * @internal
      */
-    virtual int32_t handleGetYearLength(int32_t eyear) const override;
+    virtual int32_t handleGetYearLength(int32_t eyear, UErrorCode& status) const override;
 
     /**
      * Subclasses may override this method to compute several fields
diff --git a/deps/icu-small/source/i18n/indiancal.cpp b/deps/icu-small/source/i18n/indiancal.cpp
index b1fd39b9927e5d..bb4b6b9939c029 100644
--- a/deps/icu-small/source/i18n/indiancal.cpp
+++ b/deps/icu-small/source/i18n/indiancal.cpp
@@ -41,7 +41,6 @@ IndianCalendar* IndianCalendar::clone() const {
 IndianCalendar::IndianCalendar(const Locale& aLocale, UErrorCode& success)
   :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success)
 {
-  setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 IndianCalendar::IndianCalendar(const IndianCalendar& other) : Calendar(other) {
@@ -129,7 +128,8 @@ int32_t IndianCalendar::handleGetMonthLength(int32_t eyear, int32_t month, UErro
  *
  * @param eyear The year in Saka Era.
  */
-int32_t IndianCalendar::handleGetYearLength(int32_t eyear) const {
+int32_t IndianCalendar::handleGetYearLength(int32_t eyear, UErrorCode& status) const {
+    if (U_FAILURE(status)) return 0;
     return isGregorianLeap(eyear + INDIAN_ERA_START) ? 366 : 365;
 }
 /*
@@ -143,18 +143,6 @@ static double gregorianToJD(int32_t year, int32_t month, int32_t date) {
    return Grego::fieldsToDay(year, month, date) + kEpochStartAsJulianDay - 0.5;
 }
 
-/*
- * Returns the Gregorian Date corresponding to a given Julian Day
- * Month is 0 based.
- * @param jd The Julian Day
- */
-static int32_t* jdToGregorian(double jd, int32_t gregorianDate[3], UErrorCode& status) {
-   int32_t gdow;
-   Grego::dayToFields(jd - kEpochStartAsJulianDay,
-                      gregorianDate[0], gregorianDate[1], gregorianDate[2], gdow, status);
-   return gregorianDate;
-}
-
    
 //-------------------------------------------------------------------------
 // Functions for converting from field values to milliseconds....
@@ -266,10 +254,9 @@ int32_t IndianCalendar::handleGetExtendedYear(UErrorCode& status) {
 void IndianCalendar::handleComputeFields(int32_t julianDay, UErrorCode&  status) {
     double jdAtStartOfGregYear;
     int32_t leapMonth, IndianYear, yday, IndianMonth, IndianDayOfMonth, mday;
-    int32_t gregorianYear;      // Stores gregorian date corresponding to Julian day;
-    int32_t gd[3];
+    // Stores gregorian date corresponding to Julian day;
+    int32_t gregorianYear = Grego::dayToYear(julianDay - kEpochStartAsJulianDay, status);
 
-    gregorianYear = jdToGregorian(julianDay, gd, status)[0];          // Gregorian date for Julian day
     if (U_FAILURE(status)) return;
     IndianYear = gregorianYear - INDIAN_ERA_START;            // Year in Saka era
     jdAtStartOfGregYear = gregorianToJD(gregorianYear, 0, 1); // JD at start of Gregorian year
diff --git a/deps/icu-small/source/i18n/indiancal.h b/deps/icu-small/source/i18n/indiancal.h
index 2062bcec9103a4..ff067d0b3c3d92 100644
--- a/deps/icu-small/source/i18n/indiancal.h
+++ b/deps/icu-small/source/i18n/indiancal.h
@@ -215,7 +215,7 @@ class U_I18N_API IndianCalendar : public Calendar {
    * Return the number of days in the given Indian year
    * @internal
    */
-  virtual int32_t handleGetYearLength(int32_t extendedYear) const override;
+  virtual int32_t handleGetYearLength(int32_t extendedYear, UErrorCode& status) const override;
 
   //-------------------------------------------------------------------------
   // Functions for converting from field values to milliseconds....
diff --git a/deps/icu-small/source/i18n/islamcal.cpp b/deps/icu-small/source/i18n/islamcal.cpp
index dfeac36a66516d..e847ac28c894cc 100644
--- a/deps/icu-small/source/i18n/islamcal.cpp
+++ b/deps/icu-small/source/i18n/islamcal.cpp
@@ -202,7 +202,6 @@ IslamicCalendar* IslamicCalendar::clone() const {
 IslamicCalendar::IslamicCalendar(const Locale& aLocale, UErrorCode& success)
 :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success)
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 IslamicCalendar::~IslamicCalendar()
@@ -444,15 +443,8 @@ int32_t yearLength(int32_t extendedYear, UErrorCode& status) {
 * Return the number of days in the given Islamic year
 * @draft ICU 2.4
 */
-int32_t IslamicCalendar::handleGetYearLength(int32_t extendedYear) const {
-    UErrorCode status = U_ZERO_ERROR;
-    int32_t length = yearLength(extendedYear, status);
-    if (U_FAILURE(status)) {
-        // fallback to normal Islamic calendar length 355 day a year if we
-        // encounter error and cannot propagate.
-        return 355;
-    }
-    return length;
+int32_t IslamicCalendar::handleGetYearLength(int32_t extendedYear, UErrorCode& status) const {
+    return yearLength(extendedYear, status);
 }
 
 //-------------------------------------------------------------------------
@@ -706,7 +698,8 @@ int32_t IslamicCivilCalendar::handleGetMonthLength(int32_t extendedYear, int32_t
 * Return the number of days in the given Islamic year
 * @draft ICU 2.4
 */
-int32_t IslamicCivilCalendar::handleGetYearLength(int32_t extendedYear) const {
+int32_t IslamicCivilCalendar::handleGetYearLength(int32_t extendedYear, UErrorCode& status) const {
+    if (U_FAILURE(status)) return 0;
     return 354 + (civilLeapYear(extendedYear) ? 1 : 0);
 }
 
@@ -872,7 +865,7 @@ int32_t IslamicUmalquraCalendar::handleGetMonthLength(int32_t extendedYear, int3
 
 int32_t IslamicUmalquraCalendar::yearLength(int32_t extendedYear, UErrorCode& status) const {
     if (extendedYear<UMALQURA_YEAR_START || extendedYear>UMALQURA_YEAR_END) {
-        return IslamicCivilCalendar::handleGetYearLength(extendedYear);
+        return IslamicCivilCalendar::handleGetYearLength(extendedYear, status);
     }
     int length = 0;
     for(int i=0; i<12; i++) {
@@ -888,15 +881,8 @@ int32_t IslamicUmalquraCalendar::yearLength(int32_t extendedYear, UErrorCode& st
 * Return the number of days in the given Islamic year
 * @draft ICU 2.4
 */
-int32_t IslamicUmalquraCalendar::handleGetYearLength(int32_t extendedYear) const {
-    UErrorCode status = U_ZERO_ERROR;
-    int32_t length = yearLength(extendedYear, status);
-    if (U_FAILURE(status)) {
-        // fallback to normal Islamic calendar length 355 day a year if we
-        // encounter error and cannot propagate.
-        return 355;
-    }
-    return length;
+int32_t IslamicUmalquraCalendar::handleGetYearLength(int32_t extendedYear, UErrorCode& status) const {
+    return yearLength(extendedYear, status);
 }
 
 /**
diff --git a/deps/icu-small/source/i18n/islamcal.h b/deps/icu-small/source/i18n/islamcal.h
index e42e681328b4d2..db90255b5b5a4f 100644
--- a/deps/icu-small/source/i18n/islamcal.h
+++ b/deps/icu-small/source/i18n/islamcal.h
@@ -235,7 +235,7 @@ class U_I18N_API IslamicCalendar : public Calendar {
    * Return the number of days in the given Islamic year
    * @internal
    */
-  virtual int32_t handleGetYearLength(int32_t extendedYear) const override;
+  virtual int32_t handleGetYearLength(int32_t extendedYear, UErrorCode& status) const override;
     
   //-------------------------------------------------------------------------
   // Functions for converting from field values to milliseconds....
@@ -438,7 +438,7 @@ class U_I18N_API IslamicCivilCalendar : public IslamicCalendar {
    * Return the number of days in the given Islamic year
    * @internal
    */
-  virtual int32_t handleGetYearLength(int32_t extendedYear) const override;
+  virtual int32_t handleGetYearLength(int32_t extendedYear, UErrorCode& status) const override;
 
   /**
    * Override Calendar to compute several fields specific to the Islamic
@@ -621,7 +621,7 @@ class U_I18N_API IslamicUmalquraCalendar : public IslamicCivilCalendar {
    * Return the number of days in the given Islamic year
    * @internal
    */
-  virtual int32_t handleGetYearLength(int32_t extendedYear) const override;
+  virtual int32_t handleGetYearLength(int32_t extendedYear, UErrorCode& status) const override;
 
   /**
    * Override Calendar to compute several fields specific to the Islamic
diff --git a/deps/icu-small/source/i18n/japancal.cpp b/deps/icu-small/source/i18n/japancal.cpp
index c0dd9fad0dc09b..b389b4530b290a 100644
--- a/deps/icu-small/source/i18n/japancal.cpp
+++ b/deps/icu-small/source/i18n/japancal.cpp
@@ -115,7 +115,6 @@ JapaneseCalendar::JapaneseCalendar(const Locale& aLocale, UErrorCode& success)
 :   GregorianCalendar(aLocale, success)
 {
     init(success);
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 JapaneseCalendar::~JapaneseCalendar()
@@ -130,12 +129,6 @@ JapaneseCalendar::JapaneseCalendar(const JapaneseCalendar& source)
     U_ASSERT(U_SUCCESS(status));
 }
 
-JapaneseCalendar& JapaneseCalendar::operator= ( const JapaneseCalendar& right)
-{
-    GregorianCalendar::operator=(right);
-    return *this;
-}
-
 JapaneseCalendar* JapaneseCalendar::clone() const
 {
     return new JapaneseCalendar(*this);
diff --git a/deps/icu-small/source/i18n/japancal.h b/deps/icu-small/source/i18n/japancal.h
index 627b12750b98a3..3271cbfb794a9e 100644
--- a/deps/icu-small/source/i18n/japancal.h
+++ b/deps/icu-small/source/i18n/japancal.h
@@ -104,13 +104,6 @@ class JapaneseCalendar : public GregorianCalendar {
      */
     JapaneseCalendar(const JapaneseCalendar& source);
 
-    /**
-     * Default assignment operator
-     * @param right    the object to be copied.
-     * @internal
-     */
-    JapaneseCalendar& operator=(const JapaneseCalendar& right);
-
     /**
      * Create and return a polymorphic copy of this calendar.
      * @return    return a polymorphic copy of this calendar.
diff --git a/deps/icu-small/source/i18n/measunit.cpp b/deps/icu-small/source/i18n/measunit.cpp
index 2741b84aabf0c4..8dda799d0165db 100644
--- a/deps/icu-small/source/i18n/measunit.cpp
+++ b/deps/icu-small/source/i18n/measunit.cpp
@@ -41,26 +41,26 @@ static const int32_t gOffsets[] = {
     2,
     7,
     17,
-    27,
-    31,
-    333,
-    344,
-    362,
-    366,
-    375,
-    378,
-    382,
-    390,
-    412,
-    416,
-    431,
+    28,
+    32,
+    334,
+    345,
+    363,
+    367,
+    376,
+    379,
+    383,
+    391,
+    413,
+    417,
     432,
-    438,
-    449,
-    455,
-    459,
-    461,
-    495
+    433,
+    439,
+    450,
+    456,
+    460,
+    462,
+    496
 };
 
 static const int32_t kCurrencyOffset = 5;
@@ -121,6 +121,7 @@ static const char * const gSubTypes[] = {
     "permille",
     "permillion",
     "permyriad",
+    "portion-per-1e9",
     "liter-per-100-kilometer",
     "liter-per-kilometer",
     "mile-per-gallon",
@@ -811,6 +812,14 @@ MeasureUnit MeasureUnit::getPermyriad() {
     return MeasureUnit(3, 9);
 }
 
+MeasureUnit *MeasureUnit::createPortionPer1E9(UErrorCode &status) {
+    return MeasureUnit::create(3, 10, status);
+}
+
+MeasureUnit MeasureUnit::getPortionPer1E9() {
+    return MeasureUnit(3, 10);
+}
+
 MeasureUnit *MeasureUnit::createLiterPer100Kilometers(UErrorCode &status) {
     return MeasureUnit::create(4, 0, status);
 }
@@ -2400,6 +2409,7 @@ MeasureUnitImpl MeasureUnitImpl::copy(UErrorCode &status) const {
     MeasureUnitImpl result;
     result.complexity = complexity;
     result.identifier.append(identifier, status);
+    result.constantDenominator = constantDenominator;
     for (int32_t i = 0; i < singleUnits.length(); i++) {
         SingleUnitImpl *item = result.singleUnits.emplaceBack(*singleUnits[i]);
         if (!item) {
diff --git a/deps/icu-small/source/i18n/measunit_extra.cpp b/deps/icu-small/source/i18n/measunit_extra.cpp
index a6348422738bfa..8eb9fe55167077 100644
--- a/deps/icu-small/source/i18n/measunit_extra.cpp
+++ b/deps/icu-small/source/i18n/measunit_extra.cpp
@@ -15,6 +15,7 @@
 #include "charstr.h"
 #include "cmemory.h"
 #include "cstring.h"
+#include "double-conversion-string-to-double.h"
 #include "measunit_impl.h"
 #include "resource.h"
 #include "uarrsort.h"
@@ -30,13 +31,15 @@
 #include "unicode/ustringtrie.h"
 #include "uresimp.h"
 #include "util.h"
+#include <limits.h>
 #include <cstdlib>
-
 U_NAMESPACE_BEGIN
 
 
 namespace {
 
+using icu::double_conversion::StringToDoubleConverter;
+
 // TODO: Propose a new error code for this?
 constexpr UErrorCode kUnitIdentifierSyntaxError = U_ILLEGAL_ARGUMENT_ERROR;
 
@@ -467,37 +470,55 @@ void U_CALLCONV initUnitExtras(UErrorCode& status) {
 
 class Token {
 public:
-    Token(int32_t match) : fMatch(match) {}
-
-    enum Type {
-        TYPE_UNDEFINED,
-        TYPE_PREFIX,
-        // Token type for "-per-", "-", and "-and-".
-        TYPE_COMPOUND_PART,
-        // Token type for "per-".
-        TYPE_INITIAL_COMPOUND_PART,
-        TYPE_POWER_PART,
-        TYPE_SIMPLE_UNIT,
-    };
-
-    // Calling getType() is invalid, resulting in an assertion failure, if Token
-    // value isn't positive.
-    Type getType() const {
-        U_ASSERT(fMatch > 0);
-        if (fMatch < kCompoundPartOffset) {
-            return TYPE_PREFIX;
-        }
-        if (fMatch < kInitialCompoundPartOffset) {
-            return TYPE_COMPOUND_PART;
-        }
-        if (fMatch < kPowerPartOffset) {
-            return TYPE_INITIAL_COMPOUND_PART;
-        }
-        if (fMatch < kSimpleUnitOffset) {
-            return TYPE_POWER_PART;
-        }
-        return TYPE_SIMPLE_UNIT;
-    }
+  Token(int64_t match) : fMatch(match) {
+      if (fMatch < kCompoundPartOffset) {
+          this->fType = TYPE_PREFIX;
+      } else if (fMatch < kInitialCompoundPartOffset) {
+          this->fType = TYPE_COMPOUND_PART;
+      } else if (fMatch < kPowerPartOffset) {
+          this->fType = TYPE_INITIAL_COMPOUND_PART;
+      } else if (fMatch < kSimpleUnitOffset) {
+          this->fType = TYPE_POWER_PART;
+      } else {
+          this->fType = TYPE_SIMPLE_UNIT;
+      }
+  }
+
+  static Token constantToken(StringPiece str, UErrorCode &status) {
+      Token result;
+      auto value = Token::parseStringToLong(str, status);
+      if (U_FAILURE(status)) {
+          return result;
+      }
+      result.fMatch = value;
+      result.fType = TYPE_CONSTANT_DENOMINATOR;
+      return result;
+  }
+
+  enum Type {
+      TYPE_UNDEFINED,
+      TYPE_PREFIX,
+      // Token type for "-per-", "-", and "-and-".
+      TYPE_COMPOUND_PART,
+      // Token type for "per-".
+      TYPE_INITIAL_COMPOUND_PART,
+      TYPE_POWER_PART,
+      TYPE_SIMPLE_UNIT,
+      TYPE_CONSTANT_DENOMINATOR,
+  };
+
+  // Calling getType() is invalid, resulting in an assertion failure, if Token
+  // value isn't positive.
+  Type getType() const {
+      U_ASSERT(fMatch >= 0);
+      return this->fType;
+  }
+
+  // Retrieve the value of the constant denominator if the token is of type TYPE_CONSTANT_DENOMINATOR.
+  uint64_t getConstantDenominator() const {
+      U_ASSERT(getType() == TYPE_CONSTANT_DENOMINATOR);
+      return static_cast<uint64_t>(fMatch);
+  }
 
     UMeasurePrefix getUnitPrefix() const {
         U_ASSERT(getType() == TYPE_PREFIX);
@@ -530,8 +551,41 @@ class Token {
         return fMatch - kSimpleUnitOffset;
     }
 
+    // TODO: Consider moving this to a separate utility class.
+    // Utility function to parse a string into an unsigned long value.
+    // The value must be a positive integer within the range [1, INT64_MAX].
+    // The input can be in integer or scientific notation.
+    static uint64_t parseStringToLong(const StringPiece strNum, UErrorCode &status) {
+        // We are processing well-formed input, so we don't need any special options to
+        // StringToDoubleConverter.
+        StringToDoubleConverter converter(0, 0, 0, "", "");
+        int32_t count;
+        double double_result = converter.StringToDouble(strNum.data(), strNum.length(), &count);
+        if (count != strNum.length()) {
+            status = kUnitIdentifierSyntaxError;
+            return 0;
+        }
+
+        if (U_FAILURE(status) || double_result < 1.0 || double_result > static_cast<double>(INT64_MAX)) {
+            status = kUnitIdentifierSyntaxError;
+            return 0;
+        }
+
+        // Check if the value is integer.
+        uint64_t int_result = static_cast<uint64_t>(double_result);
+        const double kTolerance = 1e-9;
+        if (abs(double_result - int_result) > kTolerance) {
+            status = kUnitIdentifierSyntaxError;
+            return 0;
+        }
+
+        return int_result;
+    }
+
 private:
-    int32_t fMatch;
+  Token() = default;
+  int64_t fMatch;
+  Type fType = TYPE_UNDEFINED;
 };
 
 class Parser {
@@ -555,6 +609,50 @@ class Parser {
         return {source};
     }
 
+    /**
+     * A single unit or a constant denominator.
+     */
+    struct SingleUnitOrConstant {
+        enum ValueType {
+            kSingleUnit,
+            kConstantDenominator,
+        };
+
+        ValueType type = kSingleUnit;
+        SingleUnitImpl singleUnit;
+        uint64_t constantDenominator;
+
+        static SingleUnitOrConstant singleUnitValue(SingleUnitImpl singleUnit) {
+            SingleUnitOrConstant result;
+            result.type = kSingleUnit;
+            result.singleUnit = singleUnit;
+            result.constantDenominator = 0;
+            return result;
+        }
+
+        static SingleUnitOrConstant constantDenominatorValue(uint64_t constant) {
+            SingleUnitOrConstant result;
+            result.type = kConstantDenominator;
+            result.singleUnit = {};
+            result.constantDenominator = constant;
+            return result;
+        }
+
+        uint64_t getConstantDenominator() const {
+            U_ASSERT(type == kConstantDenominator);
+            return constantDenominator;
+        }
+
+        SingleUnitImpl getSingleUnit() const {
+            U_ASSERT(type == kSingleUnit);
+            return singleUnit;
+        }
+
+        bool isSingleUnit() const { return type == kSingleUnit; }
+
+        bool isConstantDenominator() const { return type == kConstantDenominator; }
+    };
+
     MeasureUnitImpl parse(UErrorCode& status) {
         MeasureUnitImpl result;
 
@@ -569,12 +667,19 @@ class Parser {
         while (hasNext()) {
             bool sawAnd = false;
 
-            SingleUnitImpl singleUnit = nextSingleUnit(sawAnd, status);
+            auto singleUnitOrConstant = nextSingleUnitOrConstant(sawAnd, status);
             if (U_FAILURE(status)) {
                 return result;
             }
 
-            bool added = result.appendSingleUnit(singleUnit, status);
+            if (singleUnitOrConstant.isConstantDenominator()) {
+                result.constantDenominator = singleUnitOrConstant.getConstantDenominator();
+                result.complexity = UMEASURE_UNIT_COMPOUND;
+                continue;
+            }
+
+            U_ASSERT(singleUnitOrConstant.isSingleUnit());
+            bool added = result.appendSingleUnit(singleUnitOrConstant.getSingleUnit(), status);
             if (U_FAILURE(status)) {
                 return result;
             }
@@ -604,6 +709,12 @@ class Parser {
             }
         }
 
+        if (result.singleUnits.length() == 0) {
+            // The identifier was empty or only had a constant denominator.
+            status = kUnitIdentifierSyntaxError;
+            return result; // add it for code consistency.
+        }
+
         return result;
     }
 
@@ -622,6 +733,10 @@ class Parser {
     // identifier is invalid pending TODO(CLDR-13701).
     bool fAfterPer = false;
 
+    // Set to true when we've just seen a "per-". This is used to determine if
+    // the next token can be a constant denominator token.
+    bool fJustSawPer = false;
+
     Parser() : fSource(""), fTrie(u"") {}
 
     Parser(StringPiece source)
@@ -640,6 +755,10 @@ class Parser {
         // Saves the position in the fSource string for the end of the most
         // recent matching token.
         int32_t previ = -1;
+
+        // Saves the position in the fSource string for later use in case of unit constant found.
+        int32_t currentFIndex = fIndex;
+
         // Find the longest token that matches a value in the trie:
         while (fIndex < fSource.length()) {
             auto result = fTrie.next(fSource.data()[fIndex++]);
@@ -658,12 +777,25 @@ class Parser {
             // continue;
         }
 
-        if (match < 0) {
-            status = kUnitIdentifierSyntaxError;
-        } else {
+        if (match >= 0) {
             fIndex = previ;
+            return {match};
+        }
+
+        // If no match was found, we check if the token is a constant denominator.
+        // 1. We find the index of the start of the next token or the end of the string.
+        int32_t endOfConstantIndex = fSource.find("-", currentFIndex);
+        endOfConstantIndex = (endOfConstantIndex == -1) ? fSource.length() : endOfConstantIndex;
+        if (endOfConstantIndex <= currentFIndex) {
+            status = kUnitIdentifierSyntaxError;
+            return {match};
         }
-        return {match};
+
+        // 2. We extract the substring from the start of the constant to the end of the constant.
+        StringPiece constantDenominatorStr =
+            fSource.substr(currentFIndex, endOfConstantIndex - currentFIndex);
+        fIndex = endOfConstantIndex;
+        return Token::constantToken(constantDenominatorStr, status);
     }
 
     /**
@@ -680,10 +812,10 @@ class Parser {
      * unit", sawAnd is set to true. If not, it is left as is.
      * @param status ICU error code.
      */
-    SingleUnitImpl nextSingleUnit(bool &sawAnd, UErrorCode &status) {
-        SingleUnitImpl result;
+    SingleUnitOrConstant nextSingleUnitOrConstant(bool &sawAnd, UErrorCode &status) {
+        SingleUnitImpl singleUnitResult;
         if (U_FAILURE(status)) {
-            return result;
+            return {};
         }
 
         // state:
@@ -695,19 +827,22 @@ class Parser {
         bool atStart = fIndex == 0;
         Token token = nextToken(status);
         if (U_FAILURE(status)) {
-            return result;
+            return {};
         }
 
+        fJustSawPer = false;
+
         if (atStart) {
             // Identifiers optionally start with "per-".
             if (token.getType() == Token::TYPE_INITIAL_COMPOUND_PART) {
                 U_ASSERT(token.getInitialCompoundPart() == INITIAL_COMPOUND_PART_PER);
                 fAfterPer = true;
-                result.dimensionality = -1;
+                fJustSawPer = true;
+                singleUnitResult.dimensionality = -1;
 
                 token = nextToken(status);
                 if (U_FAILURE(status)) {
-                    return result;
+                    return {};
                 }
             }
         } else {
@@ -715,7 +850,7 @@ class Parser {
             // via a compound part:
             if (token.getType() != Token::TYPE_COMPOUND_PART) {
                 status = kUnitIdentifierSyntaxError;
-                return result;
+                return {};
             }
 
             switch (token.getMatch()) {
@@ -724,15 +859,16 @@ class Parser {
                     // Mixed compound units not yet supported,
                     // TODO(CLDR-13701).
                     status = kUnitIdentifierSyntaxError;
-                    return result;
+                    return {};
                 }
                 fAfterPer = true;
-                result.dimensionality = -1;
+                fJustSawPer = true;
+                singleUnitResult.dimensionality = -1;
                 break;
 
             case COMPOUND_PART_TIMES:
                 if (fAfterPer) {
-                    result.dimensionality = -1;
+                    singleUnitResult.dimensionality = -1;
                 }
                 break;
 
@@ -741,7 +877,7 @@ class Parser {
                     // Can't start with "-and-", and mixed compound units
                     // not yet supported, TODO(CLDR-13701).
                     status = kUnitIdentifierSyntaxError;
-                    return result;
+                    return {};
                 }
                 sawAnd = true;
                 break;
@@ -749,52 +885,65 @@ class Parser {
 
             token = nextToken(status);
             if (U_FAILURE(status)) {
-                return result;
+                return {};
+            }
+        }
+
+        if (token.getType() == Token::TYPE_CONSTANT_DENOMINATOR) {
+            if (!fJustSawPer) {
+                status = kUnitIdentifierSyntaxError;
+                return {};
             }
+
+            return SingleUnitOrConstant::constantDenominatorValue(token.getConstantDenominator());
         }
 
         // Read tokens until we have a complete SingleUnit or we reach the end.
         while (true) {
             switch (token.getType()) {
-                case Token::TYPE_POWER_PART:
-                    if (state > 0) {
-                        status = kUnitIdentifierSyntaxError;
-                        return result;
-                    }
-                    result.dimensionality *= token.getPower();
-                    state = 1;
-                    break;
-
-                case Token::TYPE_PREFIX:
-                    if (state > 1) {
-                        status = kUnitIdentifierSyntaxError;
-                        return result;
-                    }
-                    result.unitPrefix = token.getUnitPrefix();
-                    state = 2;
-                    break;
-
-                case Token::TYPE_SIMPLE_UNIT:
-                    result.index = token.getSimpleUnitIndex();
-                    return result;
+            case Token::TYPE_POWER_PART:
+                if (state > 0) {
+                    status = kUnitIdentifierSyntaxError;
+                    return {};
+                }
+                singleUnitResult.dimensionality *= token.getPower();
+                state = 1;
+                break;
 
-                default:
+            case Token::TYPE_PREFIX:
+                if (state > 1) {
                     status = kUnitIdentifierSyntaxError;
-                    return result;
+                    return {};
+                }
+                singleUnitResult.unitPrefix = token.getUnitPrefix();
+                state = 2;
+                break;
+
+            case Token::TYPE_SIMPLE_UNIT:
+                singleUnitResult.index = token.getSimpleUnitIndex();
+                break;
+
+            default:
+                status = kUnitIdentifierSyntaxError;
+                return {};
+            }
+
+            if (token.getType() == Token::TYPE_SIMPLE_UNIT) {
+                break;
             }
 
             if (!hasNext()) {
                 // We ran out of tokens before finding a complete single unit.
                 status = kUnitIdentifierSyntaxError;
-                return result;
+                return {};
             }
             token = nextToken(status);
             if (U_FAILURE(status)) {
-                return result;
+                return {};
             }
         }
 
-        return result;
+        return SingleUnitOrConstant::singleUnitValue(singleUnitResult);
     }
 };
 
@@ -1120,6 +1269,51 @@ MeasureUnitImpl::extractIndividualUnitsWithIndices(UErrorCode &status) const {
     return result;
 }
 
+int32_t countCharacter(const CharString &str, char c) {
+    int32_t count = 0;
+    for (int32_t i = 0, n = str.length(); i < n; i++) {
+        if (str[i] == c) {
+            count++;
+        }
+    }
+    return count;
+}
+
+/**
+ * Internal function that returns a string of the constants in the correct
+ * format.
+ *
+ * Example:
+ * 1000 --> "-per-1000"
+ * 1000000 --> "-per-1e6"
+ *
+ * NOTE: this function is only used when the constant denominator is greater
+ * than 0.
+ */
+CharString getConstantsString(uint64_t constantDenominator, UErrorCode &status) {
+    U_ASSERT(constantDenominator > 0 && constantDenominator <= LLONG_MAX);
+
+    CharString result;
+    result.appendNumber(constantDenominator, status);
+    if (U_FAILURE(status)) {
+        return result;
+    }
+
+    if (constantDenominator <= 1000) {
+        return result;
+    }
+
+    // Check if the constant is a power of 10.
+    int32_t zeros = countCharacter(result, '0');
+    if (zeros == result.length() - 1 && result[0] == '1') {
+        result.clear();
+        result.append(StringPiece("1e"), status);
+        result.appendNumber(zeros, status);
+    }
+
+    return result;
+}
+
 /**
  * Normalize a MeasureUnitImpl and generate the identifier string in place.
  */
@@ -1128,7 +1322,7 @@ void MeasureUnitImpl::serialize(UErrorCode &status) {
         return;
     }
 
-    if (this->singleUnits.length() == 0) {
+    if (this->singleUnits.length() == 0 && this->constantDenominator == 0) {
         // Dimensionless, constructed by the default constructor.
         return;
     }
@@ -1145,6 +1339,7 @@ void MeasureUnitImpl::serialize(UErrorCode &status) {
     CharString result;
     bool beforePer = true;
     bool firstTimeNegativeDimension = false;
+    bool constantDenominatorAppended = false;
     for (int32_t i = 0; i < this->singleUnits.length(); i++) {
         if (beforePer && (*this->singleUnits[i]).dimensionality < 0) {
             beforePer = false;
@@ -1168,43 +1363,103 @@ void MeasureUnitImpl::serialize(UErrorCode &status) {
                 } else {
                     result.append(StringPiece("-per-"), status);
                 }
-            } else {
-                if (result.length() != 0) {
+
+                if (this->constantDenominator > 0) {
+                    result.append(getConstantsString(this->constantDenominator, status), status);
                     result.append(StringPiece("-"), status);
+                    constantDenominatorAppended = true;
                 }
+
+            } else if (result.length() != 0) {
+                result.append(StringPiece("-"), status);
             }
         }
 
         this->singleUnits[i]->appendNeutralIdentifier(result, status);
     }
 
+    if (!constantDenominatorAppended && this->constantDenominator > 0) {
+        result.append(StringPiece("-per-"), status);
+        result.append(getConstantsString(this->constantDenominator, status), status);
+    }
+
+    if (U_FAILURE(status)) {
+        return;
+    }
     this->identifier = CharString(result, status);
 }
 
-MeasureUnit MeasureUnitImpl::build(UErrorCode& status) && {
+MeasureUnit MeasureUnitImpl::build(UErrorCode &status) && {
     this->serialize(status);
     return MeasureUnit(std::move(*this));
 }
 
-MeasureUnit MeasureUnit::forIdentifier(StringPiece identifier, UErrorCode& status) {
+MeasureUnit MeasureUnit::forIdentifier(StringPiece identifier, UErrorCode &status) {
     return Parser::from(identifier, status).parse(status).build(status);
 }
 
-UMeasureUnitComplexity MeasureUnit::getComplexity(UErrorCode& status) const {
+UMeasureUnitComplexity MeasureUnit::getComplexity(UErrorCode &status) const {
     MeasureUnitImpl temp;
     return MeasureUnitImpl::forMeasureUnit(*this, temp, status).complexity;
 }
 
-UMeasurePrefix MeasureUnit::getPrefix(UErrorCode& status) const {
+UMeasurePrefix MeasureUnit::getPrefix(UErrorCode &status) const {
     return SingleUnitImpl::forMeasureUnit(*this, status).unitPrefix;
 }
 
-MeasureUnit MeasureUnit::withPrefix(UMeasurePrefix prefix, UErrorCode& status) const UPRV_NO_SANITIZE_UNDEFINED {
+MeasureUnit MeasureUnit::withPrefix(UMeasurePrefix prefix,
+                                    UErrorCode &status) const UPRV_NO_SANITIZE_UNDEFINED {
     SingleUnitImpl singleUnit = SingleUnitImpl::forMeasureUnit(*this, status);
     singleUnit.unitPrefix = prefix;
     return singleUnit.build(status);
 }
 
+uint64_t MeasureUnit::getConstantDenominator(UErrorCode &status) const {
+    auto complexity = this->getComplexity(status);
+    if (U_FAILURE(status)) {
+        return 0;
+    }
+
+    if (complexity != UMEASURE_UNIT_SINGLE && complexity != UMEASURE_UNIT_COMPOUND) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return 0;
+    }
+
+    if (this->fImpl == nullptr) {
+        return 0;
+    }
+
+    return this->fImpl->constantDenominator;
+}
+
+MeasureUnit MeasureUnit::withConstantDenominator(uint64_t denominator, UErrorCode &status) const {
+    // To match the behavior of the Java API, we do not allow a constant denominator
+    // bigger than LONG_MAX.
+    if (denominator > LONG_MAX) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return {};
+    }
+
+    auto complexity = this->getComplexity(status);
+    if (U_FAILURE(status)) {
+        return {};
+    }
+    if (complexity != UMEASURE_UNIT_SINGLE && complexity != UMEASURE_UNIT_COMPOUND) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return {};
+    }
+
+    MeasureUnitImpl impl = MeasureUnitImpl::forMeasureUnitMaybeCopy(*this, status);
+    if (U_FAILURE(status)) {
+        return {};
+    }
+
+    impl.constantDenominator = denominator;
+    impl.complexity = (impl.singleUnits.length() < 2 && denominator == 0) ? UMEASURE_UNIT_SINGLE
+                                                                          : UMEASURE_UNIT_COMPOUND;
+    return std::move(impl).build(status);
+}
+
 int32_t MeasureUnit::getDimensionality(UErrorCode& status) const {
     SingleUnitImpl singleUnit = SingleUnitImpl::forMeasureUnit(*this, status);
     if (U_FAILURE(status)) { return 0; }
@@ -1222,6 +1477,11 @@ MeasureUnit MeasureUnit::withDimensionality(int32_t dimensionality, UErrorCode&
 
 MeasureUnit MeasureUnit::reciprocal(UErrorCode& status) const {
     MeasureUnitImpl impl = MeasureUnitImpl::forMeasureUnitMaybeCopy(*this, status);
+    // The reciprocal of a unit that has a constant denominator is not allowed.
+    if (impl.constantDenominator != 0) {
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return {};
+    }
     impl.takeReciprocal(status);
     return std::move(impl).build(status);
 }
@@ -1237,9 +1497,25 @@ MeasureUnit MeasureUnit::product(const MeasureUnit& other, UErrorCode& status) c
     for (int32_t i = 0; i < otherImpl.singleUnits.length(); i++) {
         impl.appendSingleUnit(*otherImpl.singleUnits[i], status);
     }
-    if (impl.singleUnits.length() > 1) {
+
+    uint64_t currentConstatDenominator = this->getConstantDenominator(status);
+    uint64_t otherConstantDenominator = other.getConstantDenominator(status);
+
+    // TODO: we can also multiply the constant denominators instead of returning an error.
+    if (currentConstatDenominator != 0 && otherConstantDenominator != 0) {
+        // There is only `one` constant denominator in a compound unit.
+        // Therefore, we Cannot multiply units that both of them have a constant denominator
+        status = U_ILLEGAL_ARGUMENT_ERROR;
+        return {};
+    }
+
+    // Because either one of the constant denominators is zero, we can use the maximum of them.
+    impl.constantDenominator = uprv_max(currentConstatDenominator, otherConstantDenominator);
+
+    if (impl.singleUnits.length() > 1 || impl.constantDenominator > 0) {
         impl.complexity = UMEASURE_UNIT_COMPOUND;
     }
+
     return std::move(impl).build(status);
 }
 
diff --git a/deps/icu-small/source/i18n/measunit_impl.h b/deps/icu-small/source/i18n/measunit_impl.h
index f6a8f90dc94f0c..db31435944c2ec 100644
--- a/deps/icu-small/source/i18n/measunit_impl.h
+++ b/deps/icu-small/source/i18n/measunit_impl.h
@@ -328,6 +328,14 @@ class U_I18N_API MeasureUnitImpl : public UMemory {
      */
     CharString identifier;
 
+    /**
+     * Represents the unit constant denominator.
+     *
+     * NOTE:
+     *   if set to 0, it means that the constant is not set.
+     */
+    uint64_t constantDenominator = 0;
+
     // For calling serialize
     // TODO(icu-units#147): revisit serialization
     friend class number::impl::LongNameHandler;
diff --git a/deps/icu-small/source/i18n/messageformat2.cpp b/deps/icu-small/source/i18n/messageformat2.cpp
index 73f7fa45e69f81..50d87422b64c57 100644
--- a/deps/icu-small/source/i18n/messageformat2.cpp
+++ b/deps/icu-small/source/i18n/messageformat2.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -11,8 +13,10 @@
 #include "unicode/messageformat2_data_model.h"
 #include "unicode/messageformat2_formattable.h"
 #include "unicode/messageformat2.h"
+#include "unicode/normalizer2.h"
 #include "unicode/unistr.h"
 #include "messageformat2_allocation.h"
+#include "messageformat2_checker.h"
 #include "messageformat2_evaluation.h"
 #include "messageformat2_macros.h"
 
@@ -37,7 +41,7 @@ static Formattable evalLiteral(const Literal& lit) {
         // The fallback for a variable name is itself.
         UnicodeString str(DOLLAR);
         str += var;
-        const Formattable* val = context.getGlobal(var, errorCode);
+        const Formattable* val = context.getGlobal(*this, var, errorCode);
         if (U_SUCCESS(errorCode)) {
             return (FormattedPlaceholder(*val, str));
         }
@@ -51,16 +55,16 @@ static Formattable evalLiteral(const Literal& lit) {
     return FormattedPlaceholder(evalLiteral(lit), lit.quoted());
 }
 
-[[nodiscard]] FormattedPlaceholder MessageFormatter::formatOperand(const Environment& env,
-                                                             const Operand& rand,
-                                                             MessageContext& context,
-                                                             UErrorCode &status) const {
+[[nodiscard]] InternalValue* MessageFormatter::formatOperand(const Environment& env,
+                                                            const Operand& rand,
+                                                            MessageContext& context,
+                                                            UErrorCode &status) const {
     if (U_FAILURE(status)) {
         return {};
     }
 
     if (rand.isNull()) {
-        return FormattedPlaceholder();
+        return create<InternalValue>(InternalValue(FormattedPlaceholder()), status);
     }
     if (rand.isVariable()) {
         // Check if it's local or global
@@ -71,15 +75,19 @@ static Formattable evalLiteral(const Literal& lit) {
         // Eager vs. lazy evaluation is an open issue:
         // see https://github.com/unicode-org/message-format-wg/issues/299
 
+        // NFC-normalize the variable name. See
+        // https://github.com/unicode-org/message-format-wg/blob/main/spec/syntax.md#names-and-identifiers
+        const VariableName normalized = normalizeNFC(var);
+
         // Look up the variable in the environment
-        if (env.has(var)) {
+        if (env.has(normalized)) {
           // `var` is a local -- look it up
-          const Closure& rhs = env.lookup(var);
+          const Closure& rhs = env.lookup(normalized);
           // Format the expression using the environment from the closure
           return formatExpression(rhs.getEnv(), rhs.getExpr(), context, status);
         }
         // Variable wasn't found in locals -- check if it's global
-        FormattedPlaceholder result = evalArgument(var, context, status);
+        FormattedPlaceholder result = evalArgument(normalized, context, status);
         if (status == U_ILLEGAL_ARGUMENT_ERROR) {
             status = U_ZERO_ERROR;
             // Unbound variable -- set a resolution error
@@ -88,12 +96,12 @@ static Formattable evalLiteral(const Literal& lit) {
             // https://github.com/unicode-org/message-format-wg/blob/main/spec/formatting.md#fallback-resolution
             UnicodeString str(DOLLAR);
             str += var;
-            return FormattedPlaceholder(str);
+            return create<InternalValue>(InternalValue(FormattedPlaceholder(str)), status);
         }
-        return result;
+        return create<InternalValue>(InternalValue(std::move(result)), status);
     } else {
         U_ASSERT(rand.isLiteral());
-        return formatLiteral(rand.asLiteral());
+        return create<InternalValue>(InternalValue(formatLiteral(rand.asLiteral())), status);
     }
 }
 
@@ -114,28 +122,32 @@ FunctionOptions MessageFormatter::resolveOptions(const Environment& env, const O
 
         // Options are fully evaluated before calling the function
         // Format the operand
-        FormattedPlaceholder rhsVal = formatOperand(env, v, context, status);
+        LocalPointer<InternalValue> rhsVal(formatOperand(env, v, context, status));
         if (U_FAILURE(status)) {
             return {};
         }
-        if (!rhsVal.isFallback()) {
-            resolvedOpt.adoptInstead(create<ResolvedFunctionOption>(ResolvedFunctionOption(k, rhsVal.asFormattable()), status));
-            if (U_FAILURE(status)) {
-                return {};
-            }
-            optionsVector->adoptElement(resolvedOpt.orphan(), status);
+        // Note: this means option values are "eagerly" evaluated.
+        // Currently, options don't have options. This will be addressed by the
+        // full FormattedPlaceholder redesign.
+        FormattedPlaceholder optValue = rhsVal->forceFormatting(context.getErrors(), status);
+        resolvedOpt.adoptInstead(create<ResolvedFunctionOption>
+                                 (ResolvedFunctionOption(k,
+                                                         optValue.asFormattable()),
+                                  status));
+        if (U_FAILURE(status)) {
+            return {};
         }
+        optionsVector->adoptElement(resolvedOpt.orphan(), status);
     }
-
     return FunctionOptions(std::move(*optionsVector), status);
 }
 
 // Overload that dispatches on argument type. Syntax doesn't provide for options in this case.
-[[nodiscard]] FormattedPlaceholder MessageFormatter::evalFormatterCall(FormattedPlaceholder&& argument,
-                                                                       MessageContext& context,
-                                                                       UErrorCode& status) const {
+[[nodiscard]] InternalValue* MessageFormatter::evalFunctionCall(FormattedPlaceholder&& argument,
+                                                                MessageContext& context,
+                                                                UErrorCode& status) const {
     if (U_FAILURE(status)) {
-        return {};
+        return nullptr;
     }
 
     // These cases should have been checked for already
@@ -153,11 +165,11 @@ FunctionOptions MessageFormatter::resolveOptions(const Environment& env, const O
             // No formatter for this type -- follow default behavior
             break;
         }
-        return evalFormatterCall(functionName,
-                                 std::move(argument),
-                                 FunctionOptions(),
-                                 context,
-                                 status);
+        return evalFunctionCall(functionName,
+                                create<InternalValue>(std::move(argument), status),
+                                FunctionOptions(),
+                                context,
+                                status);
     }
     default: {
         // TODO: The array case isn't handled yet; not sure whether it's desirable
@@ -167,104 +179,76 @@ FunctionOptions MessageFormatter::resolveOptions(const Environment& env, const O
     }
     // No formatter for this type, or it's a primitive type (which will be formatted later)
     // -- just return the argument itself
-    return std::move(argument);
+    return create<InternalValue>(std::move(argument), status);
 }
 
 // Overload that dispatches on function name
-[[nodiscard]] FormattedPlaceholder MessageFormatter::evalFormatterCall(const FunctionName& functionName,
-                                                                 FormattedPlaceholder&& argument,
-                                                                 FunctionOptions&& options,
-                                                                 MessageContext& context,
-                                                                 UErrorCode& status) const {
+// Adopts `arg`
+[[nodiscard]] InternalValue* MessageFormatter::evalFunctionCall(const FunctionName& functionName,
+                                                                InternalValue* arg_,
+                                                                FunctionOptions&& options,
+                                                                MessageContext& context,
+                                                                UErrorCode& status) const {
     if (U_FAILURE(status)) {
         return {};
     }
 
-    DynamicErrors& errs = context.getErrors();
-
-    UnicodeString fallback(COLON);
-    fallback += functionName;
-    if (!argument.isNullOperand()) {
-        fallback = argument.fallback;
-    }
+    LocalPointer<InternalValue> arg(arg_);
 
+    // Look up the formatter or selector
+    LocalPointer<Formatter> formatterImpl(nullptr);
+    LocalPointer<Selector> selectorImpl(nullptr);
     if (isFormatter(functionName)) {
-        LocalPointer<Formatter> formatterImpl(getFormatter(functionName, status));
-        if (U_FAILURE(status)) {
-            if (status == U_MF_FORMATTING_ERROR) {
-                errs.setFormattingError(functionName, status);
-                status = U_ZERO_ERROR;
-                return {};
-            }
-            if (status == U_MF_UNKNOWN_FUNCTION_ERROR) {
-                errs.setUnknownFunction(functionName, status);
-                status = U_ZERO_ERROR;
-                return {};
-            }
-            // Other errors are non-recoverable
-            return {};
-        }
-        U_ASSERT(formatterImpl != nullptr);
-
-        UErrorCode savedStatus = status;
-        FormattedPlaceholder result = formatterImpl->format(std::move(argument), std::move(options), status);
-        // Update errors
-        if (savedStatus != status) {
-            if (U_FAILURE(status)) {
-                if (status == U_MF_OPERAND_MISMATCH_ERROR) {
-                    status = U_ZERO_ERROR;
-                    errs.setOperandMismatchError(functionName, status);
-                } else {
-                    status = U_ZERO_ERROR;
-                    // Convey any error generated by the formatter
-                    // as a formatting error, except for operand mismatch errors
-                    errs.setFormattingError(functionName, status);
-                }
-                return FormattedPlaceholder(fallback);
-            } else {
-                // Ignore warnings
-                status = savedStatus;
-            }
-        }
-        // Ignore the output if any errors occurred
-        if (errs.hasFormattingError()) {
-            return FormattedPlaceholder(fallback);
-        }
-        return result;
+        formatterImpl.adoptInstead(getFormatter(functionName, status));
+        U_ASSERT(U_SUCCESS(status));
     }
-    // No formatter with this name -- set error
     if (isSelector(functionName)) {
-        errs.setFormattingError(functionName, status);
-    } else {
-        errs.setUnknownFunction(functionName, status);
+        selectorImpl.adoptInstead(getSelector(context, functionName, status));
+        U_ASSERT(U_SUCCESS(status));
+    }
+    if (formatterImpl == nullptr && selectorImpl == nullptr) {
+        // Unknown function error
+        context.getErrors().setUnknownFunction(functionName, status);
+
+        if (arg->hasNullOperand()) {
+            // Non-selector used as selector; an error would have been recorded earlier
+            UnicodeString fallback(COLON);
+            fallback += functionName;
+            return new InternalValue(FormattedPlaceholder(fallback));
+        } else {
+            return new InternalValue(FormattedPlaceholder(arg->getFallback()));
+        }
     }
-    return FormattedPlaceholder(fallback);
+    return new InternalValue(arg.orphan(),
+                             std::move(options),
+                             functionName,
+                             formatterImpl.isValid() ? formatterImpl.orphan() : nullptr,
+                             selectorImpl.isValid() ? selectorImpl.orphan() : nullptr);
 }
 
 // Formats an expression using `globalEnv` for the values of variables
-[[nodiscard]] FormattedPlaceholder MessageFormatter::formatExpression(const Environment& globalEnv,
-                                                                const Expression& expr,
-                                                                MessageContext& context,
-                                                                UErrorCode &status) const {
+[[nodiscard]] InternalValue* MessageFormatter::formatExpression(const Environment& globalEnv,
+                                                               const Expression& expr,
+                                                               MessageContext& context,
+                                                               UErrorCode &status) const {
     if (U_FAILURE(status)) {
         return {};
     }
 
     const Operand& rand = expr.getOperand();
     // Format the operand (formatOperand handles the case of a null operand)
-    FormattedPlaceholder randVal = formatOperand(globalEnv, rand, context, status);
+    LocalPointer<InternalValue> randVal(formatOperand(globalEnv, rand, context, status));
 
-    // Don't call the function on error values
-    if (randVal.isFallback()) {
-        return randVal;
-    }
+    FormattedPlaceholder maybeRand = randVal->takeArgument(status);
 
-    if (!expr.isFunctionCall()) {
+    if (!expr.isFunctionCall() && U_SUCCESS(status)) {
         // Dispatch based on type of `randVal`
-        return evalFormatterCall(std::move(randVal),
-                                 context,
-                                 status);
-    } else {
+         if (maybeRand.isFallback()) {
+            return randVal.orphan();
+        }
+        return evalFunctionCall(std::move(maybeRand), context, status);
+    } else if (expr.isFunctionCall()) {
+        status = U_ZERO_ERROR;
         const Operator* rator = expr.getOperator(status);
         U_ASSERT(U_SUCCESS(status));
         const FunctionName& functionName = rator->getFunctionName();
@@ -273,19 +257,14 @@ FunctionOptions MessageFormatter::resolveOptions(const Environment& env, const O
         FunctionOptions resolvedOptions = resolveOptions(globalEnv, options, context, status);
 
         // Call the formatter function
-        // The fallback for a nullary function call is the function name
-        UnicodeString fallback;
-        if (rand.isNull()) {
-            fallback = UnicodeString(COLON);
-            fallback += functionName;
-        } else {
-            fallback = randVal.fallback;
-        }
-        return evalFormatterCall(functionName,
-                                 std::move(randVal),
-                                 std::move(resolvedOptions),
-                                 context,
-                                 status);
+        return evalFunctionCall(functionName,
+                                randVal.orphan(),
+                                std::move(resolvedOptions),
+                                context,
+                                status);
+    } else {
+        status = U_ZERO_ERROR;
+        return randVal.orphan();
     }
 }
 
@@ -301,11 +280,13 @@ void MessageFormatter::formatPattern(MessageContext& context, const Environment&
             // Markup is ignored
         } else {
 	      // Format the expression
-	      FormattedPlaceholder partVal = formatExpression(globalEnv, part.contents(), context, status);
-	      // Force full evaluation, e.g. applying default formatters to
+              LocalPointer<InternalValue> partVal(
+                  formatExpression(globalEnv, part.contents(), context, status));
+              FormattedPlaceholder partResult = partVal->forceFormatting(context.getErrors(),
+                                                                         status);
+              // Force full evaluation, e.g. applying default formatters to
 	      // unformatted input (or formatting numbers as strings)
-              UnicodeString partResult = partVal.formatToString(locale, status);
-              result += partResult;
+              result += partResult.formatToString(locale, status);
               // Handle formatting errors. `formatToString()` can't take a context and thus can't
               // register an error directly
               if (status == U_MF_FORMATTING_ERROR) {
@@ -328,14 +309,14 @@ void MessageFormatter::resolveSelectors(MessageContext& context, const Environme
     CHECK_ERROR(status);
     U_ASSERT(!dataModel.hasPattern());
 
-    const Expression* selectors = dataModel.getSelectorsInternal();
+    const VariableName* selectors = dataModel.getSelectorsInternal();
     // 1. Let res be a new empty list of resolved values that support selection.
     // (Implicit, since `res` is an out-parameter)
     // 2. For each expression exp of the message's selectors
     for (int32_t i = 0; i < dataModel.numSelectors(); i++) {
         // 2i. Let rv be the resolved value of exp.
-        ResolvedSelector rv = formatSelectorExpression(env, selectors[i], context, status);
-        if (rv.hasSelector()) {
+        LocalPointer<InternalValue> rv(formatOperand(env, Operand(selectors[i]), context, status));
+        if (rv->canSelect()) {
             // 2ii. If selection is supported for rv:
             // (True if this code has been reached)
         } else {
@@ -344,17 +325,17 @@ void MessageFormatter::resolveSelectors(MessageContext& context, const Environme
             // Append nomatch as the last element of the list res.
             // Emit a Selection Error.
             // (Note: in this case, rv, being a fallback, serves as `nomatch`)
-            #if U_DEBUG
-            const DynamicErrors& err = context.getErrors();
-            U_ASSERT(err.hasError());
-            U_ASSERT(rv.argument().isFallback());
-            #endif
+            DynamicErrors& err = context.getErrors();
+            err.setSelectorError(rv->getFunctionName(), status);
+            rv.adoptInstead(new InternalValue(FormattedPlaceholder(rv->getFallback())));
+            if (!rv.isValid()) {
+                status = U_MEMORY_ALLOCATION_ERROR;
+                return;
+            }
         }
         // 2ii(a). Append rv as the last element of the list res.
         // (Also fulfills 2iii)
-        LocalPointer<ResolvedSelector> v(create<ResolvedSelector>(std::move(rv), status));
-        CHECK_ERROR(status);
-        res.adoptElement(v.orphan(), status);
+        res.adoptElement(rv.orphan(), status);
     }
 }
 
@@ -362,18 +343,17 @@ void MessageFormatter::resolveSelectors(MessageContext& context, const Environme
 // `keys` and `matches` are vectors of strings
 void MessageFormatter::matchSelectorKeys(const UVector& keys,
                                          MessageContext& context,
-					 ResolvedSelector&& rv,
+					 InternalValue* rv, // Does not adopt `rv`
 					 UVector& keysOut,
 					 UErrorCode& status) const {
     CHECK_ERROR(status);
 
-    if (!rv.hasSelector()) {
+    if (U_FAILURE(status)) {
         // Return an empty list of matches
+        status = U_ZERO_ERROR;
         return;
     }
 
-    auto selectorImpl = rv.getSelector();
-    U_ASSERT(selectorImpl != nullptr);
     UErrorCode savedStatus = status;
 
     // Convert `keys` to an array
@@ -400,15 +380,17 @@ void MessageFormatter::matchSelectorKeys(const UVector& keys,
     int32_t prefsLen = 0;
 
     // Call the selector
-    selectorImpl->selectKey(rv.takeArgument(), rv.takeOptions(),
-                            adoptedKeys.getAlias(), keysLen, adoptedPrefs.getAlias(), prefsLen,
-                            status);
+    FunctionName name = rv->getFunctionName();
+    rv->forceSelection(context.getErrors(),
+                       adoptedKeys.getAlias(), keysLen,
+                       adoptedPrefs.getAlias(), prefsLen,
+                       status);
 
     // Update errors
     if (savedStatus != status) {
         if (U_FAILURE(status)) {
             status = U_ZERO_ERROR;
-            context.getErrors().setSelectorError(rv.getSelectorName(), status);
+            context.getErrors().setSelectorError(name, status);
         } else {
             // Ignore warnings
             status = savedStatus;
@@ -461,8 +443,8 @@ void MessageFormatter::resolvePreferences(MessageContext& context, UVector& res,
             if (!key.isWildcard()) {
                 // 2ii(b)(a) Assert that key is a literal.
                 // (Not needed)
-                // 2ii(b)(b) Let `ks` be the resolved value of `key`.
-                ks = key.asLiteral().unquoted();
+                // 2ii(b)(b) Let `ks` be the resolved value of `key` in Unicode Normalization Form C.
+                ks = normalizeNFC(key.asLiteral().unquoted());
                 // 2ii(b)(c) Append `ks` as the last element of the list `keys`.
                 ksP.adoptInstead(create<UnicodeString>(std::move(ks), status));
                 CHECK_ERROR(status);
@@ -471,7 +453,7 @@ void MessageFormatter::resolvePreferences(MessageContext& context, UVector& res,
         }
         // 2iii. Let `rv` be the resolved value at index `i` of `res`.
         U_ASSERT(i < res.size());
-        ResolvedSelector rv = std::move(*(static_cast<ResolvedSelector*>(res[i])));
+        InternalValue* rv = static_cast<InternalValue*>(res[i]);
         // 2iv. Let matches be the result of calling the method MatchSelectorKeys(rv, keys)
         LocalPointer<UVector> matches(createUVector(status));
         matchSelectorKeys(*keys, context, std::move(rv), *matches, status);
@@ -523,7 +505,7 @@ void MessageFormatter::filterVariants(const UVector& pref, UVector& vars, UError
             // 2i(c). Assert that `key` is a literal.
             // (Not needed)
             // 2i(d). Let `ks` be the resolved value of `key`.
-            UnicodeString ks = key.asLiteral().unquoted();
+            UnicodeString ks = normalizeNFC(key.asLiteral().unquoted());
             // 2i(e). Let `matches` be the list of strings at index `i` of `pref`.
             const UVector& matches = *(static_cast<UVector*>(pref[i])); // `matches` is a vector of strings
             // 2i(f). If `matches` includes `ks`
@@ -585,7 +567,7 @@ void MessageFormatter::sortVariants(const UVector& pref, UVector& vars, UErrorCo
                 // 5iii(c)(a). Assert that `key` is a literal.
                 // (Not needed)
                 // 5iii(c)(b). Let `ks` be the resolved value of `key`.
-                UnicodeString ks = key.asLiteral().unquoted();
+                UnicodeString ks = normalizeNFC(key.asLiteral().unquoted());
                 // 5iii(c)(c) Let matchpref be the integer position of ks in `matches`.
                 matchpref = vectorFind(matches, ks);
                 U_ASSERT(matchpref >= 0);
@@ -604,123 +586,13 @@ void MessageFormatter::sortVariants(const UVector& pref, UVector& vars, UErrorCo
     // 7. Select the pattern of `var`
 }
 
-
-// Evaluate the operand
-ResolvedSelector MessageFormatter::resolveVariables(const Environment& env, const Operand& rand, MessageContext& context, UErrorCode &status) const {
-    if (U_FAILURE(status)) {
-        return {};
-    }
-
-    if (rand.isNull()) {
-        return ResolvedSelector(FormattedPlaceholder());
-    }
-
-    if (rand.isLiteral()) {
-        return ResolvedSelector(formatLiteral(rand.asLiteral()));
-    }
-
-    // Must be variable
-    const VariableName& var = rand.asVariable();
-    // Resolve the variable
-    if (env.has(var)) {
-        const Closure& referent = env.lookup(var);
-        // Resolve the referent
-        return resolveVariables(referent.getEnv(), referent.getExpr(), context, status);
-    }
-    // Either this is a global var or an unbound var --
-    // either way, it can't be bound to a function call.
-    // Check globals
-    FormattedPlaceholder val = evalArgument(var, context, status);
-    if (status == U_ILLEGAL_ARGUMENT_ERROR) {
-        status = U_ZERO_ERROR;
-        // Unresolved variable -- could be a previous warning. Nothing to resolve
-        U_ASSERT(context.getErrors().hasUnresolvedVariableError());
-        return ResolvedSelector(FormattedPlaceholder(var));
-    }
-    // Pass through other errors
-    return ResolvedSelector(std::move(val));
-}
-
-// Evaluate the expression except for not performing the top-level function call
-// (which is expected to be a selector, but may not be, in error cases)
-ResolvedSelector MessageFormatter::resolveVariables(const Environment& env,
-                                                    const Expression& expr,
-                                                    MessageContext& context,
-                                                    UErrorCode &status) const {
-    if (U_FAILURE(status)) {
-        return {};
-    }
-
-    // Function call -- resolve the operand and options
-    if (expr.isFunctionCall()) {
-        const Operator* rator = expr.getOperator(status);
-        U_ASSERT(U_SUCCESS(status));
-        // Already checked that rator is non-reserved
-        const FunctionName& selectorName = rator->getFunctionName();
-        if (isSelector(selectorName)) {
-            auto selector = getSelector(context, selectorName, status);
-            if (U_SUCCESS(status)) {
-                FunctionOptions resolvedOptions = resolveOptions(env, rator->getOptionsInternal(), context, status);
-                // Operand may be the null argument, but resolveVariables() handles that
-                FormattedPlaceholder argument = formatOperand(env, expr.getOperand(), context, status);
-                return ResolvedSelector(selectorName, selector, std::move(resolvedOptions), std::move(argument));
-            }
-        } else if (isFormatter(selectorName)) {
-            context.getErrors().setSelectorError(selectorName, status);
-        } else {
-            context.getErrors().setUnknownFunction(selectorName, status);
-        }
-        // Non-selector used as selector; an error would have been recorded earlier
-        UnicodeString fallback(COLON);
-        fallback += selectorName;
-        if (!expr.getOperand().isNull()) {
-            fallback = formatOperand(env, expr.getOperand(), context, status).fallback;
-        }
-        return ResolvedSelector(FormattedPlaceholder(fallback));
-    } else {
-        // Might be a variable reference, so expand one more level of variable
-        return resolveVariables(env, expr.getOperand(), context, status);
-    }
-}
-
-ResolvedSelector MessageFormatter::formatSelectorExpression(const Environment& globalEnv, const Expression& expr, MessageContext& context, UErrorCode &status) const {
-    if (U_FAILURE(status)) {
-        return {};
-    }
-
-    // Resolve expression to determine if it's a function call
-    ResolvedSelector exprResult = resolveVariables(globalEnv, expr, context, status);
-
-    DynamicErrors& err = context.getErrors();
-
-    // If there is a selector, then `resolveVariables()` recorded it in the context
-    if (exprResult.hasSelector()) {
-        // Check if there was an error
-        if (exprResult.argument().isFallback()) {
-            // Use a null expression if it's a syntax or data model warning;
-            // create a valid (non-fallback) formatted placeholder from the
-            // fallback string otherwise
-            if (err.hasSyntaxError() || err.hasDataModelError()) {
-                return ResolvedSelector(FormattedPlaceholder()); // Null operand
-            } else {
-                return ResolvedSelector(exprResult.takeArgument());
-            }
-        }
-        return exprResult;
-    }
-
-    // No selector was found; error should already have been set
-    U_ASSERT(err.hasMissingSelectorAnnotationError() || err.hasUnknownFunctionError() || err.hasSelectorError());
-    return ResolvedSelector(FormattedPlaceholder(exprResult.argument().fallback));
-}
-
 void MessageFormatter::formatSelectors(MessageContext& context, const Environment& env, UErrorCode &status, UnicodeString& result) const {
     CHECK_ERROR(status);
 
     // See https://github.com/unicode-org/message-format-wg/blob/main/spec/formatting.md#pattern-selection
 
     // Resolve Selectors
-    // res is a vector of FormattedPlaceholders
+    // res is a vector of InternalValues
     LocalPointer<UVector> res(createUVector(status));
     CHECK_ERROR(status);
     resolveSelectors(context, env, status, *res);
@@ -761,28 +633,35 @@ void MessageFormatter::formatSelectors(MessageContext& context, const Environmen
 UnicodeString MessageFormatter::formatToString(const MessageArguments& arguments, UErrorCode &status) {
     EMPTY_ON_ERROR(status);
 
-    // Create a new environment that will store closures for all local variables
-    Environment* env = Environment::create(status);
     // Create a new context with the given arguments and the `errors` structure
     MessageContext context(arguments, *errors, status);
-
-    // Check for unresolved variable errors
-    checkDeclarations(context, env, status);
-    LocalPointer<Environment> globalEnv(env);
-
     UnicodeString result;
-    if (dataModel.hasPattern()) {
-        formatPattern(context, *globalEnv, dataModel.getPattern(), status, result);
-    } else {
-        // Check for errors/warnings -- if so, then the result of pattern selection is the fallback value
-        // See https://github.com/unicode-org/message-format-wg/blob/main/spec/formatting.md#pattern-selection
-        const DynamicErrors& err = context.getErrors();
-        if (err.hasSyntaxError() || err.hasDataModelError()) {
-            result += REPLACEMENT;
+
+    if (!(errors->hasSyntaxError() || errors->hasDataModelError())) {
+        // Create a new environment that will store closures for all local variables
+        // Check for unresolved variable errors
+        // checkDeclarations needs a reference to the pointer to the environment
+        // since it uses its `env` argument as an out-parameter. So it needs to be
+        // temporarily not a LocalPointer...
+        Environment* env(Environment::create(status));
+        checkDeclarations(context, env, status);
+        // ...and then it's adopted to avoid leaks
+        LocalPointer<Environment> globalEnv(env);
+
+        if (dataModel.hasPattern()) {
+            formatPattern(context, *globalEnv, dataModel.getPattern(), status, result);
         } else {
-            formatSelectors(context, *globalEnv, status, result);
+            // Check for errors/warnings -- if so, then the result of pattern selection is the fallback value
+            // See https://github.com/unicode-org/message-format-wg/blob/main/spec/formatting.md#pattern-selection
+            const DynamicErrors& err = context.getErrors();
+            if (err.hasSyntaxError() || err.hasDataModelError()) {
+                result += REPLACEMENT;
+            } else {
+                formatSelectors(context, *globalEnv, status, result);
+            }
         }
     }
+
     // Update status according to all errors seen while formatting
     if (signalErrors) {
         context.checkErrors(status);
@@ -813,12 +692,14 @@ void MessageFormatter::check(MessageContext& context, const Environment& localEn
 
     // Check that variable is in scope
     const VariableName& var = rand.asVariable();
+    UnicodeString normalized = normalizeNFC(var);
+
     // Check local scope
-    if (localEnv.has(var)) {
+    if (localEnv.has(normalized)) {
         return;
     }
     // Check global scope
-    context.getGlobal(var, status);
+    context.getGlobal(*this, normalized, status);
     if (status == U_ILLEGAL_ARGUMENT_ERROR) {
         status = U_ZERO_ERROR;
         context.getErrors().setUnresolvedVariable(var, status);
@@ -855,7 +736,10 @@ void MessageFormatter::checkDeclarations(MessageContext& context, Environment*&
         // memoizing the value of localEnv up to this point
 
         // Add the LHS to the environment for checking the next declaration
-        env = Environment::create(decl.getVariable(), Closure(rhs, *env), env, status);
+        env = Environment::create(normalizeNFC(decl.getVariable()),
+                                  Closure(rhs, *env),
+                                  env,
+                                  status);
         CHECK_ERROR(status);
     }
 }
@@ -866,3 +750,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_allocation.h b/deps/icu-small/source/i18n/messageformat2_allocation.h
index 7be27e222520d6..5b06d0851296a1 100644
--- a/deps/icu-small/source/i18n/messageformat2_allocation.h
+++ b/deps/icu-small/source/i18n/messageformat2_allocation.h
@@ -10,6 +10,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -139,6 +141,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_UTILS_H
diff --git a/deps/icu-small/source/i18n/messageformat2_arguments.cpp b/deps/icu-small/source/i18n/messageformat2_arguments.cpp
index ded3f4dda160c3..c43c600a2f402c 100644
--- a/deps/icu-small/source/i18n/messageformat2_arguments.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_arguments.cpp
@@ -3,12 +3,16 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
 
+#include "unicode/messageformat2.h"
 #include "unicode/messageformat2_arguments.h"
 #include "unicode/messageformat2_data_model_names.h"
+#include "messageformat2_evaluation.h"
 #include "uvector.h" // U_ASSERT
 
 U_NAMESPACE_BEGIN
@@ -22,11 +26,15 @@ namespace message2 {
 
     using Arguments = MessageArguments;
 
-    const Formattable* Arguments::getArgument(const VariableName& arg, UErrorCode& errorCode) const {
+    const Formattable* Arguments::getArgument(const MessageFormatter& context,
+                                              const VariableName& arg,
+                                              UErrorCode& errorCode) const {
         if (U_SUCCESS(errorCode)) {
             U_ASSERT(argsLen == 0 || arguments.isValid());
             for (int32_t i = 0; i < argsLen; i++) {
-                if (argumentNames[i] == arg) {
+                UnicodeString normalized = context.normalizeNFC(argumentNames[i]);
+                // arg already assumed to be normalized
+                if (normalized == arg) {
                     return &arguments[i];
                 }
             }
@@ -57,3 +65,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_checker.cpp b/deps/icu-small/source/i18n/messageformat2_checker.cpp
index bdc5c383b6e89f..46b25ff389b5dc 100644
--- a/deps/icu-small/source/i18n/messageformat2_checker.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_checker.cpp
@@ -3,12 +3,16 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
 
+#include "unicode/messageformat2.h"
 #include "messageformat2_allocation.h"
 #include "messageformat2_checker.h"
+#include "messageformat2_evaluation.h"
 #include "messageformat2_macros.h"
 #include "uvector.h" // U_ASSERT
 
@@ -104,6 +108,14 @@ TypeEnvironment::~TypeEnvironment() {}
 
 // ---------------------
 
+Key Checker::normalizeNFC(const Key& k) const {
+    if (k.isWildcard()) {
+        return k;
+    }
+    return Key(Literal(k.asLiteral().isQuoted(),
+                       context.normalizeNFC(k.asLiteral().unquoted())));
+}
+
 static bool areDefaultKeys(const Key* keys, int32_t len) {
     U_ASSERT(len > 0);
     for (int32_t i = 0; i < len; i++) {
@@ -185,7 +197,7 @@ void Checker::checkVariants(UErrorCode& status) {
                 // This variant was already checked,
                 // so we know keys1.len == len
                 for (int32_t kk = 0; kk < len; kk++) {
-                    if (!(keys[kk] == keys1[kk])) {
+                    if (!(normalizeNFC(keys[kk]) == normalizeNFC(keys1[kk]))) {
                         allEqual = false;
                         break;
                     }
@@ -205,18 +217,14 @@ void Checker::checkVariants(UErrorCode& status) {
     }
 }
 
-void Checker::requireAnnotated(const TypeEnvironment& t, const Expression& selectorExpr, UErrorCode& status) {
+void Checker::requireAnnotated(const TypeEnvironment& t,
+                               const VariableName& selectorVar,
+                               UErrorCode& status) {
     CHECK_ERROR(status);
 
-    if (selectorExpr.isFunctionCall()) {
+    if (t.get(selectorVar) == TypeEnvironment::Type::Annotated) {
         return; // No error
     }
-    const Operand& rand = selectorExpr.getOperand();
-    if (rand.isVariable()) {
-        if (t.get(rand.asVariable()) == TypeEnvironment::Type::Annotated) {
-            return; // No error
-        }
-    }
     // If this code is reached, an error was detected
     errors.addError(StaticErrorType::MissingSelectorAnnotation, status);
 }
@@ -226,7 +234,7 @@ void Checker::checkSelectors(const TypeEnvironment& t, UErrorCode& status) {
 
     // Check each selector; if it's not annotated, emit a
     // "missing selector annotation" error
-    const Expression* selectors = dataModel.getSelectorsInternal();
+    const VariableName* selectors = dataModel.getSelectorsInternal();
     for (int32_t i = 0; i < dataModel.numSelectors(); i++) {
         requireAnnotated(t, selectors[i], status);
     }
@@ -312,3 +320,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_checker.h b/deps/icu-small/source/i18n/messageformat2_checker.h
index 4bb0498efb9940..122f668f4b7e9b 100644
--- a/deps/icu-small/source/i18n/messageformat2_checker.h
+++ b/deps/icu-small/source/i18n/messageformat2_checker.h
@@ -10,6 +10,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -56,15 +58,20 @@ namespace message2 {
                                         // an explicit declaration
     }; // class TypeEnvironment
 
+    class MessageFormatter;
+
     // Checks a data model for semantic errors
     // (Errors are defined in https://github.com/unicode-org/message-format-wg/blob/main/spec/formatting.md       )
     class Checker {
     public:
         void check(UErrorCode&);
-        Checker(const MFDataModel& m, StaticErrors& e) : dataModel(m), errors(e) {}
+        Checker(const MFDataModel& d, StaticErrors& e, const MessageFormatter& mf)
+            : dataModel(d), errors(e), context(mf) {}
     private:
 
-        void requireAnnotated(const TypeEnvironment&, const Expression&, UErrorCode&);
+        Key normalizeNFC(const Key&) const;
+
+        void requireAnnotated(const TypeEnvironment&, const VariableName&, UErrorCode&);
         void addFreeVars(TypeEnvironment& t, const Operand&, UErrorCode&);
         void addFreeVars(TypeEnvironment& t, const Operator&, UErrorCode&);
         void addFreeVars(TypeEnvironment& t, const OptionMap&, UErrorCode&);
@@ -78,6 +85,9 @@ namespace message2 {
         void check(const Pattern&);
         const MFDataModel& dataModel;
         StaticErrors& errors;
+
+        // Used for NFC normalization
+        const MessageFormatter& context;
     }; // class Checker
 
 } // namespace message2
@@ -88,6 +98,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT_CHECKER_H
diff --git a/deps/icu-small/source/i18n/messageformat2_data_model.cpp b/deps/icu-small/source/i18n/messageformat2_data_model.cpp
index 3fe5f65b532364..3406080695c6f9 100644
--- a/deps/icu-small/source/i18n/messageformat2_data_model.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_data_model.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -691,9 +693,9 @@ Matcher::Matcher(const Matcher& other) {
     numSelectors = other.numSelectors;
     numVariants = other.numVariants;
     UErrorCode localErrorCode = U_ZERO_ERROR;
-    selectors.adoptInstead(copyArray<Expression>(other.selectors.getAlias(),
-                                                 numSelectors,
-                                                 localErrorCode));
+    selectors.adoptInstead(copyArray<VariableName>(other.selectors.getAlias(),
+                                                   numSelectors,
+                                                   localErrorCode));
     variants.adoptInstead(copyArray<Variant>(other.variants.getAlias(),
                                              numVariants,
                                              localErrorCode));
@@ -702,7 +704,7 @@ Matcher::Matcher(const Matcher& other) {
     }
 }
 
-Matcher::Matcher(Expression* ss, int32_t ns, Variant* vs, int32_t nv)
+Matcher::Matcher(VariableName* ss, int32_t ns, Variant* vs, int32_t nv)
     : selectors(ss), numSelectors(ns), variants(vs), numVariants(nv) {}
 
 Matcher::~Matcher() {}
@@ -724,7 +726,7 @@ const Binding* MFDataModel::getLocalVariablesInternal() const {
     return bindings.getAlias();
 }
 
-const Expression* MFDataModel::getSelectorsInternal() const {
+const VariableName* MFDataModel::getSelectorsInternal() const {
     U_ASSERT(!bogus);
     U_ASSERT(!hasPattern());
     return std::get_if<Matcher>(&body)->selectors.getAlias();
@@ -786,15 +788,13 @@ MFDataModel::Builder& MFDataModel::Builder::addBinding(Binding&& b, UErrorCode&
     return *this;
 }
 
-/*
-  selector must be non-null
-*/
-MFDataModel::Builder& MFDataModel::Builder::addSelector(Expression&& selector, UErrorCode& status) noexcept {
+MFDataModel::Builder& MFDataModel::Builder::addSelector(VariableName&& selector,
+                                                        UErrorCode& status) {
     THIS_ON_ERROR(status);
 
     buildSelectorsMessage(status);
     U_ASSERT(selectors != nullptr);
-    selectors->adoptElement(create<Expression>(std::move(selector), status), status);
+    selectors->adoptElement(create<VariableName>(std::move(selector), status), status);
 
     return *this;
 }
@@ -830,11 +830,11 @@ MFDataModel::MFDataModel(const MFDataModel& other) : body(Pattern()) {
     if (other.hasPattern()) {
         body = *std::get_if<Pattern>(&other.body);
     } else {
-        const Expression* otherSelectors = other.getSelectorsInternal();
+        const VariableName* otherSelectors = other.getSelectorsInternal();
         const Variant* otherVariants = other.getVariantsInternal();
         int32_t numSelectors = other.numSelectors();
         int32_t numVariants = other.numVariants();
-        Expression* copiedSelectors = copyArray(otherSelectors, numSelectors, localErrorCode);
+        VariableName* copiedSelectors = copyArray(otherSelectors, numSelectors, localErrorCode);
         Variant* copiedVariants = copyArray(otherVariants, numVariants, localErrorCode);
         if (U_FAILURE(localErrorCode)) {
             bogus = true;
@@ -863,7 +863,9 @@ MFDataModel::MFDataModel(const MFDataModel::Builder& builder, UErrorCode& errorC
         int32_t numVariants = builder.variants->size();
         int32_t numSelectors = builder.selectors->size();
         LocalArray<Variant> variants(copyVectorToArray<Variant>(*builder.variants, errorCode), errorCode);
-        LocalArray<Expression> selectors(copyVectorToArray<Expression>(*builder.selectors, errorCode), errorCode);
+        LocalArray<VariableName> selectors(copyVectorToArray<VariableName>(*builder.selectors,
+                                                                           errorCode),
+                                           errorCode);
         if (U_FAILURE(errorCode)) {
             bogus = true;
             return;
@@ -918,3 +920,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_errors.cpp b/deps/icu-small/source/i18n/messageformat2_errors.cpp
index 9d1d6bab81a1f7..5d3d938f02030e 100644
--- a/deps/icu-small/source/i18n/messageformat2_errors.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_errors.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -290,3 +292,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_errors.h b/deps/icu-small/source/i18n/messageformat2_errors.h
index f84aa736283786..085263e88b068d 100644
--- a/deps/icu-small/source/i18n/messageformat2_errors.h
+++ b/deps/icu-small/source/i18n/messageformat2_errors.h
@@ -15,6 +15,8 @@
  * \brief C++ API: Formats messages using the draft MessageFormat 2.0.
  */
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -151,6 +153,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_ERRORS_H
diff --git a/deps/icu-small/source/i18n/messageformat2_evaluation.cpp b/deps/icu-small/source/i18n/messageformat2_evaluation.cpp
index 41e4c9a8020a04..fcccbf5ae5e781 100644
--- a/deps/icu-small/source/i18n/messageformat2_evaluation.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_evaluation.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -89,34 +91,53 @@ FunctionOptions::FunctionOptions(FunctionOptions&& other) {
 FunctionOptions::~FunctionOptions() {
     if (options != nullptr) {
         delete[] options;
+        options = nullptr;
     }
 }
-// ResolvedSelector
-// ----------------
-
-ResolvedSelector::ResolvedSelector(const FunctionName& fn,
-                                   Selector* sel,
-                                   FunctionOptions&& opts,
-                                   FormattedPlaceholder&& val)
-    : selectorName(fn), selector(sel), options(std::move(opts)), value(std::move(val))  {
-    U_ASSERT(sel != nullptr);
+
+static bool containsOption(const UVector& opts, const ResolvedFunctionOption& opt) {
+    for (int32_t i = 0; i < opts.size(); i++) {
+        if (static_cast<ResolvedFunctionOption*>(opts[i])->getName()
+            == opt.getName()) {
+            return true;
+        }
+    }
+    return false;
 }
 
-ResolvedSelector::ResolvedSelector(FormattedPlaceholder&& val) : value(std::move(val)) {}
+// Options in `this` take precedence
+// `this` can't be used after mergeOptions is called
+FunctionOptions FunctionOptions::mergeOptions(FunctionOptions&& other,
+                                              UErrorCode& status) {
+    UVector mergedOptions(status);
+    mergedOptions.setDeleter(uprv_deleteUObject);
 
-ResolvedSelector& ResolvedSelector::operator=(ResolvedSelector&& other) noexcept {
-    selectorName = std::move(other.selectorName);
-    selector.adoptInstead(other.selector.orphan());
-    options = std::move(other.options);
-    value = std::move(other.value);
-    return *this;
-}
+    if (U_FAILURE(status)) {
+        return {};
+    }
 
-ResolvedSelector::ResolvedSelector(ResolvedSelector&& other) {
-    *this = std::move(other);
-}
+    // Create a new vector consisting of the options from this `FunctionOptions`
+    for (int32_t i = 0; i < functionOptionsLen; i++) {
+        mergedOptions.adoptElement(create<ResolvedFunctionOption>(std::move(options[i]), status),
+                                 status);
+    }
 
-ResolvedSelector::~ResolvedSelector() {}
+    // Add each option from `other` that doesn't appear in this `FunctionOptions`
+    for (int i = 0; i < other.functionOptionsLen; i++) {
+        // Note: this is quadratic in the length of `options`
+        if (!containsOption(mergedOptions, other.options[i])) {
+            mergedOptions.adoptElement(create<ResolvedFunctionOption>(std::move(other.options[i]),
+                                                                    status),
+                                     status);
+        }
+    }
+
+    delete[] options;
+    options = nullptr;
+    functionOptionsLen = 0;
+
+    return FunctionOptions(std::move(mergedOptions), status);
+}
 
 // PrioritizedVariant
 // ------------------
@@ -190,18 +211,210 @@ PrioritizedVariant::~PrioritizedVariant() {}
         errors.checkErrors(status);
     }
 
-    const Formattable* MessageContext::getGlobal(const VariableName& v, UErrorCode& errorCode) const {
-       return arguments.getArgument(v, errorCode);
+    const Formattable* MessageContext::getGlobal(const MessageFormatter& context,
+                                                 const VariableName& v,
+                                                 UErrorCode& errorCode) const {
+       return arguments.getArgument(context, v, errorCode);
     }
 
     MessageContext::MessageContext(const MessageArguments& args,
                                    const StaticErrors& e,
                                    UErrorCode& status) : arguments(args), errors(e, status) {}
+
     MessageContext::~MessageContext() {}
 
+    // InternalValue
+    // -------------
+
+    bool InternalValue::isFallback() const {
+        return std::holds_alternative<FormattedPlaceholder>(argument)
+            && std::get_if<FormattedPlaceholder>(&argument)->isFallback();
+    }
+
+    bool InternalValue::hasNullOperand() const {
+        return std::holds_alternative<FormattedPlaceholder>(argument)
+            && std::get_if<FormattedPlaceholder>(&argument)->isNullOperand();
+    }
+
+    FormattedPlaceholder InternalValue::takeArgument(UErrorCode& errorCode) {
+        if (U_FAILURE(errorCode)) {
+            return {};
+        }
+
+        if (std::holds_alternative<FormattedPlaceholder>(argument)) {
+            return std::move(*std::get_if<FormattedPlaceholder>(&argument));
+        }
+        errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+        return {};
+    }
+
+    const UnicodeString& InternalValue::getFallback() const {
+        if (std::holds_alternative<FormattedPlaceholder>(argument)) {
+            return std::get_if<FormattedPlaceholder>(&argument)->getFallback();
+        }
+        return (*std::get_if<InternalValue*>(&argument))->getFallback();
+    }
+
+    const Selector* InternalValue::getSelector(UErrorCode& errorCode) const {
+        if (U_FAILURE(errorCode)) {
+            return nullptr;
+        }
+
+        if (selector == nullptr) {
+            errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+        }
+        return selector;
+    }
+
+    InternalValue::InternalValue(FormattedPlaceholder&& arg) {
+        argument = std::move(arg);
+        selector = nullptr;
+        formatter = nullptr;
+    }
+
+    InternalValue::InternalValue(InternalValue* operand,
+                                 FunctionOptions&& opts,
+                                 const FunctionName& functionName,
+                                 const Formatter* f,
+                                 const Selector* s) {
+        argument = operand;
+        options = std::move(opts);
+        name = functionName;
+        selector = s;
+        formatter = f;
+        U_ASSERT(selector != nullptr || formatter != nullptr);
+    }
+
+    // `this` cannot be used after calling this method
+    void InternalValue::forceSelection(DynamicErrors& errs,
+                                       const UnicodeString* keys,
+                                       int32_t keysLen,
+                                       UnicodeString* prefs,
+                                       int32_t& prefsLen,
+                                       UErrorCode& errorCode) {
+        if (U_FAILURE(errorCode)) {
+            return;
+        }
+
+        if (!canSelect()) {
+            errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+            return;
+        }
+        // Find the argument and complete set of options by traversing `argument`
+        FunctionOptions opts;
+        InternalValue* p = this;
+        FunctionName selectorName = name;
+        while (std::holds_alternative<InternalValue*>(p->argument)) {
+            if (p->name != selectorName) {
+                // Can only compose calls to the same selector
+                errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+                return;
+            }
+            // First argument to mergeOptions takes precedence
+            opts = opts.mergeOptions(std::move(p->options), errorCode);
+            if (U_FAILURE(errorCode)) {
+                return;
+            }
+            InternalValue* next = *std::get_if<InternalValue*>(&p->argument);
+            p = next;
+        }
+        FormattedPlaceholder arg = std::move(*std::get_if<FormattedPlaceholder>(&p->argument));
+
+        selector->selectKey(std::move(arg), std::move(opts),
+                            keys, keysLen,
+                            prefs, prefsLen, errorCode);
+        if (U_FAILURE(errorCode)) {
+            errorCode = U_ZERO_ERROR;
+            errs.setSelectorError(selectorName, errorCode);
+        }
+    }
+
+    FormattedPlaceholder InternalValue::forceFormatting(DynamicErrors& errs, UErrorCode& errorCode) {
+        if (U_FAILURE(errorCode)) {
+            return {};
+        }
+
+        if (formatter == nullptr && selector == nullptr) {
+            U_ASSERT(std::holds_alternative<FormattedPlaceholder>(argument));
+            return std::move(*std::get_if<FormattedPlaceholder>(&argument));
+        }
+        if (formatter == nullptr) {
+            errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+            return {};
+        }
+
+        FormattedPlaceholder arg;
+
+        if (std::holds_alternative<FormattedPlaceholder>(argument)) {
+            arg = std::move(*std::get_if<FormattedPlaceholder>(&argument));
+        } else {
+            arg = (*std::get_if<InternalValue*>(&argument))->forceFormatting(errs,
+                                                                             errorCode);
+        }
+
+        if (U_FAILURE(errorCode)) {
+            return {};
+        }
+
+        // The fallback for a nullary function call is the function name
+        UnicodeString fallback;
+        if (arg.isNullOperand()) {
+            fallback = u":";
+            fallback += name;
+        } else {
+            fallback = arg.getFallback();
+        }
+
+        // Call the function with the argument
+        FormattedPlaceholder result = formatter->format(std::move(arg), std::move(options), errorCode);
+        if (U_FAILURE(errorCode)) {
+            if (errorCode == U_MF_OPERAND_MISMATCH_ERROR) {
+                errorCode = U_ZERO_ERROR;
+                errs.setOperandMismatchError(name, errorCode);
+            } else {
+                errorCode = U_ZERO_ERROR;
+                // Convey any error generated by the formatter
+                // as a formatting error, except for operand mismatch errors
+                errs.setFormattingError(name, errorCode);
+            }
+        }
+        // Ignore the output if any error occurred
+        if (errs.hasFormattingError()) {
+            return FormattedPlaceholder(fallback);
+        }
+
+        return result;
+    }
+
+    InternalValue& InternalValue::operator=(InternalValue&& other) noexcept {
+        argument = std::move(other.argument);
+        other.argument = nullptr;
+        options = std::move(other.options);
+        name = other.name;
+        selector = other.selector;
+        formatter = other.formatter;
+        other.selector = nullptr;
+        other.formatter = nullptr;
+
+        return *this;
+    }
+
+    InternalValue::~InternalValue() {
+        delete selector;
+        selector = nullptr;
+        delete formatter;
+        formatter = nullptr;
+        if (std::holds_alternative<InternalValue*>(argument)) {
+            delete *std::get_if<InternalValue*>(&argument);
+            argument = nullptr;
+        }
+    }
+
 } // namespace message2
 U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_evaluation.h b/deps/icu-small/source/i18n/messageformat2_evaluation.h
index b8ae0242367df6..fcb30bc3e6388c 100644
--- a/deps/icu-small/source/i18n/messageformat2_evaluation.h
+++ b/deps/icu-small/source/i18n/messageformat2_evaluation.h
@@ -14,6 +14,7 @@
  * \file
  * \brief C++ API: Formats messages using the draft MessageFormat 2.0.
  */
+#if !UCONFIG_NO_NORMALIZATION
 
 #if !UCONFIG_NO_FORMATTING
 
@@ -63,38 +64,6 @@ namespace message2 {
         return 1;
     }
 
-    // Encapsulates a value to be scrutinized by a `match` with its resolved
-    // options and the name of the selector
-    class ResolvedSelector : public UObject {
-    public:
-        ResolvedSelector() {}
-        ResolvedSelector(const FunctionName& fn,
-                         Selector* selector,
-                         FunctionOptions&& options,
-                         FormattedPlaceholder&& value);
-        // Used either for errors, or when selector isn't yet known
-        explicit ResolvedSelector(FormattedPlaceholder&& value);
-        bool hasSelector() const { return selector.isValid(); }
-        const FormattedPlaceholder& argument() const { return value; }
-        FormattedPlaceholder&& takeArgument() { return std::move(value); }
-        const Selector* getSelector() {
-            U_ASSERT(selector.isValid());
-            return selector.getAlias();
-        }
-        FunctionOptions&& takeOptions() {
-            return std::move(options);
-        }
-        const FunctionName& getSelectorName() const { return selectorName; }
-        virtual ~ResolvedSelector();
-        ResolvedSelector& operator=(ResolvedSelector&&) noexcept;
-        ResolvedSelector(ResolvedSelector&&);
-    private:
-        FunctionName selectorName; // For error reporting
-        LocalPointer<Selector> selector;
-        FunctionOptions options;
-        FormattedPlaceholder value;
-    }; // class ResolvedSelector
-
     // Closures and environments
     // -------------------------
 
@@ -174,11 +143,15 @@ namespace message2 {
     // The context contains all the information needed to process
     // an entire message: arguments, formatter cache, and error list
 
+    class MessageFormatter;
+
     class MessageContext : public UMemory {
     public:
         MessageContext(const MessageArguments&, const StaticErrors&, UErrorCode&);
 
-        const Formattable* getGlobal(const VariableName&, UErrorCode&) const;
+        const Formattable* getGlobal(const MessageFormatter&,
+                                     const VariableName&,
+                                     UErrorCode&) const;
 
         // If any errors were set, update `status` accordingly
         void checkErrors(UErrorCode& status) const;
@@ -191,8 +164,47 @@ namespace message2 {
         const MessageArguments& arguments; // External message arguments
         // Errors accumulated during parsing/formatting
         DynamicErrors errors;
+
     }; // class MessageContext
 
+    // InternalValue
+    // ----------------
+
+    class InternalValue : public UObject {
+    public:
+        const FunctionName& getFunctionName() const { return name; }
+        bool canSelect() const { return selector != nullptr; }
+        const Selector* getSelector(UErrorCode&) const;
+        FormattedPlaceholder forceFormatting(DynamicErrors& errs,
+                                             UErrorCode& errorCode);
+        void forceSelection(DynamicErrors& errs,
+                            const UnicodeString* keys,
+                            int32_t keysLen,
+                            UnicodeString* prefs,
+                            int32_t& prefsLen,
+                            UErrorCode& errorCode);
+        // Needs to be deep-copyable and movable
+        virtual ~InternalValue();
+        InternalValue(FormattedPlaceholder&&);
+        // Formatter and selector may be null
+        InternalValue(InternalValue*, FunctionOptions&&, const FunctionName&, const Formatter*,
+                      const Selector*);
+        const UnicodeString& getFallback() const;
+        bool isFallback() const;
+        bool hasNullOperand() const;
+        // Can't be used anymore after calling this
+        FormattedPlaceholder takeArgument(UErrorCode& errorCode);
+        InternalValue(InternalValue&& other) { *this = std::move(other); }
+        InternalValue& operator=(InternalValue&& other) noexcept;
+    private:
+        // InternalValue is owned (if present)
+        std::variant<InternalValue*, FormattedPlaceholder> argument;
+        FunctionOptions options;
+        FunctionName name;
+        const Selector* selector; // May be null
+        const Formatter* formatter; // May be null, but one or the other should be non-null unless argument is a FormattedPlaceholder
+    }; // class InternalValue
+
 } // namespace message2
 
 U_NAMESPACE_END
@@ -201,6 +213,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_EVALUATION_H
diff --git a/deps/icu-small/source/i18n/messageformat2_formattable.cpp b/deps/icu-small/source/i18n/messageformat2_formattable.cpp
index 3152ccb44fd8b7..4e2df49aeccf01 100644
--- a/deps/icu-small/source/i18n/messageformat2_formattable.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_formattable.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -336,3 +338,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_formatter.cpp b/deps/icu-small/source/i18n/messageformat2_formatter.cpp
index 8d17ae49b99a9a..ead6f62a7899e7 100644
--- a/deps/icu-small/source/i18n/messageformat2_formatter.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_formatter.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -43,7 +45,8 @@ namespace message2 {
 
         // Parse the pattern
         MFDataModel::Builder tree(errorCode);
-        Parser(pat, tree, *errors, normalizedInput).parse(parseError, errorCode);
+        Parser(pat, tree, *errors, normalizedInput, errorCode)
+            .parse(parseError, errorCode);
 
         // Fail on syntax errors
         if (errors->hasSyntaxError()) {
@@ -116,6 +119,24 @@ namespace message2 {
 
     // MessageFormatter
 
+    // Returns the NFC-normalized version of s, returning s itself
+    // if it's already normalized.
+    UnicodeString MessageFormatter::normalizeNFC(const UnicodeString& s) const {
+        UErrorCode status = U_ZERO_ERROR;
+        // Check if string is already normalized
+        UNormalizationCheckResult result = nfcNormalizer->quickCheck(s, status);
+        // If so, return it
+        if (U_SUCCESS(status) && result == UNORM_YES) {
+            return s;
+        }
+        // Otherwise, normalize it
+        UnicodeString normalized = nfcNormalizer->normalize(s, status);
+        if (U_FAILURE(status)) {
+            return {};
+        }
+        return normalized;
+    }
+
     MessageFormatter::MessageFormatter(const MessageFormatter::Builder& builder, UErrorCode &success) : locale(builder.locale), customMFFunctionRegistry(builder.customMFFunctionRegistry) {
         CHECK_ERROR(success);
 
@@ -132,9 +153,13 @@ namespace message2 {
             .adoptFormatter(FunctionName(UnicodeString("time")), time, success)
             .adoptFormatter(FunctionName(UnicodeString("number")), number, success)
             .adoptFormatter(FunctionName(UnicodeString("integer")), integer, success)
+            .adoptFormatter(FunctionName(UnicodeString("test:function")), new StandardFunctions::TestFormatFactory(), success)
+            .adoptFormatter(FunctionName(UnicodeString("test:format")), new StandardFunctions::TestFormatFactory(), success)
             .adoptSelector(FunctionName(UnicodeString("number")), new StandardFunctions::PluralFactory(UPLURAL_TYPE_CARDINAL), success)
             .adoptSelector(FunctionName(UnicodeString("integer")), new StandardFunctions::PluralFactory(StandardFunctions::PluralFactory::integer()), success)
-            .adoptSelector(FunctionName(UnicodeString("string")), new StandardFunctions::TextFactory(), success);
+            .adoptSelector(FunctionName(UnicodeString("string")), new StandardFunctions::TextFactory(), success)
+            .adoptSelector(FunctionName(UnicodeString("test:function")), new StandardFunctions::TestSelectFactory(), success)
+            .adoptSelector(FunctionName(UnicodeString("test:select")), new StandardFunctions::TestSelectFactory(), success);
         CHECK_ERROR(success);
         standardMFFunctionRegistry = standardFunctionsBuilder.build();
         CHECK_ERROR(success);
@@ -163,6 +188,8 @@ namespace message2 {
             errors = errorsNew.orphan();
         }
 
+        nfcNormalizer = Normalizer2::getNFCInstance(success);
+
         // Note: we currently evaluate variables lazily,
         // without memoization. This call is still necessary
         // to check out-of-scope uses of local variables in
@@ -170,7 +197,7 @@ namespace message2 {
         // only be checked when arguments are known)
 
         // Check for resolution errors
-        Checker(dataModel, *errors).check(success);
+        Checker(dataModel, *errors, *this).check(success);
     }
 
     void MessageFormatter::cleanup() noexcept {
@@ -191,6 +218,7 @@ namespace message2 {
         signalErrors = other.signalErrors;
         errors = other.errors;
         other.errors = nullptr;
+        nfcNormalizer = other.nfcNormalizer;
         return *this;
     }
 
@@ -256,8 +284,11 @@ namespace message2 {
         return formatter;
     }
 
-    bool MessageFormatter::getDefaultFormatterNameByType(const UnicodeString& type, FunctionName& name) const {
-        U_ASSERT(hasCustomMFFunctionRegistry());
+    bool MessageFormatter::getDefaultFormatterNameByType(const UnicodeString& type,
+                                                         FunctionName& name) const {
+        if (!hasCustomMFFunctionRegistry()) {
+            return false;
+        }
         const MFFunctionRegistry& reg = getCustomMFFunctionRegistry();
         return reg.getDefaultFormatterNameByType(type, name);
     }
@@ -352,3 +383,5 @@ U_NAMESPACE_END
 #endif /* #if !UCONFIG_NO_MF2 */
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
+
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_function_registry.cpp b/deps/icu-small/source/i18n/messageformat2_function_registry.cpp
index 17955760ecfb44..e45fb3544ec524 100644
--- a/deps/icu-small/source/i18n/messageformat2_function_registry.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_function_registry.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -85,10 +87,11 @@ MFFunctionRegistry::Builder::Builder(UErrorCode& errorCode) {
     formattersByType = new Hashtable();
     if (!(formatters != nullptr && selectors != nullptr && formattersByType != nullptr)) {
         errorCode = U_MEMORY_ALLOCATION_ERROR;
+    } else {
+        formatters->setValueDeleter(uprv_deleteUObject);
+        selectors->setValueDeleter(uprv_deleteUObject);
+        formattersByType->setValueDeleter(uprv_deleteUObject);
     }
-    formatters->setValueDeleter(uprv_deleteUObject);
-    selectors->setValueDeleter(uprv_deleteUObject);
-    formattersByType->setValueDeleter(uprv_deleteUObject);
 }
 
 MFFunctionRegistry::Builder::~Builder() {
@@ -158,9 +161,13 @@ void MFFunctionRegistry::checkStandard() const {
     checkFormatter("time");
     checkFormatter("number");
     checkFormatter("integer");
+    checkFormatter("test:function");
+    checkFormatter("test:format");
     checkSelector("number");
     checkSelector("integer");
     checkSelector("string");
+    checkSelector("test:function");
+    checkSelector("test:select");
 }
 
 // Formatter/selector helpers
@@ -424,14 +431,14 @@ static FormattedPlaceholder notANumber(const FormattedPlaceholder& input) {
     return FormattedPlaceholder(input, FormattedValue(UnicodeString("NaN")));
 }
 
-static double parseNumberLiteral(const FormattedPlaceholder& input, UErrorCode& errorCode) {
+static double parseNumberLiteral(const Formattable& input, UErrorCode& errorCode) {
     if (U_FAILURE(errorCode)) {
         return {};
     }
 
     // Copying string to avoid GCC dangling-reference warning
     // (although the reference is safe)
-    UnicodeString inputStr = input.asFormattable().getString(errorCode);
+    UnicodeString inputStr = input.getString(errorCode);
     // Precondition: `input`'s source Formattable has type string
     if (U_FAILURE(errorCode)) {
         return {};
@@ -463,8 +470,42 @@ static double parseNumberLiteral(const FormattedPlaceholder& input, UErrorCode&
     return result;
 }
 
+static UChar32 digitToChar(int32_t val, UErrorCode errorCode) {
+    if (U_FAILURE(errorCode)) {
+        return '0';
+    }
+    if (val < 0 || val > 9) {
+        errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+    }
+    switch(val) {
+        case 0:
+            return '0';
+        case 1:
+            return '1';
+        case 2:
+            return '2';
+        case 3:
+            return '3';
+        case 4:
+            return '4';
+        case 5:
+            return '5';
+        case 6:
+            return '6';
+        case 7:
+            return '7';
+        case 8:
+            return '8';
+        case 9:
+            return '9';
+        default:
+            errorCode = U_ILLEGAL_ARGUMENT_ERROR;
+            return '0';
+    }
+}
+
 static FormattedPlaceholder tryParsingNumberLiteral(const number::LocalizedNumberFormatter& nf, const FormattedPlaceholder& input, UErrorCode& errorCode) {
-    double numberValue = parseNumberLiteral(input, errorCode);
+    double numberValue = parseNumberLiteral(input.asFormattable(), errorCode);
     if (U_FAILURE(errorCode)) {
         return notANumber(input);
     }
@@ -1235,6 +1276,273 @@ void StandardFunctions::TextSelector::selectKey(FormattedPlaceholder&& toFormat,
 StandardFunctions::TextFactory::~TextFactory() {}
 StandardFunctions::TextSelector::~TextSelector() {}
 
+// ------------ TestFormatFactory
+
+Formatter* StandardFunctions::TestFormatFactory::createFormatter(const Locale& locale, UErrorCode& errorCode) {
+    NULL_ON_ERROR(errorCode);
+
+    // Results are not locale-dependent
+    (void) locale;
+
+    Formatter* result = new TestFormat();
+    if (result == nullptr) {
+        errorCode = U_MEMORY_ALLOCATION_ERROR;
+    }
+    return result;
+}
+
+StandardFunctions::TestFormatFactory::~TestFormatFactory() {}
+StandardFunctions::TestFormat::~TestFormat() {}
+
+// Extract numeric value from a Formattable or, if it's a string,
+// parse it as a number according to the MF2 `number-literal` grammar production
+double formattableToNumber(const Formattable& arg, UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return 0;
+    }
+
+    double result = 0;
+
+    switch (arg.getType()) {
+        case UFMT_DOUBLE: {
+            result = arg.getDouble(status);
+            U_ASSERT(U_SUCCESS(status));
+            break;
+        }
+        case UFMT_LONG: {
+            result = (double) arg.getLong(status);
+            U_ASSERT(U_SUCCESS(status));
+            break;
+        }
+        case UFMT_INT64: {
+            result = (double) arg.getInt64(status);
+            U_ASSERT(U_SUCCESS(status));
+            break;
+        }
+        case UFMT_STRING: {
+            // Try to parse the string as a number
+            result = parseNumberLiteral(arg, status);
+            if (U_FAILURE(status)) {
+                status = U_MF_OPERAND_MISMATCH_ERROR;
+            }
+            break;
+        }
+        default: {
+            // Other types can't be parsed as a number
+            status = U_MF_OPERAND_MISMATCH_ERROR;
+            break;
+        }
+        }
+    return result;
+}
+
+
+/* static */ void StandardFunctions::TestFormat::testFunctionParameters(const FormattedPlaceholder& arg,
+                                                                        const FunctionOptions& options,
+                                                                        int32_t& decimalPlaces,
+                                                                        bool& failsFormat,
+                                                                        bool& failsSelect,
+                                                                        double& input,
+                                                                        UErrorCode& status) {
+    CHECK_ERROR(status);
+
+    // 1. Let DecimalPlaces be 0.
+    decimalPlaces = 0;
+
+    // 2. Let FailsFormat be false.
+    failsFormat = false;
+
+    // 3. Let FailsSelect be false.
+    failsSelect = false;
+
+    // 4. Let arg be the resolved value of the expression operand.
+    // (already true)
+
+    // Step 5 omitted because composition isn't fully implemented yet
+    // 6. Else if arg is a numerical value or a string matching the number-literal production, then
+    input = formattableToNumber(arg.asFormattable(), status);
+    if (U_FAILURE(status)) {
+        // 7. Else,
+        // 7i. Emit "bad-input" Resolution Error.
+        status = U_MF_OPERAND_MISMATCH_ERROR;
+        // 7ii. Use a fallback value as the resolved value of the expression.
+        // Further steps of this algorithm are not followed.
+    }
+    // 8. If the decimalPlaces option is set, then
+    Formattable opt;
+    if (options.getFunctionOption(UnicodeString("decimalPlaces"), opt)) {
+        // 8i. If its value resolves to a numerical integer value 0 or 1
+        // or their corresponding string representations '0' or '1', then
+        double decimalPlacesInput = formattableToNumber(opt, status);
+        if (U_SUCCESS(status)) {
+            if (decimalPlacesInput == 0 || decimalPlacesInput == 1) {
+                // 8ia. Set DecimalPlaces to be the numerical value of the option.
+                decimalPlaces = decimalPlacesInput;
+            }
+        }
+        // 8ii. Else if its value is not an unresolved value set by option resolution,
+        else {
+            // 8iia. Emit "bad-option" Resolution Error.
+            status = U_MF_BAD_OPTION;
+            // 8iib. Use a fallback value as the resolved value of the expression.
+        }
+    }
+    // 9. If the fails option is set, then
+    Formattable failsOpt;
+    if (options.getFunctionOption(UnicodeString("fails"), failsOpt)) {
+        UnicodeString failsString = failsOpt.getString(status);
+        if (U_SUCCESS(status)) {
+            // 9i. If its value resolves to the string 'always', then
+            if (failsString == u"always") {
+                // 9ia. Set FailsFormat to be true
+                failsFormat = true;
+                // 9ib. Set FailsSelect to be true.
+                failsSelect = true;
+            }
+            // 9ii. Else if its value resolves to the string "format", then
+            else if (failsString == u"format") {
+                // 9ia. Set FailsFormat to be true
+                failsFormat = true;
+            }
+            // 9iii. Else if its value resolves to the string "select", then
+            else if (failsString == u"select") {
+                // 9iiia. Set FailsSelect to be true.
+                failsSelect = true;
+            }
+            // 9iv. Else if its value does not resolve to the string "never", then
+            else if (failsString != u"never") {
+                // 9iv(a). Emit "bad-option" Resolution Error.
+                status = U_MF_BAD_OPTION;
+            }
+        } else {
+            // 9iv. again
+            status = U_MF_BAD_OPTION;
+        }
+    }
+}
+
+FormattedPlaceholder StandardFunctions::TestFormat::format(FormattedPlaceholder&& arg,
+                                                           FunctionOptions&& options,
+                                                           UErrorCode& status) const{
+
+    int32_t decimalPlaces;
+    bool failsFormat;
+    bool failsSelect;
+    double input;
+
+    testFunctionParameters(arg, options, decimalPlaces,
+                           failsFormat, failsSelect, input, status);
+    if (U_FAILURE(status)) {
+        return FormattedPlaceholder(arg.getFallback());
+    }
+
+    // If FailsFormat is true, attempting to format the placeholder to any
+    // formatting target will fail.
+    if (failsFormat) {
+        status = U_MF_FORMATTING_ERROR;
+        return FormattedPlaceholder(arg.getFallback());
+    }
+    UnicodeString result;
+    // When :test:function is used as a formatter, a placeholder resolving to a value
+    // with a :test:function expression is formatted as a concatenation of the following parts:
+    // 1. If Input is less than 0, the character - U+002D Hyphen-Minus.
+    if (input < 0) {
+        result += HYPHEN;
+    }
+    // 2. The truncated absolute integer value of Input, i.e. floor(abs(Input)), formatted as a
+    // sequence of decimal digit characters (U+0030...U+0039).
+    char buffer[256];
+    bool ignore;
+    int ignoreLen;
+    int ignorePoint;
+    double_conversion::DoubleToStringConverter::DoubleToAscii(floor(abs(input)),
+                                                              double_conversion::DoubleToStringConverter::DtoaMode::SHORTEST,
+                                                              0,
+                                                              buffer,
+                                                              256,
+                                                              &ignore,
+                                                              &ignoreLen,
+                                                              &ignorePoint);
+    result += UnicodeString(buffer);
+    // 3. If DecimalPlaces is 1, then
+    if (decimalPlaces == 1) {
+        // 3i. The character . U+002E Full Stop.
+        result += u".";
+        // 3ii. The single decimal digit character representing the value
+        // floor((abs(Input) - floor(abs(Input))) * 10)
+        int32_t val = floor((abs(input) - floor(abs(input)) * 10));
+        result += digitToChar(val, status);
+        U_ASSERT(U_SUCCESS(status));
+    }
+    return FormattedPlaceholder(result);
+}
+
+// ------------ TestSelectFactory
+
+StandardFunctions::TestSelectFactory::~TestSelectFactory() {}
+StandardFunctions::TestSelect::~TestSelect() {}
+
+Selector* StandardFunctions::TestSelectFactory::createSelector(const Locale& locale,
+                                                               UErrorCode& errorCode) const {
+    NULL_ON_ERROR(errorCode);
+
+    // Results are not locale-dependent
+    (void) locale;
+
+    Selector* result = new TestSelect();
+    if (result == nullptr) {
+        errorCode = U_MEMORY_ALLOCATION_ERROR;
+    }
+    return result;
+}
+
+void StandardFunctions::TestSelect::selectKey(FormattedPlaceholder&& val,
+                                              FunctionOptions&& options,
+                                              const UnicodeString* keys,
+                                              int32_t keysLen,
+                                              UnicodeString* prefs,
+                                              int32_t& prefsLen,
+                                              UErrorCode& status) const {
+    int32_t decimalPlaces;
+    bool failsFormat;
+    bool failsSelect;
+    double input;
+
+    TestFormat::testFunctionParameters(val, options, decimalPlaces,
+                                       failsFormat, failsSelect, input, status);
+
+    if (U_FAILURE(status)) {
+        return;
+    }
+
+    if (failsSelect) {
+        status = U_MF_SELECTOR_ERROR;
+        return;
+    }
+
+    // If the Input is 1 and DecimalPlaces is 1, the method will return some slice
+    // of the list « '1.0', '1' », depending on whether those values are included in keys.
+    bool include1point0 = false;
+    bool include1 = false;
+    if (input == 1 && decimalPlaces == 1) {
+        include1point0 = true;
+        include1 = true;
+    } else if (input == 1 && decimalPlaces == 0) {
+        include1 = true;
+    }
+
+    // If the Input is 1 and DecimalPlaces is 0, the method will return the list « '1' » if
+    // keys includes '1', or an empty list otherwise.
+    // If the Input is any other value, the method will return an empty list.
+    for (int32_t i = 0; i < keysLen; i++) {
+        if ((keys[i] == u"1" && include1)
+            || (keys[i] == u"1.0" && include1point0)) {
+            prefs[prefsLen] = keys[i];
+            prefsLen++;
+        }
+    }
+}
+
 } // namespace message2
 U_NAMESPACE_END
 
@@ -1242,3 +1550,4 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_function_registry_internal.h b/deps/icu-small/source/i18n/messageformat2_function_registry_internal.h
index 733fc5e945d5c8..9599b67bb2ba14 100644
--- a/deps/icu-small/source/i18n/messageformat2_function_registry_internal.h
+++ b/deps/icu-small/source/i18n/messageformat2_function_registry_internal.h
@@ -10,6 +10,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -209,6 +211,60 @@ namespace message2 {
 
             TextSelector(const Locale& l) : locale(l) {}
         };
+
+        // See https://github.com/unicode-org/message-format-wg/blob/main/test/README.md
+        class TestFormatFactory : public FormatterFactory {
+        public:
+            Formatter* createFormatter(const Locale& locale, UErrorCode& status) override;
+            TestFormatFactory() {}
+            virtual ~TestFormatFactory();
+        };
+
+        class TestSelect;
+
+        class TestFormat : public Formatter {
+        public:
+            FormattedPlaceholder format(FormattedPlaceholder&& toFormat, FunctionOptions&& options, UErrorCode& status) const override;
+            virtual ~TestFormat();
+
+        private:
+            friend class TestFormatFactory;
+            friend class TestSelect;
+            TestFormat() {}
+            static void testFunctionParameters(const FormattedPlaceholder& arg,
+                                               const FunctionOptions& options,
+                                               int32_t& decimalPlaces,
+                                               bool& failsFormat,
+                                               bool& failsSelect,
+                                               double& input,
+                                               UErrorCode& status);
+
+        };
+
+        // See https://github.com/unicode-org/message-format-wg/blob/main/test/README.md
+        class TestSelectFactory : public SelectorFactory {
+        public:
+            Selector* createSelector(const Locale& locale, UErrorCode& status) const override;
+            TestSelectFactory() {}
+            virtual ~TestSelectFactory();
+        };
+
+        class TestSelect : public Selector {
+        public:
+            void selectKey(FormattedPlaceholder&& val,
+                           FunctionOptions&& options,
+                           const UnicodeString* keys,
+                           int32_t keysLen,
+                           UnicodeString* prefs,
+                           int32_t& prefsLen,
+                           UErrorCode& status) const override;
+            virtual ~TestSelect();
+
+        private:
+            friend class TestSelectFactory;
+            TestSelect() {}
+        };
+
     };
 
     extern void formatDateWithDefaults(const Locale& locale, UDate date, UnicodeString&, UErrorCode& errorCode);
@@ -226,6 +282,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_FUNCTION_REGISTRY_INTERNAL_H
diff --git a/deps/icu-small/source/i18n/messageformat2_macros.h b/deps/icu-small/source/i18n/messageformat2_macros.h
index f06ed1a5a97746..20e81377d4d5cf 100644
--- a/deps/icu-small/source/i18n/messageformat2_macros.h
+++ b/deps/icu-small/source/i18n/messageformat2_macros.h
@@ -10,6 +10,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -97,6 +99,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_MACROS_H
diff --git a/deps/icu-small/source/i18n/messageformat2_parser.cpp b/deps/icu-small/source/i18n/messageformat2_parser.cpp
index b4768756c5ead2..9a9f8e78df03bd 100644
--- a/deps/icu-small/source/i18n/messageformat2_parser.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_parser.cpp
@@ -3,13 +3,18 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
 
+#include "unicode/uniset.h"
 #include "messageformat2_errors.h"
 #include "messageformat2_macros.h"
 #include "messageformat2_parser.h"
+#include "ucln_in.h"
+#include "umutex.h"
 #include "uvector.h" // U_ASSERT
 
 U_NAMESPACE_BEGIN
@@ -91,14 +96,282 @@ static void copyContext(const UChar in[U_PARSE_CONTEXT_LEN], UChar out[U_PARSE_C
 }
 
 // -------------------------------------
-// Predicates
+// Initialization of UnicodeSets
+
+namespace unisets {
+
+UnicodeSet* gUnicodeSets[unisets::UNISETS_KEY_COUNT] = {};
+
+inline UnicodeSet* getImpl(Key key) {
+    return gUnicodeSets[key];
+}
+
+icu::UInitOnce gMF2ParseUniSetsInitOnce {};
+}
+
+UnicodeSet* initContentChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    UnicodeSet* result = new UnicodeSet(0x0001, 0x0008); // Omit NULL, HTAB and LF
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    }
+    result->add(0x000B, 0x000C); // Omit CR
+    result->add(0x000E, 0x001F); // Omit SP
+    result->add(0x0021, 0x002D); // Omit '.'
+    result->add(0x002F, 0x003F); // Omit '@'
+    result->add(0x0041, 0x005B); // Omit '\'
+    result->add(0x005D, 0x007A); // Omit { | }
+    result->add(0x007E, 0x2FFF); // Omit IDEOGRAPHIC_SPACE
+    result->add(0x3001, 0x10FFFF); // Allowing surrogates is intentional
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initWhitespace(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    UnicodeSet* result = new UnicodeSet();
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    }
+    result->add(SPACE);
+    result->add(HTAB);
+    result->add(CR);
+    result->add(LF);
+    result->add(IDEOGRAPHIC_SPACE);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initBidiControls(UErrorCode& status) {
+    UnicodeSet* result = new UnicodeSet(UnicodeString("[\\u061C]"), status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    result->add(0x200E, 0x200F);
+    result->add(0x2066, 0x2069);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initAlpha(UErrorCode& status) {
+    UnicodeSet* result = new UnicodeSet(UnicodeString("[:letter:]"), status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initDigits(UErrorCode& status) {
+    UnicodeSet* result = new UnicodeSet(UnicodeString("[:number:]"), status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initNameStartChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    UnicodeSet* isAlpha = unisets::gUnicodeSets[unisets::ALPHA] = initAlpha(status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    UnicodeSet* result = new UnicodeSet(*isAlpha);
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    };
+    result->add(UNDERSCORE);
+    result->add(0x00C0, 0x00D6);
+    result->add(0x00D8, 0x00F6);
+    result->add(0x00F8, 0x02FF);
+    result->add(0x0370, 0x037D);
+    result->add(0x037F, 0x061B);
+    result->add(0x061D, 0x1FFF);
+    result->add(0x200C, 0x200D);
+    result->add(0x2070, 0x218F);
+    result->add(0x2C00, 0x2FEF);
+    result->add(0x3001, 0xD7FF);
+    result->add(0xF900, 0xFDCF);
+    result->add(0xFDF0, 0xFFFD);
+    result->add(0x100000, 0xEFFFF);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initNameChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    UnicodeSet* nameStart = unisets::gUnicodeSets[unisets::NAME_START] = initNameStartChars(status);
+    UnicodeSet* digit = unisets::gUnicodeSets[unisets::DIGIT] = initDigits(status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    UnicodeSet* result = new UnicodeSet();
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    };
+    result->addAll(*nameStart);
+    result->addAll(*digit);
+    result->add(HYPHEN);
+    result->add(PERIOD);
+    result->add(0x00B7);
+    result->add(0x0300, 0x036F);
+    result->add(0x203F, 0x2040);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initTextChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    UnicodeSet* content = unisets::gUnicodeSets[unisets::CONTENT] = initContentChars(status);
+    UnicodeSet* whitespace = unisets::gUnicodeSets[unisets::WHITESPACE] = initWhitespace(status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    UnicodeSet* result = new UnicodeSet();
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    };
+    result->addAll(*content);
+    result->addAll(*whitespace);
+    result->add(PERIOD);
+    result->add(AT);
+    result->add(PIPE);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initQuotedChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+
+    unisets::gUnicodeSets[unisets::TEXT] = initTextChars(status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    UnicodeSet* result = new UnicodeSet();
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    };
+    // content and whitespace were initialized by `initTextChars()`
+    UnicodeSet* content = unisets::getImpl(unisets::CONTENT);
+    if (content == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    }
+    result->addAll(*content);
+    UnicodeSet* whitespace = unisets::getImpl(unisets::WHITESPACE);
+    if (whitespace == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    }
+    result->addAll(*whitespace);
+    result->add(PERIOD);
+    result->add(AT);
+    result->add(LEFT_CURLY_BRACE);
+    result->add(RIGHT_CURLY_BRACE);
+    result->freeze();
+    return result;
+}
+
+UnicodeSet* initEscapableChars(UErrorCode& status) {
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
 
-// Returns true if `c` is in the interval [`first`, `last`]
-static bool inRange(UChar32 c, UChar32 first, UChar32 last) {
-    U_ASSERT(first < last);
-    return c >= first && c <= last;
+    UnicodeSet* result = new UnicodeSet();
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+        return nullptr;
+    }
+    result->add(PIPE);
+    result->add(BACKSLASH);
+    result->add(LEFT_CURLY_BRACE);
+    result->add(RIGHT_CURLY_BRACE);
+    result->freeze();
+    return result;
 }
 
+namespace unisets {
+
+UBool U_CALLCONV cleanupMF2ParseUniSets() {
+    for (int32_t i = 0; i < UNISETS_KEY_COUNT; i++) {
+        delete gUnicodeSets[i];
+        gUnicodeSets[i] = nullptr;
+    }
+    gMF2ParseUniSetsInitOnce.reset();
+    return true;
+}
+
+void U_CALLCONV initMF2ParseUniSets(UErrorCode& status) {
+    ucln_i18n_registerCleanup(UCLN_I18N_MF2_UNISETS, cleanupMF2ParseUniSets);
+    /*
+      Each of the init functions initializes the UnicodeSets
+      that it depends on.
+
+      initBidiControls (no dependencies)
+
+      initEscapableChars (no dependencies)
+
+      initNameChars depends on
+         initDigits
+         initNameStartChars depends on
+           initAlpha
+
+      initQuotedChars depends on
+         initTextChars depends on
+            initContentChars
+            initWhitespace
+     */
+    gUnicodeSets[unisets::BIDI] = initBidiControls(status);
+    gUnicodeSets[unisets::NAME_CHAR] = initNameChars(status);
+    gUnicodeSets[unisets::QUOTED] = initQuotedChars(status);
+    gUnicodeSets[unisets::ESCAPABLE] = initEscapableChars(status);
+
+    if (U_FAILURE(status)) {
+        cleanupMF2ParseUniSets();
+    }
+}
+
+const UnicodeSet* get(Key key, UErrorCode& status) {
+    umtx_initOnce(gMF2ParseUniSetsInitOnce, &initMF2ParseUniSets, status);
+    if (U_FAILURE(status)) {
+        return nullptr;
+    }
+    UnicodeSet* result = getImpl(key);
+    if (result == nullptr) {
+        status = U_MEMORY_ALLOCATION_ERROR;
+    }
+    return result;
+}
+
+}
+
+// -------------------------------------
+// Predicates
+
 /*
   The following helper predicates should exactly match nonterminals in the MessageFormat 2 grammar:
 
@@ -113,76 +386,50 @@ static bool inRange(UChar32 c, UChar32 first, UChar32 last) {
   `isWhitespace()`    : `s`
 */
 
-static bool isContentChar(UChar32 c) {
-    return inRange(c, 0x0001, 0x0008)    // Omit NULL, HTAB and LF
-           || inRange(c, 0x000B, 0x000C) // Omit CR
-           || inRange(c, 0x000E, 0x001F) // Omit SP
-           || inRange(c, 0x0021, 0x002D) // Omit '.'
-           || inRange(c, 0x002F, 0x003F) // Omit '@'
-           || inRange(c, 0x0041, 0x005B) // Omit '\'
-           || inRange(c, 0x005D, 0x007A) // Omit { | }
-           || inRange(c, 0x007E, 0xD7FF) // Omit surrogates
-           || inRange(c, 0xE000, 0x10FFFF);
+bool Parser::isContentChar(UChar32 c) const {
+    return contentChars->contains(c);
 }
 
-// See `s` in the MessageFormat 2 grammar
-inline bool isWhitespace(UChar32 c) {
-    switch (c) {
-    case SPACE:
-    case HTAB:
-    case CR:
-    case LF:
-    case IDEOGRAPHIC_SPACE:
-        return true;
-    default:
-        return false;
-    }
+// See `bidi` in the MF2 grammar
+bool Parser::isBidiControl(UChar32 c) const {
+    return bidiControlChars->contains(c);
 }
 
-static bool isTextChar(UChar32 c) {
-    return isContentChar(c)
-        || isWhitespace(c)
-        || c == PERIOD
-        || c == AT
-        || c == PIPE;
+// See `ws` in the MessageFormat 2 grammar
+bool Parser::isWhitespace(UChar32 c) const {
+    return whitespaceChars->contains(c);
 }
 
-static bool isAlpha(UChar32 c) { return inRange(c, 0x0041, 0x005A) || inRange(c, 0x0061, 0x007A); }
+bool Parser::isTextChar(UChar32 c) const {
+    return textChars->contains(c);
+}
+
+bool Parser::isAlpha(UChar32 c) const {
+    return alphaChars->contains(c);
+}
 
-static bool isDigit(UChar32 c) { return inRange(c, 0x0030, 0x0039); }
+bool Parser::isDigit(UChar32 c) const {
+    return digitChars->contains(c);
+}
 
-static bool isNameStart(UChar32 c) {
-    return isAlpha(c) || c == UNDERSCORE || inRange(c, 0x00C0, 0x00D6) || inRange(c, 0x00D8, 0x00F6) ||
-           inRange(c, 0x00F8, 0x02FF) || inRange(c, 0x0370, 0x037D) || inRange(c, 0x037F, 0x1FFF) ||
-           inRange(c, 0x200C, 0x200D) || inRange(c, 0x2070, 0x218F) || inRange(c, 0x2C00, 0x2FEF) ||
-           inRange(c, 0x3001, 0xD7FF) || inRange(c, 0xF900, 0xFDCF) || inRange(c, 0xFDF0, 0xFFFD) ||
-           inRange(c, 0x10000, 0xEFFFF);
+bool Parser::isNameStart(UChar32 c) const {
+    return nameStartChars->contains(c);
 }
 
-static bool isNameChar(UChar32 c) {
-    return isNameStart(c) || isDigit(c) || c == HYPHEN || c == PERIOD || c == 0x00B7 ||
-           inRange(c, 0x0300, 0x036F) || inRange(c, 0x203F, 0x2040);
+bool Parser::isNameChar(UChar32 c) const {
+    return nameChars->contains(c);
 }
 
-static bool isUnquotedStart(UChar32 c) {
-    return isNameStart(c) || isDigit(c) || c == HYPHEN || c == PERIOD || c == 0x00B7 ||
-           inRange(c, 0x0300, 0x036F) || inRange(c, 0x203F, 0x2040);
+bool Parser::isUnquotedStart(UChar32 c) const {
+    return isNameChar(c);
 }
 
-static bool isQuotedChar(UChar32 c) {
-    return isContentChar(c)
-        || isWhitespace(c)
-        || c == PERIOD
-        || c == AT
-        || c == LEFT_CURLY_BRACE
-        || c == RIGHT_CURLY_BRACE;
+bool Parser::isQuotedChar(UChar32 c) const {
+    return quotedChars->contains(c);
 }
 
-static bool isEscapableChar(UChar32 c) {
-    return c == PIPE
-        || c == BACKSLASH
-        || c == LEFT_CURLY_BRACE
-        || c == RIGHT_CURLY_BRACE;
+bool Parser::isEscapableChar(UChar32 c) const {
+    return escapableChars->contains(c);
 }
 
 // Returns true iff `c` can begin a `function` nonterminal
@@ -203,12 +450,12 @@ static bool isAnnotationStart(UChar32 c) {
 }
 
 // Returns true iff `c` can begin a `literal` nonterminal
-static bool isLiteralStart(UChar32 c) {
+bool Parser::isLiteralStart(UChar32 c) const {
     return (c == PIPE || isNameStart(c) || c == HYPHEN || isDigit(c));
 }
 
 // Returns true iff `c` can begin a `key` nonterminal
-static bool isKeyStart(UChar32 c) {
+bool Parser::isKeyStart(UChar32 c) const {
     return (c == ASTERISK || isLiteralStart(c));
 }
 
@@ -347,7 +594,7 @@ option, or the optional space before an attribute.
   No pre, no post.
   A message may end with whitespace, so `index` may equal `len()` on exit.
 */
-void Parser::parseWhitespaceMaybeRequired(bool required, UErrorCode& errorCode) {
+void Parser::parseRequiredWS(UErrorCode& errorCode) {
     bool sawWhitespace = false;
 
     // The loop exits either when we consume all the input,
@@ -358,7 +605,7 @@ void Parser::parseWhitespaceMaybeRequired(bool required, UErrorCode& errorCode)
             // If whitespace isn't required -- or if we saw it already --
             // then the caller is responsible for checking this case and
             // setting an error if necessary.
-            if (!required || sawWhitespace) {
+            if (sawWhitespace) {
                 // Not an error.
                 return;
             }
@@ -380,24 +627,51 @@ void Parser::parseWhitespaceMaybeRequired(bool required, UErrorCode& errorCode)
         }
     }
 
-    if (!sawWhitespace && required) {
+    if (!sawWhitespace) {
         ERROR(errorCode);
     }
 }
 
+void Parser::parseOptionalBidi() {
+    while (true) {
+        if (!inBounds()) {
+            return;
+        }
+        if (isBidiControl(peek())) {
+            next();
+        } else {
+            break;
+        }
+    }
+}
+
 /*
-  No pre, no post, for the same reason as `parseWhitespaceMaybeRequired()`.
+  No pre, no post, because a message may end with whitespace
+  Matches `s` in the MF2 grammar
 */
 void Parser::parseRequiredWhitespace(UErrorCode& errorCode) {
-    parseWhitespaceMaybeRequired(true, errorCode);
+    parseOptionalBidi();
+    parseRequiredWS(errorCode);
+    parseOptionalWhitespace();
     normalizedInput += SPACE;
 }
 
 /*
   No pre, no post, for the same reason as `parseWhitespaceMaybeRequired()`.
 */
-void Parser::parseOptionalWhitespace(UErrorCode& errorCode) {
-    parseWhitespaceMaybeRequired(false, errorCode);
+void Parser::parseOptionalWhitespace() {
+    while (true) {
+        if (!inBounds()) {
+            return;
+        }
+        auto cp = peek();
+        if (isWhitespace(cp) || isBidiControl(cp)) {
+            maybeAdvanceLine();
+            next();
+        } else {
+            break;
+        }
+    }
 }
 
 // Consumes a single character, signaling an error if `peek()` != `c`
@@ -442,11 +716,11 @@ void Parser::parseToken(const std::u16string_view& token, UErrorCode& errorCode)
 */
 void Parser::parseTokenWithWhitespace(const std::u16string_view& token, UErrorCode& errorCode) {
     // No need for error check or bounds check before parseOptionalWhitespace
-    parseOptionalWhitespace(errorCode);
+    parseOptionalWhitespace();
     // Establish precondition
     CHECK_BOUNDS(errorCode);
     parseToken(token, errorCode);
-    parseOptionalWhitespace(errorCode);
+    parseOptionalWhitespace();
     // Guarantee postcondition
     CHECK_BOUNDS(errorCode);
 }
@@ -458,12 +732,12 @@ void Parser::parseTokenWithWhitespace(const std::u16string_view& token, UErrorCo
    then consumes optional whitespace again
 */
 void Parser::parseTokenWithWhitespace(UChar32 c, UErrorCode& errorCode) {
-    // No need for error check or bounds check before parseOptionalWhitespace(errorCode)
-    parseOptionalWhitespace(errorCode);
+    // No need for error check or bounds check before parseOptionalWhitespace()
+    parseOptionalWhitespace();
     // Establish precondition
     CHECK_BOUNDS(errorCode);
     parseToken(c, errorCode);
-    parseOptionalWhitespace(errorCode);
+    parseOptionalWhitespace();
     // Guarantee postcondition
     CHECK_BOUNDS(errorCode);
 }
@@ -482,11 +756,17 @@ UnicodeString Parser::parseName(UErrorCode& errorCode) {
 
     U_ASSERT(inBounds());
 
-    if (!isNameStart(peek())) {
+    if (!(isNameStart(peek()) || isBidiControl(peek()))) {
         ERROR(errorCode);
         return name;
     }
 
+    // name       = [bidi] name-start *name-char [bidi]
+
+    // [bidi]
+    parseOptionalBidi();
+
+    // name-start *name-char
     while (isNameChar(peek())) {
         UChar32 c = peek();
         name += c;
@@ -497,6 +777,10 @@ UnicodeString Parser::parseName(UErrorCode& errorCode) {
             break;
         }
     }
+
+    // [bidi]
+    parseOptionalBidi();
+
     return name;
 }
 
@@ -510,21 +794,13 @@ VariableName Parser::parseVariableName(UErrorCode& errorCode) {
     VariableName result;
 
     U_ASSERT(inBounds());
-    // If the '$' is missing, we don't want a binding
-    // for this variable to be created.
-    bool valid = peek() == DOLLAR;
+
     parseToken(DOLLAR, errorCode);
     if (!inBounds()) {
         ERROR(errorCode);
         return result;
     }
-    UnicodeString varName = parseName(errorCode);
-    // Set the name to "" if the variable wasn't
-    // declared correctly
-    if (!valid) {
-        varName.remove();
-    }
-    return VariableName(varName);
+    return VariableName(parseName(errorCode));
 }
 
 /*
@@ -853,7 +1129,7 @@ void Parser::parseAttribute(AttributeAdder<T>& attrAdder, UErrorCode& errorCode)
     // about whether whitespace precedes another
     // attribute, or the '=' sign
     int32_t savedIndex = index;
-    parseOptionalWhitespace(errorCode);
+    parseOptionalWhitespace();
 
     Operand rand;
     if (peek() == EQUALS) {
@@ -861,19 +1137,9 @@ void Parser::parseAttribute(AttributeAdder<T>& attrAdder, UErrorCode& errorCode)
         parseTokenWithWhitespace(EQUALS, errorCode);
 
         UnicodeString rhsStr;
-        // Parse RHS, which is either a literal or variable
-        switch (peek()) {
-        case DOLLAR: {
-            rand = Operand(parseVariableName(errorCode));
-            break;
-        }
-        default: {
-            // Must be a literal
-            rand = Operand(parseLiteral(errorCode));
-            break;
-        }
-        }
-        U_ASSERT(!rand.isNull());
+        // Parse RHS, which must be a literal
+        // attribute = "@" identifier [o "=" o literal]
+        rand = Operand(parseLiteral(errorCode));
     } else {
         // attribute -> "@" identifier [[s] "=" [s]]
         // Use null operand, which `rand` is already set to
@@ -881,7 +1147,7 @@ void Parser::parseAttribute(AttributeAdder<T>& attrAdder, UErrorCode& errorCode)
         index = savedIndex;
     }
 
-    attrAdder.addAttribute(lhs, std::move(rand), errorCode);
+    attrAdder.addAttribute(lhs, std::move(Operand(rand)), errorCode);
 }
 
 /*
@@ -1149,7 +1415,7 @@ the comment in `parseOptions()` for details.
       // (the character is either the required space before an annotation, or optional
       // trailing space after the literal or variable). It's still ambiguous which
       // one does apply.
-      parseOptionalWhitespace(status);
+      parseOptionalWhitespace();
       // Restore precondition
       CHECK_BOUNDS(status);
 
@@ -1220,7 +1486,7 @@ Expression Parser::parseExpression(UErrorCode& status) {
     // Parse opening brace
     parseToken(LEFT_CURLY_BRACE, status);
     // Optional whitespace after opening brace
-    parseOptionalWhitespace(status);
+    parseOptionalWhitespace();
 
     Expression::Builder exprBuilder(status);
     // Restore precondition
@@ -1263,7 +1529,7 @@ Expression Parser::parseExpression(UErrorCode& status) {
 
     // Parse optional space
     // (the last [s] in e.g. "{" [s] literal [s annotation] *(s attribute) [s] "}")
-    parseOptionalWhitespace(status);
+    parseOptionalWhitespace();
 
     // Either an operand or operator (or both) must have been set already,
     // so there can't be an error
@@ -1339,7 +1605,7 @@ void Parser::parseInputDeclaration(UErrorCode& status) {
     CHECK_BOUNDS(status);
 
     parseToken(ID_INPUT, status);
-    parseOptionalWhitespace(status);
+    parseOptionalWhitespace();
 
     // Restore precondition before calling parseExpression()
     CHECK_BOUNDS(status);
@@ -1400,7 +1666,7 @@ void Parser::parseDeclarations(UErrorCode& status) {
         // Avoid looping infinitely
         CHECK_ERROR(status);
 
-        parseOptionalWhitespace(status);
+        parseOptionalWhitespace();
         // Restore precondition
         CHECK_BOUNDS(status);
     }
@@ -1510,8 +1776,8 @@ This is addressed using "backtracking" (similarly to `parseOptions()`).
 
     // We've seen at least one whitespace-key pair, so now we can parse
     // *(s key) [s]
-    while (peek() != LEFT_CURLY_BRACE || isWhitespace(peek())) { // Try to recover from errors
-        bool wasWhitespace = isWhitespace(peek());
+    while (peek() != LEFT_CURLY_BRACE || isWhitespace(peek()) || isBidiControl(peek())) {
+        bool wasWhitespace = isWhitespace(peek()) || isBidiControl(peek());
         parseRequiredWhitespace(status);
         if (!wasWhitespace) {
             // Avoid infinite loop when parsing something like:
@@ -1569,7 +1835,7 @@ Markup Parser::parseMarkup(UErrorCode& status) {
     // Consume the '{'
     next();
     normalizedInput += LEFT_CURLY_BRACE;
-    parseOptionalWhitespace(status);
+    parseOptionalWhitespace();
     bool closing = false;
     switch (peek()) {
     case NUMBER_SIGN: {
@@ -1596,19 +1862,19 @@ Markup Parser::parseMarkup(UErrorCode& status) {
 
     // Parse the options, which must begin with a ' '
     // if present
-    if (inBounds() && isWhitespace(peek())) {
+    if (inBounds() && (isWhitespace(peek()) || isBidiControl(peek()))) {
         OptionAdder<Markup::Builder> optionAdder(builder);
         parseOptions(optionAdder, status);
     }
 
     // Parse the attributes, which also must begin
     // with a ' '
-    if (inBounds() && isWhitespace(peek())) {
+    if (inBounds() && (isWhitespace(peek()) || isBidiControl(peek()))) {
         AttributeAdder<Markup::Builder> attrAdder(builder);
         parseAttributes(attrAdder, status);
     }
 
-    parseOptionalWhitespace(status);
+    parseOptionalWhitespace();
 
     bool standalone = false;
     // Check if this is a standalone or not
@@ -1656,7 +1922,7 @@ std::variant<Expression, Markup> Parser::parsePlaceholder(UErrorCode& status) {
             isMarkup = true;
             break;
         }
-        if (!isWhitespace(c)){
+        if (!(isWhitespace(c) || isBidiControl(c))) {
             break;
         }
         tempIndex++;
@@ -1712,7 +1978,7 @@ Pattern Parser::parseSimpleMessage(UErrorCode& status) {
                 break;
             }
             // Don't loop infinitely
-            if (errors.hasSyntaxError()) {
+            if (errors.hasSyntaxError() || U_FAILURE(status)) {
                 break;
             }
         }
@@ -1720,6 +1986,22 @@ Pattern Parser::parseSimpleMessage(UErrorCode& status) {
     return result.build(status);
 }
 
+void Parser::parseVariant(UErrorCode& status) {
+    CHECK_ERROR(status);
+
+    // At least one key is required
+    SelectorKeys keyList(parseNonEmptyKeys(status));
+
+    // parseNonEmptyKeys() consumes any trailing whitespace,
+    // so the pattern can be consumed next.
+
+    // Restore precondition before calling parsePattern()
+    // (which must return a non-null value)
+    CHECK_BOUNDS(status);
+    Pattern rhs = parseQuotedPattern(status);
+
+    dataModel.addVariant(std::move(keyList), std::move(rhs), status);
+}
 
 /*
   Consume a `selectors` (matching the nonterminal in the grammar),
@@ -1739,22 +2021,25 @@ void Parser::parseSelectors(UErrorCode& status) {
     // Parse selectors
     // "Backtracking" is required here. It's not clear if whitespace is
     // (`[s]` selector) or (`[s]` variant)
-    while (isWhitespace(peek()) || peek() == LEFT_CURLY_BRACE) {
-        parseOptionalWhitespace(status);
+    while (isWhitespace(peek()) || peek() == DOLLAR) {
+        int32_t whitespaceStart = index;
+        parseRequiredWhitespace(status);
         // Restore precondition
         CHECK_BOUNDS(status);
-        if (peek() != LEFT_CURLY_BRACE) {
+        if (peek() != DOLLAR) {
             // This is not necessarily an error, but rather,
             // means the whitespace we parsed was the optional
             // whitespace preceding the first variant, not the
-            // optional whitespace preceding a subsequent expression.
+            // required whitespace preceding a subsequent variable.
+            // In that case, "push back" the whitespace.
+            normalizedInput.truncate(normalizedInput.length() - 1);
+            index = whitespaceStart;
             break;
         }
-        Expression expression;
-        expression = parseExpression(status);
+        VariableName var = parseVariableName(status);
         empty = false;
 
-        dataModel.addSelector(std::move(expression), status);
+        dataModel.addSelector(std::move(var), status);
         CHECK_ERROR(status);
     }
 
@@ -1770,27 +2055,29 @@ void Parser::parseSelectors(UErrorCode& status) {
         }                                          \
 
     // Parse variants
-    while (isWhitespace(peek()) || isKeyStart(peek())) {
-        // Trailing whitespace is allowed
-        parseOptionalWhitespace(status);
+    // matcher = match-statement s variant *(o variant)
+
+    // Parse first variant
+    parseRequiredWhitespace(status);
+    if (!inBounds()) {
+        ERROR(status);
+        return;
+    }
+    parseVariant(status);
+    if (!inBounds()) {
+        // Not an error; there might be only one variant
+        return;
+    }
+
+    while (isWhitespace(peek()) || isBidiControl(peek()) || isKeyStart(peek())) {
+        parseOptionalWhitespace();
+        // Restore the precondition.
+        // Trailing whitespace is allowed.
         if (!inBounds()) {
             return;
         }
 
-        // At least one key is required
-        SelectorKeys keyList(parseNonEmptyKeys(status));
-
-        CHECK_ERROR(status);
-
-        // parseNonEmptyKeys() consumes any trailing whitespace,
-        // so the pattern can be consumed next.
-
-        // Restore precondition before calling parsePattern()
-        // (which must return a non-null value)
-        CHECK_BOUNDS(status);
-        Pattern rhs = parseQuotedPattern(status);
-
-        dataModel.addVariant(std::move(keyList), std::move(rhs), status);
+        parseVariant(status);
 
         // Restore the precondition, *without* erroring out if we've
         // reached the end of input. That's because it's valid for the
@@ -1799,6 +2086,10 @@ void Parser::parseSelectors(UErrorCode& status) {
         // Because if we don't check it here, the `isWhitespace()` call in
         // the loop head will read off the end of the input string.
         CHECK_END_OF_INPUT
+
+        if (errors.hasSyntaxError() || U_FAILURE(status)) {
+            break;
+        }
     }
 }
 
@@ -1871,7 +2162,7 @@ void Parser::parse(UParseError &parseErrorResult, UErrorCode& status) {
     bool complex = false;
     // First, "look ahead" to determine if this is a simple or complex
     // message. To do that, check the first non-whitespace character.
-    while (inBounds(index) && isWhitespace(peek())) {
+    while (inBounds(index) && (isWhitespace(peek()) || isBidiControl(peek()))) {
         next();
     }
 
@@ -1891,10 +2182,10 @@ void Parser::parse(UParseError &parseErrorResult, UErrorCode& status) {
     // Message can be empty, so we need to only look ahead
     // if we know it's non-empty
     if (complex) {
-        parseOptionalWhitespace(status);
+        parseOptionalWhitespace();
         parseDeclarations(status);
         parseBody(status);
-        parseOptionalWhitespace(status);
+        parseOptionalWhitespace();
     } else {
         // Simple message
         // For normalization, quote the pattern
@@ -1926,3 +2217,4 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_parser.h b/deps/icu-small/source/i18n/messageformat2_parser.h
index b62cbe9200b94a..62a52d8f680a1e 100644
--- a/deps/icu-small/source/i18n/messageformat2_parser.h
+++ b/deps/icu-small/source/i18n/messageformat2_parser.h
@@ -10,12 +10,15 @@
 
 #include "unicode/messageformat2_data_model.h"
 #include "unicode/parseerr.h"
+#include "unicode/uniset.h"
 
 #include "messageformat2_allocation.h"
 #include "messageformat2_errors.h"
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -54,6 +57,26 @@ namespace message2 {
             }
     };
 
+
+    // Initialization of UnicodeSets
+    namespace unisets {
+        enum Key {
+            CONTENT,
+            WHITESPACE,
+            BIDI,
+            ALPHA,
+            DIGIT,
+            NAME_START,
+            NAME_CHAR,
+            TEXT,
+            QUOTED,
+            ESCAPABLE,
+            UNISETS_KEY_COUNT
+        };
+
+    U_I18N_API const UnicodeSet* get(Key key, UErrorCode& status);
+    }
+
     // Parser class (private)
     class Parser : public UMemory {
     public:
@@ -82,8 +105,23 @@ namespace message2 {
 	    UChar   postContext[U_PARSE_CONTEXT_LEN];
 	} MessageParseError;
 
-	Parser(const UnicodeString &input, MFDataModel::Builder& dataModelBuilder, StaticErrors& e, UnicodeString& normalizedInputRef)
-	  : source(input), index(0), errors(e), normalizedInput(normalizedInputRef), dataModel(dataModelBuilder) {
+	Parser(const UnicodeString &input,
+               MFDataModel::Builder& dataModelBuilder,
+               StaticErrors& e,
+               UnicodeString& normalizedInputRef,
+               UErrorCode& status)
+            : contentChars(unisets::get(unisets::CONTENT, status)),
+              whitespaceChars(unisets::get(unisets::WHITESPACE, status)),
+              bidiControlChars(unisets::get(unisets::BIDI, status)),
+              alphaChars(unisets::get(unisets::ALPHA, status)),
+              digitChars(unisets::get(unisets::DIGIT, status)),
+              nameStartChars(unisets::get(unisets::NAME_START, status)),
+              nameChars(unisets::get(unisets::NAME_CHAR, status)),
+              textChars(unisets::get(unisets::TEXT, status)),
+              quotedChars(unisets::get(unisets::QUOTED, status)),
+              escapableChars(unisets::get(unisets::ESCAPABLE, status)),
+            source(input), index(0), errors(e), normalizedInput(normalizedInputRef), dataModel(dataModelBuilder) {
+            (void) status;
 	  parseError.line = 0;
 	  parseError.offset = 0;
 	  parseError.lengthBeforeCurrentLine = 0;
@@ -91,6 +129,20 @@ namespace message2 {
 	  parseError.postContext[0] = '\0';
 	}
 
+        bool isContentChar(UChar32) const;
+        bool isBidiControl(UChar32) const;
+        bool isWhitespace(UChar32) const;
+        bool isTextChar(UChar32) const;
+        bool isQuotedChar(UChar32) const;
+        bool isEscapableChar(UChar32) const;
+        bool isAlpha(UChar32) const;
+        bool isDigit(UChar32) const;
+        bool isNameStart(UChar32) const;
+        bool isNameChar(UChar32) const;
+        bool isUnquotedStart(UChar32) const;
+        bool isLiteralStart(UChar32) const;
+        bool isKeyStart(UChar32) const;
+
 	static void translateParseError(const MessageParseError&, UParseError&);
 	static void setParseError(MessageParseError&, uint32_t);
 	void maybeAdvanceLine();
@@ -100,11 +152,13 @@ namespace message2 {
         void parseUnsupportedStatement(UErrorCode&);
         void parseLocalDeclaration(UErrorCode&);
         void parseInputDeclaration(UErrorCode&);
-	void parseSelectors(UErrorCode&);
+        void parseSelectors(UErrorCode&);
+        void parseVariant(UErrorCode&);
 
-	void parseWhitespaceMaybeRequired(bool, UErrorCode&);
+	void parseRequiredWS(UErrorCode&);
 	void parseRequiredWhitespace(UErrorCode&);
-	void parseOptionalWhitespace(UErrorCode&);
+	void parseOptionalBidi();
+	void parseOptionalWhitespace();
 	void parseToken(UChar32, UErrorCode&);
 	void parseTokenWithWhitespace(UChar32, UErrorCode&);
 	void parseToken(const std::u16string_view&, UErrorCode&);
@@ -149,6 +203,18 @@ namespace message2 {
         bool inBounds(uint32_t i) const { return source.moveIndex32(index, i) < source.length(); }
         bool allConsumed() const { return (int32_t) index == source.length(); }
 
+        // UnicodeSets for checking character ranges
+        const UnicodeSet* contentChars;
+        const UnicodeSet* whitespaceChars;
+        const UnicodeSet* bidiControlChars;
+        const UnicodeSet* alphaChars;
+        const UnicodeSet* digitChars;
+        const UnicodeSet* nameStartChars;
+        const UnicodeSet* nameChars;
+        const UnicodeSet* textChars;
+        const UnicodeSet* quotedChars;
+        const UnicodeSet* escapableChars;
+
 	// The input string
 	const UnicodeString &source;
 	// The current position within the input string -- counting in UChar32
@@ -165,8 +231,8 @@ namespace message2 {
 
 	// The parent builder
 	MFDataModel::Builder &dataModel;
-    }; // class Parser
 
+    }; // class Parser
 } // namespace message2
 
 U_NAMESPACE_END
@@ -175,6 +241,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT_PARSER_H
diff --git a/deps/icu-small/source/i18n/messageformat2_serializer.cpp b/deps/icu-small/source/i18n/messageformat2_serializer.cpp
index b2765f5acf434f..dfae0083392bf1 100644
--- a/deps/icu-small/source/i18n/messageformat2_serializer.cpp
+++ b/deps/icu-small/source/i18n/messageformat2_serializer.cpp
@@ -3,6 +3,8 @@
 
 #include "unicode/utypes.h"
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -244,11 +246,12 @@ void Serializer::serializeDeclarations() {
 
 void Serializer::serializeSelectors() {
     U_ASSERT(!dataModel.hasPattern());
-    const Expression* selectors = dataModel.getSelectorsInternal();
+    const VariableName* selectors = dataModel.getSelectorsInternal();
 
     emit(ID_MATCH);
     for (int32_t i = 0; i < dataModel.numSelectors(); i++) {
-        // No whitespace needed here -- see `selectors` in the grammar
+        whitespace();
+        emit(DOLLAR);
         emit(selectors[i]);
     }
 }
@@ -256,6 +259,7 @@ void Serializer::serializeSelectors() {
 void Serializer::serializeVariants() {
     U_ASSERT(!dataModel.hasPattern());
     const Variant* variants = dataModel.getVariantsInternal();
+    whitespace();
     for (int32_t i = 0; i < dataModel.numVariants(); i++) {
         const Variant& v = variants[i];
         emit(v.getKeys());
@@ -285,3 +289,4 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
diff --git a/deps/icu-small/source/i18n/messageformat2_serializer.h b/deps/icu-small/source/i18n/messageformat2_serializer.h
index 1b97b3b930087d..f190b255f0d3e5 100644
--- a/deps/icu-small/source/i18n/messageformat2_serializer.h
+++ b/deps/icu-small/source/i18n/messageformat2_serializer.h
@@ -10,6 +10,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -63,6 +65,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT_SERIALIZER_H
diff --git a/deps/icu-small/source/i18n/nfrs.cpp b/deps/icu-small/source/i18n/nfrs.cpp
index be2ab2932e7a5c..b7ffb561461b98 100644
--- a/deps/icu-small/source/i18n/nfrs.cpp
+++ b/deps/icu-small/source/i18n/nfrs.cpp
@@ -152,7 +152,7 @@ NFRuleSet::NFRuleSet(RuleBasedNumberFormat *_owner, UnicodeString* descriptions,
 
     UnicodeString& description = descriptions[index]; // !!! make sure index is valid
 
-    if (description.length() == 0) {
+    if (description.isEmpty()) {
         // throw new IllegalArgumentException("Empty rule set description");
         status = U_PARSE_ERROR;
         return;
@@ -177,16 +177,16 @@ NFRuleSet::NFRuleSet(RuleBasedNumberFormat *_owner, UnicodeString* descriptions,
         name.setTo(UNICODE_STRING_SIMPLE("%default"));
     }
 
-    if (description.length() == 0) {
+    if (description.isEmpty()) {
         // throw new IllegalArgumentException("Empty rule set description");
         status = U_PARSE_ERROR;
     }
 
     fIsPublic = name.indexOf(gPercentPercent, 2, 0) != 0;
 
-    if ( name.endsWith(gNoparse,8) ) {
+    if (name.endsWith(gNoparse, 8)) {
         fIsParseable = false;
-        name.truncate(name.length()-8); // remove the @noparse from the name
+        name.truncate(name.length() - 8); // remove the @noparse from the name
     }
 
     // all of the other members of NFRuleSet are initialized
diff --git a/deps/icu-small/source/i18n/nfrule.cpp b/deps/icu-small/source/i18n/nfrule.cpp
index 264e8d79e2d968..ef7f5924c4eb55 100644
--- a/deps/icu-small/source/i18n/nfrule.cpp
+++ b/deps/icu-small/source/i18n/nfrule.cpp
@@ -19,7 +19,6 @@
 
 #if U_HAVE_RBNF
 
-#include <limits>
 #include "unicode/localpointer.h"
 #include "unicode/rbnf.h"
 #include "unicode/tblcoll.h"
@@ -65,6 +64,7 @@ NFRule::~NFRule()
 
 static const char16_t gLeftBracket = 0x005b;
 static const char16_t gRightBracket = 0x005d;
+static const char16_t gVerticalLine = 0x007C;
 static const char16_t gColon = 0x003a;
 static const char16_t gZero = 0x0030;
 static const char16_t gNine = 0x0039;
@@ -147,6 +147,7 @@ NFRule::makeRules(UnicodeString& description,
         // then it's really shorthand for two rules (with one exception)
         LocalPointer<NFRule> rule2;
         UnicodeString sbuf;
+        int32_t orElseOp = description.indexOf(gVerticalLine);
 
         // we'll actually only split the rule into two rules if its
         // base value is an even multiple of its divisor (or it's one
@@ -194,9 +195,13 @@ NFRule::makeRules(UnicodeString& description,
             rule2->radix = rule1->radix;
             rule2->exponent = rule1->exponent;
 
-            // rule2's rule text omits the stuff in brackets: initialize
-            // its rule text and substitutions accordingly
+            // By default, rule2's rule text omits the stuff in brackets,
+            // unless it contains a | between the brackets.
+            // Initialize its rule text and substitutions accordingly.
             sbuf.append(description, 0, brack1);
+            if (orElseOp >= 0) {
+                sbuf.append(description, orElseOp + 1, brack2 - orElseOp - 1);
+            }
             if (brack2 + 1 < description.length()) {
                 sbuf.append(description, brack2 + 1, description.length() - brack2 - 1);
             }
@@ -207,7 +212,12 @@ NFRule::makeRules(UnicodeString& description,
         // the brackets themselves: initialize _its_ rule text and
         // substitutions accordingly
         sbuf.setTo(description, 0, brack1);
-        sbuf.append(description, brack1 + 1, brack2 - brack1 - 1);
+        if (orElseOp >= 0) {
+            sbuf.append(description, brack1 + 1, orElseOp - brack1 - 1);
+        }
+        else {
+            sbuf.append(description, brack1 + 1, brack2 - brack1 - 1);
+        }
         if (brack2 + 1 < description.length()) {
             sbuf.append(description, brack2 + 1, description.length() - brack2 - 1);
         }
@@ -286,18 +296,17 @@ NFRule::parseRuleDescriptor(UnicodeString& description, UErrorCode& status)
             // into "tempValue", skip periods, commas, and spaces,
             // stop on a slash or > sign (or at the end of the string),
             // and throw an exception on any other character
-            int64_t ll_10 = 10;
             while (p < descriptorLength) {
                 c = descriptor.charAt(p);
                 if (c >= gZero && c <= gNine) {
-                    int32_t single_digit = static_cast<int32_t>(c - gZero);
-                    if ((val > 0 && val > (std::numeric_limits<int64_t>::max() - single_digit) / 10) ||
-                        (val < 0 && val < (std::numeric_limits<int64_t>::min() - single_digit) / 10)) {
+                    int64_t digit = static_cast<int64_t>(c - gZero);
+                    if ((val > 0 && val > (INT64_MAX - digit) / 10) ||
+                        (val < 0 && val < (INT64_MIN - digit) / 10)) {
                         // out of int64_t range
                         status = U_PARSE_ERROR;
                         return;
                     }
-                    val = val * ll_10 + single_digit;
+                    val = val * 10 + digit;
                 }
                 else if (c == gSlash || c == gGreaterThan) {
                     break;
@@ -322,11 +331,17 @@ NFRule::parseRuleDescriptor(UnicodeString& description, UErrorCode& status)
             if (c == gSlash) {
                 val = 0;
                 ++p;
-                ll_10 = 10;
                 while (p < descriptorLength) {
                     c = descriptor.charAt(p);
                     if (c >= gZero && c <= gNine) {
-                        val = val * ll_10 + static_cast<int32_t>(c - gZero);
+                        int64_t digit = static_cast<int64_t>(c - gZero);
+                        if ((val > 0 && val > (INT64_MAX - digit) / 10) ||
+                            (val < 0 && val < (INT64_MIN - digit) / 10)) {
+                            // out of int64_t range
+                            status = U_PARSE_ERROR;
+                            return;
+                        }
+                        val = val * 10 + digit;
                     }
                     else if (c == gGreaterThan) {
                         break;
@@ -400,7 +415,7 @@ NFRule::parseRuleDescriptor(UnicodeString& description, UErrorCode& status)
     // finally, if the rule body begins with an apostrophe, strip it off
     // (this is generally used to put whitespace at the beginning of
     // a rule's rule text)
-    if (description.length() > 0 && description.charAt(0) == gTick) {
+    if (!description.isEmpty() && description.charAt(0) == gTick) {
         description.removeBetween(0, 1);
     }
 
diff --git a/deps/icu-small/source/i18n/number_decimalquantity.cpp b/deps/icu-small/source/i18n/number_decimalquantity.cpp
index f9350d5d5cc8a2..ca1dacd3579d85 100644
--- a/deps/icu-small/source/i18n/number_decimalquantity.cpp
+++ b/deps/icu-small/source/i18n/number_decimalquantity.cpp
@@ -1133,7 +1133,7 @@ void DecimalQuantity::setDigitPos(int32_t position, int8_t value) {
 }
 
 void DecimalQuantity::shiftLeft(int32_t numDigits) {
-    if (!usingBytes && precision + numDigits > 16) {
+    if (!usingBytes && precision + numDigits >= 16) {
         switchStorage();
     }
     if (usingBytes) {
diff --git a/deps/icu-small/source/i18n/number_longnames.cpp b/deps/icu-small/source/i18n/number_longnames.cpp
index de6aad7c3e8af7..bf2617e773d06f 100644
--- a/deps/icu-small/source/i18n/number_longnames.cpp
+++ b/deps/icu-small/source/i18n/number_longnames.cpp
@@ -48,8 +48,12 @@ constexpr int32_t PER_INDEX = StandardPlural::Form::COUNT + 1;
  * Gender of the word, in languages with grammatical gender.
  */
 constexpr int32_t GENDER_INDEX = StandardPlural::Form::COUNT + 2;
+/**
+ *  Denominator constant of the unit.
+ */
+constexpr int32_t CONSTANT_DENOMINATOR_INDEX = StandardPlural::Form::COUNT + 3;
 // Number of keys in the array populated by PluralTableSink.
-constexpr int32_t ARRAY_LENGTH = StandardPlural::Form::COUNT + 3;
+constexpr int32_t ARRAY_LENGTH = StandardPlural::Form::COUNT + 4;
 
 // TODO(icu-units#28): load this list from resources, after creating a "&set"
 // function for use in ldml2icu rules.
@@ -1010,6 +1014,11 @@ void LongNameHandler::forArbitraryUnit(const Locale &loc,
     //    denominator (the part after the "-per-). If both are empty, fail
     MeasureUnitImpl unit;
     MeasureUnitImpl perUnit;
+
+    if (unitRef.getConstantDenominator(status) != 0) {
+        perUnit.constantDenominator = unitRef.getConstantDenominator(status);
+    }
+
     {
         MeasureUnitImpl fullUnit = MeasureUnitImpl::forMeasureUnitMaybeCopy(unitRef, status);
         if (U_FAILURE(status)) {
@@ -1196,6 +1205,12 @@ void LongNameHandler::processPatternTimes(MeasureUnitImpl &&productUnit,
     DerivedComponents derivedTimesCases(loc, "case", "times");
     DerivedComponents derivedPowerCases(loc, "case", "power");
 
+    if (productUnit.constantDenominator != 0) {
+        CharString constantString;
+        constantString.appendNumber(productUnit.constantDenominator, status);
+        outArray[CONSTANT_DENOMINATOR_INDEX] = UnicodeString::fromUTF8(constantString.toStringPiece());
+    }
+
     // 4. For each single_unit in product_unit
     for (int32_t singleUnitIndex = 0; singleUnitIndex < productUnit.singleUnits.length();
          singleUnitIndex++) {
@@ -1454,6 +1469,39 @@ void LongNameHandler::processPatternTimes(MeasureUnitImpl &&productUnit,
             }
         }
     }
+
+    // 5. Handling constant denominator if it exists.
+    if (productUnit.constantDenominator != 0) {
+        int32_t pluralIndex = -1;
+        for (int32_t index = 0; index < StandardPlural::Form::COUNT; index++) {
+            if (!outArray[index].isBogus()) {
+                pluralIndex = index;
+                break;
+            }
+        }
+
+        U_ASSERT(pluralIndex >= 0); // "No plural form found for constant denominator"
+
+        // TODO(ICU-23039):
+        // Improve the handling of constant_denominator representation.
+        // For instance, a constant_denominator of 1000000 should be adaptable to
+        // formats like
+        // 1,000,000, 1e6, or 1 million.
+        // Furthermore, ensure consistent pluralization rules for units. For example,
+        // "meter per 100 seconds" should be evaluated for correct singular/plural
+        // usage: "second" or "seconds"?
+        // Similarly, "kilogram per 1000 meters" should be checked for "meter" or
+        // "meters"?
+        if (outArray[pluralIndex].length() == 0) {
+            outArray[pluralIndex] = outArray[CONSTANT_DENOMINATOR_INDEX];
+        } else {
+            UnicodeString tmp;
+            timesPatternFormatter.format(outArray[CONSTANT_DENOMINATOR_INDEX], outArray[pluralIndex],
+                                         tmp, status);
+            outArray[pluralIndex] = tmp;
+        }
+    }
+
     for (int32_t pluralIndex = 0; pluralIndex < StandardPlural::Form::COUNT; pluralIndex++) {
         if (globalPlaceholder[pluralIndex] == PH_BEGINNING) {
             UnicodeString tmp;
diff --git a/deps/icu-small/source/i18n/number_mapper.cpp b/deps/icu-small/source/i18n/number_mapper.cpp
index 2f398d4a9392fb..457fbc0d0712a3 100644
--- a/deps/icu-small/source/i18n/number_mapper.cpp
+++ b/deps/icu-small/source/i18n/number_mapper.cpp
@@ -74,9 +74,11 @@ MacroProps NumberPropertyMapper::oldToNew(const DecimalFormatProperties& propert
             !properties.currencyPluralInfo.fPtr.isNull() ||
             !properties.currencyUsage.isNull() ||
             warehouse.affixProvider.get().hasCurrencySign());
-    CurrencyUnit currency = resolveCurrency(properties, locale, status);
-    UCurrencyUsage currencyUsage = properties.currencyUsage.getOrDefault(UCURR_USAGE_STANDARD);
+    CurrencyUnit currency;
+    UCurrencyUsage currencyUsage;
     if (useCurrency) {
+        currency = resolveCurrency(properties, locale, status);
+        currencyUsage = properties.currencyUsage.getOrDefault(UCURR_USAGE_STANDARD);
         // NOTE: Slicing is OK.
         macros.unit = currency; // NOLINT
     }
@@ -129,6 +131,7 @@ MacroProps NumberPropertyMapper::oldToNew(const DecimalFormatProperties& propert
     }
     Precision precision;
     if (!properties.currencyUsage.isNull()) {
+        U_ASSERT(useCurrency);
         precision = Precision::constructCurrency(currencyUsage).withCurrency(currency);
     } else if (roundingIncrement != 0.0) {
         if (PatternStringUtils::ignoreRoundingIncrement(roundingIncrement, maxFrac)) {
@@ -276,7 +279,7 @@ MacroProps NumberPropertyMapper::oldToNew(const DecimalFormatProperties& propert
         exportedProperties->maximumIntegerDigits = maxInt == -1 ? INT32_MAX : maxInt;
 
         Precision rounding_;
-        if (precision.fType == Precision::PrecisionType::RND_CURRENCY) {
+        if (useCurrency && precision.fType == Precision::PrecisionType::RND_CURRENCY) {
             rounding_ = precision.withCurrency(currency, status);
         } else {
             rounding_ = precision;
diff --git a/deps/icu-small/source/i18n/number_rounding.cpp b/deps/icu-small/source/i18n/number_rounding.cpp
index 8f1aa453ada44c..0f3975393c581b 100644
--- a/deps/icu-small/source/i18n/number_rounding.cpp
+++ b/deps/icu-small/source/i18n/number_rounding.cpp
@@ -278,23 +278,23 @@ Precision IncrementPrecision::withMinFraction(int32_t minFrac) const {
 }
 
 FractionPrecision Precision::constructFraction(int32_t minFrac, int32_t maxFrac) {
-    FractionSignificantSettings settings;
+    FractionSignificantSettings settings{};
     settings.fMinFrac = static_cast<digits_t>(minFrac);
     settings.fMaxFrac = static_cast<digits_t>(maxFrac);
     settings.fMinSig = -1;
     settings.fMaxSig = -1;
-    PrecisionUnion union_;
+    PrecisionUnion union_{};
     union_.fracSig = settings;
     return {RND_FRACTION, union_};
 }
 
 Precision Precision::constructSignificant(int32_t minSig, int32_t maxSig) {
-    FractionSignificantSettings settings;
+    FractionSignificantSettings settings{};
     settings.fMinFrac = -1;
     settings.fMaxFrac = -1;
     settings.fMinSig = static_cast<digits_t>(minSig);
     settings.fMaxSig = static_cast<digits_t>(maxSig);
-    PrecisionUnion union_;
+    PrecisionUnion union_{};
     union_.fracSig = settings;
     return {RND_SIGNIFICANT, union_};
 }
@@ -311,20 +311,20 @@ Precision::constructFractionSignificant(
     settings.fMaxSig = static_cast<digits_t>(maxSig);
     settings.fPriority = priority;
     settings.fRetain = retain;
-    PrecisionUnion union_;
+    PrecisionUnion union_{};
     union_.fracSig = settings;
     return {RND_FRACTION_SIGNIFICANT, union_};
 }
 
 IncrementPrecision Precision::constructIncrement(uint64_t increment, digits_t magnitude) {
-    IncrementSettings settings;
+    IncrementSettings settings{};
     // Note: For number formatting, fIncrement is used for RND_INCREMENT but not
     // RND_INCREMENT_ONE or RND_INCREMENT_FIVE. However, fIncrement is used in all
     // three when constructing a skeleton.
     settings.fIncrement = increment;
     settings.fIncrementMagnitude = magnitude;
     settings.fMinFrac = magnitude > 0 ? 0 : -magnitude;
-    PrecisionUnion union_;
+    PrecisionUnion union_{};
     union_.increment = settings;
     if (increment == 1) {
         // NOTE: In C++, we must return the correct value type with the correct union.
@@ -339,7 +339,7 @@ IncrementPrecision Precision::constructIncrement(uint64_t increment, digits_t ma
 }
 
 CurrencyPrecision Precision::constructCurrency(UCurrencyUsage usage) {
-    PrecisionUnion union_;
+    PrecisionUnion union_{};
     union_.currencyUsage = usage;
     return {RND_CURRENCY, union_};
 }
diff --git a/deps/icu-small/source/i18n/number_skeletons.cpp b/deps/icu-small/source/i18n/number_skeletons.cpp
index 562a8663d05769..67a38dad073766 100644
--- a/deps/icu-small/source/i18n/number_skeletons.cpp
+++ b/deps/icu-small/source/i18n/number_skeletons.cpp
@@ -1015,6 +1015,12 @@ blueprint_helpers::parseExponentSignOption(const StringSegment& segment, MacroPr
     return true;
 }
 
+// The function is called by skeleton::parseOption which called by skeleton::parseSkeleton
+// the data pointed in the return macros.unit is stack allocated in the parseSkeleton function.
+#if U_GCC_MAJOR_MINOR >= 1204
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wdangling-pointer"
+#endif
 void blueprint_helpers::parseCurrencyOption(const StringSegment& segment, MacroProps& macros,
                                             UErrorCode& status) {
     // Unlike ICU4J, have to check length manually because ICU4C CurrencyUnit does not check it for us
@@ -1034,6 +1040,9 @@ void blueprint_helpers::parseCurrencyOption(const StringSegment& segment, MacroP
     // Slicing is OK
     macros.unit = currency; // NOLINT
 }
+#if U_GCC_MAJOR_MINOR >= 1204
+#pragma GCC diagnostic pop
+#endif
 
 void
 blueprint_helpers::generateCurrencyOption(const CurrencyUnit& currency, UnicodeString& sb, UErrorCode&) {
diff --git a/deps/icu-small/source/i18n/olsontz.cpp b/deps/icu-small/source/i18n/olsontz.cpp
index 9d9770dd4224e4..7826d47a7d06f5 100644
--- a/deps/icu-small/source/i18n/olsontz.cpp
+++ b/deps/icu-small/source/i18n/olsontz.cpp
@@ -436,11 +436,11 @@ int32_t OlsonTimeZone::getRawOffset() const {
 
 #if defined U_DEBUG_TZ
 void printTime(double ms) {
-            int32_t year, month, dom, dow;
-            double millis=0;
-            double days = ClockMath::floorDivide(((double)ms), (double)U_MILLIS_PER_DAY, millis);
-            
-            Grego::dayToFields(days, year, month, dom, dow);
+            int32_t year;
+            int8_t month, dom, dow;
+            int32_t millis=0;
+            UErrorCode status = U_ZERO_ERROR;
+            Grego::timeToFields(ms, year, month, dom, dow, millis, status);
             U_DEBUG_TZ_MSG(("   getHistoricalOffset:  time %.1f (%04d.%02d.%02d+%.1fh)\n", ms,
                             year, month+1, dom, (millis/kOneHour)));
     }
@@ -568,9 +568,8 @@ UBool OlsonTimeZone::useDaylightTime() const {
         return finalZone->useDaylightTime();
     }
 
-    int32_t year, month, dom, dow, doy, mid;
     UErrorCode status = U_ZERO_ERROR;
-    Grego::timeToFields(current, year, month, dom, dow, doy, mid, status);
+    int32_t year = Grego::timeToYear(current, status);
     U_ASSERT(U_SUCCESS(status));
     if (U_FAILURE(status)) return false; // If error, just return false.
 
diff --git a/deps/icu-small/source/i18n/persncal.cpp b/deps/icu-small/source/i18n/persncal.cpp
index 31f7ae252b5e8a..792a88807956b8 100644
--- a/deps/icu-small/source/i18n/persncal.cpp
+++ b/deps/icu-small/source/i18n/persncal.cpp
@@ -25,6 +25,9 @@
 #include "umutex.h"
 #include "gregoimp.h" // Math
 #include <float.h>
+#include "cmemory.h"
+#include "ucln_in.h"
+#include "unicode/uniset.h"
 
 static const int16_t kPersianNumDays[]
 = {0,31,62,93,124,155,186,216,246,276,306,336}; // 0-based, for day-in-year
@@ -62,6 +65,45 @@ static const int32_t kPersianCalendarLimits[UCAL_FIELD_COUNT][4] = {
     {        0,        0,       11,       11}, // ORDINAL_MONTH
 };
 
+namespace { // anonymous
+
+static const icu::UnicodeSet *gLeapCorrection = nullptr;
+static icu::UInitOnce gCorrectionInitOnce {};
+static int32_t gMinCorrection;
+}  // namespace
+U_CDECL_BEGIN
+static UBool calendar_persian_cleanup() {
+    if (gLeapCorrection) {
+        delete gLeapCorrection;
+        gLeapCorrection = nullptr;
+    }
+    gCorrectionInitOnce.reset();
+    return true;
+}
+U_CDECL_END
+
+namespace { // anonymous
+static void U_CALLCONV initLeapCorrection() {
+    static int16_t nonLeapYears[] = {
+       1502, 1601, 1634, 1667, 1700, 1733, 1766, 1799, 1832, 1865, 1898, 1931, 1964, 1997, 2030, 2059,
+       2063, 2096, 2129, 2158, 2162, 2191, 2195, 2224, 2228, 2257, 2261, 2290, 2294, 2323, 2327, 2356,
+       2360, 2389, 2393, 2422, 2426, 2455, 2459, 2488, 2492, 2521, 2525, 2554, 2558, 2587, 2591, 2620,
+       2624, 2653, 2657, 2686, 2690, 2719, 2723, 2748, 2752, 2756, 2781, 2785, 2789, 2818, 2822, 2847,
+       2851, 2855, 2880, 2884, 2888, 2913, 2917, 2921, 2946, 2950, 2954, 2979, 2983, 2987,
+    };
+    gMinCorrection = nonLeapYears[0];
+    icu::UnicodeSet prefab;
+    for (auto year : nonLeapYears) {
+        prefab.add(year);
+    }
+    gLeapCorrection = prefab.cloneAsThawed()->freeze();
+    ucln_i18n_registerCleanup(UCLN_I18N_PERSIAN_CALENDAR, calendar_persian_cleanup);
+}
+const icu::UnicodeSet* getLeapCorrection() {
+    umtx_initOnce(gCorrectionInitOnce, &initLeapCorrection);
+    return gLeapCorrection;
+}
+} // namespace anonymous
 U_NAMESPACE_BEGIN
 
 static const int32_t PERSIAN_EPOCH = 1948320;
@@ -83,7 +125,6 @@ PersianCalendar* PersianCalendar::clone() const {
 PersianCalendar::PersianCalendar(const Locale& aLocale, UErrorCode& success)
   :   Calendar(TimeZone::forLocaleOrDefault(aLocale), aLocale, success)
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 PersianCalendar::PersianCalendar(const PersianCalendar& other) : Calendar(other) {
@@ -111,8 +152,15 @@ int32_t PersianCalendar::handleGetLimit(UCalendarDateFields field, ELimitType li
  */
 UBool PersianCalendar::isLeapYear(int32_t year)
 {
+    if (year >= gMinCorrection && getLeapCorrection()->contains(year)) {
+        return false;
+    }
+    if (year > gMinCorrection && getLeapCorrection()->contains(year-1)) {
+        return true;
+    }
     int64_t y = static_cast<int64_t>(year) * 25LL + 11LL;
-    return (y % 33L < 8);
+    bool res = (y % 33L < 8);
+    return res;
 }
     
 /**
@@ -157,7 +205,8 @@ int32_t PersianCalendar::handleGetMonthLength(int32_t extendedYear, int32_t mont
 /**
  * Return the number of days in the given Persian year
  */
-int32_t PersianCalendar::handleGetYearLength(int32_t extendedYear) const {
+int32_t PersianCalendar::handleGetYearLength(int32_t extendedYear, UErrorCode& status) const {
+    if (U_FAILURE(status)) return 0;
     return isLeapYear(extendedYear) ? 366 : 365;
 }
     
@@ -165,6 +214,15 @@ int32_t PersianCalendar::handleGetYearLength(int32_t extendedYear) const {
 // Functions for converting from field values to milliseconds....
 //-------------------------------------------------------------------------
 
+static int64_t firstJulianOfYear(int64_t year) {
+    int64_t julianDay = 365LL * (year - 1LL) + ClockMath::floorDivide(8LL * year + 21, 33);
+    if (year > gMinCorrection && getLeapCorrection()->contains(year-1)) {
+        julianDay--;
+    }
+    return julianDay;
+}
+
+
 // Return JD of start of given month/year
 int64_t PersianCalendar::handleComputeMonthStart(int32_t eyear, int32_t month, UBool /*useMonth*/, UErrorCode& status) const {
     if (U_FAILURE(status)) {
@@ -179,7 +237,7 @@ int64_t PersianCalendar::handleComputeMonthStart(int32_t eyear, int32_t month, U
         }
     }
 
-    int64_t julianDay = PERSIAN_EPOCH - 1LL + 365LL * (eyear - 1LL) + ClockMath::floorDivide(8LL * eyear + 21, 33);
+    int64_t julianDay = PERSIAN_EPOCH - 1LL + firstJulianOfYear(eyear);
 
     if (month != 0) {
         julianDay += kPersianNumDays[month];
@@ -219,6 +277,7 @@ int32_t PersianCalendar::handleGetExtendedYear(UErrorCode& status) {
 void PersianCalendar::handleComputeFields(int32_t julianDay, UErrorCode& status) {
     int64_t daysSinceEpoch = julianDay;
     daysSinceEpoch -= PERSIAN_EPOCH;
+
     int64_t year = ClockMath::floorDivideInt64(
         33LL * daysSinceEpoch + 3LL, 12053LL) + 1LL;
     if (year > INT32_MAX || year < INT32_MIN) {
@@ -226,11 +285,16 @@ void PersianCalendar::handleComputeFields(int32_t julianDay, UErrorCode& status)
         return;
     }
 
-    int64_t farvardin1 = 365LL * (year - 1) + ClockMath::floorDivide(8LL * year + 21, 33);
+    int64_t farvardin1 = firstJulianOfYear(year);
+
     int32_t dayOfYear = daysSinceEpoch - farvardin1; // 0-based
     U_ASSERT(dayOfYear >= 0);
     U_ASSERT(dayOfYear < 366);
-                                                     //
+
+    if (dayOfYear == 365 && year >= gMinCorrection && getLeapCorrection()->contains(year)) {
+        year++;
+        dayOfYear = 0;
+    }
     int32_t month;
     if (dayOfYear < 216) { // Compute 0-based month
         month = dayOfYear / 31;
@@ -240,11 +304,11 @@ void PersianCalendar::handleComputeFields(int32_t julianDay, UErrorCode& status)
     U_ASSERT(month >= 0);
     U_ASSERT(month < 12);
 
-    int32_t dayOfMonth = dayOfYear - kPersianNumDays[month] + 1;
+    ++dayOfYear; // Make it 1-based now
+    int32_t dayOfMonth = dayOfYear - kPersianNumDays[month];
     U_ASSERT(dayOfMonth > 0);
     U_ASSERT(dayOfMonth <= 31);
 
-    ++dayOfYear; // Make it 1-based now
 
     internalSet(UCAL_ERA, 0);
     internalSet(UCAL_YEAR, year);
diff --git a/deps/icu-small/source/i18n/persncal.h b/deps/icu-small/source/i18n/persncal.h
index daf7508b702ae4..d5bff7b232740b 100644
--- a/deps/icu-small/source/i18n/persncal.h
+++ b/deps/icu-small/source/i18n/persncal.h
@@ -209,7 +209,7 @@ class PersianCalendar : public Calendar {
    * Return the number of days in the given Persian year
    * @internal
    */
-  virtual int32_t handleGetYearLength(int32_t extendedYear) const override;
+  virtual int32_t handleGetYearLength(int32_t extendedYear, UErrorCode& status) const override;
     
   //-------------------------------------------------------------------------
   // Functions for converting from field values to milliseconds....
diff --git a/deps/icu-small/source/i18n/rbnf.cpp b/deps/icu-small/source/i18n/rbnf.cpp
index c4e8ff73a7cc0c..5b6b5e2c189282 100644
--- a/deps/icu-small/source/i18n/rbnf.cpp
+++ b/deps/icu-small/source/i18n/rbnf.cpp
@@ -1568,12 +1568,12 @@ RuleBasedNumberFormat::init(const UnicodeString& rules, LocalizationInfo* locali
 
     // divide up the descriptions into individual rule-set descriptions
     // and store them in a temporary array.  At each step, we also
-    // new up a rule set, but all this does is initialize its name
+    // create a rule set, but all this does is initialize its name
     // and remove it from its description.  We can't actually parse
     // the rest of the descriptions and finish initializing everything
     // because we have to know the names and locations of all the rule
     // sets before we can actually set everything up
-    if(!numRuleSets) {
+    if (!numRuleSets) {
         status = U_ILLEGAL_ARGUMENT_ERROR;
         return;
     }
@@ -1616,9 +1616,9 @@ RuleBasedNumberFormat::init(const UnicodeString& rules, LocalizationInfo* locali
     // last public rule set, no matter what the localization data says.
     initDefaultRuleSet();
 
-    // finally, we can go back through the temporary descriptions
-    // list and finish setting up the substructure (and we throw
-    // away the temporary descriptions as we go)
+    // Now that we know all the rule names, we can go back through
+    // the temporary descriptions list and finish setting up the substructure
+    // (and we throw away the temporary descriptions as we go)
     {
         for (int i = 0; i < numRuleSets; i++) {
             fRuleSets[i]->parseRules(ruleSetDescriptions[i], status);
@@ -1706,10 +1706,13 @@ RuleBasedNumberFormat::stripWhitespace(UnicodeString& description)
     UnicodeString result;
 
     int start = 0;
-    while (start != -1 && start < description.length()) {
-        // seek to the first non-whitespace character...
+    UChar ch;
+    while (start < description.length()) {
+        // Seek to the first non-whitespace character...
+        // If the first non-whitespace character is semicolon, skip it and continue
         while (start < description.length()
-            && PatternProps::isWhiteSpace(description.charAt(start))) {
+            && (PatternProps::isWhiteSpace(ch = description.charAt(start)) || ch == gSemiColon))
+        {
             ++start;
         }
 
@@ -1720,20 +1723,16 @@ RuleBasedNumberFormat::stripWhitespace(UnicodeString& description)
             // or if we don't find a semicolon, just copy the rest of
             // the string into the result
             result.append(description, start, description.length() - start);
-            start = -1;
+            break;
         }
         else if (p < description.length()) {
             result.append(description, start, p + 1 - start);
             start = p + 1;
         }
-
-        // when we get here, we've seeked off the end of the string, and
+        // when we get here from the else, we've seeked off the end of the string, and
         // we terminate the loop (we continue until *start* is -1 rather
         // than until *p* is -1, because otherwise we'd miss the last
         // rule in the description)
-        else {
-            start = -1;
-        }
     }
 
     description.setTo(result);
diff --git a/deps/icu-small/source/i18n/scriptset.cpp b/deps/icu-small/source/i18n/scriptset.cpp
index eec1eeb37dafbf..576917e81c4196 100644
--- a/deps/icu-small/source/i18n/scriptset.cpp
+++ b/deps/icu-small/source/i18n/scriptset.cpp
@@ -285,19 +285,19 @@ uhash_equalsScriptSet(const UElement key1, const UElement key2) {
     return (*s1 == *s2);
 }
 
-U_CAPI int8_t U_EXPORT2
+U_CAPI int32_t U_EXPORT2
 uhash_compareScriptSet(UElement key0, UElement key1) {
     icu::ScriptSet *s0 = static_cast<icu::ScriptSet *>(key0.pointer);
     icu::ScriptSet *s1 = static_cast<icu::ScriptSet *>(key1.pointer);
     int32_t diff = s0->countMembers() - s1->countMembers();
-    if (diff != 0) return static_cast<UBool>(diff);
+    if (diff != 0) return diff;
     int32_t i0 = s0->nextSetBit(0);
     int32_t i1 = s1->nextSetBit(0);
     while ((diff = i0-i1) == 0 && i0 > 0) {
         i0 = s0->nextSetBit(i0+1);
         i1 = s1->nextSetBit(i1+1);
     }
-    return (int8_t)diff;
+    return diff;
 }
 
 U_CAPI int32_t U_EXPORT2
diff --git a/deps/icu-small/source/i18n/scriptset.h b/deps/icu-small/source/i18n/scriptset.h
index df5cfdc7486890..d21d0db8a0144b 100644
--- a/deps/icu-small/source/i18n/scriptset.h
+++ b/deps/icu-small/source/i18n/scriptset.h
@@ -74,7 +74,7 @@ class U_I18N_API ScriptSet: public UMemory {
 
 U_NAMESPACE_END
 
-U_CAPI UBool U_EXPORT2
+U_CAPI int32_t U_EXPORT2
 uhash_compareScriptSet(const UElement key1, const UElement key2);
 
 U_CAPI int32_t U_EXPORT2
diff --git a/deps/icu-small/source/i18n/simpletz.cpp b/deps/icu-small/source/i18n/simpletz.cpp
index cbefc29830ffd9..3f3b236ea45ca9 100644
--- a/deps/icu-small/source/i18n/simpletz.cpp
+++ b/deps/icu-small/source/i18n/simpletz.cpp
@@ -518,15 +518,10 @@ SimpleTimeZone::getOffsetFromLocal(UDate date, UTimeZoneLocalOption nonExistingT
     }
 
     rawOffsetGMT = getRawOffset();
-    int32_t year, month, dom, dow, millis;
-    double dday = ClockMath::floorDivide(date, U_MILLIS_PER_DAY, &millis);
-    if (dday > INT32_MAX || dday < INT32_MIN) {
-        status = U_ILLEGAL_ARGUMENT_ERROR;
-        return;
-    }
-    int32_t day = dday;
+    int32_t year, millis;
+    int8_t month, dom, dow;
 
-    Grego::dayToFields(day, year, month, dom, dow, status);
+    Grego::timeToFields(date, year, month, dom, dow, millis, status);
     if (U_FAILURE(status)) return;
 
     savingsDST = getOffset(GregorianCalendar::AD, year, month, dom,
@@ -554,8 +549,7 @@ SimpleTimeZone::getOffsetFromLocal(UDate date, UTimeZoneLocalOption nonExistingT
         }
     }
     if (recalc) {
-        day = ClockMath::floorDivide(date, U_MILLIS_PER_DAY, &millis);
-        Grego::dayToFields(day, year, month, dom, dow, status);
+        Grego::timeToFields(date, year, month, dom, dow, millis, status);
         if (U_FAILURE(status)) return;
         savingsDST = getOffset(GregorianCalendar::AD, year, month, dom,
                           static_cast<uint8_t>(dow), millis,
diff --git a/deps/icu-small/source/i18n/smpdtfmt.cpp b/deps/icu-small/source/i18n/smpdtfmt.cpp
index f79d4ae49532b0..3c13d5a413fa54 100644
--- a/deps/icu-small/source/i18n/smpdtfmt.cpp
+++ b/deps/icu-small/source/i18n/smpdtfmt.cpp
@@ -77,6 +77,10 @@
 #include "dayperiodrules.h"
 #include "tznames_impl.h"   // ZONE_NAME_U16_MAX
 #include "number_utypes.h"
+#include "chnsecal.h"
+#include "dangical.h"
+#include "japancal.h"
+#include <typeinfo>
 
 #if defined( U_DEBUG_CALSVC ) || defined (U_DEBUG_CAL)
 #include <stdio.h>
@@ -945,7 +949,8 @@ SimpleDateFormat::initialize(const Locale& locale,
     // if format is non-numeric (includes 年) and fDateOverride is not already specified.
     // Now this does get updated if applyPattern subsequently changes the pattern type.
     if (fDateOverride.isBogus() && fHasHanYearChar &&
-            fCalendar != nullptr && uprv_strcmp(fCalendar->getType(),"japanese") == 0 &&
+            fCalendar != nullptr &&
+            typeid(*fCalendar) == typeid(JapaneseCalendar) &&
             uprv_strcmp(fLocale.getLanguage(),"ja") == 0) {
         fDateOverride.setTo(u"y=jpanyear", -1);
     }
@@ -1050,7 +1055,7 @@ SimpleDateFormat::_format(Calendar& cal, UnicodeString& appendTo,
     }
     Calendar* workCal = &cal;
     Calendar* calClone = nullptr;
-    if (&cal != fCalendar && uprv_strcmp(cal.getType(), fCalendar->getType()) != 0) {
+    if (&cal != fCalendar && typeid(cal) != typeid(*fCalendar)) {
         // Different calendar type
         // We use the time and time zone from the input calendar, but
         // do not use the input calendar for field calculation.
@@ -1523,8 +1528,8 @@ SimpleDateFormat::subFormat(UnicodeString &appendTo,
     // "GGGG" is wide era name, "GGGGG" is narrow era name, anything else is abbreviated name
     case UDAT_ERA_FIELD:
         {
-            const auto* calType = cal.getType();
-            if (uprv_strcmp(calType,"chinese") == 0 || uprv_strcmp(calType,"dangi") == 0) {
+            if (typeid(cal) == typeid(ChineseCalendar) ||
+                typeid(cal) == typeid(DangiCalendar)) {
                 zeroPaddingNumber(currentNumberFormat,appendTo, value, 1, 9); // as in ICU4J
             } else {
                 if (count == 5) {
@@ -1575,7 +1580,7 @@ SimpleDateFormat::subFormat(UnicodeString &appendTo,
     // for "MMMMM"/"LLLLL", use the narrow form
     case UDAT_MONTH_FIELD:
     case UDAT_STANDALONE_MONTH_FIELD:
-        if (uprv_strcmp(cal.getType(),"hebrew") == 0) {
+        if (typeid(cal) == typeid(HebrewCalendar)) {
            if (HebrewCalendar::isLeapYear(cal.get(UCAL_YEAR,status)) && value == 6 && count >= 3 )
                value = 13; // Show alternate form for Adar II in leap years in Hebrew calendar.
            if (!HebrewCalendar::isLeapYear(cal.get(UCAL_YEAR,status)) && value >= 6 && count < 3 )
@@ -2272,7 +2277,7 @@ SimpleDateFormat::parse(const UnicodeString& text, Calendar& cal, ParsePosition&
 
     Calendar* calClone = nullptr;
     Calendar *workCal = &cal;
-    if (&cal != fCalendar && uprv_strcmp(cal.getType(), fCalendar->getType()) != 0) {
+    if (&cal != fCalendar && typeid(cal) != typeid(*fCalendar)) {
         // Different calendar type
         // We use the time/zone from the input calendar, but
         // do not use the input calendar for field calculation.
@@ -2903,7 +2908,7 @@ int32_t SimpleDateFormat::matchAlphaMonthStrings(const UnicodeString& text,
 
     if (bestMatch >= 0) { 
         // Adjustment for Hebrew Calendar month Adar II
-        if (!strcmp(cal.getType(),"hebrew") && bestMatch==13) {
+        if (typeid(cal) == typeid(HebrewCalendar) && bestMatch==13) {
             cal.set(UCAL_MONTH,6);
         } else {
             cal.set(UCAL_MONTH, bestMatch);
@@ -2963,7 +2968,7 @@ int32_t SimpleDateFormat::matchString(const UnicodeString& text,
     if (bestMatch >= 0) {
         if (field < UCAL_FIELD_COUNT) {
             // Adjustment for Hebrew Calendar month Adar II
-            if (!strcmp(cal.getType(),"hebrew") && field==UCAL_MONTH && bestMatch==13) {
+            if (typeid(cal) == typeid(HebrewCalendar) && field==UCAL_MONTH && bestMatch==13) {
                 cal.set(field,6);
             } else {
                 if (field == UCAL_YEAR) {
@@ -3052,7 +3057,6 @@ int32_t SimpleDateFormat::subParse(const UnicodeString& text, int32_t& start, ch
     if (numericLeapMonthFormatter != nullptr) {
         numericLeapMonthFormatter->setFormats(reinterpret_cast<const Format**>(&currentNumberFormat), 1);
     }
-    UBool isChineseCalendar = (uprv_strcmp(cal.getType(),"chinese") == 0 || uprv_strcmp(cal.getType(),"dangi") == 0);
 
     // If there are any spaces here, skip over them.  If we hit the end
     // of the string, then fail.
@@ -3068,6 +3072,8 @@ int32_t SimpleDateFormat::subParse(const UnicodeString& text, int32_t& start, ch
     }
     pos.setIndex(start);
 
+    UBool isChineseCalendar = typeid(cal) == typeid(ChineseCalendar) ||
+            typeid(cal) == typeid(DangiCalendar);
     // We handle a few special cases here where we need to parse
     // a number value.  We handle further, more generic cases below.  We need
     // to handle some of them here because some fields require extra processing on
@@ -3289,7 +3295,7 @@ int32_t SimpleDateFormat::subParse(const UnicodeString& text, int32_t& start, ch
             // When parsing month numbers from the Hebrew Calendar, we might need to adjust the month depending on whether
             // or not it was a leap year.  We may or may not yet know what year it is, so might have to delay checking until
             // the year is parsed.
-            if (!strcmp(cal.getType(),"hebrew")) {
+            if (typeid(cal) == typeid(HebrewCalendar)) {
                 HebrewCalendar *hc = (HebrewCalendar*)&cal;
                 if (cal.isSet(UCAL_YEAR)) {
                    UErrorCode monthStatus = U_ZERO_ERROR;
@@ -3852,7 +3858,7 @@ int32_t SimpleDateFormat::subParse(const UnicodeString& text, int32_t& start, ch
         switch (patternCharIndex) {
         case UDAT_MONTH_FIELD:
             // See notes under UDAT_MONTH_FIELD case above
-            if (!strcmp(cal.getType(),"hebrew")) {
+            if (typeid(cal) == typeid(HebrewCalendar)) {
                 HebrewCalendar *hc = (HebrewCalendar*)&cal;
                 if (cal.isSet(UCAL_YEAR)) {
                    UErrorCode monthStatus = U_ZERO_ERROR;
@@ -4034,7 +4040,7 @@ SimpleDateFormat::applyPattern(const UnicodeString& pattern)
 
     // Hack to update use of Gannen year numbering for ja@calendar=japanese -
     // use only if format is non-numeric (includes 年) and no other fDateOverride.
-    if (fCalendar != nullptr && uprv_strcmp(fCalendar->getType(),"japanese") == 0 &&
+    if (fCalendar != nullptr && typeid(*fCalendar) == typeid(JapaneseCalendar) &&
             uprv_strcmp(fLocale.getLanguage(),"ja") == 0) {
         if (fDateOverride==UnicodeString(u"y=jpanyear") && !fHasHanYearChar) {
             // Gannen numbering is set but new pattern should not use it, unset;
diff --git a/deps/icu-small/source/i18n/taiwncal.cpp b/deps/icu-small/source/i18n/taiwncal.cpp
index e6ffd71ba3089a..1f948031002205 100644
--- a/deps/icu-small/source/i18n/taiwncal.cpp
+++ b/deps/icu-small/source/i18n/taiwncal.cpp
@@ -36,7 +36,6 @@ static const int32_t kGregorianEpoch = 1970;
 TaiwanCalendar::TaiwanCalendar(const Locale& aLocale, UErrorCode& success)
 :   GregorianCalendar(aLocale, success)
 {
-    setTimeInMillis(getNow(), success); // Call this again now that the vtable is set up properly.
 }
 
 TaiwanCalendar::~TaiwanCalendar()
@@ -48,12 +47,6 @@ TaiwanCalendar::TaiwanCalendar(const TaiwanCalendar& source)
 {
 }
 
-TaiwanCalendar& TaiwanCalendar::operator= ( const TaiwanCalendar& right)
-{
-    GregorianCalendar::operator=(right);
-    return *this;
-}
-
 TaiwanCalendar* TaiwanCalendar::clone() const
 {
     return new TaiwanCalendar(*this);
diff --git a/deps/icu-small/source/i18n/taiwncal.h b/deps/icu-small/source/i18n/taiwncal.h
index a8fa3d1829ed2e..d12ab1c4a2ee6d 100644
--- a/deps/icu-small/source/i18n/taiwncal.h
+++ b/deps/icu-small/source/i18n/taiwncal.h
@@ -79,13 +79,6 @@ class TaiwanCalendar : public GregorianCalendar {
      */
     TaiwanCalendar(const TaiwanCalendar& source);
 
-    /**
-     * Default assignment operator
-     * @param right    the object to be copied.
-     * @internal
-     */
-    TaiwanCalendar& operator=(const TaiwanCalendar& right);
-
     /**
      * Create and return a polymorphic copy of this calendar.
      * @return    return a polymorphic copy of this calendar.
diff --git a/deps/icu-small/source/i18n/timezone.cpp b/deps/icu-small/source/i18n/timezone.cpp
index 118dfe2f2af629..8028c4ecf5cb96 100644
--- a/deps/icu-small/source/i18n/timezone.cpp
+++ b/deps/icu-small/source/i18n/timezone.cpp
@@ -730,15 +730,9 @@ void TimeZone::getOffset(UDate date, UBool local, int32_t& rawOffset,
     // (with 7 args) twice when local == true and DST is
     // detected in the initial call.
     for (int32_t pass=0; ; ++pass) {
-        int32_t year, month, dom, dow, millis;
-        double day = ClockMath::floorDivide(date, U_MILLIS_PER_DAY, &millis);
-
-        // out of the range
-        if (day < INT32_MIN || day > INT32_MAX) {
-            ec = U_ILLEGAL_ARGUMENT_ERROR;
-            return;
-        }
-        Grego::dayToFields(day, year, month, dom, dow, ec);
+        int32_t year, millis;
+        int8_t month, dom, dow;
+        Grego::timeToFields(date, year, month, dom, dow, millis, ec);
         if (U_FAILURE(ec)) return;
 
         dstOffset = getOffset(GregorianCalendar::AD, year, month, dom,
@@ -1057,7 +1051,7 @@ TimeZone::countEquivalentIDs(const UnicodeString& id) {
 
 // ---------------------------------------
 
-const UnicodeString U_EXPORT2
+UnicodeString U_EXPORT2
 TimeZone::getEquivalentID(const UnicodeString& id, int32_t index) {
     U_DEBUG_TZ_MSG(("gEI(%d)\n", index));
     UnicodeString result;
diff --git a/deps/icu-small/source/i18n/tzgnames.cpp b/deps/icu-small/source/i18n/tzgnames.cpp
index 57ee984ee9ba35..e7b09efd01a76e 100644
--- a/deps/icu-small/source/i18n/tzgnames.cpp
+++ b/deps/icu-small/source/i18n/tzgnames.cpp
@@ -406,7 +406,7 @@ TZGNCore::initialize(const Locale& locale, UErrorCode& status) {
     int32_t regionLen = static_cast<int32_t>(uprv_strlen(region));
     if (regionLen == 0) {
         CharString loc = ulocimp_addLikelySubtags(fLocale.getName(), status);
-        ulocimp_getSubtags(loc.data(), nullptr, nullptr, &fTargetRegion, nullptr, nullptr, status);
+        ulocimp_getSubtags(loc.toStringPiece(), nullptr, nullptr, &fTargetRegion, nullptr, nullptr, status);
         if (U_FAILURE(status)) {
             cleanup();
             return;
diff --git a/deps/icu-small/source/i18n/tznames_impl.cpp b/deps/icu-small/source/i18n/tznames_impl.cpp
index 9b7ade7f0bb12f..769c9a6526c587 100644
--- a/deps/icu-small/source/i18n/tznames_impl.cpp
+++ b/deps/icu-small/source/i18n/tznames_impl.cpp
@@ -2149,7 +2149,7 @@ TZDBTimeZoneNames::TZDBTimeZoneNames(const Locale& locale)
     if (regionLen == 0) {
         UErrorCode status = U_ZERO_ERROR;
         CharString loc = ulocimp_addLikelySubtags(fLocale.getName(), status);
-        ulocimp_getSubtags(loc.data(), nullptr, nullptr, &fRegion, nullptr, nullptr, status);
+        ulocimp_getSubtags(loc.toStringPiece(), nullptr, nullptr, &fRegion, nullptr, nullptr, status);
         if (U_SUCCESS(status)) {
             useWorld = false;
         }
diff --git a/deps/icu-small/source/i18n/tzrule.cpp b/deps/icu-small/source/i18n/tzrule.cpp
index 7507068c8807d8..8d6a37a844cbd8 100644
--- a/deps/icu-small/source/i18n/tzrule.cpp
+++ b/deps/icu-small/source/i18n/tzrule.cpp
@@ -355,9 +355,8 @@ AnnualTimeZoneRule::getNextStart(UDate base,
                                  int32_t prevDSTSavings,
                                  UBool inclusive,
                                  UDate& result) const {
-    int32_t year, month, dom, dow, doy, mid;
     UErrorCode status = U_ZERO_ERROR;
-    Grego::timeToFields(base, year, month, dom, dow, doy, mid, status);
+    int32_t year = Grego::timeToYear(base, status);
     U_ASSERT(U_SUCCESS(status));
     if (year < fStartYear) {
         return getFirstStart(prevRawOffset, prevDSTSavings, result);
@@ -381,9 +380,8 @@ AnnualTimeZoneRule::getPreviousStart(UDate base,
                                      int32_t prevDSTSavings,
                                      UBool inclusive,
                                      UDate& result) const {
-    int32_t year, month, dom, dow, doy, mid;
     UErrorCode status = U_ZERO_ERROR;
-    Grego::timeToFields(base, year, month, dom, dow, doy, mid, status);
+    int32_t year = Grego::timeToYear(base, status);
     U_ASSERT(U_SUCCESS(status));
     if (year > fEndYear) {
         return getFinalStart(prevRawOffset, prevDSTSavings, result);
diff --git a/deps/icu-small/source/i18n/ucln_in.h b/deps/icu-small/source/i18n/ucln_in.h
index 765cdd559fb4e2..74868891c83744 100644
--- a/deps/icu-small/source/i18n/ucln_in.h
+++ b/deps/icu-small/source/i18n/ucln_in.h
@@ -39,6 +39,7 @@ typedef enum ECleanupI18NType {
     UCLN_I18N_HEBREW_CALENDAR,
     UCLN_I18N_ASTRO_CALENDAR,
     UCLN_I18N_DANGI_CALENDAR,
+    UCLN_I18N_PERSIAN_CALENDAR,
     UCLN_I18N_CALENDAR,
     UCLN_I18N_TIMEZONEFORMAT,
     UCLN_I18N_TZDBTIMEZONENAMES,
@@ -62,6 +63,7 @@ typedef enum ECleanupI18NType {
     UCLN_I18N_REGION,
     UCLN_I18N_LIST_FORMATTER,
     UCLN_I18N_NUMSYS,
+    UCLN_I18N_MF2_UNISETS,
     UCLN_I18N_COUNT /* This must be last */
 } ECleanupI18NType;
 
diff --git a/deps/icu-small/source/i18n/ucol_sit.cpp b/deps/icu-small/source/i18n/ucol_sit.cpp
index 87387f879d8392..f8fa02fad71997 100644
--- a/deps/icu-small/source/i18n/ucol_sit.cpp
+++ b/deps/icu-small/source/i18n/ucol_sit.cpp
@@ -450,7 +450,7 @@ ucol_prepareShortStringOpen( const char *definition,
     ucol_sit_readSpecs(&s, definition, parseError, status);
     ucol_sit_calculateWholeLocale(&s, *status);
 
-    CharString buffer = ulocimp_canonicalize(s.locale.data(), *status);
+    CharString buffer = ulocimp_canonicalize(s.locale.toStringPiece(), *status);
 
     UResourceBundle *b = ures_open(U_ICUDATA_COLL, buffer.data(), status);
     /* we try to find stuff from keyword */
@@ -514,7 +514,7 @@ ucol_openFromShortString( const char *definition,
 #ifdef UCOL_TRACE_SIT
     fprintf(stderr, "DEF %s, DATA %s, ERR %s\n", definition, s.locale.data(), u_errorName(*status));
 #endif
-    CharString buffer = ulocimp_canonicalize(s.locale.data(), *status);
+    CharString buffer = ulocimp_canonicalize(s.locale.toStringPiece(), *status);
 
     UCollator *result = ucol_open(buffer.data(), status);
     int32_t i = 0;
diff --git a/deps/icu-small/source/i18n/unicode/calendar.h b/deps/icu-small/source/i18n/unicode/calendar.h
index a04f5b65bd5ba6..4499e281f9c55e 100644
--- a/deps/icu-small/source/i18n/unicode/calendar.h
+++ b/deps/icu-small/source/i18n/unicode/calendar.h
@@ -55,6 +55,7 @@ class ICUServiceFactory;
 typedef int32_t UFieldResolutionTable[12][8];
 
 class BasicTimeZone;
+class CharString;
 /**
  * `Calendar` is an abstract base class for converting between
  * a `UDate` object and a set of integer fields such as
@@ -1692,10 +1693,9 @@ class U_I18N_API Calendar : public UObject {
      * calendar system.  Subclasses should override this method if they can
      * provide a more correct or more efficient implementation than the
      * default implementation in Calendar.
-     * @stable ICU 2.0
+     * @internal
      */
-    virtual int32_t handleGetYearLength(int32_t eyear) const;
-
+    virtual int32_t handleGetYearLength(int32_t eyear, UErrorCode& status) const;
 
     /**
      * Return the extended year defined by the current fields.  This will
@@ -1881,42 +1881,7 @@ class U_I18N_API Calendar : public UObject {
      */
     int32_t getActualHelper(UCalendarDateFields field, int32_t startValue, int32_t endValue, UErrorCode &status) const;
 
-
 protected:
-    /**
-     * The flag which indicates if the current time is set in the calendar.
-     * @stable ICU 2.0
-     */
-    UBool      fIsTimeSet;
-
-    /**
-     * True if the fields are in sync with the currently set time of this Calendar.
-     * If false, then the next attempt to get the value of a field will
-     * force a recomputation of all fields from the current value of the time
-     * field.
-     * <P>
-     * This should really be named areFieldsInSync, but the old name is retained
-     * for backward compatibility.
-     * @stable ICU 2.0
-     */
-    UBool      fAreFieldsSet;
-
-    /**
-     * True if all of the fields have been set.  This is initially false, and set to
-     * true by computeFields().
-     * @stable ICU 2.0
-     */
-    UBool      fAreAllFieldsSet;
-
-    /**
-     * True if all fields have been virtually set, but have not yet been
-     * computed.  This occurs only in setTimeInMillis().  A calendar set
-     * to this state will compute all fields from the time if it becomes
-     * necessary, but otherwise will delay such computation.
-     * @stable ICU 3.0
-     */
-    UBool fAreFieldsVirtuallySet;
-
     /**
      * Get the current time without recomputing.
      *
@@ -1940,14 +1905,7 @@ class U_I18N_API Calendar : public UObject {
      */
     int32_t     fFields[UCAL_FIELD_COUNT];
 
-#ifndef U_FORCE_HIDE_DEPRECATED_API
-    /**
-     * The flags which tell if a specified time field for the calendar is set.
-     * @deprecated ICU 2.8 use (fStamp[n]!=kUnset)
-     */
-    UBool      fIsSet[UCAL_FIELD_COUNT];
-#endif  // U_FORCE_HIDE_DEPRECATED_API
-
+protected:
     /** Special values of stamp[]
      * @stable ICU 2.0
      */
@@ -1957,14 +1915,15 @@ class U_I18N_API Calendar : public UObject {
         kMinimumUserStamp
     };
 
+private:
     /**
      * Pseudo-time-stamps which specify when each field was set. There
      * are two special values, UNSET and INTERNALLY_SET. Values from
-     * MINIMUM_USER_SET to Integer.MAX_VALUE are legal user set values.
-     * @stable ICU 2.0
+     * MINIMUM_USER_SET to STAMP_MAX are legal user set values.
      */
-    int32_t        fStamp[UCAL_FIELD_COUNT];
+    int8_t        fStamp[UCAL_FIELD_COUNT];
 
+protected:
     /**
      * Subclasses may override this method to compute several fields
      * specific to each calendar system.  These are:
@@ -2178,7 +2137,7 @@ class U_I18N_API Calendar : public UObject {
     /**
      * The next available value for fStamp[]
      */
-    int32_t fNextStamp;// = MINIMUM_USER_STAMP;
+    int8_t fNextStamp = kMinimumUserStamp;
 
     /**
      * Recalculates the time stamp array (fStamp).
@@ -2189,30 +2148,60 @@ class U_I18N_API Calendar : public UObject {
     /**
      * The current time set for the calendar.
      */
-    UDate        fTime;
+    UDate        fTime = 0;
 
     /**
-     * @see   #setLenient
+     * Time zone affects the time calculation done by Calendar. Calendar subclasses use
+     * the time zone data to produce the local time. Always set; never nullptr.
      */
-    UBool      fLenient;
+    TimeZone*   fZone = nullptr;
 
     /**
-     * Time zone affects the time calculation done by Calendar. Calendar subclasses use
-     * the time zone data to produce the local time. Always set; never nullptr.
+     * The flag which indicates if the current time is set in the calendar.
+     */
+    bool      fIsTimeSet:1;
+
+    /**
+     * True if the fields are in sync with the currently set time of this Calendar.
+     * If false, then the next attempt to get the value of a field will
+     * force a recomputation of all fields from the current value of the time
+     * field.
+     * <P>
+     * This should really be named areFieldsInSync, but the old name is retained
+     * for backward compatibility.
+     */
+    bool      fAreFieldsSet:1;
+
+    /**
+     * True if all of the fields have been set.  This is initially false, and set to
+     * true by computeFields().
+     */
+    bool      fAreAllFieldsSet:1;
+
+    /**
+     * True if all fields have been virtually set, but have not yet been
+     * computed.  This occurs only in setTimeInMillis().  A calendar set
+     * to this state will compute all fields from the time if it becomes
+     * necessary, but otherwise will delay such computation.
      */
-    TimeZone*   fZone;
+    bool      fAreFieldsVirtuallySet:1;
+
+    /**
+     * @see   #setLenient
+     */
+    bool      fLenient:1;
 
     /**
      * Option for repeated wall time
      * @see #setRepeatedWallTimeOption
      */
-    UCalendarWallTimeOption fRepeatedWallTime;
+    UCalendarWallTimeOption fRepeatedWallTime:3; // Somehow MSVC need 3 bits for UCalendarWallTimeOption
 
     /**
      * Option for skipped wall time
      * @see #setSkippedWallTimeOption
      */
-    UCalendarWallTimeOption fSkippedWallTime;
+    UCalendarWallTimeOption fSkippedWallTime:3; // Somehow MSVC need 3 bits for UCalendarWallTimeOption
 
     /**
      * Both firstDayOfWeek and minimalDaysInFirstWeek are locale-dependent. They are
@@ -2222,11 +2211,14 @@ class U_I18N_API Calendar : public UObject {
      * out the week count for a specific date for a given locale. These must be set when
      * a Calendar is constructed.
      */
-    UCalendarDaysOfWeek fFirstDayOfWeek;
-    uint8_t     fMinimalDaysInFirstWeek;
-    UCalendarDaysOfWeek fWeekendOnset;
+    UCalendarDaysOfWeek fFirstDayOfWeek:4; // Somehow MSVC need 4 bits for
+                                           // UCalendarDaysOfWeek
+    UCalendarDaysOfWeek fWeekendOnset:4; // Somehow MSVC need 4 bits for
+                                         // UCalendarDaysOfWeek
+    UCalendarDaysOfWeek fWeekendCease:4; // Somehow MSVC need 4 bits for
+                                         // UCalendarDaysOfWeek
+    uint8_t fMinimalDaysInFirstWeek;
     int32_t fWeekendOnsetMillis;
-    UCalendarDaysOfWeek fWeekendCease;
     int32_t fWeekendCeaseMillis;
 
     /**
@@ -2264,32 +2256,24 @@ class U_I18N_API Calendar : public UObject {
      * returned by getGregorianMonth().
      * @see #computeGregorianFields
      */
-    int32_t fGregorianMonth;
+    int8_t fGregorianMonth;
 
     /**
-     * The Gregorian day of the year, as computed by
-     * computeGregorianFields() and returned by getGregorianDayOfYear().
+     * The Gregorian day of the month, as computed by
+     * computeGregorianFields() and returned by getGregorianDayOfMonth().
      * @see #computeGregorianFields
      */
-    int32_t fGregorianDayOfYear;
+    int8_t fGregorianDayOfMonth;
 
     /**
-     * The Gregorian day of the month, as computed by
-     * computeGregorianFields() and returned by getGregorianDayOfMonth().
+     * The Gregorian day of the year, as computed by
+     * computeGregorianFields() and returned by getGregorianDayOfYear().
      * @see #computeGregorianFields
      */
-    int32_t fGregorianDayOfMonth;
+    int16_t fGregorianDayOfYear;
 
     /* calculations */
 
-    /**
-     * Compute the Gregorian calendar year, month, and day of month from
-     * the given Julian day.  These values are not stored in fields, but in
-     * member variables gregorianXxx.  Also compute the DAY_OF_WEEK and
-     * DOW_LOCAL fields.
-     */
-    void computeGregorianAndDOWFields(int32_t julianDay, UErrorCode &ec);
-
 protected:
 
     /**
@@ -2359,8 +2343,8 @@ class U_I18N_API Calendar : public UObject {
 #endif  /* U_HIDE_INTERNAL_API */
 
  private:
-    char validLocale[ULOC_FULLNAME_CAPACITY];
-    char actualLocale[ULOC_FULLNAME_CAPACITY];
+    CharString* validLocale = nullptr;
+    CharString* actualLocale = nullptr;
 
  public:
 #if !UCONFIG_NO_SERVICE
@@ -2563,7 +2547,6 @@ Calendar::internalSet(UCalendarDateFields field, int32_t value)
 {
     fFields[field] = value;
     fStamp[field] = kInternallySet;
-    fIsSet[field]     = true; // Remove later
 }
 
 /**
diff --git a/deps/icu-small/source/i18n/unicode/dcfmtsym.h b/deps/icu-small/source/i18n/unicode/dcfmtsym.h
index 02e12f9c39b2dc..3aecceda03a6ee 100644
--- a/deps/icu-small/source/i18n/unicode/dcfmtsym.h
+++ b/deps/icu-small/source/i18n/unicode/dcfmtsym.h
@@ -48,6 +48,7 @@
 
 U_NAMESPACE_BEGIN
 
+class CharString;
 /**
  * This class represents the set of symbols needed by DecimalFormat
  * to format numbers. DecimalFormat creates for itself an instance of
@@ -504,8 +505,8 @@ class U_I18N_API DecimalFormatSymbols : public UObject {
 
     Locale locale;
 
-    char actualLocale[ULOC_FULLNAME_CAPACITY];
-    char validLocale[ULOC_FULLNAME_CAPACITY];
+    CharString* actualLocale = nullptr;
+    CharString* validLocale = nullptr;
     const char16_t* currPattern = nullptr;
 
     UnicodeString currencySpcBeforeSym[UNUM_CURRENCY_SPACING_COUNT];
diff --git a/deps/icu-small/source/i18n/unicode/dtfmtsym.h b/deps/icu-small/source/i18n/unicode/dtfmtsym.h
index df8da36d8153fb..18e2641b588fcd 100644
--- a/deps/icu-small/source/i18n/unicode/dtfmtsym.h
+++ b/deps/icu-small/source/i18n/unicode/dtfmtsym.h
@@ -43,6 +43,7 @@ U_NAMESPACE_BEGIN
 /* forward declaration */
 class SimpleDateFormat;
 class Hashtable;
+class CharString;
 
 /**
  * DateFormatSymbols is a public class for encapsulating localizable date-time
@@ -917,8 +918,8 @@ class U_I18N_API DateFormatSymbols final : public UObject  {
     /** valid/actual locale information 
      *  these are always ICU locales, so the length should not be a problem
      */
-    char validLocale[ULOC_FULLNAME_CAPACITY];
-    char actualLocale[ULOC_FULLNAME_CAPACITY];
+    CharString* validLocale = nullptr;
+    CharString* actualLocale = nullptr;
 
     DateFormatSymbols() = delete; // default constructor not implemented
 
diff --git a/deps/icu-small/source/i18n/unicode/format.h b/deps/icu-small/source/i18n/unicode/format.h
index a21e61ad56d85c..3d435f0de7e367 100644
--- a/deps/icu-small/source/i18n/unicode/format.h
+++ b/deps/icu-small/source/i18n/unicode/format.h
@@ -45,6 +45,7 @@
 
 U_NAMESPACE_BEGIN
 
+class CharString;
 /**
  * Base class for all formats.  This is an abstract base class which
  * specifies the protocol for classes which convert other objects or
@@ -297,8 +298,8 @@ class U_I18N_API Format : public UObject {
                             UParseError& parseError);
 
  private:
-    char actualLocale[ULOC_FULLNAME_CAPACITY];
-    char validLocale[ULOC_FULLNAME_CAPACITY];
+    CharString* actualLocale = nullptr;
+    CharString* validLocale = nullptr;
 };
 
 U_NAMESPACE_END
diff --git a/deps/icu-small/source/i18n/unicode/gregocal.h b/deps/icu-small/source/i18n/unicode/gregocal.h
index 5112548522d557..cd84471c9ba1a6 100644
--- a/deps/icu-small/source/i18n/unicode/gregocal.h
+++ b/deps/icu-small/source/i18n/unicode/gregocal.h
@@ -518,7 +518,7 @@ class U_I18N_API GregorianCalendar: public Calendar {
      * default implementation in Calendar.
      * @stable ICU 2.0
      */
-    virtual int32_t handleGetYearLength(int32_t eyear) const override;
+    virtual int32_t handleGetYearLength(int32_t eyear, UErrorCode& status) const override;
 
     /**
      * return the length of the given month.
diff --git a/deps/icu-small/source/i18n/unicode/measunit.h b/deps/icu-small/source/i18n/unicode/measunit.h
index b23897192eb4cb..f0abd4f4f92f44 100644
--- a/deps/icu-small/source/i18n/unicode/measunit.h
+++ b/deps/icu-small/source/i18n/unicode/measunit.h
@@ -105,21 +105,19 @@ typedef enum UMeasurePrefix {
      */
     UMEASURE_PREFIX_YOTTA = UMEASURE_PREFIX_ONE + 24,
 
-#ifndef U_HIDE_DRAFT_API
     /**
      * SI prefix: ronna, 10^27.
      *
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UMEASURE_PREFIX_RONNA = UMEASURE_PREFIX_ONE + 27,
 
     /**
      * SI prefix: quetta, 10^30.
      *
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UMEASURE_PREFIX_QUETTA = UMEASURE_PREFIX_ONE + 30,
-#endif  /* U_HIDE_DRAFT_API */
 
 #ifndef U_HIDE_INTERNAL_API
     /**
@@ -268,21 +266,19 @@ typedef enum UMeasurePrefix {
      */
     UMEASURE_PREFIX_YOCTO = UMEASURE_PREFIX_ONE + -24,
 
-#ifndef U_HIDE_DRAFT_API
     /**
      * SI prefix: ronto, 10^-27.
      *
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UMEASURE_PREFIX_RONTO = UMEASURE_PREFIX_ONE + -27,
 
     /**
      * SI prefix: quecto, 10^-30.
      *
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UMEASURE_PREFIX_QUECTO = UMEASURE_PREFIX_ONE + -30,
-#endif  /* U_HIDE_DRAFT_API */
 
 #ifndef U_HIDE_INTERNAL_API
     /**
@@ -430,16 +426,19 @@ class U_I18N_API MeasureUnit: public UObject {
     MeasureUnit(MeasureUnit &&other) noexcept;
 
     /**
-     * Construct a MeasureUnit from a CLDR Core Unit Identifier, defined in UTS
-     * 35. (Core unit identifiers and mixed unit identifiers are supported, long
-     * unit identifiers are not.) Validates and canonicalizes the identifier.
+     * Constructs a MeasureUnit from a CLDR Core Unit Identifier, as defined in UTS 35.
+     * This method supports core unit identifiers and mixed unit identifiers.
+     * It validates and canonicalizes the given identifier.
+     *
      *
+     * Example usage:
      * <pre>
-     * MeasureUnit example = MeasureUnit::forIdentifier("furlong-per-nanosecond")
+     * MeasureUnit example = MeasureUnit::forIdentifier("meter-per-second", status);
      * </pre>
      *
-     * @param identifier The CLDR Unit Identifier.
-     * @param status Set if the identifier is invalid.
+     * @param identifier the CLDR Unit Identifier
+     * @param status Set error if the identifier is invalid.
+     * @return the corresponding MeasureUnit
      * @stable ICU 67
      */
     static MeasureUnit forIdentifier(StringPiece identifier, UErrorCode& status);
@@ -552,6 +551,44 @@ class U_I18N_API MeasureUnit: public UObject {
      */
     UMeasurePrefix getPrefix(UErrorCode& status) const;
 
+#ifndef U_HIDE_DRAFT_API
+
+    /**
+     * Creates a new MeasureUnit with a specified constant denominator.
+     *
+     * This method is applicable only to COMPOUND and SINGLE units. If invoked on a
+     * MIXED unit, an error will be set in the status.
+     *
+     * NOTE: If the constant denominator is set to 0, it means that you are removing
+     * the constant denominator.
+     *
+     * @param denominator The constant denominator to set.
+     * @param status Set if this is not a COMPOUND or SINGLE unit or if another error occurs.
+     * @return A new MeasureUnit with the specified constant denominator.
+     * @draft ICU 77
+     */
+    MeasureUnit withConstantDenominator(uint64_t denominator, UErrorCode &status) const;
+
+    /**
+     * Retrieves the constant denominator for this COMPOUND unit.
+     *
+     * Examples:
+     * - For the unit "liter-per-1000-kiloliter", the constant denominator is 1000.
+     * - For the unit "liter-per-kilometer", the constant denominator is zero.
+     *
+     * This method is applicable only to COMPOUND and SINGLE units. If invoked on
+     * a MIXED unit, an error will be set in the status.
+     *
+     * NOTE: If no constant denominator exists, the method returns 0.
+     *
+     * @param status Set if this is not a COMPOUND or SINGLE unit or if another error occurs.
+     * @return The value of the constant denominator.
+     * @draft ICU 77
+     */
+    uint64_t getConstantDenominator(UErrorCode &status) const;
+
+#endif /* U_HIDE_DRAFT_API */
+
     /**
      * Creates a MeasureUnit which is this SINGLE unit augmented with the specified dimensionality
      * (power). For example, if dimensionality is 2, the unit will be squared.
@@ -591,7 +628,9 @@ class U_I18N_API MeasureUnit: public UObject {
      * NOTE: Only works on SINGLE and COMPOUND units. If this is a MIXED unit, an error will
      * occur. For more information, see UMeasureUnitComplexity.
      *
-     * @param status Set if this is a MIXED unit or if another error occurs.
+     * NOTE: An Error will be returned for units that have a constant denominator.
+     *
+     * @param status Set if this is a MIXED unit, has a constant denominator or if another error occurs.
      * @return The reciprocal of the target unit.
      * @stable ICU 67
      */
@@ -627,6 +666,10 @@ class U_I18N_API MeasureUnit: public UObject {
      *
      * If this is a SINGLE unit, an array of length 1 will be returned.
      *
+     * NOTE: For units with a constant denominator, the returned single units will
+     * not include the constant denominator. To obtain the constant denominator,
+     * retrieve it from the original unit.
+     *
      * @param status Set if an error occurs.
      * @return A pair with the list of units as a LocalArray and the number of units in the list.
      * @stable ICU 68
@@ -1152,6 +1195,24 @@ class U_I18N_API MeasureUnit: public UObject {
      */
     static MeasureUnit getPermyriad();
 
+#ifndef U_HIDE_DRAFT_API
+    /**
+     * Returns by pointer, unit of concentr: portion-per-1e9.
+     * Caller owns returned value and must free it.
+     * Also see {@link #getPortionPer1E9()}.
+     * @param status ICU error code.
+     * @draft ICU 77
+     */
+    static MeasureUnit *createPortionPer1E9(UErrorCode &status);
+
+    /**
+     * Returns by value, unit of concentr: portion-per-1e9.
+     * Also see {@link #createPortionPer1E9()}.
+     * @draft ICU 77
+     */
+    static MeasureUnit getPortionPer1E9();
+#endif /* U_HIDE_DRAFT_API */
+
     /**
      * Returns by pointer, unit of consumption: liter-per-100-kilometer.
      * Caller owns returned value and must free it.
diff --git a/deps/icu-small/source/i18n/unicode/messageformat2.h b/deps/icu-small/source/i18n/unicode/messageformat2.h
index c5459f042f40f8..926d14318d17eb 100644
--- a/deps/icu-small/source/i18n/unicode/messageformat2.h
+++ b/deps/icu-small/source/i18n/unicode/messageformat2.h
@@ -8,6 +8,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -20,6 +22,7 @@
 #include "unicode/messageformat2_arguments.h"
 #include "unicode/messageformat2_data_model.h"
 #include "unicode/messageformat2_function_registry.h"
+#include "unicode/normalizer2.h"
 #include "unicode/unistr.h"
 
 #ifndef U_HIDE_DEPRECATED_API
@@ -30,8 +33,8 @@ namespace message2 {
 
     class Environment;
     class MessageContext;
-    class ResolvedSelector;
     class StaticErrors;
+    class InternalValue;
 
     /**
      * <p>MessageFormatter is a Technical Preview API implementing MessageFormat 2.0.
@@ -325,6 +328,8 @@ namespace message2 {
 
     private:
         friend class Builder;
+        friend class Checker;
+        friend class MessageArguments;
         friend class MessageContext;
 
         MessageFormatter(const MessageFormatter::Builder& builder, UErrorCode &status);
@@ -334,9 +339,6 @@ namespace message2 {
         // Do not define default assignment operator
         const MessageFormatter &operator=(const MessageFormatter &) = delete;
 
-        ResolvedSelector resolveVariables(const Environment& env, const data_model::Operand&, MessageContext&, UErrorCode &) const;
-        ResolvedSelector resolveVariables(const Environment& env, const data_model::Expression&, MessageContext&, UErrorCode &) const;
-
         // Selection methods
 
         // Takes a vector of FormattedPlaceholders
@@ -346,31 +348,35 @@ namespace message2 {
         // Takes a vector of vectors of strings (input) and a vector of PrioritizedVariants (input/output)
         void sortVariants(const UVector&, UVector&, UErrorCode&) const;
         // Takes a vector of strings (input) and a vector of strings (output)
-        void matchSelectorKeys(const UVector&, MessageContext&, ResolvedSelector&& rv, UVector&, UErrorCode&) const;
+        void matchSelectorKeys(const UVector&, MessageContext&, InternalValue* rv, UVector&, UErrorCode&) const;
         // Takes a vector of FormattedPlaceholders (input),
         // and a vector of vectors of strings (output)
         void resolvePreferences(MessageContext&, UVector&, UVector&, UErrorCode&) const;
 
         // Formatting methods
+
+        // Used for normalizing variable names and keys for comparison
+        UnicodeString normalizeNFC(const UnicodeString&) const;
         [[nodiscard]] FormattedPlaceholder formatLiteral(const data_model::Literal&) const;
         void formatPattern(MessageContext&, const Environment&, const data_model::Pattern&, UErrorCode&, UnicodeString&) const;
-        // Formats a call to a formatting function
+        // Evaluates a function call
         // Dispatches on argument type
-        [[nodiscard]] FormattedPlaceholder evalFormatterCall(FormattedPlaceholder&& argument,
-                                                       MessageContext& context,
-                                                       UErrorCode& status) const;
+        [[nodiscard]] InternalValue* evalFunctionCall(FormattedPlaceholder&& argument,
+                                                     MessageContext& context,
+                                                     UErrorCode& status) const;
         // Dispatches on function name
-        [[nodiscard]] FormattedPlaceholder evalFormatterCall(const FunctionName& functionName,
-                                                       FormattedPlaceholder&& argument,
-                                                       FunctionOptions&& options,
-                                                       MessageContext& context,
-                                                       UErrorCode& status) const;
-        // Formats an expression that appears as a selector
-        ResolvedSelector formatSelectorExpression(const Environment& env, const data_model::Expression&, MessageContext&, UErrorCode&) const;
+        [[nodiscard]] InternalValue* evalFunctionCall(const FunctionName& functionName,
+                                                     InternalValue* argument,
+                                                     FunctionOptions&& options,
+                                                     MessageContext& context,
+                                                     UErrorCode& status) const;
         // Formats an expression that appears in a pattern or as the definition of a local variable
-        [[nodiscard]] FormattedPlaceholder formatExpression(const Environment&, const data_model::Expression&, MessageContext&, UErrorCode&) const;
+        [[nodiscard]] InternalValue* formatExpression(const Environment&,
+                                                     const data_model::Expression&,
+                                                     MessageContext&,
+                                                     UErrorCode&) const;
         [[nodiscard]] FunctionOptions resolveOptions(const Environment& env, const OptionMap&, MessageContext&, UErrorCode&) const;
-        [[nodiscard]] FormattedPlaceholder formatOperand(const Environment&, const data_model::Operand&, MessageContext&, UErrorCode&) const;
+        [[nodiscard]] InternalValue* formatOperand(const Environment&, const data_model::Operand&, MessageContext&, UErrorCode&) const;
         [[nodiscard]] FormattedPlaceholder evalArgument(const data_model::VariableName&, MessageContext&, UErrorCode&) const;
         void formatSelectors(MessageContext& context, const Environment& env, UErrorCode &status, UnicodeString& result) const;
 
@@ -445,6 +451,10 @@ namespace message2 {
         // formatting methods return best-effort output.
         // The default is false.
         bool signalErrors = false;
+
+        // Used for implementing normalizeNFC()
+        const Normalizer2* nfcNormalizer = nullptr;
+
     }; // class MessageFormatter
 
 } // namespace message2
@@ -457,6 +467,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_H
diff --git a/deps/icu-small/source/i18n/unicode/messageformat2_arguments.h b/deps/icu-small/source/i18n/unicode/messageformat2_arguments.h
index c43d96191f16f1..07c96f892bcaff 100644
--- a/deps/icu-small/source/i18n/unicode/messageformat2_arguments.h
+++ b/deps/icu-small/source/i18n/unicode/messageformat2_arguments.h
@@ -8,6 +8,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -43,7 +45,7 @@ template class U_I18N_API LocalArray<message2::Formattable>;
 
 namespace message2 {
 
-    class MessageContext;
+    class MessageFormatter;
 
     // Arguments
     // ----------
@@ -112,7 +114,9 @@ namespace message2 {
     private:
         friend class MessageContext;
 
-        const Formattable* getArgument(const data_model::VariableName&, UErrorCode&) const;
+        const Formattable* getArgument(const MessageFormatter&,
+                                       const data_model::VariableName&,
+                                       UErrorCode&) const;
 
         // Avoids using Hashtable so that code constructing a Hashtable
         // doesn't have to appear in this header file
@@ -131,6 +135,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_ARGUMENTS_H
diff --git a/deps/icu-small/source/i18n/unicode/messageformat2_data_model.h b/deps/icu-small/source/i18n/unicode/messageformat2_data_model.h
index 0c836af1bdfd35..fd9b6432a5d1ee 100644
--- a/deps/icu-small/source/i18n/unicode/messageformat2_data_model.h
+++ b/deps/icu-small/source/i18n/unicode/messageformat2_data_model.h
@@ -8,6 +8,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -2211,7 +2213,7 @@ namespace message2 {
 
         friend class MFDataModel;
 
-        Matcher(Expression* ss, int32_t ns, Variant* vs, int32_t nv);
+        Matcher(VariableName* ss, int32_t ns, Variant* vs, int32_t nv);
         Matcher() {}
 
         // A Matcher may have numSelectors=0 and numVariants=0
@@ -2219,8 +2221,8 @@ namespace message2 {
         // So we have to keep a separate flag to track failed copies.
         bool bogus = false;
 
-        // The expressions that are being matched on.
-        LocalArray<Expression> selectors;
+        // The variables that are being matched on.
+        LocalArray<VariableName> selectors;
         // The number of selectors
         int32_t numSelectors = 0;
         // The list of `when` clauses (case arms).
@@ -2328,13 +2330,13 @@ namespace message2 {
          * @internal ICU 75 technology preview
          * @deprecated This API is for technology preview only.
          */
-        const std::vector<Expression> getSelectors() const {
+        std::vector<VariableName> getSelectors() const {
             if (std::holds_alternative<Pattern>(body)) {
                 return {};
             }
             const Matcher* match = std::get_if<Matcher>(&body);
             // match must be non-null, given the previous check
-            return toStdVector<Expression>(match->selectors.getAlias(), match->numSelectors);
+            return toStdVector<VariableName>(match->selectors.getAlias(), match->numSelectors);
         }
         /**
          * Accesses the variants. Returns an empty vector if this is a pattern message.
@@ -2462,17 +2464,17 @@ namespace message2 {
              */
             Builder& addBinding(Binding&& b, UErrorCode& status);
             /**
-             * Adds a selector expression. Copies `expression`.
+             * Adds a selector variable.
              * If a pattern was previously set, clears the pattern.
              *
-             * @param selector Expression to add as a selector. Passed by move.
+             * @param selector Variable to add as a selector. Passed by move.
              * @param errorCode Input/output error code
              * @return A reference to the builder.
              *
              * @internal ICU 75 technology preview
              * @deprecated This API is for technology preview only.
              */
-            Builder& addSelector(Expression&& selector, UErrorCode& errorCode) noexcept;
+            Builder& addSelector(VariableName&& selector, UErrorCode& errorCode);
             /**
              * Adds a single variant.
              * If a pattern was previously set using `setPattern()`, clears the pattern.
@@ -2564,7 +2566,7 @@ namespace message2 {
         int32_t bindingsLen = 0;
 
         const Binding* getLocalVariablesInternal() const;
-        const Expression* getSelectorsInternal() const;
+        const VariableName* getSelectorsInternal() const;
         const Variant* getVariantsInternal() const;
 
         int32_t numSelectors() const {
@@ -2592,6 +2594,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT_DATA_MODEL_H
diff --git a/deps/icu-small/source/i18n/unicode/messageformat2_formattable.h b/deps/icu-small/source/i18n/unicode/messageformat2_formattable.h
index 8a779adb9ab348..d7f4130f493b99 100644
--- a/deps/icu-small/source/i18n/unicode/messageformat2_formattable.h
+++ b/deps/icu-small/source/i18n/unicode/messageformat2_formattable.h
@@ -8,6 +8,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -549,6 +551,7 @@ class U_I18N_API FunctionOptions : public UObject {
      */
     FunctionOptions& operator=(const FunctionOptions&) = delete;
  private:
+    friend class InternalValue;
     friend class MessageFormatter;
     friend class StandardFunctions;
 
@@ -566,12 +569,10 @@ class U_I18N_API FunctionOptions : public UObject {
     // that code in the header because it would have to call internal Hashtable methods.
     ResolvedFunctionOption* options;
     int32_t functionOptionsLen = 0;
-}; // class FunctionOptions
-
 
-    // TODO doc comments
-    // Encapsulates either a formatted string or formatted number;
-    // more output types could be added in the future.
+    // Returns a new FunctionOptions
+    FunctionOptions mergeOptions(FunctionOptions&& other, UErrorCode&);
+}; // class FunctionOptions
 
     /**
      * A `FormattedValue` represents the result of formatting a `message2::Formattable`.
@@ -1010,6 +1011,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_FORMATTABLE_H
diff --git a/deps/icu-small/source/i18n/unicode/messageformat2_function_registry.h b/deps/icu-small/source/i18n/unicode/messageformat2_function_registry.h
index b8429e3b83aa91..37690d5e04a1e3 100644
--- a/deps/icu-small/source/i18n/unicode/messageformat2_function_registry.h
+++ b/deps/icu-small/source/i18n/unicode/messageformat2_function_registry.h
@@ -8,6 +8,8 @@
 
 #if U_SHOW_CPLUSPLUS_API
 
+#if !UCONFIG_NO_NORMALIZATION
+
 #if !UCONFIG_NO_FORMATTING
 
 #if !UCONFIG_NO_MF2
@@ -422,6 +424,8 @@ U_NAMESPACE_END
 
 #endif /* #if !UCONFIG_NO_FORMATTING */
 
+#endif /* #if !UCONFIG_NO_NORMALIZATION */
+
 #endif /* U_SHOW_CPLUSPLUS_API */
 
 #endif // MESSAGEFORMAT2_FUNCTION_REGISTRY_H
diff --git a/deps/icu-small/source/i18n/unicode/numberformatter.h b/deps/icu-small/source/i18n/unicode/numberformatter.h
index b02d987ce2bcba..28531afeb10fd3 100644
--- a/deps/icu-small/source/i18n/unicode/numberformatter.h
+++ b/deps/icu-small/source/i18n/unicode/numberformatter.h
@@ -2609,12 +2609,11 @@ class U_I18N_API LocalizedNumberFormatter
      */
     Format* toFormat(UErrorCode& status) const;
 
-#ifndef U_HIDE_DRAFT_API
     /**
      * Disassociate the locale from this formatter.
      *
      * @return The fluent chain.
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UnlocalizedNumberFormatter withoutLocale() const &;
 
@@ -2623,10 +2622,9 @@ class U_I18N_API LocalizedNumberFormatter
      *
      * @return The fluent chain.
      * @see #withoutLocale
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UnlocalizedNumberFormatter withoutLocale() &&;
-#endif // U_HIDE_DRAFT_API
 
     /**
      * Default constructor: puts the formatter into a valid but undefined state.
diff --git a/deps/icu-small/source/i18n/unicode/numberrangeformatter.h b/deps/icu-small/source/i18n/unicode/numberrangeformatter.h
index b8bbc1ba072d5b..bb6e43bed8cf25 100644
--- a/deps/icu-small/source/i18n/unicode/numberrangeformatter.h
+++ b/deps/icu-small/source/i18n/unicode/numberrangeformatter.h
@@ -503,12 +503,11 @@ class U_I18N_API LocalizedNumberRangeFormatter
     FormattedNumberRange formatFormattableRange(
         const Formattable& first, const Formattable& second, UErrorCode& status) const;
 
-#ifndef U_HIDE_DRAFT_API
     /**
      * Disassociate the locale from this formatter.
      *
      * @return The fluent chain.
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UnlocalizedNumberRangeFormatter withoutLocale() const &;
 
@@ -517,10 +516,9 @@ class U_I18N_API LocalizedNumberRangeFormatter
      *
      * @return The fluent chain.
      * @see #withoutLocale
-     * @draft ICU 75
+     * @stable ICU 75
      */
     UnlocalizedNumberRangeFormatter withoutLocale() &&;
-#endif // U_HIDE_DRAFT_API
 
     /**
      * Default constructor: puts the formatter into a valid but undefined state.
diff --git a/deps/icu-small/source/i18n/unicode/rbnf.h b/deps/icu-small/source/i18n/unicode/rbnf.h
index f42d91d776fed9..5a23f723363c9b 100644
--- a/deps/icu-small/source/i18n/unicode/rbnf.h
+++ b/deps/icu-small/source/i18n/unicode/rbnf.h
@@ -88,23 +88,24 @@ enum URBNFRuleSetTag {
 };
 
 /**
- * The RuleBasedNumberFormat class formats numbers according to a set of rules. This number formatter is
- * typically used for spelling out numeric values in words (e.g., 25,3476 as
- * &quot;twenty-five thousand three hundred seventy-six&quot; or &quot;vingt-cinq mille trois
+ * The RuleBasedNumberFormat class formats numbers according to a set of rules.
+ *
+ * <p>This number formatter is typically used for spelling out numeric values in words (e.g., 25,3476
+ * as &quot;twenty-five thousand three hundred seventy-six&quot; or &quot;vingt-cinq mille trois
  * cents soixante-seize&quot; or
  * &quot;f&uuml;nfundzwanzigtausenddreihundertsechsundsiebzig&quot;), but can also be used for
  * other complicated formatting tasks, such as formatting a number of seconds as hours,
- * minutes and seconds (e.g., 3,730 as &quot;1:02:10&quot;).
+ * minutes and seconds (e.g., 3,730 as &quot;1:02:10&quot;).</p>
  *
  * <p>The resources contain three predefined formatters for each locale: spellout, which
  * spells out a value in words (123 is &quot;one hundred twenty-three&quot;); ordinal, which
  * appends an ordinal suffix to the end of a numeral (123 is &quot;123rd&quot;); and
  * duration, which shows a duration in seconds as hours, minutes, and seconds (123 is
- * &quot;2:03&quot;).&nbsp; The client can also define more specialized <tt>RuleBasedNumberFormat</tt>s
+ * &quot;2:03&quot;).&nbsp; The client can also define more specialized <code>RuleBasedNumberFormat</code>s
  * by supplying programmer-defined rule sets.</p>
  *
- * <p>The behavior of a <tt>RuleBasedNumberFormat</tt> is specified by a textual description
- * that is either passed to the constructor as a <tt>String</tt> or loaded from a resource
+ * <p>The behavior of a <code>RuleBasedNumberFormat</code> is specified by a textual description
+ * that is either passed to the constructor as a <code>String</code> or loaded from a resource
  * bundle. In its simplest form, the description consists of a semicolon-delimited list of <em>rules.</em>
  * Each rule has a string of output text and a value or range of values it is applicable to.
  * In a typical spellout rule set, the first twenty rules are the words for the numbers from
@@ -116,7 +117,8 @@ enum URBNFRuleSetTag {
  * <p>For larger numbers, we can use the preceding set of rules to format the ones place, and
  * we only have to supply the words for the multiples of 10:</p>
  *
- * <pre> 20: twenty[-&gt;&gt;];
+ * <pre>
+ * 20: twenty[-&gt;&gt;];
  * 30: thirty[-&gt;&gt;];
  * 40: forty[-&gt;&gt;];
  * 50: fifty[-&gt;&gt;];
@@ -137,7 +139,8 @@ enum URBNFRuleSetTag {
  * <p>For even larger numbers, we can actually look up several parts of the number in the
  * list:</p>
  *
- * <pre>100: &lt;&lt; hundred[ &gt;&gt;];</pre>
+ * <pre>
+ * 100: &lt;&lt; hundred[ &gt;&gt;];</pre>
  *
  * <p>The &quot;&lt;&lt;&quot; represents a new kind of substitution. The &lt;&lt; isolates
  * the hundreds digit (and any digits to its left), formats it using this same rule set, and
@@ -155,13 +158,15 @@ enum URBNFRuleSetTag {
  *
  * <p>This rule covers values up to 999, at which point we add another rule:</p>
  *
- * <pre>1000: &lt;&lt; thousand[ &gt;&gt;];</pre>
+ * <pre>
+ * 1000: &lt;&lt; thousand[ &gt;&gt;];</pre>
  *
  * <p>Again, the meanings of the brackets and substitution tokens shift because the rule's
  * base value is a higher power of 10, changing the rule's divisor. This rule can actually be
  * used all the way up to 999,999. This allows us to finish out the rules as follows:</p>
  *
- * <pre> 1,000,000: &lt;&lt; million[ &gt;&gt;];
+ * <pre>
+ * 1,000,000: &lt;&lt; million[ &gt;&gt;];
  * 1,000,000,000: &lt;&lt; billion[ &gt;&gt;];
  * 1,000,000,000,000: &lt;&lt; trillion[ &gt;&gt;];
  * 1,000,000,000,000,000: OUT OF RANGE!;</pre>
@@ -177,30 +182,30 @@ enum URBNFRuleSetTag {
  * <p>To see how these rules actually work in practice, consider the following example:
  * Formatting 25,430 with this rule set would work like this:</p>
  *
- * <table border="0" width="100%">
+ * <table style="border-collapse: collapse;">
  *   <tr>
- *     <td><strong>&lt;&lt; thousand &gt;&gt;</strong></td>
- *     <td>[the rule whose base value is 1,000 is applicable to 25,340]</td>
+ *     <td style="width: 257; vertical-align: top;"><strong>&lt;&lt; thousand &gt;&gt;</strong></td>
+ *     <td style="width: 340; vertical-align: top;">[the rule whose base value is 1,000 is applicable to 25,340]</td>
  *   </tr>
  *   <tr>
- *     <td><strong>twenty-&gt;&gt;</strong> thousand &gt;&gt;</td>
- *     <td>[25,340 over 1,000 is 25. The rule for 20 applies.]</td>
+ *     <td style="width: 257; vertical-align: top;"><strong>twenty-&gt;&gt;</strong> thousand &gt;&gt;</td>
+ *     <td style="width: 340; vertical-align: top;">[25,340 over 1,000 is 25. The rule for 20 applies.]</td>
  *   </tr>
  *   <tr>
- *     <td>twenty-<strong>five</strong> thousand &gt;&gt;</td>
- *     <td>[25 mod 10 is 5. The rule for 5 is &quot;five.&quot;</td>
+ *     <td style="width: 257; vertical-align: top;">twenty-<strong>five</strong> thousand &gt;&gt;</td>
+ *     <td style="width: 340; vertical-align: top;">[25 mod 10 is 5. The rule for 5 is &quot;five.&quot;</td>
  *   </tr>
  *   <tr>
- *     <td>twenty-five thousand <strong>&lt;&lt; hundred &gt;&gt;</strong></td>
- *     <td>[25,340 mod 1,000 is 340. The rule for 100 applies.]</td>
+ *     <td style="width: 257; vertical-align: top;">twenty-five thousand <strong>&lt;&lt; hundred &gt;&gt;</strong></td>
+ *     <td style="width: 340; vertical-align: top;">[25,340 mod 1,000 is 340. The rule for 100 applies.]</td>
  *   </tr>
  *   <tr>
- *     <td>twenty-five thousand <strong>three</strong> hundred &gt;&gt;</td>
- *     <td>[340 over 100 is 3. The rule for 3 is &quot;three.&quot;]</td>
+ *     <td style="width: 257; vertical-align: top;">twenty-five thousand <strong>three</strong> hundred &gt;&gt;</td>
+ *     <td style="width: 340; vertical-align: top;">[340 over 100 is 3. The rule for 3 is &quot;three.&quot;]</td>
  *   </tr>
  *   <tr>
- *     <td>twenty-five thousand three hundred <strong>forty</strong></td>
- *     <td>[340 mod 100 is 40. The rule for 40 applies. Since 40 divides
+ *     <td style="width: 257; vertical-align: top;">twenty-five thousand three hundred <strong>forty</strong></td>
+ *     <td style="width: 340; vertical-align: top;">[340 mod 100 is 40. The rule for 40 applies. Since 40 divides
  *     evenly by 10, the hyphen and substitution in the brackets are omitted.]</td>
  *   </tr>
  * </table>
@@ -237,20 +242,20 @@ enum URBNFRuleSetTag {
  *
  * <hr>
  *
- * <p>The description of a <tt>RuleBasedNumberFormat</tt>'s behavior consists of one or more <em>rule
+ * <p>The description of a <code>RuleBasedNumberFormat</code>'s behavior consists of one or more <em>rule
  * sets.</em> Each rule set consists of a name, a colon, and a list of <em>rules.</em> A rule
  * set name must begin with a % sign. Rule sets with names that begin with a single % sign
  * are <em>public:</em> the caller can specify that they be used to format and parse numbers.
  * Rule sets with names that begin with %% are <em>private:</em> they exist only for the use
  * of other rule sets. If a formatter only has one rule set, the name may be omitted.</p>
  *
- * <p>The user can also specify a special &quot;rule set&quot; named <tt>%%lenient-parse</tt>.
- * The body of <tt>%%lenient-parse</tt> isn't a set of number-formatting rules, but a <tt>RuleBasedCollator</tt>
+ * <p>The user can also specify a special &quot;rule set&quot; named <code>%%lenient-parse</code>.
+ * The body of <code>%%lenient-parse</code> isn't a set of number-formatting rules, but a <code>RuleBasedCollator</code>
  * description which is used to define equivalences for lenient parsing. For more information
- * on the syntax, see <tt>RuleBasedCollator</tt>. For more information on lenient parsing,
- * see <tt>setLenientParse()</tt>.  <em>Note:</em> symbols that have syntactic meaning
+ * on the syntax, see <code>RuleBasedCollator</code>. For more information on lenient parsing,
+ * see <code>setLenientParse()</code>. <em>Note:</em> symbols that have syntactic meaning
  * in collation rules, such as '&amp;', have no particular meaning when appearing outside
- * of the <tt>lenient-parse</tt> rule set.</p>
+ * of the <code>lenient-parse</code> rule set.</p>
  *
  * <p>The body of a rule set consists of an ordered, semicolon-delimited list of <em>rules.</em>
  * Internally, every rule has a base value, a divisor, rule text, and zero, one, or two <em>substitutions.</em>
@@ -260,42 +265,46 @@ enum URBNFRuleSetTag {
  * <p>A rule descriptor can take one of the following forms (text in <em>italics</em> is the
  * name of a token):</p>
  *
- * <table border="0" width="100%">
+ * <table style="border-collapse: collapse;">
  *   <tr>
- *     <td><em>bv</em>:</td>
- *     <td><em>bv</em> specifies the rule's base value. <em>bv</em> is a decimal
+ *     <th style="padding-left: 1em; padding-right: 1em;">Descriptor</th>
+ *     <th>Description</th>
+ *   </tr>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;"><em>bv</em>:</td>
+ *     <td style="vertical-align: top;"><em>bv</em> specifies the rule's base value. <em>bv</em> is a decimal
  *     number expressed using ASCII digits. <em>bv</em> may contain spaces, period, and commas,
  *     which are ignored. The rule's divisor is the highest power of 10 less than or equal to
  *     the base value.</td>
  *   </tr>
- *   <tr>
- *     <td><em>bv</em>/<em>rad</em>:</td>
- *     <td><em>bv</em> specifies the rule's base value. The rule's divisor is the
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;"><em>bv</em>/<em>rad</em>:</td>
+ *     <td style="vertical-align: top;"><em>bv</em> specifies the rule's base value. The rule's divisor is the
  *     highest power of <em>rad</em> less than or equal to the base value.</td>
  *   </tr>
- *   <tr>
- *     <td><em>bv</em>&gt;:</td>
- *     <td><em>bv</em> specifies the rule's base value. To calculate the divisor,
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;"><em>bv</em>&gt;:</td>
+ *     <td style="vertical-align: top;"><em>bv</em> specifies the rule's base value. To calculate the divisor,
  *     let the radix be 10, and the exponent be the highest exponent of the radix that yields a
  *     result less than or equal to the base value. Every &gt; character after the base value
  *     decreases the exponent by 1. If the exponent is positive or 0, the divisor is the radix
  *     raised to the power of the exponent; otherwise, the divisor is 1.</td>
  *   </tr>
- *   <tr>
- *     <td><em>bv</em>/<em>rad</em>&gt;:</td>
- *     <td><em>bv</em> specifies the rule's base value. To calculate the divisor,
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;"><em>bv</em>/<em>rad</em>&gt;:</td>
+ *     <td style="vertical-align: top;"><em>bv</em> specifies the rule's base value. To calculate the divisor,
  *     let the radix be <em>rad</em>, and the exponent be the highest exponent of the radix that
  *     yields a result less than or equal to the base value. Every &gt; character after the radix
  *     decreases the exponent by 1. If the exponent is positive or 0, the divisor is the radix
  *     raised to the power of the exponent; otherwise, the divisor is 1.</td>
  *   </tr>
- *   <tr>
- *     <td>-x:</td>
- *     <td>The rule is a negative-number rule.</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">-x:</td>
+ *     <td style="vertical-align: top;">The rule is a negative-number rule.</td>
  *   </tr>
- *   <tr>
- *     <td>x.x:</td>
- *     <td>The rule is an <em>improper fraction rule</em>. If the full stop in
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">x.x:</td>
+ *     <td style="vertical-align: top;">The rule is an <em>improper fraction rule</em>. If the full stop in
  *     the middle of the rule name is replaced with the decimal point
  *     that is used in the language or DecimalFormatSymbols, then that rule will
  *     have precedence when formatting and parsing this rule. For example, some
@@ -304,39 +313,39 @@ enum URBNFRuleSetTag {
  *     handle the decimal point that matches the language's natural spelling of
  *     the punctuation of either the full stop or comma.</td>
  *   </tr>
- *   <tr>
- *     <td>0.x:</td>
- *     <td>The rule is a <em>proper fraction rule</em>. If the full stop in
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">0.x:</td>
+ *     <td style="vertical-align: top;">The rule is a <em>proper fraction rule</em>. If the full stop in
  *     the middle of the rule name is replaced with the decimal point
  *     that is used in the language or DecimalFormatSymbols, then that rule will
  *     have precedence when formatting and parsing this rule. For example, some
  *     languages use the comma, and can thus be written as 0,x instead. For example,
  *     you can use "0.x: point &gt;&gt;;0,x: comma &gt;&gt;;" to
  *     handle the decimal point that matches the language's natural spelling of
- *     the punctuation of either the full stop or comma.</td>
+ *     the punctuation of either the full stop or comma</td>
  *   </tr>
- *   <tr>
- *     <td>x.0:</td>
- *     <td>The rule is a <em>default rule</em>. If the full stop in
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">x.0:</td>
+ *     <td style="vertical-align: top;">The rule is a <em>default rule</em>. If the full stop in
  *     the middle of the rule name is replaced with the decimal point
  *     that is used in the language or DecimalFormatSymbols, then that rule will
  *     have precedence when formatting and parsing this rule. For example, some
  *     languages use the comma, and can thus be written as x,0 instead. For example,
  *     you can use "x.0: &lt;&lt; point;x,0: &lt;&lt; comma;" to
  *     handle the decimal point that matches the language's natural spelling of
- *     the punctuation of either the full stop or comma.</td>
+ *     the punctuation of either the full stop or comma</td>
  *   </tr>
- *   <tr>
- *     <td>Inf:</td>
- *     <td>The rule for infinity.</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">Inf:</td>
+ *     <td style="vertical-align: top;">The rule for infinity.</td>
  *   </tr>
- *   <tr>
- *     <td>NaN:</td>
- *     <td>The rule for an IEEE 754 NaN (not a number).</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">NaN:</td>
+ *     <td style="vertical-align: top;">The rule for an IEEE 754 NaN (not a number).</td>
  *   </tr>
- *   <tr>
- *     <td><em>nothing</em></td>
- *     <td>If the rule's rule descriptor is left out, the base value is one plus the
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;"><em>nothing</em></td>
+ *     <td style="vertical-align: top;">If the rule's rule descriptor is left out, the base value is one plus the
  *     preceding rule's base value (or zero if this is the first rule in the list) in a normal
  *     rule set.&nbsp; In a fraction rule set, the base value is the same as the preceding rule's
  *     base value.</td>
@@ -352,8 +361,8 @@ enum URBNFRuleSetTag {
  * algorithms: If the rule set is a regular rule set, do the following:
  *
  * <ul>
- *   <li>If the rule set includes a default rule (and the number was passed in as a <tt>double</tt>),
- *     use the default rule.&nbsp; (If the number being formatted was passed in as a <tt>long</tt>,
+ *   <li>If the rule set includes a default rule (and the number was passed in as a <code>double</code>),
+ *     use the default rule.&nbsp; (If the number being formatted was passed in as a <code>long</code>,
  *     the default rule is ignored.)</li>
  *   <li>If the number is negative, use the negative-number rule.</li>
  *   <li>If the number has a fractional part and is greater than 1, use the improper fraction
@@ -400,42 +409,43 @@ enum URBNFRuleSetTag {
  *
  * <p>The meanings of the substitution token characters are as follows:</p>
  *
- * <table border="0" width="100%">
+ * <table style="border-collapse: collapse;">
  *   <tr>
- *     <td>&gt;&gt;</td>
- *     <td>in normal rule</td>
+ *     <th>Syntax</th>
+ *     <th>Usage</th>
+ *     <th>Description</th>
+ *   </tr>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;" rowspan="4">&gt;&gt;</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in normal rule</td>
  *     <td>Divide the number by the rule's divisor and format the remainder</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in negative-number rule</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in negative-number rule</td>
  *     <td>Find the absolute value of the number and format the result</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in fraction or default rule</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in fraction or default rule</td>
  *     <td>Isolate the number's fractional part and format it.</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in rule in fraction rule set</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in rule in fraction rule set</td>
  *     <td>Not allowed.</td>
  *   </tr>
- *   <tr>
- *     <td>&gt;&gt;&gt;</td>
- *     <td>in normal rule</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;" rowspan="2">&gt;&gt;&gt;</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in normal rule</td>
  *     <td>Divide the number by the rule's divisor and format the remainder,
  *       but bypass the normal rule-selection process and just use the
  *       rule that precedes this one in this rule list.</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in all other rules</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in all other rules</td>
  *     <td>Not allowed.</td>
  *   </tr>
- *   <tr>
- *     <td>&lt;&lt;</td>
- *     <td>in normal rule</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;" rowspan="4">&lt;&lt;</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in normal rule</td>
  *     <td>Divide the number by the rule's divisor, perform floor() on the quotient,
  *         and format the resulting value.<br>
  *         If there is a DecimalFormat pattern between the &lt; characters and the
@@ -448,73 +458,93 @@ enum URBNFRuleSetTag {
  *     </td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in negative-number rule</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in negative-number rule</td>
  *     <td>Not allowed.</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in fraction or default rule</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in fraction or default rule</td>
  *     <td>Isolate the number's integral part and format it.</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in rule in fraction rule set</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in rule in fraction rule set</td>
  *     <td>Multiply the number by the rule's base value and format the result.</td>
  *   </tr>
- *   <tr>
- *     <td>==</td>
- *     <td>in all rule sets</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;">==</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in all rule sets</td>
  *     <td>Format the number unchanged</td>
  *   </tr>
- *   <tr>
- *     <td>[]</td>
- *     <td>in normal rule</td>
- *     <td>Omit the optional text if the number is an even multiple of the rule's divisor</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;" rowspan="6">[]<br/>[|]</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in normal rule</td>
+ *     <td>
+ *       <ul>
+ *         <li>When the number is not an even multiple of the rule's divisor, use the text and rules between the beginning square bracket,
+ *         and the end square bracket or the | symbol.</li>
+ *         <li>When the number is an even multiple of the rule's divisor, and no | symbol is used, omit the text.</li>
+ *         <li>When the number is an even multiple of the rule's divisor, and | symbol is used, use the text and rules between the | symbol,
+ *         and the end square bracket.</li>
+ *       </ul>
+ *     </td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in negative-number rule</td>
- *     <td>Not allowed.</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in improper-fraction rule</td>
+ *     <td>This syntax is the same as specifying both an x.x rule and a 0.x rule.
+ *       <ul>
+ *         <li>When the number is not between 0 and 1, use the text and rules between the beginning square bracket,
+ *         and the end square bracket or the | symbol.</li>
+ *         <li>When the number is between 0 and 1, and no | symbol is used, omit the text.</li>
+ *         <li>When the number is between 0 and 1, and | symbol is used, use the text and rules between the | symbol,
+ *         and the end square bracket.</li>
+ *       </ul>
+ *     </td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in improper-fraction rule</td>
- *     <td>Omit the optional text if the number is between 0 and 1 (same as specifying both an
- *     x.x rule and a 0.x rule)</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in default rule</td>
+ *     <td>This syntax is the same as specifying both an x.x rule and an x.0 rule.
+ *       <ul>
+ *         <li>When the number is not an integer, use the text and rules between the beginning square bracket,
+ *         and the end square bracket or the | symbol.</li>
+ *         <li>When the number is an integer, and no | symbol is used, omit the text.</li>
+ *         <li>When the number is an integer, and | symbol is used, use the text and rules between the | symbol,
+ *         and the end square bracket.</li>
+ *       </ul>
+ *     </td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in default rule</td>
- *     <td>Omit the optional text if the number is an integer (same as specifying both an x.x
- *     rule and an x.0 rule)</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in rule in fraction rule set</td>
+ *     <td>
+ *       <ul>
+ *         <li>When multiplying the number by the rule's base value does not yield 1, use the text and rules between the beginning square bracket,
+ *         and the end square bracket or the | symbol.</li>
+ *         <li>When multiplying the number by the rule's base value yields 1, and no | symbol is used, omit the text.</li>
+ *         <li>When multiplying the number by the rule's base value yields 1, and | symbol is used, use the text and rules between the | symbol,
+ *         and the end square bracket.</li>
+ *       </ul>
+ *     </td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in proper-fraction rule</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in proper-fraction rule</td>
  *     <td>Not allowed.</td>
  *   </tr>
  *   <tr>
- *     <td></td>
- *     <td>in rule in fraction rule set</td>
- *     <td>Omit the optional text if multiplying the number by the rule's base value yields 1.</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in negative-number rule</td>
+ *     <td>Not allowed.</td>
  *   </tr>
- *   <tr>
- *     <td width="37">$(cardinal,<i>plural syntax</i>)$</td>
- *     <td width="23"></td>
- *     <td width="165" valign="top">in all rule sets</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;">$(cardinal,<i>plural syntax</i>)$</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in all rule sets</td>
  *     <td>This provides the ability to choose a word based on the number divided by the radix to the power of the
  *     exponent of the base value for the specified locale, which is normally equivalent to the &lt;&lt; value.
- *     This uses the cardinal plural rules from PluralFormat. All strings used in the plural format are treated
+ *     This uses the cardinal plural rules from {@link PluralFormat}. All strings used in the plural format are treated
  *     as the same base value for parsing.</td>
  *   </tr>
- *   <tr>
- *     <td width="37">$(ordinal,<i>plural syntax</i>)$</td>
- *     <td width="23"></td>
- *     <td width="165" valign="top">in all rule sets</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="white-space: nowrap;">$(ordinal,<i>plural syntax</i>)$</td>
+ *     <td style="white-space: nowrap; vertical-align: top; padding-left: 1em; padding-right: 1em;">in all rule sets</td>
  *     <td>This provides the ability to choose a word based on the number divided by the radix to the power of the
  *     exponent of the base value for the specified locale, which is normally equivalent to the &lt;&lt; value.
- *     This uses the ordinal plural rules from PluralFormat. All strings used in the plural format are treated
+ *     This uses the ordinal plural rules from {@link PluralFormat}. All strings used in the plural format are treated
  *     as the same base value for parsing.</td>
  *   </tr>
  * </table>
@@ -522,22 +552,25 @@ enum URBNFRuleSetTag {
  * <p>The substitution descriptor (i.e., the text between the token characters) may take one
  * of three forms:</p>
  *
- * <table border="0" width="100%">
+ * <table style="border-collapse: collapse;">
  *   <tr>
- *     <td>a rule set name</td>
+ *     <th>Descriptor</th>
+ *     <th>Description</th>
+ *   </tr>
+ *   <tr>
+ *     <td style="vertical-align: top;">a rule set name</td>
  *     <td>Perform the mathematical operation on the number, and format the result using the
  *     named rule set.</td>
  *   </tr>
- *   <tr>
- *     <td>a DecimalFormat pattern</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">a DecimalFormat pattern</td>
  *     <td>Perform the mathematical operation on the number, and format the result using a
  *     DecimalFormat with the specified pattern.&nbsp; The pattern must begin with 0 or #.</td>
  *   </tr>
- *   <tr>
- *     <td>nothing</td>
+ *   <tr style="border-top: 1px solid black;">
+ *     <td style="vertical-align: top;">nothing</td>
  *     <td>Perform the mathematical operation on the number, and format the result using the rule
- *     set containing the current rule, except:
- *     <ul>
+ *     set containing the current rule, except:<ul>
  *       <li>You can't have an empty substitution descriptor with a == substitution.</li>
  *       <li>If you omit the substitution descriptor in a &gt;&gt; substitution in a fraction rule,
  *         format the result one digit at a time using the rule set containing the current rule.</li>
diff --git a/deps/icu-small/source/i18n/unicode/simplenumberformatter.h b/deps/icu-small/source/i18n/unicode/simplenumberformatter.h
index d0121c2b97ec52..83fa8a8b1c92ca 100644
--- a/deps/icu-small/source/i18n/unicode/simplenumberformatter.h
+++ b/deps/icu-small/source/i18n/unicode/simplenumberformatter.h
@@ -78,14 +78,12 @@ class U_I18N_API SimpleNumber : public UMemory {
      */
     void roundTo(int32_t power, UNumberFormatRoundingMode roundingMode, UErrorCode& status);
 
-#ifndef U_HIDE_DRAFT_API
     /**
      * Sets the number of integer digits to the given amount, truncating if necessary.
      *
-     * @draft ICU 75
+     * @stable ICU 75
      */
     void setMaximumIntegerDigits(uint32_t maximumIntegerDigits, UErrorCode& status);
-#endif // U_HIDE_DRAFT_API
 
     /**
      * Pads the beginning of the number with zeros up to the given minimum number of integer digits.
diff --git a/deps/icu-small/source/i18n/unicode/timezone.h b/deps/icu-small/source/i18n/unicode/timezone.h
index b19900b76712c4..9591b7d90affad 100644
--- a/deps/icu-small/source/i18n/unicode/timezone.h
+++ b/deps/icu-small/source/i18n/unicode/timezone.h
@@ -323,7 +323,7 @@ class U_I18N_API TimeZone : public UObject {
      * @see #countEquivalentIDs
      * @stable ICU 2.0
      */
-    static const UnicodeString U_EXPORT2 getEquivalentID(const UnicodeString& id,
+    static UnicodeString U_EXPORT2 getEquivalentID(const UnicodeString& id,
                                                int32_t index);
 
     /**
diff --git a/deps/icu-small/source/i18n/unicode/ucol.h b/deps/icu-small/source/i18n/unicode/ucol.h
index ae4f29c3c6c4e6..9f721f67864255 100644
--- a/deps/icu-small/source/i18n/unicode/ucol.h
+++ b/deps/icu-small/source/i18n/unicode/ucol.h
@@ -1526,7 +1526,6 @@ ucol_openBinary(const uint8_t *bin, int32_t length,
 #include <type_traits>
 
 #include "unicode/char16ptr.h"
-#include "unicode/stringpiece.h"
 #include "unicode/unistr.h"
 
 namespace U_HEADER_ONLY_NAMESPACE {
@@ -1547,6 +1546,7 @@ class Predicate {
     /** @internal */
     explicit Predicate(const UCollator* ucol) : collator(ucol) {}
 
+#if U_SHOW_CPLUSPLUS_API
     /** @internal */
     template <
         typename T, typename U,
@@ -1554,6 +1554,28 @@ class Predicate {
     bool operator()(const T& lhs, const U& rhs) const {
         return match(UnicodeString::readOnlyAlias(lhs), UnicodeString::readOnlyAlias(rhs));
     }
+#else
+    /** @internal */
+    bool operator()(std::u16string_view lhs, std::u16string_view rhs) const {
+        return match(lhs, rhs);
+    }
+
+#if !U_CHAR16_IS_TYPEDEF && (!defined(_LIBCPP_VERSION) || _LIBCPP_VERSION < 180000)
+    /** @internal */
+    bool operator()(std::basic_string_view<uint16_t> lhs, std::basic_string_view<uint16_t> rhs) const {
+        return match({uprv_char16PtrFromUint16(lhs.data()), lhs.length()},
+                     {uprv_char16PtrFromUint16(rhs.data()), rhs.length()});
+    }
+#endif
+
+#if U_SIZEOF_WCHAR_T==2
+    /** @internal */
+    bool operator()(std::wstring_view lhs, std::wstring_view rhs) const {
+        return match({uprv_char16PtrFromWchar(lhs.data()), lhs.length()},
+                     {uprv_char16PtrFromWchar(rhs.data()), rhs.length()});
+    }
+#endif
+#endif
 
     /** @internal */
     bool operator()(std::string_view lhs, std::string_view rhs) const {
@@ -1563,27 +1585,28 @@ class Predicate {
 #if defined(__cpp_char8_t)
     /** @internal */
     bool operator()(std::u8string_view lhs, std::u8string_view rhs) const {
-        return match(lhs, rhs);
+        return match({reinterpret_cast<const char*>(lhs.data()), lhs.length()},
+                     {reinterpret_cast<const char*>(rhs.data()), rhs.length()});
     }
 #endif
 
   private:
-    bool match(UnicodeString lhs, UnicodeString rhs) const {
+    bool match(std::u16string_view lhs, std::u16string_view rhs) const {
         return compare(
             ucol_strcoll(
                 collator,
-                toUCharPtr(lhs.getBuffer()), lhs.length(),
-                toUCharPtr(rhs.getBuffer()), rhs.length()),
+                toUCharPtr(lhs.data()), static_cast<int32_t>(lhs.length()),
+                toUCharPtr(rhs.data()), static_cast<int32_t>(rhs.length())),
             result);
     }
 
-    bool match(StringPiece lhs, StringPiece rhs) const {
+    bool match(std::string_view lhs, std::string_view rhs) const {
         UErrorCode status = U_ZERO_ERROR;
         return compare(
             ucol_strcollUTF8(
                 collator,
-                lhs.data(), lhs.length(),
-                rhs.data(), rhs.length(),
+                lhs.data(), static_cast<int32_t>(lhs.length()),
+                rhs.data(), static_cast<int32_t>(rhs.length()),
                 &status),
             result);
     }
diff --git a/deps/icu-small/source/i18n/unicode/usimplenumberformatter.h b/deps/icu-small/source/i18n/unicode/usimplenumberformatter.h
index 22e81ba2c939b0..ccc5754b4348bb 100644
--- a/deps/icu-small/source/i18n/unicode/usimplenumberformatter.h
+++ b/deps/icu-small/source/i18n/unicode/usimplenumberformatter.h
@@ -156,15 +156,13 @@ U_CAPI void U_EXPORT2
 usnum_setMinimumFractionDigits(USimpleNumber* unumber, int32_t minimumFractionDigits, UErrorCode* ec);
 
 
-#ifndef U_HIDE_DRAFT_API
 /**
  * Sets the number of integer digits to the given amount, truncating if necessary.
  *
- * @draft ICU 75
+ * @stable ICU 75
  */
 U_CAPI void U_EXPORT2
 usnum_setMaximumIntegerDigits(USimpleNumber* unumber, int32_t maximumIntegerDigits, UErrorCode* ec);
-#endif // U_HIDE_DRAFT_API
 
 
 /**
diff --git a/deps/icu-small/source/i18n/units_converter.cpp b/deps/icu-small/source/i18n/units_converter.cpp
index 3ccb0065bfca34..1ab60bd4c06064 100644
--- a/deps/icu-small/source/i18n/units_converter.cpp
+++ b/deps/icu-small/source/i18n/units_converter.cpp
@@ -49,6 +49,8 @@ void U_I18N_API Factor::divideBy(const Factor &rhs) {
     offset = std::max(rhs.offset, offset);
 }
 
+void U_I18N_API Factor::divideBy(const uint64_t constant) { factorDen *= constant; }
+
 void U_I18N_API Factor::power(int32_t power) {
     // multiply all the constant by the power.
     for (int i = 0; i < CONSTANTS_COUNT; i++) {
@@ -239,6 +241,12 @@ Factor loadCompoundFactor(const MeasureUnitImpl &source, const ConversionRates &
         result.multiplyBy(singleFactor);
     }
 
+    // If the source has a constant denominator, then we need to divide the
+    // factor by the constant denominator.
+    if (source.constantDenominator != 0) {
+        result.divideBy(source.constantDenominator);
+    }
+
     return result;
 }
 
diff --git a/deps/icu-small/source/i18n/units_converter.h b/deps/icu-small/source/i18n/units_converter.h
index 01fa557062feae..6f4b55f81d9bff 100644
--- a/deps/icu-small/source/i18n/units_converter.h
+++ b/deps/icu-small/source/i18n/units_converter.h
@@ -82,6 +82,7 @@ struct U_I18N_API Factor {
 
     void multiplyBy(const Factor &rhs);
     void divideBy(const Factor &rhs);
+    void divideBy(const uint64_t constant);
 
     // Apply the power to the factor.
     void power(int32_t power);
diff --git a/deps/icu-small/source/i18n/vtzone.cpp b/deps/icu-small/source/i18n/vtzone.cpp
index 6067c5490e8d46..8c0295bdcbe8b5 100644
--- a/deps/icu-small/source/i18n/vtzone.cpp
+++ b/deps/icu-small/source/i18n/vtzone.cpp
@@ -182,8 +182,9 @@ static UnicodeString& appendMillis(UDate date, UnicodeString& str) {
  */
 static UnicodeString& getDateTimeString(UDate time, UnicodeString& str, UErrorCode& status) {
     if (U_FAILURE(status)) {return str;}
-    int32_t year, month, dom, dow, doy, mid;
-    Grego::timeToFields(time, year, month, dom, dow, doy, mid, status);
+    int32_t year, mid;
+    int8_t month, dom, dow;
+    Grego::timeToFields(time, year, month, dom, dow, mid, status);
     if (U_FAILURE(status)) {return str;}
 
     str.remove();
@@ -675,9 +676,10 @@ static TimeZoneRule* createRuleByRRULE(const UnicodeString& zonename, int rawOff
     }
 
     // Calculate start/end year and missing fields
-    int32_t startYear, startMonth, startDOM, startDOW, startDOY, startMID;
+    int32_t startYear, startMID;
+    int8_t startMonth, startDOM;
     Grego::timeToFields(start + fromOffset, startYear, startMonth, startDOM,
-        startDOW, startDOY, startMID, status);
+        startMID, status);
     if (U_FAILURE(status)) {
         return nullptr;
     }
@@ -692,8 +694,7 @@ static TimeZoneRule* createRuleByRRULE(const UnicodeString& zonename, int rawOff
 
     int32_t endYear;
     if (until != MIN_MILLIS) {
-        int32_t endMonth, endDOM, endDOW, endDOY, endMID;
-        Grego::timeToFields(until, endYear, endMonth, endDOM, endDOW, endDOY, endMID, status);
+        endYear = Grego::timeToYear(until, status);
         if (U_FAILURE(status)) return nullptr;
     } else {
         endYear = AnnualTimeZoneRule::MAX_YEAR;
@@ -1674,8 +1675,7 @@ VTimeZone::parse(UErrorCode& status) {
                     status);
             } else {
                 // Update the end year
-                int32_t y, m, d, dow, doy, mid;
-                Grego::timeToFields(start, y, m, d, dow, doy, mid, status);
+                int32_t y = Grego::timeToYear(start, status);
                 if (U_FAILURE(status)) return;
                 newRule.adoptInsteadAndCheckErrorCode(
                     new AnnualTimeZoneRule(
@@ -1902,7 +1902,8 @@ VTimeZone::writeZone(VTZWriter& w, BasicTimeZone& basictz,
     int32_t stdCount = 0;
     AnnualTimeZoneRule *finalStdRule = nullptr;
 
-    int32_t year, month, dom, dow, doy, mid;
+    int32_t year, mid;
+    int8_t month, dom, dow;
     UBool hasTransitions = false;
     TimeZoneTransition tzt;
     UBool tztAvail;
@@ -1922,7 +1923,7 @@ VTimeZone::writeZone(VTZWriter& w, BasicTimeZone& basictz,
         int32_t fromOffset = tzt.getFrom()->getRawOffset() + tzt.getFrom()->getDSTSavings();
         int32_t fromDSTSavings = tzt.getFrom()->getDSTSavings();
         int32_t toOffset = tzt.getTo()->getRawOffset() + tzt.getTo()->getDSTSavings();
-        Grego::timeToFields(tzt.getTime() + fromOffset, year, month, dom, dow, doy, mid, status);
+        Grego::timeToFields(tzt.getTime() + fromOffset, year, month, dom, dow, mid, status);
         if (U_FAILURE(status)) return;
         int32_t weekInMonth = Grego::dayOfWeekInMonth(year, month, dom);
         UBool sameRule = false;
diff --git a/deps/icu-small/source/i18n/windtfmt.cpp b/deps/icu-small/source/i18n/windtfmt.cpp
index 0241ec3b4424d0..f1d5c25f3ead9b 100644
--- a/deps/icu-small/source/i18n/windtfmt.cpp
+++ b/deps/icu-small/source/i18n/windtfmt.cpp
@@ -46,6 +46,7 @@
 #   define NOIME
 #   define NOMCX
 #include <windows.h>
+#include <typeinfo>
 
 U_NAMESPACE_BEGIN
 
@@ -251,7 +252,7 @@ UnicodeString &Win32DateFormat::format(Calendar &cal, UnicodeString &appendTo, F
         formatDate(&st_local, date);
         formatTime(&st_local, time);
 
-        if (strcmp(fCalendar->getType(), cal.getType()) != 0) {
+        if (typeid(cal) != typeid(*fCalendar)) {
             pattern = getTimeDateFormat(&cal, &fLocale, status);
         }
 
@@ -272,7 +273,7 @@ void Win32DateFormat::parse(const UnicodeString& /* text */, Calendar& /* cal */
 
 void Win32DateFormat::adoptCalendar(Calendar *newCalendar)
 {
-    if (fCalendar == nullptr || strcmp(fCalendar->getType(), newCalendar->getType()) != 0) {
+    if (fCalendar == nullptr || typeid(*fCalendar) != typeid(*newCalendar)) {
         UErrorCode status = U_ZERO_ERROR;
 
         if (fDateStyle != DateFormat::kNone && fTimeStyle != DateFormat::kNone) {
diff --git a/deps/icu-small/source/tools/genrb/parse.cpp b/deps/icu-small/source/tools/genrb/parse.cpp
index f487241cc18990..eb85d5157a6f9a 100644
--- a/deps/icu-small/source/tools/genrb/parse.cpp
+++ b/deps/icu-small/source/tools/genrb/parse.cpp
@@ -1153,7 +1153,7 @@ addCollation(ParseState* state, TableResource  *result, const char *collationTyp
     struct UString    *tokenValue;
     struct UString     comment;
     enum   ETokenType  token;
-    char               subtag[1024];
+    CharString         subtag;
     UnicodeString      rules;
     UBool              haveRules = false;
     UVersionInfo       version;
@@ -1189,15 +1189,15 @@ addCollation(ParseState* state, TableResource  *result, const char *collationTyp
             return nullptr;
         }
 
-        u_UCharsToChars(tokenValue->fChars, subtag, u_strlen(tokenValue->fChars) + 1);
-
+        subtag.clear();
+        subtag.appendInvariantChars(tokenValue->fChars, u_strlen(tokenValue->fChars), *status);
         if (U_FAILURE(*status))
         {
             res_close(result);
             return nullptr;
         }
 
-        member = parseResource(state, subtag, nullptr, status);
+        member = parseResource(state, subtag.data(), nullptr, status);
 
         if (U_FAILURE(*status))
         {
@@ -1208,7 +1208,7 @@ addCollation(ParseState* state, TableResource  *result, const char *collationTyp
         {
             // Ignore the parsed resources, continue parsing.
         }
-        else if (uprv_strcmp(subtag, "Version") == 0 && member->isString())
+        else if (uprv_strcmp(subtag.data(), "Version") == 0 && member->isString())
         {
             StringResource *sr = static_cast<StringResource *>(member);
             char     ver[40];
@@ -1225,11 +1225,11 @@ addCollation(ParseState* state, TableResource  *result, const char *collationTyp
             result->add(member, line, *status);
             member = nullptr;
         }
-        else if(uprv_strcmp(subtag, "%%CollationBin")==0)
+        else if(uprv_strcmp(subtag.data(), "%%CollationBin")==0)
         {
             /* discard duplicate %%CollationBin if any*/
         }
-        else if (uprv_strcmp(subtag, "Sequence") == 0 && member->isString())
+        else if (uprv_strcmp(subtag.data(), "Sequence") == 0 && member->isString())
         {
             StringResource *sr = static_cast<StringResource *>(member);
             rules = sr->fString;
@@ -1395,7 +1395,7 @@ parseCollationElements(ParseState* state, char *tag, uint32_t startline, UBool n
     struct UString    *tokenValue;
     struct UString     comment;
     enum   ETokenType  token;
-    char               subtag[1024], typeKeyword[1024];
+    CharString         subtag, typeKeyword;
     uint32_t           line;
 
     result = table_open(state->bundle, tag, nullptr, status);
@@ -1437,7 +1437,8 @@ parseCollationElements(ParseState* state, char *tag, uint32_t startline, UBool n
                 return nullptr;
             }
 
-            u_UCharsToChars(tokenValue->fChars, subtag, u_strlen(tokenValue->fChars) + 1);
+            subtag.clear();
+            subtag.appendInvariantChars(tokenValue->fChars, u_strlen(tokenValue->fChars), *status);
 
             if (U_FAILURE(*status))
             {
@@ -1445,9 +1446,9 @@ parseCollationElements(ParseState* state, char *tag, uint32_t startline, UBool n
                 return nullptr;
             }
 
-            if (uprv_strcmp(subtag, "default") == 0)
+            if (uprv_strcmp(subtag.data(), "default") == 0)
             {
-                member = parseResource(state, subtag, nullptr, status);
+                member = parseResource(state, subtag.data(), nullptr, status);
 
                 if (U_FAILURE(*status))
                 {
@@ -1466,22 +1467,29 @@ parseCollationElements(ParseState* state, char *tag, uint32_t startline, UBool n
                 if(token == TOK_OPEN_BRACE) {
                     token = getToken(state, &tokenValue, &comment, &line, status);
                     TableResource *collationRes;
-                    if (keepCollationType(subtag)) {
-                        collationRes = table_open(state->bundle, subtag, nullptr, status);
+                    if (keepCollationType(subtag.data())) {
+                        collationRes = table_open(state->bundle, subtag.data(), nullptr, status);
                     } else {
                         collationRes = nullptr;
                     }
                     // need to parse the collation data regardless
-                    collationRes = addCollation(state, collationRes, subtag, startline, status);
+                    collationRes = addCollation(state, collationRes, subtag.data(), startline, status);
                     if (collationRes != nullptr) {
                         result->add(collationRes, startline, *status);
                     }
                 } else if(token == TOK_COLON) { /* right now, we'll just try to see if we have aliases */
                     /* we could have a table too */
                     token = peekToken(state, 1, &tokenValue, &line, &comment, status);
-                    u_UCharsToChars(tokenValue->fChars, typeKeyword, u_strlen(tokenValue->fChars) + 1);
-                    if(uprv_strcmp(typeKeyword, "alias") == 0) {
-                        member = parseResource(state, subtag, nullptr, status);
+                    typeKeyword.clear();
+                    typeKeyword.appendInvariantChars(tokenValue->fChars, u_strlen(tokenValue->fChars), *status);
+                    if (U_FAILURE(*status))
+                    {
+                        res_close(result);
+                        return nullptr;
+                    }
+
+                    if(uprv_strcmp(typeKeyword.data(), "alias") == 0) {
+                        member = parseResource(state, subtag.data(), nullptr, status);
                         if (U_FAILURE(*status))
                         {
                             res_close(result);
@@ -1523,7 +1531,7 @@ realParseTable(ParseState* state, TableResource *table, char *tag, uint32_t star
     struct UString    *tokenValue=nullptr;
     struct UString    comment;
     enum   ETokenType token;
-    char              subtag[1024];
+    CharString        subtag;
     uint32_t          line;
     UBool             readToken = false;
 
@@ -1562,7 +1570,8 @@ realParseTable(ParseState* state, TableResource *table, char *tag, uint32_t star
         }
 
         if(uprv_isInvariantUString(tokenValue->fChars, -1)) {
-            u_UCharsToChars(tokenValue->fChars, subtag, u_strlen(tokenValue->fChars) + 1);
+            subtag.clear();
+            subtag.appendInvariantChars(tokenValue->fChars, u_strlen(tokenValue->fChars), *status);
         } else {
             *status = U_INVALID_FORMAT_ERROR;
             error(line, "invariant characters required for table keys");
@@ -1575,7 +1584,7 @@ realParseTable(ParseState* state, TableResource *table, char *tag, uint32_t star
             return nullptr;
         }
 
-        member = parseResource(state, subtag, &comment, status);
+        member = parseResource(state, subtag.data(), &comment, status);
 
         if (member == nullptr || U_FAILURE(*status))
         {
diff --git a/deps/icu-small/source/tools/icuexportdata/icuexportdata.cpp b/deps/icu-small/source/tools/icuexportdata/icuexportdata.cpp
index 987a6bc7f6f8f1..2856873f977400 100644
--- a/deps/icu-small/source/tools/icuexportdata/icuexportdata.cpp
+++ b/deps/icu-small/source/tools/icuexportdata/icuexportdata.cpp
@@ -47,68 +47,88 @@ int16_t DATAEXPORT_SCRIPT_X_WITH_INHERITED = 0x0800;
 int16_t DATAEXPORT_SCRIPT_X_WITH_OTHER     = 0x0c00;
 
 // TODO(ICU-21821): Replace this with a call to a library function
+// This is an array of all code points with explicit scx values, and can be generated the quick and dirty
+// way with this script:
+//
+// # <ScriptExtensions.txt python script.py
+//
+// import sys
+// for line in sys.stdin:
+//   line = line.strip()
+//   if len(line) == 0 or line.startswith("#"):
+//     continue
+//   entry = line.split(" ")[0]
+//   # Either it is a range
+//   if ".." in entry:
+//     split = entry.split("..")
+//     start = int(split[0], 16)
+//     end = int(split[1], 16)
+//     # +
+//     for ch in range(start, end + 1):
+//       print("0x%04x, " % ch, end="")
+//   # or a single code point
+//   else:
+//     print("0x%s, " % entry.lower(), end="")
+
 int32_t scxCodePoints[] = {
-      7415, 7377, 7380, 7387, 7390, 7391, 7394, 7395, 7396, 7397,
-      7398, 7399, 7400, 7403, 7404, 7406, 7407, 7408, 7409, 113824,
-      113825, 113826, 113827, 834, 837, 7616, 7617, 12294, 12350, 12351,
-      12688, 12689, 12690, 12691, 12692, 12693, 12694, 12695, 12696, 12697,
-      12698, 12699, 12700, 12701, 12702, 12703, 12736, 12737, 12738, 12739,
-      12740, 12741, 12742, 12743, 12744, 12745, 12746, 12747, 12748, 12749,
-      12750, 12751, 12752, 12753, 12754, 12755, 12756, 12757, 12758, 12759,
-      12760, 12761, 12762, 12763, 12764, 12765, 12766, 12767, 12768, 12769,
-      12770, 12771, 12832, 12833, 12834, 12835, 12836, 12837, 12838, 12839,
-      12840, 12841, 12842, 12843, 12844, 12845, 12846, 12847, 12848, 12849,
-      12850, 12851, 12852, 12853, 12854, 12855, 12856, 12857, 12858, 12859,
-      12860, 12861, 12862, 12863, 12864, 12865, 12866, 12867, 12868, 12869,
-      12870, 12871, 12928, 12929, 12930, 12931, 12932, 12933, 12934, 12935,
-      12936, 12937, 12938, 12939, 12940, 12941, 12942, 12943, 12944, 12945,
-      12946, 12947, 12948, 12949, 12950, 12951, 12952, 12953, 12954, 12955,
-      12956, 12957, 12958, 12959, 12960, 12961, 12962, 12963, 12964, 12965,
-      12966, 12967, 12968, 12969, 12970, 12971, 12972, 12973, 12974, 12975,
-      12976, 12992, 12993, 12994, 12995, 12996, 12997, 12998, 12999, 13000,
-      13001, 13002, 13003, 13055, 13144, 13145, 13146, 13147, 13148, 13149,
-      13150, 13151, 13152, 13153, 13154, 13155, 13156, 13157, 13158, 13159,
-      13160, 13161, 13162, 13163, 13164, 13165, 13166, 13167, 13168, 13179,
-      13180, 13181, 13182, 13183, 13280, 13281, 13282, 13283, 13284, 13285,
-      13286, 13287, 13288, 13289, 13290, 13291, 13292, 13293, 13294, 13295,
-      13296, 13297, 13298, 13299, 13300, 13301, 13302, 13303, 13304, 13305,
-      13306, 13307, 13308, 13309, 13310, 119648, 119649, 119650, 119651, 119652,
-      119653, 119654, 119655, 119656, 119657, 119658, 119659, 119660, 119661, 119662,
-      119663, 119664, 119665, 127568, 127569, 867, 868, 869, 870, 871,
-      872, 873, 874, 875, 876, 877, 878, 879, 7418, 7674,
-      66272, 66273, 66274, 66275, 66276, 66277, 66278, 66279, 66280, 66281,
-      66282, 66283, 66284, 66285, 66286, 66287, 66288, 66289, 66290, 66291,
-      66292, 66293, 66294, 66295, 66296, 66297, 66298, 66299, 1748, 64830,
-      64831, 1611, 1612, 1613, 1614, 1615, 1616, 1617, 1618, 1619,
-      1620, 1621, 1648, 65010, 65021, 7381, 7382, 7384, 7393, 7402,
-      7405, 7413, 7414, 43249, 12330, 12331, 12332, 12333, 43471, 65794,
-      65847, 65848, 65849, 65850, 65851, 65852, 65853, 65854, 65855, 1156,
-      1159, 11843, 42607, 1157, 1158, 1155, 7672, 7379, 7411, 7416,
-      7417, 7401, 7383, 7385, 7388, 7389, 7392, 43251, 4347, 3046,
-      3047, 3048, 3049, 3050, 3051, 3052, 3053, 3054, 3055, 3056,
-      3057, 3058, 3059, 70401, 70403, 70459, 70460, 73680, 73681, 73683,
-      2790, 2791, 2792, 2793, 2794, 2795, 2796, 2797, 2798, 2799,
-      2662, 2663, 2664, 2665, 2666, 2667, 2668, 2669, 2670, 2671,
-      42752, 42753, 42754, 42755, 42756, 42757, 42758, 42759, 12337, 12338,
-      12339, 12340, 12341, 12441, 12442, 12443, 12444, 12448, 12540, 65392,
-      65438, 65439, 3302, 3303, 3304, 3305, 3306, 3307, 3308, 3309,
-      3310, 3311, 8239, 68338, 6146, 6147, 6149, 1564, 1632, 1633,
-      1634, 1635, 1636, 1637, 1638, 1639, 1640, 1641, 2534, 2535,
-      2536, 2537, 2538, 2539, 2540, 2541, 2542, 2543, 4160, 4161,
-      4162, 4163, 4164, 4165, 4166, 4167, 4168, 4169, 65792, 65793,
-      65799, 65800, 65801, 65802, 65803, 65804, 65805, 65806, 65807, 65808,
-      65809, 65810, 65811, 65812, 65813, 65814, 65815, 65816, 65817, 65818,
-      65819, 65820, 65821, 65822, 65823, 65824, 65825, 65826, 65827, 65828,
-      65829, 65830, 65831, 65832, 65833, 65834, 65835, 65836, 65837, 65838,
-      65839, 65840, 65841, 65842, 65843, 7412, 8432, 12348, 12349, 43310,
-      7376, 7378, 5941, 5942, 2406, 2407, 2408, 2409, 2410, 2411,
-      2412, 2413, 2414, 2415, 12291, 12307, 12316, 12317, 12318, 12319,
-      12336, 12343, 65093, 65094, 1548, 1563, 12289, 12290, 12296, 12297,
-      12298, 12299, 12300, 12301, 12302, 12303, 12304, 12305, 12308, 12309,
-      12310, 12311, 12312, 12313, 12314, 12315, 12539, 65377, 65378, 65379,
-      65380, 65381, 7386, 1567, 7410, 1600, 43062, 43063, 43064, 43065,
-      2386, 2385, 43059, 43060, 43061, 43056, 43057, 43058, 2404, 2405
-    };
+    0x00b7, 0x02bc, 0x02c7, 0x02c9, 0x02ca, 0x02cb, 0x02cd, 0x02d7, 0x02d9, 0x0300, 0x0301, 0x0302,
+    0x0303, 0x0304, 0x0305, 0x0306, 0x0307, 0x0308, 0x0309, 0x030a, 0x030b, 0x030c, 0x030d, 0x030e,
+    0x0310, 0x0311, 0x0313, 0x0320, 0x0323, 0x0324, 0x0325, 0x032d, 0x032e, 0x0330, 0x0331, 0x0342,
+    0x0345, 0x0358, 0x035e, 0x0363, 0x0364, 0x0365, 0x0366, 0x0367, 0x0368, 0x0369, 0x036a, 0x036b,
+    0x036c, 0x036d, 0x036e, 0x036f, 0x0374, 0x0375, 0x0483, 0x0484, 0x0485, 0x0486, 0x0487, 0x0589,
+    0x060c, 0x061b, 0x061c, 0x061f, 0x0640, 0x064b, 0x064c, 0x064d, 0x064e, 0x064f, 0x0650, 0x0651,
+    0x0652, 0x0653, 0x0654, 0x0655, 0x0660, 0x0661, 0x0662, 0x0663, 0x0664, 0x0665, 0x0666, 0x0667,
+    0x0668, 0x0669, 0x0670, 0x06d4, 0x0951, 0x0952, 0x0964, 0x0965, 0x0966, 0x0967, 0x0968, 0x0969,
+    0x096a, 0x096b, 0x096c, 0x096d, 0x096e, 0x096f, 0x09e6, 0x09e7, 0x09e8, 0x09e9, 0x09ea, 0x09eb,
+    0x09ec, 0x09ed, 0x09ee, 0x09ef, 0x0a66, 0x0a67, 0x0a68, 0x0a69, 0x0a6a, 0x0a6b, 0x0a6c, 0x0a6d,
+    0x0a6e, 0x0a6f, 0x0ae6, 0x0ae7, 0x0ae8, 0x0ae9, 0x0aea, 0x0aeb, 0x0aec, 0x0aed, 0x0aee, 0x0aef,
+    0x0be6, 0x0be7, 0x0be8, 0x0be9, 0x0bea, 0x0beb, 0x0bec, 0x0bed, 0x0bee, 0x0bef, 0x0bf0, 0x0bf1,
+    0x0bf2, 0x0bf3, 0x0ce6, 0x0ce7, 0x0ce8, 0x0ce9, 0x0cea, 0x0ceb, 0x0cec, 0x0ced, 0x0cee, 0x0cef,
+    0x1040, 0x1041, 0x1042, 0x1043, 0x1044, 0x1045, 0x1046, 0x1047, 0x1048, 0x1049, 0x10fb, 0x16eb,
+    0x16ec, 0x16ed, 0x1735, 0x1736, 0x1802, 0x1803, 0x1805, 0x1cd0, 0x1cd1, 0x1cd2, 0x1cd3, 0x1cd4,
+    0x1cd5, 0x1cd6, 0x1cd7, 0x1cd8, 0x1cd9, 0x1cda, 0x1cdb, 0x1cdc, 0x1cdd, 0x1cde, 0x1cdf, 0x1ce0,
+    0x1ce1, 0x1ce2, 0x1ce3, 0x1ce4, 0x1ce5, 0x1ce6, 0x1ce7, 0x1ce8, 0x1ce9, 0x1cea, 0x1ceb, 0x1cec,
+    0x1ced, 0x1cee, 0x1cef, 0x1cf0, 0x1cf1, 0x1cf2, 0x1cf3, 0x1cf4, 0x1cf5, 0x1cf6, 0x1cf7, 0x1cf8,
+    0x1cf9, 0x1cfa, 0x1dc0, 0x1dc1, 0x1df8, 0x1dfa, 0x202f, 0x204f, 0x205a, 0x205d, 0x20f0, 0x2e17,
+    0x2e30, 0x2e31, 0x2e3c, 0x2e41, 0x2e43, 0x2ff0, 0x2ff1, 0x2ff2, 0x2ff3, 0x2ff4, 0x2ff5, 0x2ff6,
+    0x2ff7, 0x2ff8, 0x2ff9, 0x2ffa, 0x2ffb, 0x2ffc, 0x2ffd, 0x2ffe, 0x2fff, 0x3001, 0x3002, 0x3003,
+    0x3006, 0x3008, 0x3009, 0x300a, 0x300b, 0x300c, 0x300d, 0x300e, 0x300f, 0x3010, 0x3011, 0x3013,
+    0x3014, 0x3015, 0x3016, 0x3017, 0x3018, 0x3019, 0x301a, 0x301b, 0x301c, 0x301d, 0x301e, 0x301f,
+    0x302a, 0x302b, 0x302c, 0x302d, 0x3030, 0x3031, 0x3032, 0x3033, 0x3034, 0x3035, 0x3037, 0x303c,
+    0x303d, 0x303e, 0x303f, 0x3099, 0x309a, 0x309b, 0x309c, 0x30a0, 0x30fb, 0x30fc, 0x3190, 0x3191,
+    0x3192, 0x3193, 0x3194, 0x3195, 0x3196, 0x3197, 0x3198, 0x3199, 0x319a, 0x319b, 0x319c, 0x319d,
+    0x319e, 0x319f, 0x31c0, 0x31c1, 0x31c2, 0x31c3, 0x31c4, 0x31c5, 0x31c6, 0x31c7, 0x31c8, 0x31c9,
+    0x31ca, 0x31cb, 0x31cc, 0x31cd, 0x31ce, 0x31cf, 0x31d0, 0x31d1, 0x31d2, 0x31d3, 0x31d4, 0x31d5,
+    0x31d6, 0x31d7, 0x31d8, 0x31d9, 0x31da, 0x31db, 0x31dc, 0x31dd, 0x31de, 0x31df, 0x31e0, 0x31e1,
+    0x31e2, 0x31e3, 0x31e4, 0x31e5, 0x31ef, 0x3220, 0x3221, 0x3222, 0x3223, 0x3224, 0x3225, 0x3226,
+    0x3227, 0x3228, 0x3229, 0x322a, 0x322b, 0x322c, 0x322d, 0x322e, 0x322f, 0x3230, 0x3231, 0x3232,
+    0x3233, 0x3234, 0x3235, 0x3236, 0x3237, 0x3238, 0x3239, 0x323a, 0x323b, 0x323c, 0x323d, 0x323e,
+    0x323f, 0x3240, 0x3241, 0x3242, 0x3243, 0x3244, 0x3245, 0x3246, 0x3247, 0x3280, 0x3281, 0x3282,
+    0x3283, 0x3284, 0x3285, 0x3286, 0x3287, 0x3288, 0x3289, 0x328a, 0x328b, 0x328c, 0x328d, 0x328e,
+    0x328f, 0x3290, 0x3291, 0x3292, 0x3293, 0x3294, 0x3295, 0x3296, 0x3297, 0x3298, 0x3299, 0x329a,
+    0x329b, 0x329c, 0x329d, 0x329e, 0x329f, 0x32a0, 0x32a1, 0x32a2, 0x32a3, 0x32a4, 0x32a5, 0x32a6,
+    0x32a7, 0x32a8, 0x32a9, 0x32aa, 0x32ab, 0x32ac, 0x32ad, 0x32ae, 0x32af, 0x32b0, 0x32c0, 0x32c1,
+    0x32c2, 0x32c3, 0x32c4, 0x32c5, 0x32c6, 0x32c7, 0x32c8, 0x32c9, 0x32ca, 0x32cb, 0x32ff, 0x3358,
+    0x3359, 0x335a, 0x335b, 0x335c, 0x335d, 0x335e, 0x335f, 0x3360, 0x3361, 0x3362, 0x3363, 0x3364,
+    0x3365, 0x3366, 0x3367, 0x3368, 0x3369, 0x336a, 0x336b, 0x336c, 0x336d, 0x336e, 0x336f, 0x3370,
+    0x337b, 0x337c, 0x337d, 0x337e, 0x337f, 0x33e0, 0x33e1, 0x33e2, 0x33e3, 0x33e4, 0x33e5, 0x33e6,
+    0x33e7, 0x33e8, 0x33e9, 0x33ea, 0x33eb, 0x33ec, 0x33ed, 0x33ee, 0x33ef, 0x33f0, 0x33f1, 0x33f2,
+    0x33f3, 0x33f4, 0x33f5, 0x33f6, 0x33f7, 0x33f8, 0x33f9, 0x33fa, 0x33fb, 0x33fc, 0x33fd, 0x33fe,
+    0xa66f, 0xa700, 0xa701, 0xa702, 0xa703, 0xa704, 0xa705, 0xa706, 0xa707, 0xa830, 0xa831, 0xa832,
+    0xa833, 0xa834, 0xa835, 0xa836, 0xa837, 0xa838, 0xa839, 0xa8f1, 0xa8f3, 0xa92e, 0xa9cf, 0xfd3e,
+    0xfd3f, 0xfdf2, 0xfdfd, 0xfe45, 0xfe46, 0xff61, 0xff62, 0xff63, 0xff64, 0xff65, 0xff70, 0xff9e,
+    0xff9f, 0x10100, 0x10101, 0x10102, 0x10107, 0x10108, 0x10109, 0x1010a, 0x1010b, 0x1010c, 0x1010d,
+    0x1010e, 0x1010f, 0x10110, 0x10111, 0x10112, 0x10113, 0x10114, 0x10115, 0x10116, 0x10117, 0x10118,
+    0x10119, 0x1011a, 0x1011b, 0x1011c, 0x1011d, 0x1011e, 0x1011f, 0x10120, 0x10121, 0x10122, 0x10123,
+    0x10124, 0x10125, 0x10126, 0x10127, 0x10128, 0x10129, 0x1012a, 0x1012b, 0x1012c, 0x1012d, 0x1012e,
+    0x1012f, 0x10130, 0x10131, 0x10132, 0x10133, 0x10137, 0x10138, 0x10139, 0x1013a, 0x1013b, 0x1013c,
+    0x1013d, 0x1013e, 0x1013f, 0x102e0, 0x102e1, 0x102e2, 0x102e3, 0x102e4, 0x102e5, 0x102e6, 0x102e7,
+    0x102e8, 0x102e9, 0x102ea, 0x102eb, 0x102ec, 0x102ed, 0x102ee, 0x102ef, 0x102f0, 0x102f1, 0x102f2,
+    0x102f3, 0x102f4, 0x102f5, 0x102f6, 0x102f7, 0x102f8, 0x102f9, 0x102fa, 0x102fb, 0x10af2, 0x11301,
+    0x11303, 0x1133b, 0x1133c, 0x11fd0, 0x11fd1, 0x11fd3, 0x1bca0, 0x1bca1, 0x1bca2, 0x1bca3, 0x1d360,
+    0x1d361, 0x1d362, 0x1d363, 0x1d364, 0x1d365, 0x1d366, 0x1d367, 0x1d368, 0x1d369, 0x1d36a, 0x1d36b,
+    0x1d36c, 0x1d36d, 0x1d36e, 0x1d36f, 0x1d370, 0x1d371, 0x1f250, 0x1f251,
+};
 
 void handleError(ErrorCode& status, int line, const char* context) {
     if (status.isFailure()) {
@@ -331,13 +351,13 @@ void dumpGeneralCategoryMask(FILE* f) {
 
 
     fprintf(f, "mask_for = \"General_Category\"\n");
-    uint32_t minValue = u_getIntPropertyMinValue(UCHAR_GENERAL_CATEGORY);
+    int32_t minValue = u_getIntPropertyMinValue(UCHAR_GENERAL_CATEGORY);
     U_ASSERT(minValue >= 0);
-    uint32_t maxValue = u_getIntPropertyMaxValue(UCHAR_GENERAL_CATEGORY);
+    int32_t maxValue = u_getIntPropertyMaxValue(UCHAR_GENERAL_CATEGORY);
     U_ASSERT(maxValue >= 0);
 
     fprintf(f, "values = [\n");
-    for (uint32_t v = minValue; v <= maxValue; v++) {
+    for (int32_t v = minValue; v <= maxValue; v++) {
         dumpValueEntry(uproperty, U_MASK(v), true, f);
 
         // We want to dump these masks "in order", which means they
@@ -489,12 +509,29 @@ FILE* prepareOutputFile(const char* basename) {
 
 #if !UCONFIG_NO_NORMALIZATION
 
-struct PendingDescriptor {
+class PendingDescriptor {
+public:
     UChar32 scalar;
-    uint32_t descriptor;
+    uint32_t descriptorOrFlags;
+    // If false, we use the above fields only. If true, descriptor only
+    // contains the two highest-bit flags and the rest is computed later
+    // from the fields below.
+    UBool complex;
     UBool supplementary;
+    UBool onlyNonStartersInTrail;
+    uint32_t len;
+    uint32_t offset;
+
+    PendingDescriptor(UChar32 scalar, uint32_t descriptor);
+    PendingDescriptor(UChar32 scalar, uint32_t flags, UBool supplementary, UBool onlyNonStartersInTrail, uint32_t len, uint32_t offset);
 };
 
+PendingDescriptor::PendingDescriptor(UChar32 scalar, uint32_t descriptor)
+ : scalar(scalar), descriptorOrFlags(descriptor), complex(false), supplementary(false), onlyNonStartersInTrail(false), len(0), offset(0) {}
+
+PendingDescriptor::PendingDescriptor(UChar32 scalar, uint32_t flags, UBool supplementary, UBool onlyNonStartersInTrail, uint32_t len, uint32_t offset)
+ : scalar(scalar), descriptorOrFlags(flags), complex(true), supplementary(supplementary), onlyNonStartersInTrail(onlyNonStartersInTrail), len(len), offset(offset) {}
+
 void writeCanonicalCompositions(USet* backwardCombiningStarters) {
     IcuToolErrorCode status("icuexportdata: computeCanonicalCompositions");
     const char* basename = "compositions";
@@ -557,21 +594,18 @@ void writeDecompositionTables(const char* basename, const uint16_t* ptr16, size_
     fclose(f);
 }
 
-void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t baseSize32, uint32_t supplementSize16, USet* uset, USet* reference, const std::vector<PendingDescriptor>& pendingTrieInsertions, char16_t passthroughCap) {
-    IcuToolErrorCode status("icuexportdata: writeDecompositionData");
-    FILE* f = prepareOutputFile(basename);
-
-    // Zero is a magic number that means the character decomposes to itself.
-    LocalUMutableCPTriePointer builder(umutablecptrie_open(0, 0, status));
-
+void pendingInsertionsToTrie(const char* basename, UMutableCPTrie* trie, const std::vector<PendingDescriptor>& pendingTrieInsertions, uint32_t baseSize16, uint32_t baseSize32, uint32_t supplementSize16) {
+    IcuToolErrorCode status("icuexportdata: pendingInsertionsToTrie");
     // Iterate backwards to insert lower code points in the trie first in case it matters
     // for trie block allocation.
     for (int32_t i = pendingTrieInsertions.size() - 1; i >= 0; --i) {
         const PendingDescriptor& pending = pendingTrieInsertions[i];
-        uint32_t additional = 0;
-        if (!(pending.descriptor & 0xFFFC0000)) {
-            uint32_t offset = pending.descriptor & 0xFFF;
+        if (pending.complex) {
+            uint32_t additional = 0;
+            uint32_t offset = pending.offset;
+            uint32_t len = pending.len;
             if (!pending.supplementary) {
+                len -= 2;
                 if (offset >= baseSize16) {
                     // This is a offset to supplementary 16-bit data. We have
                     // 16-bit base data and 32-bit base data before. However,
@@ -579,6 +613,7 @@ void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t
                     additional = baseSize32;
                 }
             } else {
+                len -= 1;
                 if (offset >= baseSize32) {
                     // This is an offset to supplementary 32-bit data. We have 16-bit
                     // base data, 32-bit base data, and 16-bit supplementary data before.
@@ -591,21 +626,55 @@ void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t
                     additional = baseSize16;
                 }
             }
+            // +1 to make offset always non-zero
+            offset += 1;
             if (offset + additional > 0xFFF) {
                 status.set(U_INTERNAL_PROGRAM_ERROR);
                 handleError(status, __LINE__, basename);
             }
+            if (len > 7) {
+                status.set(U_INTERNAL_PROGRAM_ERROR);
+                handleError(status, __LINE__, basename);
+            }
+            umutablecptrie_set(trie, pending.scalar, pending.descriptorOrFlags | (uint32_t(pending.onlyNonStartersInTrail) << 4) | len | (offset + additional) << 16, status);
+        } else {
+            umutablecptrie_set(trie, pending.scalar, pending.descriptorOrFlags, status);
         }
-        // It turns out it's better to swap the halves compared to the initial
-        // idea in order to put special marker values close to zero so that
-        // an important marker value becomes 1, so it's efficient to compare
-        // "1 or 0". Unfortunately, going through all the code to swap
-        // things is too error prone, so let's do the swapping here in one
-        // place.
-        uint32_t oldTrieValue = pending.descriptor + additional;
-        uint32_t swappedTrieValue = (oldTrieValue >> 16) | (oldTrieValue << 16);
-        umutablecptrie_set(builder.getAlias(), pending.scalar, swappedTrieValue, status);
     }
+}
+
+/// Marker that the decomposition does not round trip via NFC.
+const uint32_t NON_ROUND_TRIP_MASK = (1 << 30);
+
+/// Marker that the first character of the decomposition can combine
+/// backwards.
+const uint32_t BACKWARD_COMBINING_MASK = (1 << 31);
+
+void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t baseSize32, uint32_t supplementSize16, USet* uset, USet* reference, const std::vector<PendingDescriptor>& pendingTrieInsertions, const std::vector<PendingDescriptor>& nfdPendingTrieInsertions, char16_t passthroughCap) {
+    IcuToolErrorCode status("icuexportdata: writeDecompositionData");
+    FILE* f = prepareOutputFile(basename);
+
+    // Zero is a magic number that means the character decomposes to itself.
+    LocalUMutableCPTriePointer builder(umutablecptrie_open(0, 0, status));
+
+    if (uprv_strcmp(basename, "uts46d") != 0) {
+        // Make surrogates decompose to U+FFFD. Don't do this for UTS 46, since this
+        // optimization is only used by the UTF-16 slice mode, and UTS 46 is not
+        // supported in slice modes (which do not support ignorables).
+        // Mark these as potentially backward-combining, to make lead surrogates
+        // for non-BMP characters that are backward-combining count as
+        // backward-combining just in case, though the backward-combiningness
+        // is not actually being looked at today.
+        umutablecptrie_setRange(builder.getAlias(), 0xD800, 0xDFFF, NON_ROUND_TRIP_MASK | BACKWARD_COMBINING_MASK | 0xFFFD, status);
+    }
+
+    // Add a marker value for Hangul syllables
+    umutablecptrie_setRange(builder.getAlias(), 0xAC00, 0xD7A3, 1, status);
+
+    // First put the NFD data in the trie, to be partially overwritten in the NFKD and UTS 46 cases.
+    // This is easier that changing the logic that computes the pending insertions.
+    pendingInsertionsToTrie(basename, builder.getAlias(), nfdPendingTrieInsertions, baseSize16, baseSize32, supplementSize16);
+    pendingInsertionsToTrie(basename, builder.getAlias(), pendingTrieInsertions, baseSize16, baseSize32, supplementSize16);
     LocalUCPTriePointer utrie(umutablecptrie_buildImmutable(
         builder.getAlias(),
         trieType,
@@ -613,6 +682,7 @@ void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t
         status));
     handleError(status, __LINE__, basename);
 
+    // The ICU4X side has changed enough this whole block of expectation checking might be more appropriate to remove.
     if (reference) {
         if (uset_contains(reference, 0xFF9E) || uset_contains(reference, 0xFF9F) || !uset_contains(reference, 0x0345)) {
             // NFD expectations don't hold. The set must not contain the half-width
@@ -628,13 +698,9 @@ void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t
         USet* iotaSubscript = uset_openEmpty();
         uset_add(iotaSubscript, 0x0345);
 
-        uint8_t flags = 0;
-
         USet* halfWidthCheck = uset_cloneAsThawed(uset);
         uset_removeAll(halfWidthCheck, reference);
-        if (uset_equals(halfWidthCheck, halfWidthVoicing)) {
-            flags |= 1;
-        } else if (!uset_isEmpty(halfWidthCheck)) {
+        if (!uset_equals(halfWidthCheck, halfWidthVoicing) && !uset_isEmpty(halfWidthCheck)) {
             // The result was neither empty nor contained exactly
             // the two half-width voicing marks. The ICU4X
             // normalizer doesn't know how to deal with this case.
@@ -655,72 +721,14 @@ void writeDecompositionData(const char* basename, uint32_t baseSize16, uint32_t
 
         uset_close(iotaSubscript);
         uset_close(halfWidthVoicing);
-
-        fprintf(f, "flags = 0x%X\n", flags);
-        fprintf(f, "cap = 0x%X\n", passthroughCap);
     }
+    fprintf(f, "cap = 0x%X\n", passthroughCap);
     fprintf(f, "[trie]\n");
     usrc_writeUCPTrie(f, "trie", utrie.getAlias(), UPRV_TARGET_SYNTAX_TOML);
     fclose(f);
     handleError(status, __LINE__, basename);
 }
 
-// Special marker for the NFKD form of U+FDFA
-const int32_t FDFA_MARKER = 3;
-
-// Special marker for characters whose decomposition starts with a non-starter
-// and the decomposition isn't the character itself.
-const int32_t SPECIAL_NON_STARTER_DECOMPOSITION_MARKER = 2;
-
-// Special marker for starters that decompose to themselves but that may
-// combine backwards under canonical composition
-const int32_t BACKWARD_COMBINING_STARTER_MARKER = 1;
-
-/// Marker that a complex decomposition isn't round-trippable
-/// under re-composition.
-///
-/// TODO: When taking a data format break, swap this around with
-/// `BACKWARD_COMBINING_STARTER_DECOMPOSITION_MARKER`.
-const uint32_t NON_ROUND_TRIP_MARKER = 1;
-
-/// Marker that a complex decomposition starts with a starter
-/// that can combine backwards.
-///
-/// TODO: When taking a data format break, swap this around with
-/// `NON_ROUND_TRIP_MARKER` to use the same bit as with characters
-/// that decompose to self but can combine backwards.
-const uint32_t BACKWARD_COMBINING_STARTER_DECOMPOSITION_MARKER = 2;
-
-UBool permissibleBmpPair(UBool knownToRoundTrip, UChar32 c, UChar32 second) {
-    if (knownToRoundTrip) {
-        return true;
-    }
-    // Nuktas, Hebrew presentation forms and polytonic Greek with oxia
-    // are special-cased in ICU4X.
-    if (c >= 0xFB1D && c <= 0xFB4E) {
-        // Hebrew presentation forms
-        return true;
-    }
-    if (c >= 0x1F71 && c <= 0x1FFB) {
-        // Polytonic Greek with oxia
-        return true;
-    }
-    if ((second & 0x7F) == 0x3C && second >= 0x0900 && second <= 0x0BFF) {
-        // Nukta
-        return true;
-    }
-    // To avoid more branchiness, 4 characters that decompose to
-    // a BMP starter followed by a BMP non-starter are excluded
-    // from being encoded directly into the trie value and are
-    // handled as complex decompositions instead. These are:
-    // U+0F76 TIBETAN VOWEL SIGN VOCALIC R
-    // U+0F78 TIBETAN VOWEL SIGN VOCALIC L
-    // U+212B ANGSTROM SIGN
-    // U+2ADC FORKING
-    return false;
-}
-
-
 // Find the slice `needle` within `storage` and return its index, failing which,
 // append all elements of `needle` to `storage` and return the index of it at the end.
 template<typename T>
@@ -749,6 +757,8 @@ size_t findOrAppend(std::vector<T>& storage, const UChar32* needle, size_t needl
 
 
 // Computes data for canonical decompositions
+// See components/normalizer/trie-value-format.md in the ICU4X repo
+// for documentation of the trie value format.
 void computeDecompositions(const char* basename,
                            const USet* backwardCombiningStarters,
                            std::vector<uint16_t>& storage16,
@@ -814,12 +824,23 @@ void computeDecompositions(const char* basename,
             // Surrogate
             continue;
         }
+        if (c == 0xFFFD) {
+            // REPLACEMENT CHARACTER
+            // This character is a starter that decomposes to self,
+            // so without a special case here it would end up as
+            // passthrough-eligible in all normalizations forms.
+            // However, in the potentially-ill-formed UTF-8 case
+            // UTF-8 errors return U+FFFD from the iterator, and
+            // errors need to be treated as ineligible for
+            // passthrough on the slice fast path. By giving
+            // U+FFFD a trie value whose flags make it ineligible
+            // for passthrough avoids a specific U+FFFD branch on
+            // the passthrough fast path.
+            pendingTrieInsertions.push_back({c, NON_ROUND_TRIP_MASK | BACKWARD_COMBINING_MASK});
+            continue;
+        }
         UnicodeString src;
         UnicodeString dst;
-        // True if we're building non-NFD or we're building NFD but
-        // the `c` round trips to NFC.
-        // False if we're building NFD and `c` does not round trip to NFC.
-        UBool nonNfdOrRoundTrips = true;
         src.append(c);
         if (mainNormalizer != nfdNormalizer) {
             UnicodeString inter;
@@ -827,39 +848,12 @@ void computeDecompositions(const char* basename,
             nfdNormalizer->normalize(inter, dst, status);
         } else {
             nfdNormalizer->normalize(src, dst, status);
-            UnicodeString nfc;
-            nfcNormalizer->normalize(dst, nfc, status);
-            nonNfdOrRoundTrips = (src == nfc);
-        }
-        if (uts46) {
-            // Work around https://unicode-org.atlassian.net/browse/ICU-22658
-            // TODO: Remove the workaround after data corresponding to
-            // https://www.unicode.org/L2/L2024/24061.htm#179-C36 lands
-            // for Unicode 16.
-            switch (c) {
-                case 0x2F868:
-                    dst.truncate(0);
-                    dst.append(static_cast<UChar32>(0x36FC));
-                    break;
-                case 0x2F874:
-                    dst.truncate(0);
-                    dst.append(static_cast<UChar32>(0x5F53));
-                    break;
-                case 0x2F91F:
-                    dst.truncate(0);
-                    dst.append(static_cast<UChar32>(0x243AB));
-                    break;
-                case 0x2F95F:
-                    dst.truncate(0);
-                    dst.append(static_cast<UChar32>(0x7AEE));
-                    break;
-                case 0x2F9BF:
-                    dst.truncate(0);
-                    dst.append(static_cast<UChar32>(0x45D7));
-                    break;
-            }
         }
 
+        UnicodeString nfc;
+        nfcNormalizer->normalize(dst, nfc, status);
+        UBool roundTripsViaCanonicalComposition = (src == nfc);
+
         int32_t len = dst.toUTF32(utf32, DECOMPOSITION_BUFFER_SIZE, status);
 
         if (!len || (len == 1 && utf32[0] == 0xFFFD && c != 0xFFFD)) {
@@ -880,7 +874,7 @@ void computeDecompositions(const char* basename,
             compositionPassthroughBound = c;
             uset_add(decompositionStartsWithNonStarter, c);
             if (src != dst) {
-                if (c == 0x0340 || c == 0x0341 || c == 0x0343 || c == 0x0344 || c == 0x0F73 || c == 0x0F75 || c == 0x0F81 || c == 0xFF9E || c == 0xFF9F) {
+                if (c == 0x0340 || c == 0x0341 || c == 0x0343 || c == 0x0344 || c == 0x0F73 || c == 0x0F75 || c == 0x0F81 || (c == 0xFF9E && utf32[0] == 0x3099) || (c == 0xFF9F && utf32[0] == 0x309A)) {
                     specialNonStarterDecomposition = true;
                 } else {
                     // A character whose decomposition starts with a non-starter and isn't the same as the character itself and isn't already hard-coded into ICU4X.
@@ -893,18 +887,6 @@ void computeDecompositions(const char* basename,
             startsWithBackwardCombiningStarter = true;
             uset_add(decompositionStartsWithBackwardCombiningStarter, c);
         }
-        if (c != BACKWARD_COMBINING_STARTER_MARKER && len == 1 && utf32[0] == BACKWARD_COMBINING_STARTER_MARKER) {
-            status.set(U_INTERNAL_PROGRAM_ERROR);
-            handleError(status, __LINE__, basename);
-        }
-        if (c != SPECIAL_NON_STARTER_DECOMPOSITION_MARKER && len == 1 && utf32[0] == SPECIAL_NON_STARTER_DECOMPOSITION_MARKER) {
-            status.set(U_INTERNAL_PROGRAM_ERROR);
-            handleError(status, __LINE__, basename);
-        }
-        if (c != FDFA_MARKER && len == 1 && utf32[0] == FDFA_MARKER) {
-            status.set(U_INTERNAL_PROGRAM_ERROR);
-            handleError(status, __LINE__, basename);
-        }
         if (mainNormalizer != nfdNormalizer) {
             UnicodeString nfd;
             nfdNormalizer->normalize(src, nfd, status);
@@ -913,24 +895,29 @@ void computeDecompositions(const char* basename,
             }
             decompositionPassthroughBound = c;
             compositionPassthroughBound = c;
-        } else if (firstCombiningClass) {
+        }
+        if (firstCombiningClass) {
             len = 1;
             if (specialNonStarterDecomposition) {
-                utf32[0] = SPECIAL_NON_STARTER_DECOMPOSITION_MARKER; // magic value
+                // Special marker
+                pendingTrieInsertions.push_back({c, NON_ROUND_TRIP_MASK | BACKWARD_COMBINING_MASK | 0xD900 | u_getCombiningClass(c)});
             } else {
                 // Use the surrogate range to store the canonical combining class
-                utf32[0] = 0xD800 | static_cast<UChar32>(firstCombiningClass);
+                // XXX: Should non-started that decompose to self be marked as non-round-trippable in
+                // case such semantics turn out to be more useful for `NON_ROUND_TRIP_MASK`?
+                pendingTrieInsertions.push_back({c, BACKWARD_COMBINING_MASK | 0xD800 | static_cast<uint32_t>(firstCombiningClass)});
             }
+            continue;
         } else {
             if (src == dst) {
                 if (startsWithBackwardCombiningStarter) {
-                    pendingTrieInsertions.push_back({c, BACKWARD_COMBINING_STARTER_MARKER << 16, false});
+                    pendingTrieInsertions.push_back({c, BACKWARD_COMBINING_MASK});
                 }
                 continue;
             }
             decompositionPassthroughBound = c;
             // ICU4X hard-codes ANGSTROM SIGN
-            if (c != 0x212B) {
+            if (c != 0x212B && mainNormalizer == nfdNormalizer) {
                 UnicodeString raw;
                 if (!nfdNormalizer->getRawDecomposition(c, raw)) {
                     // We're always supposed to have a non-recursive decomposition
@@ -978,7 +965,7 @@ void computeDecompositions(const char* basename,
                 }
             }
         }
-        if (!nonNfdOrRoundTrips) {
+        if (!roundTripsViaCanonicalComposition) {
             compositionPassthroughBound = c;
         }
         if (!len) {
@@ -986,7 +973,7 @@ void computeDecompositions(const char* basename,
                 status.set(U_INTERNAL_PROGRAM_ERROR);
                 handleError(status, __LINE__, basename);
             }
-            pendingTrieInsertions.push_back({c, 0xFFFFFFFF, false});
+            pendingTrieInsertions.push_back({c, uint32_t(0xFFFFFFFF)});
         } else if (len == 1 && ((utf32[0] >= 0x1161 && utf32[0] <= 0x1175) || (utf32[0] >= 0x11A8 && utf32[0] <= 0x11C2))) {
             // Singleton decompositions to conjoining jamo.
             if (mainNormalizer == nfdNormalizer) {
@@ -994,16 +981,18 @@ void computeDecompositions(const char* basename,
                 status.set(U_INTERNAL_PROGRAM_ERROR);
                 handleError(status, __LINE__, basename);
             }
-            pendingTrieInsertions.push_back({c, static_cast<uint32_t>(utf32[0]) << 16, false});
+            pendingTrieInsertions.push_back({c, static_cast<uint32_t>(utf32[0]) | NON_ROUND_TRIP_MASK | (startsWithBackwardCombiningStarter ? BACKWARD_COMBINING_MASK : 0)});
         } else if (!startsWithBackwardCombiningStarter && len == 1 && utf32[0] <= 0xFFFF) {
-            pendingTrieInsertions.push_back({c, static_cast<uint32_t>(utf32[0]) << 16, false});
-        } else if (!startsWithBackwardCombiningStarter &&
+            pendingTrieInsertions.push_back({c, static_cast<uint32_t>(utf32[0]) | NON_ROUND_TRIP_MASK | (startsWithBackwardCombiningStarter ? BACKWARD_COMBINING_MASK : 0)});
+        } else if (c != 0x212B && // ANGSTROM SIGN is special to make the Harfbuzz case branch less in the more common case.
+                   !startsWithBackwardCombiningStarter &&
                    len == 2 &&
-                   utf32[0] <= 0xFFFF &&
-                   utf32[1] <= 0xFFFF &&
+                   utf32[0] <= 0x7FFF &&
+                   utf32[1] <= 0x7FFF &&
+                   utf32[0] > 0x1F &&
+                   utf32[1] > 0x1F &&
                    !u_getCombiningClass(utf32[0]) &&
-                   u_getCombiningClass(utf32[1]) &&
-                   permissibleBmpPair(nonNfdOrRoundTrips, c, utf32[1])) {
+                   u_getCombiningClass(utf32[1])) {
             for (int32_t i = 0; i < len; ++i) {
                 if (((utf32[i] == 0x0345) && (uprv_strcmp(basename, "uts46d") == 0)) || utf32[i] == 0xFF9E || utf32[i] == 0xFF9F) {
                     // Assert that iota subscript and half-width voicing marks never occur in these
@@ -1012,7 +1001,7 @@ void computeDecompositions(const char* basename,
                     handleError(status, __LINE__, basename);
                 }
             }
-            pendingTrieInsertions.push_back({c, (static_cast<uint32_t>(utf32[0]) << 16) | static_cast<uint32_t>(utf32[1]), false});
+            pendingTrieInsertions.push_back({c, static_cast<uint32_t>(utf32[0]) | (static_cast<uint32_t>(utf32[1]) << 15) | (roundTripsViaCanonicalComposition ? 0 : NON_ROUND_TRIP_MASK)});
         } else {
             UBool supplementary = false;
             UBool nonInitialStarter = false;
@@ -1046,73 +1035,38 @@ void computeDecompositions(const char* basename,
                 if (len > LONGEST_ENCODABLE_LENGTH_16 || !len || len == 1) {
                     if (len == 18 && c == 0xFDFA) {
                         // Special marker for the one character whose decomposition
-                        // is too long.
-                        pendingTrieInsertions.push_back({c, FDFA_MARKER << 16, supplementary});
+                        // is too long. (Too long even if we took the fourth bit into use!)
+                        pendingTrieInsertions.push_back({c, NON_ROUND_TRIP_MASK | 1});
                         continue;
                     } else {
+                        // Note: There's a fourth bit available, but let's error out
+                        // if it's ever needed so that it doesn't get used without
+                        // updating docs.
                         status.set(U_INTERNAL_PROGRAM_ERROR);
                         handleError(status, __LINE__, basename);
                     }
                 }
             } else if (len > LONGEST_ENCODABLE_LENGTH_32 || !len) {
+                // Note: There's a fourth bit available, but let's error out
+                // if it's ever needed so that it doesn't get used without
+                // updating docs.
                 status.set(U_INTERNAL_PROGRAM_ERROR);
                 handleError(status, __LINE__, basename);
             }
-            // Complex decomposition
-            // Format for 16-bit value:
-            // 15..13: length minus two for 16-bit case and length minus one for
-            //         the 32-bit case. Length 8 needs to fit in three bits in
-            //         the 16-bit case, and this way the value is future-proofed
-            //         up to 9 in the 16-bit case. Zero is unused and length one
-            //         in the 16-bit case goes directly into the trie.
-            //     12: 1 if all trailing characters are guaranteed non-starters,
-            //         0 if no guarantees about non-starterness.
-            //         Note: The bit choice is this way around to allow for
-            //         dynamically falling back to not having this but instead
-            //         having one more bit for length by merely choosing
-            //         different masks.
-            //  11..0: Start offset in storage. The offset is to the logical
-            //         sequence of scalars16, scalars32, supplementary_scalars16,
-            //         supplementary_scalars32.
-            uint32_t descriptor = static_cast<uint32_t>(!nonInitialStarter) << 12;
-            if (!supplementary) {
-                descriptor |= (static_cast<uint32_t>(len) - 2) << 13;
-            } else {
-                descriptor |= (static_cast<uint32_t>(len) - 1) << 13;
-            }
-            if (descriptor & 0xFFF) {
-                status.set(U_INTERNAL_PROGRAM_ERROR);
-                handleError(status, __LINE__, basename);
-            }
+
             size_t index = 0;
             if (!supplementary) {
                 index = findOrAppend(storage16, utf32, len);
             } else {
                 index = findOrAppend(storage32, utf32, len);
             }
-            if (index > 0xFFF) {
-                status.set(U_INTERNAL_PROGRAM_ERROR);
-                handleError(status, __LINE__, basename);
-            }
-            descriptor |= static_cast<uint32_t>(index);
-            if (!descriptor || descriptor > 0xFFFF) {
-                // > 0xFFFF should never happen if the code above is correct.
-                // == 0 should not happen due to the nature of the data.
-                status.set(U_INTERNAL_PROGRAM_ERROR);
-                handleError(status, __LINE__, basename);
-            }
-            uint32_t nonRoundTripMarker = 0;
-            if (!nonNfdOrRoundTrips) {
-                nonRoundTripMarker = (NON_ROUND_TRIP_MARKER << 16);
-            }
-            uint32_t canCombineBackwardsMarker = 0;
-            if (startsWithBackwardCombiningStarter) {
-                canCombineBackwardsMarker = (BACKWARD_COMBINING_STARTER_DECOMPOSITION_MARKER << 16);
-            }
-            pendingTrieInsertions.push_back({c, descriptor | nonRoundTripMarker | canCombineBackwardsMarker, supplementary});
+            pendingTrieInsertions.push_back({c, (startsWithBackwardCombiningStarter ? BACKWARD_COMBINING_MASK : 0) | (roundTripsViaCanonicalComposition ? 0 : NON_ROUND_TRIP_MASK), supplementary, !nonInitialStarter, uint32_t(len), uint32_t(index)});
         }
     }
     if (storage16.size() + storage32.size() > 0xFFF) {
+        // We actually have 14 bits available, but let's error out so
+        // that docs can be updated when taking a reserved bit out of
+        // potential future flag usage.
         status.set(U_INTERNAL_PROGRAM_ERROR);
     }
     if (f) {
@@ -1489,9 +1443,9 @@ int exportNorm() {
     uint32_t supplementSize16 = storage16.size() - baseSize16;
     uint32_t supplementSize32 = storage32.size() - baseSize32;
 
-    writeDecompositionData("nfd", baseSize16, baseSize32, supplementSize16, nfdDecompositionStartsWithNonStarter, nullptr, nfdPendingTrieInsertions, static_cast<char16_t>(nfcBound));
-    writeDecompositionData("nfkd", baseSize16, baseSize32, supplementSize16, nfkdDecompositionStartsWithNonStarter, nfdDecompositionStartsWithNonStarter, nfkdPendingTrieInsertions, static_cast<char16_t>(nfkcBound));
-    writeDecompositionData("uts46d", baseSize16, baseSize32, supplementSize16, uts46DecompositionStartsWithNonStarter, nfdDecompositionStartsWithNonStarter, uts46PendingTrieInsertions, static_cast<char16_t>(uts46Bound));
+    writeDecompositionData("nfd", baseSize16, baseSize32, supplementSize16, nfdDecompositionStartsWithNonStarter, nullptr, nfdPendingTrieInsertions, nfdPendingTrieInsertions, static_cast<char16_t>(nfcBound));
+    writeDecompositionData("nfkd", baseSize16, baseSize32, supplementSize16, nfkdDecompositionStartsWithNonStarter, nfdDecompositionStartsWithNonStarter, nfkdPendingTrieInsertions, nfdPendingTrieInsertions, static_cast<char16_t>(nfkcBound));
+    writeDecompositionData("uts46d", baseSize16, baseSize32, supplementSize16, uts46DecompositionStartsWithNonStarter, nfdDecompositionStartsWithNonStarter, uts46PendingTrieInsertions, nfdPendingTrieInsertions, static_cast<char16_t>(uts46Bound));
 
     writeDecompositionTables("nfdex", storage16.data(), baseSize16, storage32.data(), baseSize32);
     writeDecompositionTables("nfkdex", storage16.data() + baseSize16, supplementSize16, storage32.data() + baseSize32, supplementSize32);
diff --git a/deps/icu-small/source/tools/toolutil/ucm.cpp b/deps/icu-small/source/tools/toolutil/ucm.cpp
index 923041a53f607b..824362a6939205 100644
--- a/deps/icu-small/source/tools/toolutil/ucm.cpp
+++ b/deps/icu-small/source/tools/toolutil/ucm.cpp
@@ -310,7 +310,7 @@ enum {
 
 static uint8_t
 checkBaseExtUnicode(UCMStates *baseStates, UCMTable *base, UCMTable *ext,
-                    UBool moveToExt, UBool intersectBase) {
+                    UBool moveToExt, int8_t intersectBase) {
     (void)baseStates;
 
     UCMapping *mb, *me, *mbLimit, *meLimit;
@@ -416,7 +416,7 @@ checkBaseExtUnicode(UCMStates *baseStates, UCMTable *base, UCMTable *ext,
 
 static uint8_t
 checkBaseExtBytes(UCMStates *baseStates, UCMTable *base, UCMTable *ext,
-                  UBool moveToExt, UBool intersectBase) {
+                  UBool moveToExt, int8_t intersectBase) {
     UCMapping *mb, *me;
     int32_t *baseMap, *extMap;
     int32_t b, e, bLimit, eLimit, cmp;
@@ -556,7 +556,7 @@ ucm_checkValidity(UCMTable *table, UCMStates *baseStates) {
 U_CAPI UBool U_EXPORT2
 ucm_checkBaseExt(UCMStates *baseStates,
                  UCMTable *base, UCMTable *ext, UCMTable *moveTarget,
-                 UBool intersectBase) {
+                 int8_t intersectBase) {
     uint8_t result;
 
     /* if we have an extension table, we must always use precision flags */
@@ -735,7 +735,7 @@ ucm_separateMappings(UCMFile *ucm, UBool isSISO) {
     }
     if(needsMove) {
         ucm_moveMappings(ucm->base, ucm->ext);
-        return ucm_checkBaseExt(&ucm->states, ucm->base, ucm->ext, ucm->ext, false);
+        return ucm_checkBaseExt(&ucm->states, ucm->base, ucm->ext, ucm->ext, 0);
     } else {
         ucm_sortTable(ucm->base);
         return true;
diff --git a/deps/icu-small/source/tools/toolutil/ucm.h b/deps/icu-small/source/tools/toolutil/ucm.h
index 8ea90604d475ed..8f78b52e9687cd 100644
--- a/deps/icu-small/source/tools/toolutil/ucm.h
+++ b/deps/icu-small/source/tools/toolutil/ucm.h
@@ -227,7 +227,7 @@ ucm_checkValidity(UCMTable *ext, UCMStates *baseStates);
  */
 U_CAPI UBool U_EXPORT2
 ucm_checkBaseExt(UCMStates *baseStates, UCMTable *base, UCMTable *ext,
-                 UCMTable *moveTarget, UBool intersectBase);
+                 UCMTable *moveTarget, int8_t intersectBase);
 
 U_CAPI void U_EXPORT2
 ucm_printTable(UCMTable *table, FILE *f, UBool byUnicode);
diff --git a/deps/simdutf/simdutf.cpp b/deps/simdutf/simdutf.cpp
index 21962c3bad378d..8991fbe7b57877 100644
--- a/deps/simdutf/simdutf.cpp
+++ b/deps/simdutf/simdutf.cpp
@@ -1,7 +1,9 @@
-/* auto-generated on 2025-01-08 17:51:07 -0500. Do not edit! */
+/* auto-generated on 2025-04-03 18:47:20 -0400. Do not edit! */
 /* begin file src/simdutf.cpp */
 #include "simdutf.h"
-// We include base64_tables once.
+
+#if SIMDUTF_FEATURE_BASE64
+  // We include base64_tables once.
 /* begin file src/tables/base64_tables.h */
 #ifndef SIMDUTF_BASE64_TABLES_H
 #define SIMDUTF_BASE64_TABLES_H
@@ -692,11934 +694,112 @@ static_assert(to_base64_url_value[uint8_t('_')] == 63,
 
 #endif // SIMDUTF_BASE64_TABLES_H
 /* end file src/tables/base64_tables.h */
-/* begin file src/implementation.cpp */
-#include <initializer_list>
-#include <climits>
-#include <type_traits>
+#endif // SIMDUTF_FEATURE_BASE64
 
-static_assert(sizeof(uint8_t) == sizeof(char),
-              "simdutf requires that uint8_t be a char");
-static_assert(sizeof(uint16_t) == sizeof(char16_t),
-              "simdutf requires that char16_t be 16 bits");
-static_assert(sizeof(uint32_t) == sizeof(char32_t),
-              "simdutf requires that char32_t be 32 bits");
-// next line is redundant, but it is kept to catch defective systems.
-static_assert(CHAR_BIT == 8, "simdutf requires 8-bit bytes");
+/* begin file src/encoding_types.cpp */
 
-// Useful for debugging purposes
 namespace simdutf {
-namespace {
-
-template <typename T> std::string toBinaryString(T b) {
-  std::string binary = "";
-  T mask = T(1) << (sizeof(T) * CHAR_BIT - 1);
-  while (mask > 0) {
-    binary += ((b & mask) == 0) ? '0' : '1';
-    mask >>= 1;
-  }
-  return binary;
-}
-} // namespace
-} // namespace simdutf
-
-// Implementations
-// The best choice should always come first!
-/* begin file src/simdutf/arm64.h */
-#ifndef SIMDUTF_ARM64_H
-#define SIMDUTF_ARM64_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "arm64.h must be included before fallback.h"
-#endif
-
-
-#ifndef SIMDUTF_IMPLEMENTATION_ARM64
-  #define SIMDUTF_IMPLEMENTATION_ARM64 (SIMDUTF_IS_ARM64)
-#endif
-#if SIMDUTF_IMPLEMENTATION_ARM64 && SIMDUTF_IS_ARM64
-  #define SIMDUTF_CAN_ALWAYS_RUN_ARM64 1
+bool match_system(endianness e) {
+#if SIMDUTF_IS_BIG_ENDIAN
+  return e == endianness::BIG;
 #else
-  #define SIMDUTF_CAN_ALWAYS_RUN_ARM64 0
+  return e == endianness::LITTLE;
 #endif
-
-
-#if SIMDUTF_IMPLEMENTATION_ARM64
-
-namespace simdutf {
-/**
- * Implementation for NEON (ARMv8).
- */
-namespace arm64 {} // namespace arm64
-} // namespace simdutf
-
-/* begin file src/simdutf/arm64/implementation.h */
-#ifndef SIMDUTF_ARM64_IMPLEMENTATION_H
-#define SIMDUTF_ARM64_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace arm64 {
-
-namespace {
-using namespace simdutf;
 }
 
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("arm64", "ARM NEON",
-                                internal::instruction_set::NEON) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-
-} // namespace arm64
-} // namespace simdutf
-
-#endif // SIMDUTF_ARM64_IMPLEMENTATION_H
-/* end file src/simdutf/arm64/implementation.h */
-
-/* begin file src/simdutf/arm64/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "arm64"
-// #define SIMDUTF_IMPLEMENTATION arm64
-/* end file src/simdutf/arm64/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/arm64/intrinsics.h */
-#ifndef SIMDUTF_ARM64_INTRINSICS_H
-#define SIMDUTF_ARM64_INTRINSICS_H
-
-
-// This should be the correct header whether
-// you use visual studio or other compilers.
-#include <arm_neon.h>
-
-#endif //  SIMDUTF_ARM64_INTRINSICS_H
-/* end file src/simdutf/arm64/intrinsics.h */
-/* begin file src/simdutf/arm64/bitmanipulation.h */
-#ifndef SIMDUTF_ARM64_BITMANIPULATION_H
-#define SIMDUTF_ARM64_BITMANIPULATION_H
+std::string to_string(encoding_type bom) {
+  switch (bom) {
+  case UTF16_LE:
+    return "UTF16 little-endian";
+  case UTF16_BE:
+    return "UTF16 big-endian";
+  case UTF32_LE:
+    return "UTF32 little-endian";
+  case UTF32_BE:
+    return "UTF32 big-endian";
+  case UTF8:
+    return "UTF8";
+  case unspecified:
+    return "unknown";
+  default:
+    return "error";
+  }
+}
 
-namespace simdutf {
-namespace arm64 {
-namespace {
+namespace BOM {
+// Note that BOM for UTF8 is discouraged.
+encoding_type check_bom(const uint8_t *byte, size_t length) {
+  if (length >= 2 && byte[0] == 0xff and byte[1] == 0xfe) {
+    if (length >= 4 && byte[2] == 0x00 and byte[3] == 0x0) {
+      return encoding_type::UTF32_LE;
+    } else {
+      return encoding_type::UTF16_LE;
+    }
+  } else if (length >= 2 && byte[0] == 0xfe and byte[1] == 0xff) {
+    return encoding_type::UTF16_BE;
+  } else if (length >= 4 && byte[0] == 0x00 and byte[1] == 0x00 and
+             byte[2] == 0xfe and byte[3] == 0xff) {
+    return encoding_type::UTF32_BE;
+  } else if (length >= 4 && byte[0] == 0xef and byte[1] == 0xbb and
+             byte[2] == 0xbf) {
+    return encoding_type::UTF8;
+  }
+  return encoding_type::unspecified;
+}
 
-/* result might be undefined when input_num is zero */
-simdutf_really_inline int count_ones(uint64_t input_num) {
-  return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
+encoding_type check_bom(const char *byte, size_t length) {
+  return check_bom(reinterpret_cast<const uint8_t *>(byte), length);
 }
 
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
-  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  unsigned long ret;
-  // Search the mask data from least significant bit (LSB)
-  // to the most significant bit (MSB) for a set bit (1).
-  _BitScanForward64(&ret, input_num);
-  return (int)ret;
-  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
-  return __builtin_ctzll(input_num);
-  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+size_t bom_byte_size(encoding_type bom) {
+  switch (bom) {
+  case UTF16_LE:
+    return 2;
+  case UTF16_BE:
+    return 2;
+  case UTF32_LE:
+    return 4;
+  case UTF32_BE:
+    return 4;
+  case UTF8:
+    return 3;
+  case unspecified:
+    return 0;
+  default:
+    return 0;
+  }
 }
-#endif
 
-} // unnamed namespace
-} // namespace arm64
+} // namespace BOM
 } // namespace simdutf
-
-#endif // SIMDUTF_ARM64_BITMANIPULATION_H
-/* end file src/simdutf/arm64/bitmanipulation.h */
-/* begin file src/simdutf/arm64/simd.h */
-#ifndef SIMDUTF_ARM64_SIMD_H
-#define SIMDUTF_ARM64_SIMD_H
-
-#include <type_traits>
-
+/* end file src/encoding_types.cpp */
+/* begin file src/error.cpp */
 namespace simdutf {
-namespace arm64 {
-namespace {
-namespace simd {
+// deliberately empty
+}
+/* end file src/error.cpp */
+// The large tables should be included once and they
+// should not depend on a kernel.
+/* begin file src/tables/utf8_to_utf16_tables.h */
+#ifndef SIMDUTF_UTF8_TO_UTF16_TABLES_H
+#define SIMDUTF_UTF8_TO_UTF16_TABLES_H
+#include <cstdint>
 
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+namespace simdutf {
 namespace {
-  // Start of private section with Visual Studio workaround
-
-  #ifndef simdutf_make_uint8x16_t
-    #define simdutf_make_uint8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10,   \
-                                    x11, x12, x13, x14, x15, x16)              \
-      ([=]() {                                                                 \
-        uint8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,             \
-                             x9, x10, x11, x12, x13, x14, x15, x16};           \
-        return vld1q_u8(array);                                                \
-      }())
-  #endif
-  #ifndef simdutf_make_int8x16_t
-    #define simdutf_make_int8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10,    \
-                                   x11, x12, x13, x14, x15, x16)               \
-      ([=]() {                                                                 \
-        int8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,              \
-                            x9, x10, x11, x12, x13, x14, x15, x16};            \
-        return vld1q_s8(array);                                                \
-      }())
-  #endif
-
-  #ifndef simdutf_make_uint8x8_t
-    #define simdutf_make_uint8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)             \
-      ([=]() {                                                                 \
-        uint8_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                   \
-        return vld1_u8(array);                                                 \
-      }())
-  #endif
-  #ifndef simdutf_make_int8x8_t
-    #define simdutf_make_int8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)              \
-      ([=]() {                                                                 \
-        int8_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                    \
-        return vld1_s8(array);                                                 \
-      }())
-  #endif
-  #ifndef simdutf_make_uint16x8_t
-    #define simdutf_make_uint16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)            \
-      ([=]() {                                                                 \
-        uint16_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                  \
-        return vld1q_u16(array);                                               \
-      }())
-  #endif
-  #ifndef simdutf_make_int16x8_t
-    #define simdutf_make_int16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)             \
-      ([=]() {                                                                 \
-        int16_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                   \
-        return vld1q_s16(array);                                               \
-      }())
-  #endif
-
-// End of private section with Visual Studio workaround
-} // namespace
-#endif // SIMDUTF_REGULAR_VISUAL_STUDIO
-
-template <typename T> struct simd8;
-
-//
-// Base class of simd8<uint8_t> and simd8<bool>, both of which use uint8x16_t
-// internally.
-//
-template <typename T, typename Mask = simd8<bool>> struct base_u8 {
-  uint8x16_t value;
-  static const int SIZE = sizeof(value);
-
-  // Conversion from/to SIMD register
-  simdutf_really_inline base_u8(const uint8x16_t _value) : value(_value) {}
-  simdutf_really_inline operator const uint8x16_t &() const {
-    return this->value;
-  }
-  simdutf_really_inline operator uint8x16_t &() { return this->value; }
-  simdutf_really_inline T first() const { return vgetq_lane_u8(*this, 0); }
-  simdutf_really_inline T last() const { return vgetq_lane_u8(*this, 15); }
-
-  // Bit operations
-  simdutf_really_inline simd8<T> operator|(const simd8<T> other) const {
-    return vorrq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<T> operator&(const simd8<T> other) const {
-    return vandq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<T> operator^(const simd8<T> other) const {
-    return veorq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<T> bit_andnot(const simd8<T> other) const {
-    return vbicq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-  simdutf_really_inline simd8<T> &operator|=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd8<T> &operator&=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd8<T> &operator^=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-
-  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return vceqq_u8(lhs, rhs);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
-    return vextq_u8(prev_chunk, *this, 16 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base_u8<bool> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return vmovq_n_u8(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8(const uint8x16_t _value)
-      : base_u8<bool>(_value) {}
-  // False constructor
-  simdutf_really_inline simd8() : simd8(vdupq_n_u8(0)) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
-  simdutf_really_inline void store(uint8_t dst[16]) const {
-    return vst1q_u8(dst, *this);
-  }
-
-  // We return uint32_t instead of uint16_t because that seems to be more
-  // efficient for most purposes (cutting it down to uint16_t costs performance
-  // in some compilers).
-  simdutf_really_inline uint32_t to_bitmask() const {
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-    const uint8x16_t bit_mask =
-        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
-#else
-    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
-#endif
-    auto minput = *this & bit_mask;
-    uint8x16_t tmp = vpaddq_u8(minput, minput);
-    tmp = vpaddq_u8(tmp, tmp);
-    tmp = vpaddq_u8(tmp, tmp);
-    return vgetq_lane_u16(vreinterpretq_u16_u8(tmp), 0);
-  }
-
-  // Returns 4-bit out of each byte, alternating between the high 4 bits and low
-  // bits result it is 64 bit. This method is expected to be faster than none()
-  // and is equivalent when the vector register is the result of a comparison,
-  // with byte values 0xff and 0x00.
-  simdutf_really_inline uint64_t to_bitmask64() const {
-    return vget_lane_u64(
-        vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
-  }
-
-  simdutf_really_inline bool any() const {
-    return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0;
-  }
-  simdutf_really_inline bool none() const {
-    return vmaxvq_u32(vreinterpretq_u32_u8(*this)) == 0;
-  }
-  simdutf_really_inline bool all() const {
-    return vminvq_u32(vreinterpretq_u32_u8(*this)) == 0xFFFFF;
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base_u8<uint8_t> {
-  static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) {
-    return vmovq_n_u8(_value);
-  }
-  static simdutf_really_inline simd8<uint8_t> zero() { return vdupq_n_u8(0); }
-  static simdutf_really_inline simd8<uint8_t> load(const uint8_t *values) {
-    return vld1q_u8(values);
-  }
-  simdutf_really_inline simd8(const uint8x16_t _value)
-      : base_u8<uint8_t>(_value) {}
-  // Zero constructor
-  simdutf_really_inline simd8() : simd8(zero()) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Member-by-member initialization
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
-      : simd8(simdutf_make_uint8x16_t(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                                      v10, v11, v12, v13, v14, v15)) {}
-#else
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
-      : simd8(uint8x16_t{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15}) {}
-#endif
-
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15);
-  }
-
-  // Store to array
-  simdutf_really_inline void store(uint8_t dst[16]) const {
-    return vst1q_u8(dst, *this);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return vqaddq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return vqsubq_u8(*this, other);
-  }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<uint8_t>
-  operator+(const simd8<uint8_t> other) const {
-    return vaddq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  operator-(const simd8<uint8_t> other) const {
-    return vsubq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t> &operator+=(const simd8<uint8_t> other) {
-    *this = *this + other;
-    return *this;
-  }
-  simdutf_really_inline simd8<uint8_t> &operator-=(const simd8<uint8_t> other) {
-    *this = *this - other;
-    return *this;
-  }
-
-  // Order-specific operations
-  simdutf_really_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
-  simdutf_really_inline uint8_t min_val() const { return vminvq_u8(*this); }
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return vmaxq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return vminq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return vcleq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return vcgeq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return vcltq_u8(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return vcgtq_u8(*this, other);
-  }
-  // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true
-  // = nonzero. For ARM, returns all 1's.
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return simd8<uint8_t>(*this > other);
-  }
-  // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true
-  // = nonzero. For ARM, returns all 1's.
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return simd8<uint8_t>(*this < other);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return vtstq_u8(*this, bits);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    return this->max_val() < 0b10000000u;
-  }
-
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return this->max_val() != 0;
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return (*this & bits).any_bits_set_anywhere();
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return vshrq_n_u8(*this, N);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return vshlq_n_u8(*this, N);
-  }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return lookup_table.apply_lookup_16_to(*this);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-
-  template <typename T>
-  simdutf_really_inline simd8<uint8_t>
-  apply_lookup_16_to(const simd8<T> original) const {
-    return vqtbl1q_u8(*this, simd8<uint8_t>(original));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> {
-  int8x16_t value;
-
-  static simdutf_really_inline simd8<int8_t> splat(int8_t _value) {
-    return vmovq_n_s8(_value);
-  }
-  static simdutf_really_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
-  static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) {
-    return vld1q_s8(values);
-  }
-
-  // Use ST2 instead of UXTL+UXTL2 to interleave zeroes. UXTL is actually a
-  // USHLL #0, and shifting in NEON is actually quite slow.
-  //
-  // While this needs the registers to be in a specific order, bigger cores can
-  // interleave these with no overhead, and it still performs decently on little
-  // cores.
-  //    movi  v1.3d, #0
-  //      mov   v0.16b, value[0]
-  //    st2   {v0.16b, v1.16b}, [ptr], #32
-  //      mov   v0.16b, value[1]
-  //    st2   {v0.16b, v1.16b}, [ptr], #32
-  //    ...
-  template <endianness big_endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
-    int8x16x2_t pair = match_system(big_endian)
-                           ? int8x16x2_t{{this->value, vmovq_n_s8(0)}}
-                           : int8x16x2_t{{vmovq_n_s8(0), this->value}};
-    vst2q_s8(reinterpret_cast<int8_t *>(p), pair);
-  }
-
-  // currently unused
-  // Technically this could be done with ST4 like in store_ascii_as_utf16, but
-  // it is very much not worth it, as explicitly mentioned in the ARM Cortex-X1
-  // Core Software Optimization Guide:
-  //   4.18 Complex ASIMD instructions
-  //     The bandwidth of [ST4 with element size less than 64b] is limited by
-  //     decode constraints and it is advisable to avoid them when high
-  //     performing code is desired.
-  // Instead, it is better to use ZIP1+ZIP2 and two ST2.
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
-    const uint16x8_t low =
-        vreinterpretq_u16_s8(vzip1q_s8(this->value, vmovq_n_s8(0)));
-    const uint16x8_t high =
-        vreinterpretq_u16_s8(vzip2q_s8(this->value, vmovq_n_s8(0)));
-    const uint16x8x2_t low_pair{{low, vmovq_n_u16(0)}};
-    vst2q_u16(reinterpret_cast<uint16_t *>(p), low_pair);
-    const uint16x8x2_t high_pair{{high, vmovq_n_u16(0)}};
-    vst2q_u16(reinterpret_cast<uint16_t *>(p + 8), high_pair);
-  }
-
-  // In places where the table can be reused, which is most uses in simdutf, it
-  // is worth it to do 4 table lookups, as there is no direct zero extension
-  // from u8 to u32.
-  simdutf_really_inline void store_ascii_as_utf32_tbl(char32_t *p) const {
-    const simd8<uint8_t> tb1{0, 255, 255, 255, 1, 255, 255, 255,
-                             2, 255, 255, 255, 3, 255, 255, 255};
-    const simd8<uint8_t> tb2{4, 255, 255, 255, 5, 255, 255, 255,
-                             6, 255, 255, 255, 7, 255, 255, 255};
-    const simd8<uint8_t> tb3{8,  255, 255, 255, 9,  255, 255, 255,
-                             10, 255, 255, 255, 11, 255, 255, 255};
-    const simd8<uint8_t> tb4{12, 255, 255, 255, 13, 255, 255, 255,
-                             14, 255, 255, 255, 15, 255, 255, 255};
-
-    // encourage store pairing and interleaving
-    const auto shuf1 = this->apply_lookup_16_to(tb1);
-    const auto shuf2 = this->apply_lookup_16_to(tb2);
-    shuf1.store(reinterpret_cast<int8_t *>(p));
-    shuf2.store(reinterpret_cast<int8_t *>(p + 4));
-
-    const auto shuf3 = this->apply_lookup_16_to(tb3);
-    const auto shuf4 = this->apply_lookup_16_to(tb4);
-    shuf3.store(reinterpret_cast<int8_t *>(p + 8));
-    shuf4.store(reinterpret_cast<int8_t *>(p + 12));
-  }
-  // Conversion from/to SIMD register
-  simdutf_really_inline simd8(const int8x16_t _value) : value{_value} {}
-  simdutf_really_inline operator const int8x16_t &() const {
-    return this->value;
-  }
-#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
-  simdutf_really_inline operator const uint8x16_t() const {
-    return vreinterpretq_u8_s8(this->value);
-  }
-#endif
-  simdutf_really_inline operator int8x16_t &() { return this->value; }
-
-  // Zero constructor
-  simdutf_really_inline simd8() : simd8(zero()) {}
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
-                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
-                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
-      : simd8(simdutf_make_int8x16_t(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                                     v10, v11, v12, v13, v14, v15)) {}
-#else
-  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
-                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
-                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
-      : simd8(int8x16_t{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                        v13, v14, v15}) {}
-#endif
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15);
-  }
-
-  // Store to array
-  simdutf_really_inline void store(int8_t dst[16]) const {
-    return vst1q_s8(dst, value);
-  }
-  // Explicit conversion to/from unsigned
-  //
-  // Under Visual Studio/ARM64 uint8x16_t and int8x16_t are apparently the same
-  // type. In theory, we could check this occurrence with std::same_as and
-  // std::enabled_if but it is C++14 and relatively ugly and hard to read.
-#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
-  simdutf_really_inline explicit simd8(const uint8x16_t other)
-      : simd8(vreinterpretq_s8_u8(other)) {}
-#endif
-  simdutf_really_inline operator simd8<uint8_t>() const {
-    return vreinterpretq_u8_s8(this->value);
-  }
-
-  simdutf_really_inline simd8<int8_t>
-  operator|(const simd8<int8_t> other) const {
-    return vorrq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator&(const simd8<int8_t> other) const {
-    return vandq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator^(const simd8<int8_t> other) const {
-    return veorq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  bit_andnot(const simd8<int8_t> other) const {
-    return vbicq_s8(value, other.value);
-  }
-
-  // Math
-  simdutf_really_inline simd8<int8_t>
-  operator+(const simd8<int8_t> other) const {
-    return vaddq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator-(const simd8<int8_t> other) const {
-    return vsubq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t> &operator+=(const simd8<int8_t> other) {
-    *this = *this + other;
-    return *this;
-  }
-  simdutf_really_inline simd8<int8_t> &operator-=(const simd8<int8_t> other) {
-    *this = *this - other;
-    return *this;
-  }
-
-  simdutf_really_inline int8_t max_val() const { return vmaxvq_s8(value); }
-  simdutf_really_inline int8_t min_val() const { return vminvq_s8(value); }
-  simdutf_really_inline bool is_ascii() const { return this->min_val() >= 0; }
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return vmaxq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return vminq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return vcgtq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return vcltq_s8(value, other.value);
-  }
-  simdutf_really_inline simd8<bool>
-  operator==(const simd8<int8_t> other) const {
-    return vceqq_s8(value, other.value);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd8<int8_t>
-  prev(const simd8<int8_t> prev_chunk) const {
-    return vextq_s8(prev_chunk, *this, 16 - N);
-  }
-
-  // Perform a lookup assuming no value is larger than 16
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return lookup_table.apply_lookup_16_to(*this);
-  }
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-
-  template <typename T>
-  simdutf_really_inline simd8<int8_t>
-  apply_lookup_16_to(const simd8<T> original) const {
-    return vqtbl1q_s8(*this, simd8<uint8_t>(original));
-  }
-};
-
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(NUM_CHUNKS == 4,
-                "ARM kernel should use four registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
-                                 const simd8<T> chunk2, const simd8<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    this->chunks[2] |= other.chunks[2];
-    this->chunks[3] |= other.chunks[3];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
-
-  template <endianness endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 0);
-    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 1);
-    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 2);
-    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 0);
-    this->chunks[1].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 1);
-    this->chunks[2].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 2);
-    this->chunks[3].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-    const uint8x16_t bit_mask =
-        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
-#else
-    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
-#endif
-    // Add each of the elements next to each other, successively, to stuff each
-    // 8 byte mask into one.
-    uint8x16_t sum0 =
-        vpaddq_u8(vandq_u8(uint8x16_t(this->chunks[0]), bit_mask),
-                  vandq_u8(uint8x16_t(this->chunks[1]), bit_mask));
-    uint8x16_t sum1 =
-        vpaddq_u8(vandq_u8(uint8x16_t(this->chunks[2]), bit_mask),
-                  vandq_u8(uint8x16_t(this->chunks[3]), bit_mask));
-    sum0 = vpaddq_u8(sum0, sum1);
-    sum0 = vpaddq_u8(sum0, sum0);
-    return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                          this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                          this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-    return simd8x64<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
-               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
-               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                          this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
-                          this->chunks[2] > mask, this->chunks[3] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
-                          this->chunks[2] >= mask, this->chunks[3] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>(simd8<uint8_t>(uint8x16_t(this->chunks[0])) >= mask,
-                          simd8<uint8_t>(uint8x16_t(this->chunks[1])) >= mask,
-                          simd8<uint8_t>(uint8x16_t(this->chunks[2])) >= mask,
-                          simd8<uint8_t>(uint8x16_t(this->chunks[3])) >= mask)
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-/* begin file src/simdutf/arm64/simd16-inl.h */
-template <typename T> struct simd16;
-
-template <typename T, typename Mask = simd16<bool>> struct base_u16 {
-  uint16x8_t value;
-  static const int SIZE = sizeof(value);
-
-  // Conversion from/to SIMD register
-  simdutf_really_inline base_u16() = default;
-  simdutf_really_inline base_u16(const uint16x8_t _value) : value(_value) {}
-  simdutf_really_inline operator const uint16x8_t &() const {
-    return this->value;
-  }
-  simdutf_really_inline operator uint16x8_t &() { return this->value; }
-  // Bit operations
-  simdutf_really_inline simd16<T> operator|(const simd16<T> other) const {
-    return vorrq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator&(const simd16<T> other) const {
-    return vandq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator^(const simd16<T> other) const {
-    return veorq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<T> bit_andnot(const simd16<T> other) const {
-    return vbicq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
-  simdutf_really_inline simd16<T> &operator|=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd16<T> &operator&=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd16<T> &operator^=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-
-  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
-                                               const simd16<T> rhs) {
-    return vceqq_u16(lhs, rhs);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return vextq_u18(prev_chunk, *this, 8 - N);
-  }
-};
-
-template <typename T, typename Mask = simd16<bool>>
-struct base16 : base_u16<T> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  simdutf_really_inline base16() : base_u16<T>() {}
-  simdutf_really_inline base16(const uint16x8_t _value) : base_u16<T>(_value) {}
-  template <typename Pointer>
-  simdutf_really_inline base16(const Pointer *ptr) : base16(vld1q_u16(ptr)) {}
-
-  static const int SIZE = sizeof(base_u16<T>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return vextq_u18(prev_chunk, *this, 8 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd16<bool> : base16<bool> {
-  static simdutf_really_inline simd16<bool> splat(bool _value) {
-    return vmovq_n_u16(uint16_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd16() : base16() {}
-  simdutf_really_inline simd16(const uint16x8_t _value)
-      : base16<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
-};
-
-template <typename T> struct base16_numeric : base16<T> {
-  static simdutf_really_inline simd16<T> splat(T _value) {
-    return vmovq_n_u16(_value);
-  }
-  static simdutf_really_inline simd16<T> zero() { return vdupq_n_u16(0); }
-  static simdutf_really_inline simd16<T> load(const T values[8]) {
-    return vld1q_u16(reinterpret_cast<const uint16_t *>(values));
-  }
-
-  simdutf_really_inline base16_numeric() : base16<T>() {}
-  simdutf_really_inline base16_numeric(const uint16x8_t _value)
-      : base16<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[8]) const {
-    return vst1q_u16(dst, *this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
-    return vaddq_u8(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
-    return vsubq_u8(*this, other);
-  }
-  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
-    *this = *this + other;
-    return *static_cast<simd16<T> *>(this);
-  }
-  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
-    *this = *this - other;
-    return *static_cast<simd16<T> *>(this);
-  }
-};
-
-// Signed code units
-template <> struct simd16<int16_t> : base16_numeric<int16_t> {
-  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
-#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
-  simdutf_really_inline simd16(const uint16x8_t _value)
-      : base16_numeric<int16_t>(_value) {}
-#endif
-  simdutf_really_inline simd16(const int16x8_t _value)
-      : base16_numeric<int16_t>(vreinterpretq_u16_s16(_value)) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
-  simdutf_really_inline operator simd16<uint16_t>() const;
-  simdutf_really_inline operator const uint16x8_t &() const {
-    return this->value;
-  }
-  simdutf_really_inline operator const int16x8_t() const {
-    return vreinterpretq_s16_u16(this->value);
-  }
-
-  simdutf_really_inline int16_t max_val() const {
-    return vmaxvq_s16(vreinterpretq_s16_u16(this->value));
-  }
-  simdutf_really_inline int16_t min_val() const {
-    return vminvq_s16(vreinterpretq_s16_u16(this->value));
-  }
-  // Order-sensitive comparisons
-  simdutf_really_inline simd16<int16_t>
-  max_val(const simd16<int16_t> other) const {
-    return vmaxq_s16(vreinterpretq_s16_u16(this->value),
-                     vreinterpretq_s16_u16(other.value));
-  }
-  simdutf_really_inline simd16<int16_t>
-  min_val(const simd16<int16_t> other) const {
-    return vmaxq_s16(vreinterpretq_s16_u16(this->value),
-                     vreinterpretq_s16_u16(other.value));
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<int16_t> other) const {
-    return vcgtq_s16(vreinterpretq_s16_u16(this->value),
-                     vreinterpretq_s16_u16(other.value));
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<int16_t> other) const {
-    return vcltq_s16(vreinterpretq_s16_u16(this->value),
-                     vreinterpretq_s16_u16(other.value));
-  }
-};
-
-// Unsigned code units
-template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
-  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
-  simdutf_really_inline simd16(const uint16x8_t _value)
-      : base16_numeric<uint16_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
-
-  simdutf_really_inline int16_t max_val() const { return vmaxvq_u16(*this); }
-  simdutf_really_inline int16_t min_val() const { return vminvq_u16(*this); }
-  // Saturated math
-  simdutf_really_inline simd16<uint16_t>
-  saturating_add(const simd16<uint16_t> other) const {
-    return vqaddq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  saturating_sub(const simd16<uint16_t> other) const {
-    return vqsubq_u16(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd16<uint16_t>
-  max_val(const simd16<uint16_t> other) const {
-    return vmaxq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  min_val(const simd16<uint16_t> other) const {
-    return vminq_u16(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  gt_bits(const simd16<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  lt_bits(const simd16<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<=(const simd16<uint16_t> other) const {
-    return vcleq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>=(const simd16<uint16_t> other) const {
-    return vcgeq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<uint16_t> other) const {
-    return vcgtq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<uint16_t> other) const {
-    return vcltq_u16(*this, other);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd16<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
-    return simd16<uint16_t>(vshrq_n_u16(*this, N));
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
-    return simd16<uint16_t>(vshlq_n_u16(*this, N));
-  }
-
-  // logical operations
-  simdutf_really_inline simd16<uint16_t>
-  operator|(const simd16<uint16_t> other) const {
-    return vorrq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  operator&(const simd16<uint16_t> other) const {
-    return vandq_u16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  operator^(const simd16<uint16_t> other) const {
-    return veorq_u16(*this, other);
-  }
-
-  // Pack with the unsigned saturation of two uint16_t code units into single
-  // uint8_t vector
-  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
-                                                   const simd16<uint16_t> &v1) {
-    return vqmovn_high_u16(vqmovn_u16(v0), v1);
-  }
-
-  // Change the endianness
-  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
-    return vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(*this)));
-  }
-};
-simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
-  return this->value;
-}
-
-template <typename T> struct simd16x32 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
-  static_assert(NUM_CHUNKS == 4,
-                "ARM kernel should use four registers per 64-byte block.");
-  simd16<T> chunks[NUM_CHUNKS];
-
-  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
-  simd16x32<T> &
-  operator=(const simd16<T> other) = delete; // no assignment allowed
-  simd16x32() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline
-  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
-            const simd16<T> chunk2, const simd16<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd16x32(const T *ptr)
-      : chunks{simd16<T>::load(ptr),
-               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd16<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
-
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
-    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
-    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
-    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-    const uint8x16_t bit_mask =
-        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
-#else
-    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
-#endif
-    // Add each of the elements next to each other, successively, to stuff each
-    // 8 byte mask into one.
-    uint8x16_t sum0 = vpaddq_u8(
-        vreinterpretq_u8_u16(this->chunks[0] & vreinterpretq_u16_u8(bit_mask)),
-        vreinterpretq_u8_u16(this->chunks[1] & vreinterpretq_u16_u8(bit_mask)));
-    uint8x16_t sum1 = vpaddq_u8(
-        vreinterpretq_u8_u16(this->chunks[2] & vreinterpretq_u16_u8(bit_mask)),
-        vreinterpretq_u8_u16(this->chunks[3] & vreinterpretq_u16_u8(bit_mask)));
-    sum0 = vpaddq_u8(sum0, sum1);
-    sum0 = vpaddq_u8(sum0, sum0);
-    return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
-  }
-
-  simdutf_really_inline void swap_bytes() {
-    this->chunks[0] = this->chunks[0].swap_bytes();
-    this->chunks[1] = this->chunks[1].swap_bytes();
-    this->chunks[2] = this->chunks[2].swap_bytes();
-    this->chunks[3] = this->chunks[3].swap_bytes();
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                           this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                           this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-
-    return simd16x32<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-    return simd16x32<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
-               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
-               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                           this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-
-}; // struct simd16x32<T>
-template <>
-simdutf_really_inline uint64_t simd16x32<uint16_t>::not_in_range(
-    const uint16_t low, const uint16_t high) const {
-  const simd16<uint16_t> mask_low = simd16<uint16_t>::splat(low);
-  const simd16<uint16_t> mask_high = simd16<uint16_t>::splat(high);
-  simd16x32<uint16_t> x(simd16<uint16_t>((this->chunks[0] > mask_high) |
-                                         (this->chunks[0] < mask_low)),
-                        simd16<uint16_t>((this->chunks[1] > mask_high) |
-                                         (this->chunks[1] < mask_low)),
-                        simd16<uint16_t>((this->chunks[2] > mask_high) |
-                                         (this->chunks[2] < mask_low)),
-                        simd16<uint16_t>((this->chunks[3] > mask_high) |
-                                         (this->chunks[3] < mask_low)));
-  return x.to_bitmask();
-}
-/* end file src/simdutf/arm64/simd16-inl.h */
-} // namespace simd
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-
-#endif // SIMDUTF_ARM64_SIMD_H
-/* end file src/simdutf/arm64/simd.h */
-
-/* begin file src/simdutf/arm64/end.h */
-/* end file src/simdutf/arm64/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_ARM64
-
-#endif // SIMDUTF_ARM64_H
-/* end file src/simdutf/arm64.h */
-/* begin file src/simdutf/icelake.h */
-#ifndef SIMDUTF_ICELAKE_H
-#define SIMDUTF_ICELAKE_H
-
-
-#ifdef __has_include
-  // How do we detect that a compiler supports vbmi2?
-  // For sure if the following header is found, we are ok?
-  #if __has_include(<avx512vbmi2intrin.h>)
-    #define SIMDUTF_COMPILER_SUPPORTS_VBMI2 1
-  #endif
-#endif
-
-#ifdef _MSC_VER
-  #if _MSC_VER >= 1930
-    // Visual Studio 2022 and up support VBMI2 under x64 even if the header
-    // avx512vbmi2intrin.h is not found.
-    // Visual Studio 2019 technically supports VBMI2, but the implementation
-    // might be unreliable. Search for visualstudio2019icelakeissue in our
-    // tests.
-    #define SIMDUTF_COMPILER_SUPPORTS_VBMI2 1
-  #endif
-#endif
-
-// We allow icelake on x64 as long as the compiler is known to support VBMI2.
-#ifndef SIMDUTF_IMPLEMENTATION_ICELAKE
-  #define SIMDUTF_IMPLEMENTATION_ICELAKE                                       \
-    ((SIMDUTF_IS_X86_64) && (SIMDUTF_COMPILER_SUPPORTS_VBMI2))
-#endif
-
-// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
-// https://github.com/simdutf/simdutf/issues/1247
-#if ((SIMDUTF_IMPLEMENTATION_ICELAKE) && (SIMDUTF_IS_X86_64) && (__AVX2__) &&  \
-     (SIMDUTF_HAS_AVX512F && SIMDUTF_HAS_AVX512DQ && SIMDUTF_HAS_AVX512VL &&   \
-      SIMDUTF_HAS_AVX512VBMI2) &&                                              \
-     (!SIMDUTF_IS_32BITS))
-  #define SIMDUTF_CAN_ALWAYS_RUN_ICELAKE 1
-#else
-  #define SIMDUTF_CAN_ALWAYS_RUN_ICELAKE 0
-#endif
-
-#if SIMDUTF_IMPLEMENTATION_ICELAKE
-  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-    #define SIMDUTF_TARGET_ICELAKE
-  #else
-    #define SIMDUTF_TARGET_ICELAKE                                             \
-      SIMDUTF_TARGET_REGION(                                                   \
-          "avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,"         \
-          "avx512vl,avx2,bmi,bmi2,pclmul,lzcnt,popcnt,avx512vpopcntdq")
-  #endif
-
-namespace simdutf {
-namespace icelake {} // namespace icelake
-} // namespace simdutf
-
-  //
-  // These two need to be included outside SIMDUTF_TARGET_REGION
-  //
-/* begin file src/simdutf/icelake/intrinsics.h */
-#ifndef SIMDUTF_ICELAKE_INTRINSICS_H
-#define SIMDUTF_ICELAKE_INTRINSICS_H
-
-
-#ifdef SIMDUTF_VISUAL_STUDIO
-  // under clang within visual studio, this will include <x86intrin.h>
-  #include <intrin.h> // visual studio or clang
-  #include <immintrin.h>
-#else
-
-  #if SIMDUTF_GCC11ORMORE
-// We should not get warnings while including <x86intrin.h> yet we do
-// under some versions of GCC.
-// If the x86intrin.h header has uninitialized values that are problematic,
-// it is a GCC issue, we want to ignore these warnings.
-SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
-  #endif
-
-  #include <x86intrin.h> // elsewhere
-
-  #if SIMDUTF_GCC11ORMORE
-// cancels the suppression of the -Wuninitialized
-SIMDUTF_POP_DISABLE_WARNINGS
-  #endif
-
-  #ifndef _tzcnt_u64
-    #define _tzcnt_u64(x) __tzcnt_u64(x)
-  #endif // _tzcnt_u64
-#endif   // SIMDUTF_VISUAL_STUDIO
-
-#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
-  /**
-   * You are not supposed, normally, to include these
-   * headers directly. Instead you should either include intrin.h
-   * or x86intrin.h. However, when compiling with clang
-   * under Windows (i.e., when _MSC_VER is set), these headers
-   * only get included *if* the corresponding features are detected
-   * from macros:
-   * e.g., if __AVX2__ is set... in turn,  we normally set these
-   * macros by compiling against the corresponding architecture
-   * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
-   * software with these advanced instructions. In simdutf, we
-   * want to compile the whole program for a generic target,
-   * and only target our specific kernels. As a workaround,
-   * we directly include the needed headers. These headers would
-   * normally guard against such usage, but we carefully included
-   * <x86intrin.h>  (or <intrin.h>) before, so the headers
-   * are fooled.
-   */
-  #include <bmiintrin.h>   // for _blsr_u64
-  #include <bmi2intrin.h>  // for _pext_u64, _pdep_u64
-  #include <lzcntintrin.h> // for  __lzcnt64
-  #include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
-  #include <smmintrin.h>
-  #include <tmmintrin.h>
-  #include <avxintrin.h>
-  #include <avx2intrin.h>
-  // Important: we need the AVX-512 headers:
-  #include <avx512fintrin.h>
-  #include <avx512dqintrin.h>
-  #include <avx512cdintrin.h>
-  #include <avx512bwintrin.h>
-  #include <avx512vlintrin.h>
-  #include <avx512vlbwintrin.h>
-  #include <avx512vbmiintrin.h>
-  #include <avx512vbmi2intrin.h>
-  #include <avx512vpopcntdqintrin.h>
-  #include <avx512vpopcntdqvlintrin.h>
-  // unfortunately, we may not get _blsr_u64, but, thankfully, clang
-  // has it as a macro.
-  #ifndef _blsr_u64
-    // we roll our own
-    #define _blsr_u64(n) ((n - 1) & n)
-  #endif //  _blsr_u64
-#endif   // SIMDUTF_CLANG_VISUAL_STUDIO
-
-#if defined(__GNUC__) && !defined(__clang__)
-
-  #if __GNUC__ == 8
-    #define SIMDUTF_GCC8 1
-  #elif __GNUC__ == 9
-    #define SIMDUTF_GCC9 1
-  #endif //  __GNUC__ == 8 || __GNUC__ == 9
-
-#endif // defined(__GNUC__) && !defined(__clang__)
-
-#if SIMDUTF_GCC8
-  #pragma GCC push_options
-  #pragma GCC target("avx512f")
-/**
- * GCC 8 fails to provide _mm512_set_epi8. We roll our own.
- */
-inline __m512i
-_mm512_set_epi8(uint8_t a0, uint8_t a1, uint8_t a2, uint8_t a3, uint8_t a4,
-                uint8_t a5, uint8_t a6, uint8_t a7, uint8_t a8, uint8_t a9,
-                uint8_t a10, uint8_t a11, uint8_t a12, uint8_t a13, uint8_t a14,
-                uint8_t a15, uint8_t a16, uint8_t a17, uint8_t a18, uint8_t a19,
-                uint8_t a20, uint8_t a21, uint8_t a22, uint8_t a23, uint8_t a24,
-                uint8_t a25, uint8_t a26, uint8_t a27, uint8_t a28, uint8_t a29,
-                uint8_t a30, uint8_t a31, uint8_t a32, uint8_t a33, uint8_t a34,
-                uint8_t a35, uint8_t a36, uint8_t a37, uint8_t a38, uint8_t a39,
-                uint8_t a40, uint8_t a41, uint8_t a42, uint8_t a43, uint8_t a44,
-                uint8_t a45, uint8_t a46, uint8_t a47, uint8_t a48, uint8_t a49,
-                uint8_t a50, uint8_t a51, uint8_t a52, uint8_t a53, uint8_t a54,
-                uint8_t a55, uint8_t a56, uint8_t a57, uint8_t a58, uint8_t a59,
-                uint8_t a60, uint8_t a61, uint8_t a62, uint8_t a63) {
-  return _mm512_set_epi64(
-      uint64_t(a7) + (uint64_t(a6) << 8) + (uint64_t(a5) << 16) +
-          (uint64_t(a4) << 24) + (uint64_t(a3) << 32) + (uint64_t(a2) << 40) +
-          (uint64_t(a1) << 48) + (uint64_t(a0) << 56),
-      uint64_t(a15) + (uint64_t(a14) << 8) + (uint64_t(a13) << 16) +
-          (uint64_t(a12) << 24) + (uint64_t(a11) << 32) +
-          (uint64_t(a10) << 40) + (uint64_t(a9) << 48) + (uint64_t(a8) << 56),
-      uint64_t(a23) + (uint64_t(a22) << 8) + (uint64_t(a21) << 16) +
-          (uint64_t(a20) << 24) + (uint64_t(a19) << 32) +
-          (uint64_t(a18) << 40) + (uint64_t(a17) << 48) + (uint64_t(a16) << 56),
-      uint64_t(a31) + (uint64_t(a30) << 8) + (uint64_t(a29) << 16) +
-          (uint64_t(a28) << 24) + (uint64_t(a27) << 32) +
-          (uint64_t(a26) << 40) + (uint64_t(a25) << 48) + (uint64_t(a24) << 56),
-      uint64_t(a39) + (uint64_t(a38) << 8) + (uint64_t(a37) << 16) +
-          (uint64_t(a36) << 24) + (uint64_t(a35) << 32) +
-          (uint64_t(a34) << 40) + (uint64_t(a33) << 48) + (uint64_t(a32) << 56),
-      uint64_t(a47) + (uint64_t(a46) << 8) + (uint64_t(a45) << 16) +
-          (uint64_t(a44) << 24) + (uint64_t(a43) << 32) +
-          (uint64_t(a42) << 40) + (uint64_t(a41) << 48) + (uint64_t(a40) << 56),
-      uint64_t(a55) + (uint64_t(a54) << 8) + (uint64_t(a53) << 16) +
-          (uint64_t(a52) << 24) + (uint64_t(a51) << 32) +
-          (uint64_t(a50) << 40) + (uint64_t(a49) << 48) + (uint64_t(a48) << 56),
-      uint64_t(a63) + (uint64_t(a62) << 8) + (uint64_t(a61) << 16) +
-          (uint64_t(a60) << 24) + (uint64_t(a59) << 32) +
-          (uint64_t(a58) << 40) + (uint64_t(a57) << 48) +
-          (uint64_t(a56) << 56));
-}
-  #pragma GCC pop_options
-#endif // SIMDUTF_GCC8
-
-#endif // SIMDUTF_HASWELL_INTRINSICS_H
-/* end file src/simdutf/icelake/intrinsics.h */
-/* begin file src/simdutf/icelake/implementation.h */
-#ifndef SIMDUTF_ICELAKE_IMPLEMENTATION_H
-#define SIMDUTF_ICELAKE_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace icelake {
-
-namespace {
-using namespace simdutf;
-}
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation(
-            "icelake",
-            "Intel AVX512 (AVX-512BW, AVX-512CD, AVX-512VL, AVX-512VBMI2 "
-            "extensions)",
-            internal::instruction_set::AVX2 | internal::instruction_set::BMI1 |
-                internal::instruction_set::BMI2 |
-                internal::instruction_set::AVX512BW |
-                internal::instruction_set::AVX512CD |
-                internal::instruction_set::AVX512VL |
-                internal::instruction_set::AVX512VBMI2 |
-                internal::instruction_set::AVX512VPOPCNTDQ) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-
-} // namespace icelake
-} // namespace simdutf
-
-#endif // SIMDUTF_ICELAKE_IMPLEMENTATION_H
-/* end file src/simdutf/icelake/implementation.h */
-
-  //
-  // The rest need to be inside the region
-  //
-/* begin file src/simdutf/icelake/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "icelake"
-// #define SIMDUTF_IMPLEMENTATION icelake
-
-#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-// nothing needed.
-#else
-SIMDUTF_TARGET_ICELAKE
-#endif
-
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-// clang-format off
-SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
-// clang-format on
-#endif // end of workaround
-/* end file src/simdutf/icelake/begin.h */
-  // Declarations
-/* begin file src/simdutf/icelake/bitmanipulation.h */
-#ifndef SIMDUTF_ICELAKE_BITMANIPULATION_H
-#define SIMDUTF_ICELAKE_BITMANIPULATION_H
-
-namespace simdutf {
-namespace icelake {
-namespace {
-
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
-  // note: we do not support legacy 32-bit Windows
-  return __popcnt64(input_num); // Visual Studio wants two underscores
-}
-#else
-simdutf_really_inline long long int count_ones(uint64_t input_num) {
-  return _popcnt64(input_num);
-}
-#endif
-
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
-  #if SIMDUTF_REGULAR_VISUAL_STUDIO
-  return (int)_tzcnt_u64(input_num);
-  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
-  return __builtin_ctzll(input_num);
-  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
-}
-#endif
-
-} // unnamed namespace
-} // namespace icelake
-} // namespace simdutf
-
-#endif // SIMDUTF_ICELAKE_BITMANIPULATION_H
-/* end file src/simdutf/icelake/bitmanipulation.h */
-/* begin file src/simdutf/icelake/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
-
-
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-SIMDUTF_POP_DISABLE_WARNINGS
-#endif // end of workaround
-/* end file src/simdutf/icelake/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_ICELAKE
-#endif // SIMDUTF_ICELAKE_H
-/* end file src/simdutf/icelake.h */
-/* begin file src/simdutf/haswell.h */
-#ifndef SIMDUTF_HASWELL_H
-#define SIMDUTF_HASWELL_H
-
-#ifdef SIMDUTF_WESTMERE_H
-  #error "haswell.h must be included before westmere.h"
-#endif
-#ifdef SIMDUTF_FALLBACK_H
-  #error "haswell.h must be included before fallback.h"
-#endif
-
-
-// Default Haswell to on if this is x86-64. Even if we are not compiled for it,
-// it could be selected at runtime.
-#ifndef SIMDUTF_IMPLEMENTATION_HASWELL
-  //
-  // You do not want to restrict it like so: SIMDUTF_IS_X86_64 && __AVX2__
-  // because we want to rely on *runtime dispatch*.
-  //
-  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-    #define SIMDUTF_IMPLEMENTATION_HASWELL 0
-  #else
-    #define SIMDUTF_IMPLEMENTATION_HASWELL (SIMDUTF_IS_X86_64)
-  #endif
-
-#endif
-// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
-// https://github.com/simdutf/simdutf/issues/1247
-#if ((SIMDUTF_IMPLEMENTATION_HASWELL) && (SIMDUTF_IS_X86_64) && (__AVX2__))
-  #define SIMDUTF_CAN_ALWAYS_RUN_HASWELL 1
-#else
-  #define SIMDUTF_CAN_ALWAYS_RUN_HASWELL 0
-#endif
-
-#if SIMDUTF_IMPLEMENTATION_HASWELL
-
-  #define SIMDUTF_TARGET_HASWELL SIMDUTF_TARGET_REGION("avx2,bmi,lzcnt,popcnt")
-
-namespace simdutf {
-/**
- * Implementation for Haswell (Intel AVX2).
- */
-namespace haswell {} // namespace haswell
-} // namespace simdutf
-
-  //
-  // These two need to be included outside SIMDUTF_TARGET_REGION
-  //
-/* begin file src/simdutf/haswell/implementation.h */
-#ifndef SIMDUTF_HASWELL_IMPLEMENTATION_H
-#define SIMDUTF_HASWELL_IMPLEMENTATION_H
-
-
-// The constructor may be executed on any host, so we take care not to use
-// SIMDUTF_TARGET_REGION
-namespace simdutf {
-namespace haswell {
-
-using namespace simdutf;
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("haswell", "Intel/AMD AVX2",
-                                internal::instruction_set::AVX2 |
-                                    internal::instruction_set::BMI1 |
-                                    internal::instruction_set::BMI2) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused virtual size_t
-  maximal_binary_length_from_base64(const char *input,
-                                    size_t length) const noexcept;
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual size_t
-  maximal_binary_length_from_base64(const char16_t *input,
-                                    size_t length) const noexcept;
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual size_t
-  base64_length_from_binary(size_t length,
-                            base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-
-} // namespace haswell
-} // namespace simdutf
-
-#endif // SIMDUTF_HASWELL_IMPLEMENTATION_H
-/* end file src/simdutf/haswell/implementation.h */
-/* begin file src/simdutf/haswell/intrinsics.h */
-#ifndef SIMDUTF_HASWELL_INTRINSICS_H
-#define SIMDUTF_HASWELL_INTRINSICS_H
-
-
-#ifdef SIMDUTF_VISUAL_STUDIO
-  // under clang within visual studio, this will include <x86intrin.h>
-  #include <intrin.h> // visual studio or clang
-#else
-
-  #if SIMDUTF_GCC11ORMORE
-// We should not get warnings while including <x86intrin.h> yet we do
-// under some versions of GCC.
-// If the x86intrin.h header has uninitialized values that are problematic,
-// it is a GCC issue, we want to ignore these warnings.
-SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
-  #endif
-
-  #include <x86intrin.h> // elsewhere
-
-  #if SIMDUTF_GCC11ORMORE
-// cancels the suppression of the -Wuninitialized
-SIMDUTF_POP_DISABLE_WARNINGS
-  #endif
-
-#endif // SIMDUTF_VISUAL_STUDIO
-
-#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
-  /**
-   * You are not supposed, normally, to include these
-   * headers directly. Instead you should either include intrin.h
-   * or x86intrin.h. However, when compiling with clang
-   * under Windows (i.e., when _MSC_VER is set), these headers
-   * only get included *if* the corresponding features are detected
-   * from macros:
-   * e.g., if __AVX2__ is set... in turn,  we normally set these
-   * macros by compiling against the corresponding architecture
-   * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
-   * software with these advanced instructions. In simdutf, we
-   * want to compile the whole program for a generic target,
-   * and only target our specific kernels. As a workaround,
-   * we directly include the needed headers. These headers would
-   * normally guard against such usage, but we carefully included
-   * <x86intrin.h>  (or <intrin.h>) before, so the headers
-   * are fooled.
-   */
-  #include <bmiintrin.h>   // for _blsr_u64
-  #include <lzcntintrin.h> // for  __lzcnt64
-  #include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
-  #include <smmintrin.h>
-  #include <tmmintrin.h>
-  #include <avxintrin.h>
-  #include <avx2intrin.h>
-  // unfortunately, we may not get _blsr_u64, but, thankfully, clang
-  // has it as a macro.
-  #ifndef _blsr_u64
-    // we roll our own
-    #define _blsr_u64(n) ((n - 1) & n)
-  #endif //  _blsr_u64
-#endif   // SIMDUTF_CLANG_VISUAL_STUDIO
-
-#endif // SIMDUTF_HASWELL_INTRINSICS_H
-/* end file src/simdutf/haswell/intrinsics.h */
-
-  //
-  // The rest need to be inside the region
-  //
-/* begin file src/simdutf/haswell/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "haswell"
-// #define SIMDUTF_IMPLEMENTATION haswell
-
-#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
-// nothing needed.
-#else
-SIMDUTF_TARGET_HASWELL
-#endif
-
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-// clang-format off
-SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
-// clang-format on
-#endif // end of workaround
-/* end file src/simdutf/haswell/begin.h */
-  // Declarations
-/* begin file src/simdutf/haswell/bitmanipulation.h */
-#ifndef SIMDUTF_HASWELL_BITMANIPULATION_H
-#define SIMDUTF_HASWELL_BITMANIPULATION_H
-
-namespace simdutf {
-namespace haswell {
-namespace {
-
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
-  // note: we do not support legacy 32-bit Windows
-  return __popcnt64(input_num); // Visual Studio wants two underscores
-}
-#else
-simdutf_really_inline long long int count_ones(uint64_t input_num) {
-  return _popcnt64(input_num);
-}
-#endif
-
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_inline int trailing_zeroes(uint64_t input_num) {
-  #if SIMDUTF_REGULAR_VISUAL_STUDIO
-  return (int)_tzcnt_u64(input_num);
-  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
-  return __builtin_ctzll(input_num);
-  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
-}
-#endif
-
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-
-#endif // SIMDUTF_HASWELL_BITMANIPULATION_H
-/* end file src/simdutf/haswell/bitmanipulation.h */
-/* begin file src/simdutf/haswell/simd.h */
-#ifndef SIMDUTF_HASWELL_SIMD_H
-#define SIMDUTF_HASWELL_SIMD_H
-
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace simd {
-
-// Forward-declared so they can be used by splat and friends.
-template <typename Child> struct base {
-  __m256i value;
-
-  // Zero constructor
-  simdutf_really_inline base() : value{__m256i()} {}
-
-  // Conversion from SIMD register
-  simdutf_really_inline base(const __m256i _value) : value(_value) {}
-  // Conversion to SIMD register
-  simdutf_really_inline operator const __m256i &() const { return this->value; }
-  simdutf_really_inline operator __m256i &() { return this->value; }
-  template <endianness big_endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this));
-    __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1));
-    if (big_endian) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      first = _mm256_shuffle_epi8(first, swap);
-      second = _mm256_shuffle_epi8(second, swap);
-    }
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first);
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second);
-  }
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr),
-                        _mm256_cvtepu8_epi32(_mm256_castsi256_si128(*this)));
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 8),
-                        _mm256_cvtepu8_epi32(_mm256_castsi256_si128(
-                            _mm256_srli_si256(*this, 8))));
-    _mm256_storeu_si256(
-        reinterpret_cast<__m256i *>(ptr + 16),
-        _mm256_cvtepu8_epi32(_mm256_extractf128_si256(*this, 1)));
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 24),
-                        _mm256_cvtepu8_epi32(_mm_srli_si128(
-                            _mm256_extractf128_si256(*this, 1), 8)));
-  }
-  // Bit operations
-  simdutf_really_inline Child operator|(const Child other) const {
-    return _mm256_or_si256(*this, other);
-  }
-  simdutf_really_inline Child operator&(const Child other) const {
-    return _mm256_and_si256(*this, other);
-  }
-  simdutf_really_inline Child operator^(const Child other) const {
-    return _mm256_xor_si256(*this, other);
-  }
-  simdutf_really_inline Child bit_andnot(const Child other) const {
-    return _mm256_andnot_si256(other, *this);
-  }
-  simdutf_really_inline Child &operator|=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator&=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator^=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-};
-
-// Forward-declared so they can be used by splat and friends.
-template <typename T> struct simd8;
-
-template <typename T, typename Mask = simd8<bool>>
-struct base8 : base<simd8<T>> {
-  typedef uint32_t bitmask_t;
-  typedef uint64_t bitmask2_t;
-
-  simdutf_really_inline base8() : base<simd8<T>>() {}
-  simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
-  simdutf_really_inline T first() const {
-    return _mm256_extract_epi8(*this, 0);
-  }
-  simdutf_really_inline T last() const {
-    return _mm256_extract_epi8(*this, 31);
-  }
-  friend simdutf_always_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return _mm256_cmpeq_epi8(lhs, rhs);
-  }
-
-  static const int SIZE = sizeof(base<T>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
-    return _mm256_alignr_epi8(
-        *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base8<bool> {
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return _mm256_set1_epi8(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8() : base8() {}
-  simdutf_really_inline simd8(const __m256i _value) : base8<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
-
-  simdutf_really_inline uint32_t to_bitmask() const {
-    return uint32_t(_mm256_movemask_epi8(*this));
-  }
-  simdutf_really_inline bool any() const {
-    return !_mm256_testz_si256(*this, *this);
-  }
-  simdutf_really_inline bool none() const {
-    return _mm256_testz_si256(*this, *this);
-  }
-  simdutf_really_inline bool all() const {
-    return static_cast<uint32_t>(_mm256_movemask_epi8(*this)) == 0xFFFFFFFF;
-  }
-  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base8_numeric : base8<T> {
-  static simdutf_really_inline simd8<T> splat(T _value) {
-    return _mm256_set1_epi8(_value);
-  }
-  static simdutf_really_inline simd8<T> zero() {
-    return _mm256_setzero_si256();
-  }
-  static simdutf_really_inline simd8<T> load(const T values[32]) {
-    return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
-  }
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
-                                                  T v5, T v6, T v7, T v8, T v9,
-                                                  T v10, T v11, T v12, T v13,
-                                                  T v14, T v15) {
-    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
-                    v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                    v12, v13, v14, v15);
-  }
-
-  simdutf_really_inline base8_numeric() : base8<T>() {}
-  simdutf_really_inline base8_numeric(const __m256i _value)
-      : base8<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[32]) const {
-    return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this);
-  }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
-    return _mm256_add_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
-    return _mm256_sub_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
-    *this = *this + other;
-    return *static_cast<simd8<T> *>(this);
-  }
-  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
-    *this = *this - other;
-    return *static_cast<simd8<T> *>(this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return _mm256_shuffle_epi8(lookup_table, *this);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> : base8_numeric<int8_t> {
-  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
-  simdutf_really_inline simd8(const __m256i _value)
-      : base8_numeric<int8_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
-  simdutf_really_inline operator simd8<uint8_t>() const;
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-        int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-        int8_t v12, int8_t v13, int8_t v14, int8_t v15, int8_t v16, int8_t v17,
-        int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
-        int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29,
-        int8_t v30, int8_t v31)
-      : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                               v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
-                               v22, v23, v24, v25, v26, v27, v28, v29, v30,
-                               v31)) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                         v10, v11, v12, v13, v14, v15);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    return _mm256_movemask_epi8(*this) == 0;
-  }
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return _mm256_max_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return _mm256_min_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return _mm256_cmpgt_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return _mm256_cmpgt_epi8(other, *this);
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
-  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
-  simdutf_really_inline simd8(const __m256i _value)
-      : base8_numeric<uint8_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
-        uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20,
-        uint8_t v21, uint8_t v22, uint8_t v23, uint8_t v24, uint8_t v25,
-        uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30,
-        uint8_t v31)
-      : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                               v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
-                               v22, v23, v24, v25, v26, v27, v28, v29, v30,
-                               v31)) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                          v10, v11, v12, v13, v14, v15);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return _mm256_adds_epu8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return _mm256_subs_epu8(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return _mm256_max_epu8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return _mm256_min_epu8(other, *this);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return this->lt_bits(other).any_bits_set();
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> bits_not_set() const {
-    return *this == uint8_t(0);
-  }
-  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    return _mm256_movemask_epi8(*this) == 0;
-  }
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return _mm256_testz_si256(*this, *this);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
-    return _mm256_testz_si256(*this, bits);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N);
-  }
-  // Get one of the bits and make a bitmask out of it.
-  // e.g. value.get_bit<7>() gets the high bit
-  template <int N> simdutf_really_inline int get_bit() const {
-    return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7 - N));
-  }
-};
-simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
-  return this->value;
-}
-
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(NUM_CHUNKS == 2,
-                "Haswell kernel should use two registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1)
-      : chunks{chunk0, chunk1} {}
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r_hi = this->chunks[1].to_bitmask();
-    return r_lo | (r_hi << 32);
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return this->chunks[0] | this->chunks[1];
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  template <endianness endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 0);
-    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 1);
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
-    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
-  }
-
-  simdutf_really_inline simd8x64<T> bit_or(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<T>(this->chunks[0] | mask, this->chunks[1] | mask);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
-    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
-                          this->chunks[1] == other.chunks[1])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-    return simd8x64<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>((simd8<uint8_t>(__m256i(this->chunks[0])) >= mask),
-                          (simd8<uint8_t>(__m256i(this->chunks[1])) >= mask))
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-
-/* begin file src/simdutf/haswell/simd16-inl.h */
-#ifdef __GNUC__
-  #if __GNUC__ < 8
-    #define _mm256_set_m128i(xmm1, xmm2)                                       \
-      _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1),                  \
-                                _mm256_castsi128_si256(xmm2), 2)
-    #define _mm256_setr_m128i(xmm2, xmm1)                                      \
-      _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1),                  \
-                                _mm256_castsi128_si256(xmm2), 2)
-  #endif
-#endif
-
-template <typename T> struct simd16;
-
-template <typename T, typename Mask = simd16<bool>>
-struct base16 : base<simd16<T>> {
-  using bitmask_type = uint32_t;
-
-  simdutf_really_inline base16() : base<simd16<T>>() {}
-  simdutf_really_inline base16(const __m256i _value)
-      : base<simd16<T>>(_value) {}
-  template <typename Pointer>
-  simdutf_really_inline base16(const Pointer *ptr)
-      : base16(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(ptr))) {}
-  friend simdutf_always_inline Mask operator==(const simd16<T> lhs,
-                                               const simd16<T> rhs) {
-    return _mm256_cmpeq_epi16(lhs, rhs);
-  }
-
-  /// the size of vector in bytes
-  static const int SIZE = sizeof(base<simd16<T>>::value);
-
-  /// the number of elements of type T a vector can hold
-  static const int ELEMENTS = SIZE / sizeof(T);
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return _mm256_alignr_epi8(*this, prev_chunk, 16 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd16<bool> : base16<bool> {
-  static simdutf_really_inline simd16<bool> splat(bool _value) {
-    return _mm256_set1_epi16(uint16_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd16() : base16() {}
-  simdutf_really_inline simd16(const __m256i _value) : base16<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
-
-  simdutf_really_inline bitmask_type to_bitmask() const {
-    return _mm256_movemask_epi8(*this);
-  }
-  simdutf_really_inline bool any() const {
-    return !_mm256_testz_si256(*this, *this);
-  }
-  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base16_numeric : base16<T> {
-  static simdutf_really_inline simd16<T> splat(T _value) {
-    return _mm256_set1_epi16(_value);
-  }
-  static simdutf_really_inline simd16<T> zero() {
-    return _mm256_setzero_si256();
-  }
-  static simdutf_really_inline simd16<T> load(const T values[8]) {
-    return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
-  }
-
-  simdutf_really_inline base16_numeric() : base16<T>() {}
-  simdutf_really_inline base16_numeric(const __m256i _value)
-      : base16<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[8]) const {
-    return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
-    return _mm256_add_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
-    return _mm256_sub_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
-    *this = *this + other;
-    return *static_cast<simd16<T> *>(this);
-  }
-  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
-    *this = *this - other;
-    return *static_cast<simd16<T> *>(this);
-  }
-};
-
-// Signed code units
-template <> struct simd16<int16_t> : base16_numeric<int16_t> {
-  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
-  simdutf_really_inline simd16(const __m256i _value)
-      : base16_numeric<int16_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
-  // Order-sensitive comparisons
-  simdutf_really_inline simd16<int16_t>
-  max_val(const simd16<int16_t> other) const {
-    return _mm256_max_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<int16_t>
-  min_val(const simd16<int16_t> other) const {
-    return _mm256_min_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<int16_t> other) const {
-    return _mm256_cmpgt_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<int16_t> other) const {
-    return _mm256_cmpgt_epi16(other, *this);
-  }
-};
-
-// Unsigned code units
-template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
-  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
-  simdutf_really_inline simd16(const __m256i _value)
-      : base16_numeric<uint16_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
-
-  // Saturated math
-  simdutf_really_inline simd16<uint16_t>
-  saturating_add(const simd16<uint16_t> other) const {
-    return _mm256_adds_epu16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  saturating_sub(const simd16<uint16_t> other) const {
-    return _mm256_subs_epu16(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd16<uint16_t>
-  max_val(const simd16<uint16_t> other) const {
-    return _mm256_max_epu16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  min_val(const simd16<uint16_t> other) const {
-    return _mm256_min_epu16(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  gt_bits(const simd16<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  lt_bits(const simd16<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<=(const simd16<uint16_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd16<bool>
-  operator>=(const simd16<uint16_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<uint16_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<uint16_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd16<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return _mm256_testz_si256(*this, *this);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool
-  bits_not_set_anywhere(simd16<uint16_t> bits) const {
-    return _mm256_testz_si256(*this, bits);
-  }
-  simdutf_really_inline bool
-  any_bits_set_anywhere(simd16<uint16_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
-    return simd16<uint16_t>(_mm256_srli_epi16(*this, N));
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
-    return simd16<uint16_t>(_mm256_slli_epi16(*this, N));
-  }
-  // Get one of the bits and make a bitmask out of it.
-  // e.g. value.get_bit<7>() gets the high bit
-  template <int N> simdutf_really_inline int get_bit() const {
-    return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 15 - N));
-  }
-
-  // Change the endianness
-  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
-    const __m256i swap = _mm256_setr_epi8(
-        1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-        21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-    return _mm256_shuffle_epi8(*this, swap);
-  }
-
-  // Pack with the unsigned saturation of two uint16_t code units into single
-  // uint8_t vector
-  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
-                                                   const simd16<uint16_t> &v1) {
-    // Note: the AVX2 variant of pack operates on 128-bit lanes, thus
-    //       we have to shuffle lanes in order to produce bytes in the
-    //       correct order.
-
-    // get the 0th lanes
-    const __m128i lo_0 = _mm256_extracti128_si256(v0, 0);
-    const __m128i lo_1 = _mm256_extracti128_si256(v1, 0);
-
-    // get the 1st lanes
-    const __m128i hi_0 = _mm256_extracti128_si256(v0, 1);
-    const __m128i hi_1 = _mm256_extracti128_si256(v1, 1);
-
-    // build new vectors (shuffle lanes)
-    const __m256i t0 = _mm256_set_m128i(lo_1, lo_0);
-    const __m256i t1 = _mm256_set_m128i(hi_1, hi_0);
-
-    // pack code units in linear order from v0 and v1
-    return _mm256_packus_epi16(t0, t1);
-  }
-};
-
-template <typename T> struct simd16x32 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
-  static_assert(NUM_CHUNKS == 2,
-                "Haswell kernel should use two registers per 64-byte block.");
-  simd16<T> chunks[NUM_CHUNKS];
-
-  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
-  simd16x32<T> &
-  operator=(const simd16<T> other) = delete; // no assignment allowed
-  simd16x32() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd16x32(const simd16<T> chunk0,
-                                  const simd16<T> chunk1)
-      : chunks{chunk0, chunk1} {}
-  simdutf_really_inline simd16x32(const T *ptr)
-      : chunks{simd16<T>::load(ptr),
-               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r_hi = this->chunks[1].to_bitmask();
-    return r_lo | (r_hi << 32);
-  }
-
-  simdutf_really_inline simd16<T> reduce_or() const {
-    return this->chunks[0] | this->chunks[1];
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
-    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>));
-  }
-
-  simdutf_really_inline simd16x32<T> bit_or(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<T>(this->chunks[0] | mask, this->chunks[1] | mask);
-  }
-
-  simdutf_really_inline void swap_bytes() {
-    this->chunks[0] = this->chunks[0].swap_bytes();
-    this->chunks[1] = this->chunks[1].swap_bytes();
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
-    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
-                           this->chunks[1] == other.chunks[1])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-
-    return simd16x32<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
-    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
-    return simd16x32<bool>(
-               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
-               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
-        .to_bitmask();
-  }
-}; // struct simd16x32<T>
-/* end file src/simdutf/haswell/simd16-inl.h */
-
-} // namespace simd
-
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-
-#endif // SIMDUTF_HASWELL_SIMD_H
-/* end file src/simdutf/haswell/simd.h */
-
-/* begin file src/simdutf/haswell/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
-
-
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-SIMDUTF_POP_DISABLE_WARNINGS
-#endif // end of workaround
-/* end file src/simdutf/haswell/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_HASWELL
-#endif // SIMDUTF_HASWELL_COMMON_H
-/* end file src/simdutf/haswell.h */
-/* begin file src/simdutf/westmere.h */
-#ifndef SIMDUTF_WESTMERE_H
-#define SIMDUTF_WESTMERE_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "westmere.h must be included before fallback.h"
-#endif
-
-
-// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
-#ifndef SIMDUTF_IMPLEMENTATION_WESTMERE
-  //
-  // You do not want to set it to (SIMDUTF_IS_X86_64 &&
-  // !SIMDUTF_REQUIRES_HASWELL) because you want to rely on runtime dispatch!
-  //
-  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE || SIMDUTF_CAN_ALWAYS_RUN_HASWELL
-    #define SIMDUTF_IMPLEMENTATION_WESTMERE 0
-  #else
-    #define SIMDUTF_IMPLEMENTATION_WESTMERE (SIMDUTF_IS_X86_64)
-  #endif
-
-#endif
-
-#if (SIMDUTF_IMPLEMENTATION_WESTMERE && SIMDUTF_IS_X86_64 && __SSE4_2__)
-  #define SIMDUTF_CAN_ALWAYS_RUN_WESTMERE 1
-#else
-  #define SIMDUTF_CAN_ALWAYS_RUN_WESTMERE 0
-#endif
-
-#if SIMDUTF_IMPLEMENTATION_WESTMERE
-
-  #define SIMDUTF_TARGET_WESTMERE SIMDUTF_TARGET_REGION("sse4.2,popcnt")
-
-namespace simdutf {
-/**
- * Implementation for Westmere (Intel SSE4.2).
- */
-namespace westmere {} // namespace westmere
-} // namespace simdutf
-
-  //
-  // These two need to be included outside SIMDUTF_TARGET_REGION
-  //
-/* begin file src/simdutf/westmere/implementation.h */
-#ifndef SIMDUTF_WESTMERE_IMPLEMENTATION_H
-#define SIMDUTF_WESTMERE_IMPLEMENTATION_H
-
-
-// The constructor may be executed on any host, so we take care not to use
-// SIMDUTF_TARGET_REGION
-namespace simdutf {
-namespace westmere {
-
-namespace {
-using namespace simdutf;
-}
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("westmere", "Intel/AMD SSE4.2",
-                                internal::instruction_set::SSE42) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-
-} // namespace westmere
-} // namespace simdutf
-
-#endif // SIMDUTF_WESTMERE_IMPLEMENTATION_H
-/* end file src/simdutf/westmere/implementation.h */
-/* begin file src/simdutf/westmere/intrinsics.h */
-#ifndef SIMDUTF_WESTMERE_INTRINSICS_H
-#define SIMDUTF_WESTMERE_INTRINSICS_H
-
-#ifdef SIMDUTF_VISUAL_STUDIO
-  // under clang within visual studio, this will include <x86intrin.h>
-  #include <intrin.h> // visual studio or clang
-#else
-
-  #if SIMDUTF_GCC11ORMORE
-// We should not get warnings while including <x86intrin.h> yet we do
-// under some versions of GCC.
-// If the x86intrin.h header has uninitialized values that are problematic,
-// it is a GCC issue, we want to ignore these warnings.
-SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
-  #endif
-
-  #include <x86intrin.h> // elsewhere
-
-  #if SIMDUTF_GCC11ORMORE
-// cancels the suppression of the -Wuninitialized
-SIMDUTF_POP_DISABLE_WARNINGS
-  #endif
-
-#endif // SIMDUTF_VISUAL_STUDIO
-
-#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
-  /**
-   * You are not supposed, normally, to include these
-   * headers directly. Instead you should either include intrin.h
-   * or x86intrin.h. However, when compiling with clang
-   * under Windows (i.e., when _MSC_VER is set), these headers
-   * only get included *if* the corresponding features are detected
-   * from macros:
-   */
-  #include <smmintrin.h> // for _mm_alignr_epi8
-#endif
-
-#endif // SIMDUTF_WESTMERE_INTRINSICS_H
-/* end file src/simdutf/westmere/intrinsics.h */
-
-  //
-  // The rest need to be inside the region
-  //
-/* begin file src/simdutf/westmere/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "westmere"
-// #define SIMDUTF_IMPLEMENTATION westmere
-
-#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
-// nothing needed.
-#else
-SIMDUTF_TARGET_WESTMERE
-#endif
-/* end file src/simdutf/westmere/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/westmere/bitmanipulation.h */
-#ifndef SIMDUTF_WESTMERE_BITMANIPULATION_H
-#define SIMDUTF_WESTMERE_BITMANIPULATION_H
-
-namespace simdutf {
-namespace westmere {
-namespace {
-
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
-  // note: we do not support legacy 32-bit Windows
-  return __popcnt64(input_num); // Visual Studio wants two underscores
-}
-#else
-simdutf_really_inline long long int count_ones(uint64_t input_num) {
-  return _popcnt64(input_num);
-}
-#endif
-
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
-  #if SIMDUTF_REGULAR_VISUAL_STUDIO
-  unsigned long ret;
-  _BitScanForward64(&ret, input_num);
-  return (int)ret;
-  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
-  return __builtin_ctzll(input_num);
-  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
-}
-#endif
-
-} // unnamed namespace
-} // namespace westmere
-} // namespace simdutf
-
-#endif // SIMDUTF_WESTMERE_BITMANIPULATION_H
-/* end file src/simdutf/westmere/bitmanipulation.h */
-/* begin file src/simdutf/westmere/simd.h */
-#ifndef SIMDUTF_WESTMERE_SIMD_H
-#define SIMDUTF_WESTMERE_SIMD_H
-
-namespace simdutf {
-namespace westmere {
-namespace {
-namespace simd {
-
-template <typename Child> struct base {
-  __m128i value;
-
-  // Zero constructor
-  simdutf_really_inline base() : value{__m128i()} {}
-
-  // Conversion from SIMD register
-  simdutf_really_inline base(const __m128i _value) : value(_value) {}
-  // Conversion to SIMD register
-  simdutf_really_inline operator const __m128i &() const { return this->value; }
-  simdutf_really_inline operator __m128i &() { return this->value; }
-  template <endianness big_endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
-    __m128i first = _mm_cvtepu8_epi16(*this);
-    __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8));
-    if (big_endian) {
-      const __m128i swap =
-          _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-      first = _mm_shuffle_epi8(first, swap);
-      second = _mm_shuffle_epi8(second, swap);
-    }
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first);
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second);
-  }
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p), _mm_cvtepu8_epi32(*this));
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 4),
-                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 4)));
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8),
-                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 8)));
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 12),
-                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 12)));
-  }
-  // Bit operations
-  simdutf_really_inline Child operator|(const Child other) const {
-    return _mm_or_si128(*this, other);
-  }
-  simdutf_really_inline Child operator&(const Child other) const {
-    return _mm_and_si128(*this, other);
-  }
-  simdutf_really_inline Child operator^(const Child other) const {
-    return _mm_xor_si128(*this, other);
-  }
-  simdutf_really_inline Child bit_andnot(const Child other) const {
-    return _mm_andnot_si128(other, *this);
-  }
-  simdutf_really_inline Child &operator|=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator&=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator^=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-};
-
-// Forward-declared so they can be used by splat and friends.
-template <typename T> struct simd8;
-
-template <typename T, typename Mask = simd8<bool>>
-struct base8 : base<simd8<T>> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  simdutf_really_inline T first() const { return _mm_extract_epi8(*this, 0); }
-  simdutf_really_inline T last() const { return _mm_extract_epi8(*this, 15); }
-  simdutf_really_inline base8() : base<simd8<T>>() {}
-  simdutf_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
-
-  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return _mm_cmpeq_epi8(lhs, rhs);
-  }
-
-  static const int SIZE = sizeof(base<simd8<T>>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
-    return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base8<bool> {
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return _mm_set1_epi8(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8() : base8() {}
-  simdutf_really_inline simd8(const __m128i _value) : base8<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
-
-  simdutf_really_inline int to_bitmask() const {
-    return _mm_movemask_epi8(*this);
-  }
-  simdutf_really_inline bool any() const {
-    return !_mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline bool none() const {
-    return _mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline bool all() const {
-    return _mm_movemask_epi8(*this) == 0xFFFF;
-  }
-  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base8_numeric : base8<T> {
-  static simdutf_really_inline simd8<T> splat(T _value) {
-    return _mm_set1_epi8(_value);
-  }
-  static simdutf_really_inline simd8<T> zero() { return _mm_setzero_si128(); }
-  static simdutf_really_inline simd8<T> load(const T values[16]) {
-    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
-  }
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
-                                                  T v5, T v6, T v7, T v8, T v9,
-                                                  T v10, T v11, T v12, T v13,
-                                                  T v14, T v15) {
-    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
-                    v14, v15);
-  }
-
-  simdutf_really_inline base8_numeric() : base8<T>() {}
-  simdutf_really_inline base8_numeric(const __m128i _value)
-      : base8<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[16]) const {
-    return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
-    return _mm_add_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
-    return _mm_sub_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
-    *this = *this + other;
-    return *static_cast<simd8<T> *>(this);
-  }
-  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
-    *this = *this - other;
-    return *static_cast<simd8<T> *>(this);
-  }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return _mm_shuffle_epi8(lookup_table, *this);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> : base8_numeric<int8_t> {
-  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
-  simdutf_really_inline simd8(const __m128i _value)
-      : base8_numeric<int8_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
-                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
-                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
-      : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                            v12, v13, v14, v15)) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15);
-  }
-  simdutf_really_inline operator simd8<uint8_t>() const;
-  simdutf_really_inline bool is_ascii() const {
-    return _mm_movemask_epi8(*this) == 0;
-  }
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return _mm_max_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return _mm_min_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return _mm_cmpgt_epi8(*this, other);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return _mm_cmpgt_epi8(other, *this);
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
-  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
-  simdutf_really_inline simd8(const __m128i _value)
-      : base8_numeric<uint8_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
-      : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                            v12, v13, v14, v15)) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return _mm_adds_epu8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return _mm_subs_epu8(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return _mm_max_epu8(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return _mm_min_epu8(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> bits_not_set() const {
-    return *this == uint8_t(0);
-  }
-  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    return _mm_movemask_epi8(*this) == 0;
-  }
-
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return _mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
-    return _mm_testz_si128(*this, bits);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N);
-  }
-  // Get one of the bits and make a bitmask out of it.
-  // e.g. value.get_bit<7>() gets the high bit
-  template <int N> simdutf_really_inline int get_bit() const {
-    return _mm_movemask_epi8(_mm_slli_epi16(*this, 7 - N));
-  }
-};
-simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
-  return this->value;
-}
-
-// Unsigned bytes
-template <> struct simd8<uint16_t> : base<uint16_t> {
-  static simdutf_really_inline simd8<uint16_t> splat(uint16_t _value) {
-    return _mm_set1_epi16(_value);
-  }
-  static simdutf_really_inline simd8<uint16_t> load(const uint16_t values[8]) {
-    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
-  }
-
-  simdutf_really_inline simd8() : base<uint16_t>() {}
-  simdutf_really_inline simd8(const __m128i _value) : base<uint16_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint16_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint16_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline simd8(uint16_t v0, uint16_t v1, uint16_t v2,
-                              uint16_t v3, uint16_t v4, uint16_t v5,
-                              uint16_t v6, uint16_t v7)
-      : simd8(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
-
-  // Saturated math
-  simdutf_really_inline simd8<uint16_t>
-  saturating_add(const simd8<uint16_t> other) const {
-    return _mm_adds_epu16(*this, other);
-  }
-  simdutf_really_inline simd8<uint16_t>
-  saturating_sub(const simd8<uint16_t> other) const {
-    return _mm_subs_epu16(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint16_t>
-  max_val(const simd8<uint16_t> other) const {
-    return _mm_max_epu16(*this, other);
-  }
-  simdutf_really_inline simd8<uint16_t>
-  min_val(const simd8<uint16_t> other) const {
-    return _mm_min_epu16(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint16_t>
-  gt_bits(const simd8<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint16_t>
-  lt_bits(const simd8<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint16_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint16_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator==(const simd8<uint16_t> other) const {
-    return _mm_cmpeq_epi16(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator&(const simd8<uint16_t> other) const {
-    return _mm_and_si128(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator|(const simd8<uint16_t> other) const {
-    return _mm_or_si128(*this, other);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  simdutf_really_inline simd8<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return _mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint16_t> bits) const {
-    return _mm_testz_si128(*this, bits);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint16_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-};
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(NUM_CHUNKS == 4,
-                "Westmere kernel should use four registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
-                                 const simd8<T> chunk2, const simd8<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    this->chunks[2] |= other.chunks[2];
-    this->chunks[3] |= other.chunks[3];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  template <endianness endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 0);
-    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 1);
-    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 2);
-    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
-    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
-    this->chunks[2].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 2);
-    this->chunks[3].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r1 = this->chunks[1].to_bitmask();
-    uint64_t r2 = this->chunks[2].to_bitmask();
-    uint64_t r3 = this->chunks[3].to_bitmask();
-    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                          this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
-    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
-                          this->chunks[1] == other.chunks[1],
-                          this->chunks[2] == other.chunks[2],
-                          this->chunks[3] == other.chunks[3])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                          this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low - 1);
-    const simd8<T> mask_high = simd8<T>::splat(high + 1);
-    return simd8x64<bool>(
-               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
-               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
-               (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
-               (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                          this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
-                          this->chunks[2] > mask, this->chunks[3] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
-                          this->chunks[2] >= mask, this->chunks[3] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>(simd8<uint8_t>(__m128i(this->chunks[0])) >= mask,
-                          simd8<uint8_t>(__m128i(this->chunks[1])) >= mask,
-                          simd8<uint8_t>(__m128i(this->chunks[2])) >= mask,
-                          simd8<uint8_t>(__m128i(this->chunks[3])) >= mask)
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-
-/* begin file src/simdutf/westmere/simd16-inl.h */
-template <typename T> struct simd16;
-
-template <typename T, typename Mask = simd16<bool>>
-struct base16 : base<simd16<T>> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  simdutf_really_inline base16() : base<simd16<T>>() {}
-  simdutf_really_inline base16(const __m128i _value)
-      : base<simd16<T>>(_value) {}
-  template <typename Pointer>
-  simdutf_really_inline base16(const Pointer *ptr)
-      : base16(_mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr))) {}
-
-  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
-                                               const simd16<T> rhs) {
-    return _mm_cmpeq_epi16(lhs, rhs);
-  }
-
-  static const int SIZE = sizeof(base<simd16<T>>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd16<bool> : base16<bool> {
-  static simdutf_really_inline simd16<bool> splat(bool _value) {
-    return _mm_set1_epi16(uint16_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd16() : base16() {}
-  simdutf_really_inline simd16(const __m128i _value) : base16<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
-
-  simdutf_really_inline int to_bitmask() const {
-    return _mm_movemask_epi8(*this);
-  }
-  simdutf_really_inline bool any() const {
-    return !_mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base16_numeric : base16<T> {
-  static simdutf_really_inline simd16<T> splat(T _value) {
-    return _mm_set1_epi16(_value);
-  }
-  static simdutf_really_inline simd16<T> zero() { return _mm_setzero_si128(); }
-  static simdutf_really_inline simd16<T> load(const T values[8]) {
-    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
-  }
-
-  simdutf_really_inline base16_numeric() : base16<T>() {}
-  simdutf_really_inline base16_numeric(const __m128i _value)
-      : base16<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[8]) const {
-    return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
-    return _mm_add_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
-    return _mm_sub_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
-    *this = *this + other;
-    return *static_cast<simd16<T> *>(this);
-  }
-  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
-    *this = *this - other;
-    return *static_cast<simd16<T> *>(this);
-  }
-};
-
-// Signed code units
-template <> struct simd16<int16_t> : base16_numeric<int16_t> {
-  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
-  simdutf_really_inline simd16(const __m128i _value)
-      : base16_numeric<int16_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
-  // Member-by-member initialization
-  simdutf_really_inline simd16(int16_t v0, int16_t v1, int16_t v2, int16_t v3,
-                               int16_t v4, int16_t v5, int16_t v6, int16_t v7)
-      : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
-  simdutf_really_inline operator simd16<uint16_t>() const;
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd16<int16_t>
-  max_val(const simd16<int16_t> other) const {
-    return _mm_max_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<int16_t>
-  min_val(const simd16<int16_t> other) const {
-    return _mm_min_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<int16_t> other) const {
-    return _mm_cmpgt_epi16(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<int16_t> other) const {
-    return _mm_cmpgt_epi16(other, *this);
-  }
-};
-
-// Unsigned code units
-template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
-  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
-  simdutf_really_inline simd16(const __m128i _value)
-      : base16_numeric<uint16_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
-  // Member-by-member initialization
-  simdutf_really_inline simd16(uint16_t v0, uint16_t v1, uint16_t v2,
-                               uint16_t v3, uint16_t v4, uint16_t v5,
-                               uint16_t v6, uint16_t v7)
-      : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd16<uint16_t>
-  repeat_16(uint16_t v0, uint16_t v1, uint16_t v2, uint16_t v3, uint16_t v4,
-            uint16_t v5, uint16_t v6, uint16_t v7) {
-    return simd16<uint16_t>(v0, v1, v2, v3, v4, v5, v6, v7);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd16<uint16_t>
-  saturating_add(const simd16<uint16_t> other) const {
-    return _mm_adds_epu16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  saturating_sub(const simd16<uint16_t> other) const {
-    return _mm_subs_epu16(*this, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd16<uint16_t>
-  max_val(const simd16<uint16_t> other) const {
-    return _mm_max_epu16(*this, other);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  min_val(const simd16<uint16_t> other) const {
-    return _mm_min_epu16(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  gt_bits(const simd16<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  lt_bits(const simd16<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<=(const simd16<uint16_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd16<bool>
-  operator>=(const simd16<uint16_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<uint16_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<uint16_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd16<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return _mm_testz_si128(*this, *this);
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool
-  bits_not_set_anywhere(simd16<uint16_t> bits) const {
-    return _mm_testz_si128(*this, bits);
-  }
-  simdutf_really_inline bool
-  any_bits_set_anywhere(simd16<uint16_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
-    return simd16<uint16_t>(_mm_srli_epi16(*this, N));
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
-    return simd16<uint16_t>(_mm_slli_epi16(*this, N));
-  }
-  // Get one of the bits and make a bitmask out of it.
-  // e.g. value.get_bit<7>() gets the high bit
-  template <int N> simdutf_really_inline int get_bit() const {
-    return _mm_movemask_epi8(_mm_slli_epi16(*this, 7 - N));
-  }
-
-  // Change the endianness
-  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
-    const __m128i swap =
-        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-    return _mm_shuffle_epi8(*this, swap);
-  }
-
-  // Pack with the unsigned saturation of two uint16_t code units into single
-  // uint8_t vector
-  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
-                                                   const simd16<uint16_t> &v1) {
-    return _mm_packus_epi16(v0, v1);
-  }
-};
-simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
-  return this->value;
-}
-
-template <typename T> struct simd16x32 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
-  static_assert(NUM_CHUNKS == 4,
-                "Westmere kernel should use four registers per 64-byte block.");
-  simd16<T> chunks[NUM_CHUNKS];
-
-  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
-  simd16x32<T> &
-  operator=(const simd16<T> other) = delete; // no assignment allowed
-  simd16x32() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline
-  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
-            const simd16<T> chunk2, const simd16<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd16x32(const T *ptr)
-      : chunks{simd16<T>::load(ptr),
-               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd16<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
-    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
-    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
-    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r1 = this->chunks[1].to_bitmask();
-    uint64_t r2 = this->chunks[2].to_bitmask();
-    uint64_t r3 = this->chunks[3].to_bitmask();
-    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
-  }
-
-  simdutf_really_inline void swap_bytes() {
-    this->chunks[0] = this->chunks[0].swap_bytes();
-    this->chunks[1] = this->chunks[1].swap_bytes();
-    this->chunks[2] = this->chunks[2].swap_bytes();
-    this->chunks[3] = this->chunks[3].swap_bytes();
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                           this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
-    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
-                           this->chunks[1] == other.chunks[1],
-                           this->chunks[2] == other.chunks[2],
-                           this->chunks[3] == other.chunks[3])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                           this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-
-    return simd16x32<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
-    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
-    return simd16x32<bool>(
-               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
-               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
-               (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
-               (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                           this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-}; // struct simd16x32<T>
-/* end file src/simdutf/westmere/simd16-inl.h */
-
-} // namespace simd
-} // unnamed namespace
-} // namespace westmere
-} // namespace simdutf
-
-#endif // SIMDUTF_WESTMERE_SIMD_INPUT_H
-/* end file src/simdutf/westmere/simd.h */
-
-/* begin file src/simdutf/westmere/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
-
-/* end file src/simdutf/westmere/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_WESTMERE
-#endif // SIMDUTF_WESTMERE_COMMON_H
-/* end file src/simdutf/westmere.h */
-/* begin file src/simdutf/ppc64.h */
-#ifndef SIMDUTF_PPC64_H
-#define SIMDUTF_PPC64_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "ppc64.h must be included before fallback.h"
-#endif
-
-
-#ifndef SIMDUTF_IMPLEMENTATION_PPC64
-  #define SIMDUTF_IMPLEMENTATION_PPC64 (SIMDUTF_IS_PPC64)
-#endif
-#define SIMDUTF_CAN_ALWAYS_RUN_PPC64                                           \
-  SIMDUTF_IMPLEMENTATION_PPC64 &&SIMDUTF_IS_PPC64
-
-
-#if SIMDUTF_IMPLEMENTATION_PPC64
-
-namespace simdutf {
-/**
- * Implementation for ALTIVEC (PPC64).
- */
-namespace ppc64 {} // namespace ppc64
-} // namespace simdutf
-
-/* begin file src/simdutf/ppc64/implementation.h */
-#ifndef SIMDUTF_PPC64_IMPLEMENTATION_H
-#define SIMDUTF_PPC64_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace ppc64 {
-
-namespace {
-using namespace simdutf;
-} // namespace
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("ppc64", "PPC64 ALTIVEC",
-                                internal::instruction_set::ALTIVEC) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-
-} // namespace ppc64
-} // namespace simdutf
-
-#endif // SIMDUTF_PPC64_IMPLEMENTATION_H
-/* end file src/simdutf/ppc64/implementation.h */
-
-/* begin file src/simdutf/ppc64/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
-// #define SIMDUTF_IMPLEMENTATION ppc64
-/* end file src/simdutf/ppc64/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/ppc64/intrinsics.h */
-#ifndef SIMDUTF_PPC64_INTRINSICS_H
-#define SIMDUTF_PPC64_INTRINSICS_H
-
-
-// This should be the correct header whether
-// you use visual studio or other compilers.
-#include <altivec.h>
-
-// These are defined by altivec.h in GCC toolchain, it is safe to undef them.
-#ifdef bool
-  #undef bool
-#endif
-
-#ifdef vector
-  #undef vector
-#endif
-
-#endif //  SIMDUTF_PPC64_INTRINSICS_H
-/* end file src/simdutf/ppc64/intrinsics.h */
-/* begin file src/simdutf/ppc64/bitmanipulation.h */
-#ifndef SIMDUTF_PPC64_BITMANIPULATION_H
-#define SIMDUTF_PPC64_BITMANIPULATION_H
-
-namespace simdutf {
-namespace ppc64 {
-namespace {
-
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-simdutf_really_inline int count_ones(uint64_t input_num) {
-  // note: we do not support legacy 32-bit Windows
-  return __popcnt64(input_num); // Visual Studio wants two underscores
-}
-#else
-simdutf_really_inline int count_ones(uint64_t input_num) {
-  return __builtin_popcountll(input_num);
-}
-#endif
-
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdutf
-
-#endif // SIMDUTF_PPC64_BITMANIPULATION_H
-/* end file src/simdutf/ppc64/bitmanipulation.h */
-/* begin file src/simdutf/ppc64/simd.h */
-#ifndef SIMDUTF_PPC64_SIMD_H
-#define SIMDUTF_PPC64_SIMD_H
-
-#include <type_traits>
-
-namespace simdutf {
-namespace ppc64 {
-namespace {
-namespace simd {
-
-using __m128i = __vector unsigned char;
-
-template <typename Child> struct base {
-  __m128i value;
-
-  // Zero constructor
-  simdutf_really_inline base() : value{__m128i()} {}
-
-  // Conversion from SIMD register
-  simdutf_really_inline base(const __m128i _value) : value(_value) {}
-
-  // Conversion to SIMD register
-  simdutf_really_inline operator const __m128i &() const { return this->value; }
-  simdutf_really_inline operator __m128i &() { return this->value; }
-
-  // Bit operations
-  simdutf_really_inline Child operator|(const Child other) const {
-    return vec_or(this->value, (__m128i)other);
-  }
-  simdutf_really_inline Child operator&(const Child other) const {
-    return vec_and(this->value, (__m128i)other);
-  }
-  simdutf_really_inline Child operator^(const Child other) const {
-    return vec_xor(this->value, (__m128i)other);
-  }
-  simdutf_really_inline Child bit_andnot(const Child other) const {
-    return vec_andc(this->value, (__m128i)other);
-  }
-  simdutf_really_inline Child &operator|=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator&=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator^=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-};
-
-// Forward-declared so they can be used by splat and friends.
-template <typename T> struct simd8;
-
-template <typename T, typename Mask = simd8<bool>>
-struct base8 : base<simd8<T>> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  simdutf_really_inline base8() : base<simd8<T>>() {}
-  simdutf_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
-
-  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return (__m128i)vec_cmpeq(lhs.value, (__m128i)rhs);
-  }
-
-  static const int SIZE = sizeof(base<simd8<T>>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(simd8<T> prev_chunk) const {
-    __m128i chunk = this->value;
-#ifdef __LITTLE_ENDIAN__
-    chunk = (__m128i)vec_reve(this->value);
-    prev_chunk = (__m128i)vec_reve((__m128i)prev_chunk);
-#endif
-    chunk = (__m128i)vec_sld((__m128i)prev_chunk, (__m128i)chunk, 16 - N);
-#ifdef __LITTLE_ENDIAN__
-    chunk = (__m128i)vec_reve((__m128i)chunk);
-#endif
-    return chunk;
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base8<bool> {
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return (__m128i)vec_splats((unsigned char)(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8() : base8() {}
-  simdutf_really_inline simd8(const __m128i _value) : base8<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
-
-  simdutf_really_inline int to_bitmask() const {
-    __vector unsigned long long result;
-    const __m128i perm_mask = {0x78, 0x70, 0x68, 0x60, 0x58, 0x50, 0x48, 0x40,
-                               0x38, 0x30, 0x28, 0x20, 0x18, 0x10, 0x08, 0x00};
-
-    result = ((__vector unsigned long long)vec_vbpermq((__m128i)this->value,
-                                                       (__m128i)perm_mask));
-#ifdef __LITTLE_ENDIAN__
-    return static_cast<int>(result[1]);
-#else
-    return static_cast<int>(result[0]);
-#endif
-  }
-  simdutf_really_inline bool any() const {
-    return !vec_all_eq(this->value, (__m128i)vec_splats(0));
-  }
-  simdutf_really_inline simd8<bool> operator~() const {
-    return this->value ^ (__m128i)splat(true);
-  }
-};
-
-template <typename T> struct base8_numeric : base8<T> {
-  static simdutf_really_inline simd8<T> splat(T value) {
-    (void)value;
-    return (__m128i)vec_splats(value);
-  }
-  static simdutf_really_inline simd8<T> zero() { return splat(0); }
-  static simdutf_really_inline simd8<T> load(const T values[16]) {
-    return (__m128i)(vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(values)));
-  }
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
-                                                  T v5, T v6, T v7, T v8, T v9,
-                                                  T v10, T v11, T v12, T v13,
-                                                  T v14, T v15) {
-    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
-                    v14, v15);
-  }
-
-  simdutf_really_inline base8_numeric() : base8<T>() {}
-  simdutf_really_inline base8_numeric(const __m128i _value)
-      : base8<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[16]) const {
-    vec_vsx_st(this->value, 0, reinterpret_cast<__m128i *>(dst));
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
-    return (__m128i)((__m128i)this->value + (__m128i)other);
-  }
-  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
-    return (__m128i)((__m128i)this->value - (__m128i)other);
-  }
-  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
-    *this = *this + other;
-    return *static_cast<simd8<T> *>(this);
-  }
-  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
-    *this = *this - other;
-    return *static_cast<simd8<T> *>(this);
-  }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return (__m128i)vec_perm((__m128i)lookup_table, (__m128i)lookup_table,
-                             this->value);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> : base8_numeric<int8_t> {
-  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
-  simdutf_really_inline simd8(const __m128i _value)
-      : base8_numeric<int8_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
-                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
-                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
-      : simd8((__m128i)(__vector signed char){v0, v1, v2, v3, v4, v5, v6, v7,
-                                              v8, v9, v10, v11, v12, v13, v14,
-                                              v15}) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15);
-  }
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return (__m128i)vec_max((__vector signed char)this->value,
-                            (__vector signed char)(__m128i)other);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return (__m128i)vec_min((__vector signed char)this->value,
-                            (__vector signed char)(__m128i)other);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return (__m128i)vec_cmpgt((__vector signed char)this->value,
-                              (__vector signed char)(__m128i)other);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return (__m128i)vec_cmplt((__vector signed char)this->value,
-                              (__vector signed char)(__m128i)other);
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
-  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
-  simdutf_really_inline simd8(const __m128i _value)
-      : base8_numeric<uint8_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
-      : simd8((__m128i){v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                        v13, v14, v15}) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return (__m128i)vec_adds(this->value, (__m128i)other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return (__m128i)vec_subs(this->value, (__m128i)other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return (__m128i)vec_max(this->value, (__m128i)other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return (__m128i)vec_min(this->value, (__m128i)other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return other.max_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return other.min_val(*this) == other;
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return this->gt_bits(other).any_bits_set();
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> bits_not_set() const {
-    return (__m128i)vec_cmpeq(this->value, (__m128i)vec_splats(uint8_t(0)));
-  }
-  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->saturating_sub(0b01111111u).bits_not_set_anywhere();
-  }
-
-  simdutf_really_inline bool bits_not_set_anywhere() const {
-    return vec_all_eq(this->value, (__m128i)vec_splats(0));
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return !bits_not_set_anywhere();
-  }
-  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
-    return vec_all_eq(vec_and(this->value, (__m128i)bits),
-                      (__m128i)vec_splats(0));
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return !bits_not_set_anywhere(bits);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return simd8<uint8_t>(
-        (__m128i)vec_sr(this->value, (__m128i)vec_splat_u8(N)));
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return simd8<uint8_t>(
-        (__m128i)vec_sl(this->value, (__m128i)vec_splat_u8(N)));
-  }
-};
-
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(NUM_CHUNKS == 4,
-                "PPC64 kernel should use four registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
-                                 const simd8<T> chunk2, const simd8<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    this->chunks[2] |= other.chunks[2];
-    this->chunks[3] |= other.chunks[3];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return input.reduce_or().is_ascii();
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r1 = this->chunks[1].to_bitmask();
-    uint64_t r2 = this->chunks[2].to_bitmask();
-    uint64_t r3 = this->chunks[3].to_bitmask();
-    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                          this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
-    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
-                          this->chunks[1] == other.chunks[1],
-                          this->chunks[2] == other.chunks[2],
-                          this->chunks[3] == other.chunks[3])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                          this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-    return simd8x64<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
-               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
-               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                          this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
-                          this->chunks[2] > mask, this->chunks[3] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
-                          this->chunks[2] >= mask, this->chunks[3] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>(simd8<uint8_t>(this->chunks[0]) >= mask,
-                          simd8<uint8_t>(this->chunks[1]) >= mask,
-                          simd8<uint8_t>(this->chunks[2]) >= mask,
-                          simd8<uint8_t>(this->chunks[3]) >= mask)
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-
-} // namespace simd
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdutf
-
-#endif // SIMDUTF_PPC64_SIMD_INPUT_H
-/* end file src/simdutf/ppc64/simd.h */
-
-/* begin file src/simdutf/ppc64/end.h */
-/* end file src/simdutf/ppc64/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_PPC64
-
-#endif // SIMDUTF_PPC64_H
-/* end file src/simdutf/ppc64.h */
-/* begin file src/simdutf/rvv.h */
-#ifndef SIMDUTF_RVV_H
-#define SIMDUTF_RVV_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "rvv.h must be included before fallback.h"
-#endif
-
-
-#define SIMDUTF_CAN_ALWAYS_RUN_RVV SIMDUTF_IS_RVV
-
-#ifndef SIMDUTF_IMPLEMENTATION_RVV
-  #define SIMDUTF_IMPLEMENTATION_RVV                                           \
-    (SIMDUTF_CAN_ALWAYS_RUN_RVV ||                                             \
-     (SIMDUTF_IS_RISCV64 && SIMDUTF_HAS_RVV_INTRINSICS &&                      \
-      SIMDUTF_HAS_RVV_TARGET_REGION))
-#endif
-
-#if SIMDUTF_IMPLEMENTATION_RVV
-
-  #if SIMDUTF_CAN_ALWAYS_RUN_RVV
-    #define SIMDUTF_TARGET_RVV
-  #else
-    #define SIMDUTF_TARGET_RVV SIMDUTF_TARGET_REGION("arch=+v")
-  #endif
-  #if !SIMDUTF_IS_ZVBB && SIMDUTF_HAS_ZVBB_INTRINSICS
-    #define SIMDUTF_TARGET_ZVBB SIMDUTF_TARGET_REGION("arch=+v,+zvbb")
-  #endif
-
-namespace simdutf {
-namespace rvv {} // namespace rvv
-} // namespace simdutf
-
-/* begin file src/simdutf/rvv/implementation.h */
-#ifndef SIMDUTF_RVV_IMPLEMENTATION_H
-#define SIMDUTF_RVV_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace rvv {
-
-namespace {
-using namespace simdutf;
-} // namespace
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("rvv", "RISC-V Vector Extension",
-                                internal::instruction_set::RVV),
-        _supports_zvbb(internal::detect_supported_architectures() &
-                       internal::instruction_set::ZVBB) {}
-  simdutf_warn_unused int detect_encodings(const char *buf,
-                                           size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t len,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t len) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t len) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *buf, size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *buf, size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf16le(const char16_t *buf, size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf16be(const char16_t *buf, size_t len) const noexcept;
-  simdutf_warn_unused size_t utf16_length_from_utf8(const char *buf,
-                                                    size_t len) const noexcept;
-  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t *buf,
-                                                    size_t len) const noexcept;
-  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t *buf,
-                                                     size_t len) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf8(const char *buf,
-                                                    size_t len) const noexcept;
-  simdutf_warn_unused size_t latin1_length_from_utf8(const char *buf,
-                                                     size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t len) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t len) const noexcept;
-  simdutf_warn_unused size_t utf8_length_from_latin1(const char *buf,
-                                                     size_t len) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-
-private:
-  const bool _supports_zvbb;
-
-#if SIMDUTF_IS_ZVBB
-  bool supports_zvbb() const { return true; }
-#elif SIMDUTF_HAS_ZVBB_INTRINSICS
-  bool supports_zvbb() const { return _supports_zvbb; }
-#else
-  bool supports_zvbb() const { return false; }
-#endif
-};
-
-} // namespace rvv
-} // namespace simdutf
-
-#endif // SIMDUTF_RVV_IMPLEMENTATION_H
-/* end file src/simdutf/rvv/implementation.h */
-/* begin file src/simdutf/rvv/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "rvv"
-// #define SIMDUTF_IMPLEMENTATION rvv
-
-#if SIMDUTF_CAN_ALWAYS_RUN_RVV
-// nothing needed.
-#else
-SIMDUTF_TARGET_RVV
-#endif
-/* end file src/simdutf/rvv/begin.h */
-/* begin file src/simdutf/rvv/intrinsics.h */
-#ifndef SIMDUTF_RVV_INTRINSICS_H
-#define SIMDUTF_RVV_INTRINSICS_H
-
-
-#include <riscv_vector.h>
-
-#if __riscv_v_intrinsic >= 1000000 || __GCC__ >= 14
-  #define simdutf_vrgather_u8m1x2(tbl, idx)                                    \
-    __riscv_vcreate_v_u8m1_u8m2(                                               \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 0),        \
-                                 __riscv_vsetvlmax_e8m1()),                    \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),        \
-                                 __riscv_vsetvlmax_e8m1()));
-
-  #define simdutf_vrgather_u8m1x4(tbl, idx)                                    \
-    __riscv_vcreate_v_u8m1_u8m4(                                               \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),        \
-                                 __riscv_vsetvlmax_e8m1()),                    \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 1),        \
-                                 __riscv_vsetvlmax_e8m1()),                    \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),        \
-                                 __riscv_vsetvlmax_e8m1()),                    \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),        \
-                                 __riscv_vsetvlmax_e8m1()));
-#else
-  // This has worse codegen on gcc
-  #define simdutf_vrgather_u8m1x2(tbl, idx)                                    \
-    __riscv_vset_v_u8m1_u8m2(                                                  \
-        __riscv_vlmul_ext_v_u8m1_u8m2(__riscv_vrgather_vv_u8m1(                \
-            tbl, __riscv_vget_v_u8m2_u8m1(idx, 0), __riscv_vsetvlmax_e8m1())), \
-        1,                                                                     \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),        \
-                                 __riscv_vsetvlmax_e8m1()))
-
-  #define simdutf_vrgather_u8m1x4(tbl, idx)                                    \
-    __riscv_vset_v_u8m1_u8m4(                                                  \
-        __riscv_vset_v_u8m1_u8m4(                                              \
-            __riscv_vset_v_u8m1_u8m4(                                          \
-                __riscv_vlmul_ext_v_u8m1_u8m4(__riscv_vrgather_vv_u8m1(        \
-                    tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),                     \
-                    __riscv_vsetvlmax_e8m1())),                                \
-                1,                                                             \
-                __riscv_vrgather_vv_u8m1(tbl,                                  \
-                                         __riscv_vget_v_u8m4_u8m1(idx, 1),     \
-                                         __riscv_vsetvlmax_e8m1())),           \
-            2,                                                                 \
-            __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),    \
-                                     __riscv_vsetvlmax_e8m1())),               \
-        3,                                                                     \
-        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),        \
-                                 __riscv_vsetvlmax_e8m1()))
-#endif
-
-/* Zvbb adds dedicated support for endianness swaps with vrev8, but if we can't
- * use that, we have to emulate it with the standard V extension.
- * Using LMUL=1 vrgathers could be faster than the srl+macc variant, but that
- * would increase register pressure, and vrgather implementations performance
- * varies a lot. */
-enum class simdutf_ByteFlip { NONE, V, ZVBB };
-
-template <simdutf_ByteFlip method>
-simdutf_really_inline static uint16_t simdutf_byteflip(uint16_t v) {
-  if (method != simdutf_ByteFlip::NONE)
-    return (uint16_t)((v * 1u) << 8 | (v * 1u) >> 8);
-  return v;
-}
-
-#ifdef SIMDUTF_TARGET_ZVBB
-SIMDUTF_UNTARGET_REGION
-SIMDUTF_TARGET_ZVBB
-#endif
-
-template <simdutf_ByteFlip method>
-simdutf_really_inline static vuint16m1_t simdutf_byteflip(vuint16m1_t v,
-                                                          size_t vl) {
-#if SIMDUTF_HAS_ZVBB_INTRINSICS
-  if (method == simdutf_ByteFlip::ZVBB)
-    return __riscv_vrev8_v_u16m1(v, vl);
-#endif
-  if (method == simdutf_ByteFlip::V)
-    return __riscv_vmacc_vx_u16m1(__riscv_vsrl_vx_u16m1(v, 8, vl), 0x100, v,
-                                  vl);
-  return v;
-}
-
-template <simdutf_ByteFlip method>
-simdutf_really_inline static vuint16m2_t simdutf_byteflip(vuint16m2_t v,
-                                                          size_t vl) {
-#if SIMDUTF_HAS_ZVBB_INTRINSICS
-  if (method == simdutf_ByteFlip::ZVBB)
-    return __riscv_vrev8_v_u16m2(v, vl);
-#endif
-  if (method == simdutf_ByteFlip::V)
-    return __riscv_vmacc_vx_u16m2(__riscv_vsrl_vx_u16m2(v, 8, vl), 0x100, v,
-                                  vl);
-  return v;
-}
-
-template <simdutf_ByteFlip method>
-simdutf_really_inline static vuint16m4_t simdutf_byteflip(vuint16m4_t v,
-                                                          size_t vl) {
-#if SIMDUTF_HAS_ZVBB_INTRINSICS
-  if (method == simdutf_ByteFlip::ZVBB)
-    return __riscv_vrev8_v_u16m4(v, vl);
-#endif
-  if (method == simdutf_ByteFlip::V)
-    return __riscv_vmacc_vx_u16m4(__riscv_vsrl_vx_u16m4(v, 8, vl), 0x100, v,
-                                  vl);
-  return v;
-}
-
-template <simdutf_ByteFlip method>
-simdutf_really_inline static vuint16m8_t simdutf_byteflip(vuint16m8_t v,
-                                                          size_t vl) {
-#if SIMDUTF_HAS_ZVBB_INTRINSICS
-  if (method == simdutf_ByteFlip::ZVBB)
-    return __riscv_vrev8_v_u16m8(v, vl);
-#endif
-  if (method == simdutf_ByteFlip::V)
-    return __riscv_vmacc_vx_u16m8(__riscv_vsrl_vx_u16m8(v, 8, vl), 0x100, v,
-                                  vl);
-  return v;
-}
-
-#ifdef SIMDUTF_TARGET_ZVBB
-SIMDUTF_UNTARGET_REGION
-SIMDUTF_TARGET_RVV
-#endif
-
-#endif //  SIMDUTF_RVV_INTRINSICS_H
-/* end file src/simdutf/rvv/intrinsics.h */
-/* begin file src/simdutf/rvv/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_RVV
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
-
-/* end file src/simdutf/rvv/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_RVV
-
-#endif // SIMDUTF_RVV_H
-/* end file src/simdutf/rvv.h */
-/* begin file src/simdutf/lsx.h */
-#ifndef SIMDUTF_LSX_H
-#define SIMDUTF_LSX_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "lsx.h must be included before fallback.h"
-#endif
-
-
-#ifndef SIMDUTF_IMPLEMENTATION_LSX
-  #define SIMDUTF_IMPLEMENTATION_LSX (SIMDUTF_IS_LSX)
-#endif
-#if SIMDUTF_IMPLEMENTATION_LSX && SIMDUTF_IS_LSX
-  #define SIMDUTF_CAN_ALWAYS_RUN_LSX 1
-#else
-  #define SIMDUTF_CAN_ALWAYS_RUN_LSX 0
-#endif
-
-#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
-
-#if SIMDUTF_IMPLEMENTATION_LSX
-
-namespace simdutf {
-/**
- * Implementation for LoongArch SX.
- */
-namespace lsx {} // namespace lsx
-} // namespace simdutf
-
-/* begin file src/simdutf/lsx/implementation.h */
-#ifndef SIMDUTF_LSX_IMPLEMENTATION_H
-#define SIMDUTF_LSX_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace lsx {
-
-namespace {
-using namespace simdutf;
-}
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("lsx", "LOONGARCH SX",
-                                internal::instruction_set::LSX) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char *input, size_t length, char *output,
-                   base64_options options) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-};
-
-} // namespace lsx
-} // namespace simdutf
-
-#endif // SIMDUTF_LSX_IMPLEMENTATION_H
-/* end file src/simdutf/lsx/implementation.h */
-
-/* begin file src/simdutf/lsx/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "lsx"
-// #define SIMDUTF_IMPLEMENTATION lsx
-/* end file src/simdutf/lsx/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/lsx/intrinsics.h */
-#ifndef SIMDUTF_LSX_INTRINSICS_H
-#define SIMDUTF_LSX_INTRINSICS_H
-
-
-// This should be the correct header whether
-// you use visual studio or other compilers.
-#include <lsxintrin.h>
-
-#endif //  SIMDUTF_LSX_INTRINSICS_H
-/* end file src/simdutf/lsx/intrinsics.h */
-/* begin file src/simdutf/lsx/bitmanipulation.h */
-#ifndef SIMDUTF_LSX_BITMANIPULATION_H
-#define SIMDUTF_LSX_BITMANIPULATION_H
-
-#include <limits>
-
-namespace simdutf {
-namespace lsx {
-namespace {
-
-simdutf_really_inline int count_ones(uint64_t input_num) {
-  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__lsx_vreplgr2vr_d(input_num)), 0);
-}
-
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
-  return __builtin_ctzll(input_num);
-}
-#endif
-
-} // unnamed namespace
-} // namespace lsx
-} // namespace simdutf
-
-#endif // SIMDUTF_LSX_BITMANIPULATION_H
-/* end file src/simdutf/lsx/bitmanipulation.h */
-/* begin file src/simdutf/lsx/simd.h */
-#ifndef SIMDUTF_LSX_SIMD_H
-#define SIMDUTF_LSX_SIMD_H
-
-#include <type_traits>
-
-namespace simdutf {
-namespace lsx {
-namespace {
-namespace simd {
-
-template <typename T> struct simd8;
-
-//
-// Base class of simd8<uint8_t> and simd8<bool>, both of which use __m128i
-// internally.
-//
-template <typename T, typename Mask = simd8<bool>> struct base_u8 {
-  __m128i value;
-  static const int SIZE = sizeof(value);
-
-  // Conversion from/to SIMD register
-  simdutf_really_inline base_u8(const __m128i _value) : value(_value) {}
-  simdutf_really_inline operator const __m128i &() const { return this->value; }
-  simdutf_really_inline operator __m128i &() { return this->value; }
-  simdutf_really_inline T first() const {
-    return __lsx_vpickve2gr_bu(this->value, 0);
-  }
-  simdutf_really_inline T last() const {
-    return __lsx_vpickve2gr_bu(this->value, 15);
-  }
-
-  // Bit operations
-  simdutf_really_inline simd8<T> operator|(const simd8<T> other) const {
-    return __lsx_vor_v(this->value, other);
-  }
-  simdutf_really_inline simd8<T> operator&(const simd8<T> other) const {
-    return __lsx_vand_v(this->value, other);
-  }
-  simdutf_really_inline simd8<T> operator^(const simd8<T> other) const {
-    return __lsx_vxor_v(this->value, other);
-  }
-  simdutf_really_inline simd8<T> bit_andnot(const simd8<T> other) const {
-    return __lsx_vandn_v(this->value, other);
-  }
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-  simdutf_really_inline simd8<T> &operator|=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd8<T> &operator&=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd8<T> &operator^=(const simd8<T> other) {
-    auto this_cast = static_cast<simd8<T> *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-
-  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return __lsx_vseq_b(lhs, rhs);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
-    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
-                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base_u8<bool> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return __lsx_vreplgr2vr_b(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8(const __m128i _value) : base_u8<bool>(_value) {}
-  // False constructor
-  simdutf_really_inline simd8() : simd8(__lsx_vldi(0)) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
-  simdutf_really_inline void store(uint8_t dst[16]) const {
-    return __lsx_vst(this->value, dst, 0);
-  }
-
-  simdutf_really_inline uint32_t to_bitmask() const {
-    return __lsx_vpickve2gr_wu(__lsx_vmsknz_b(*this), 0);
-  }
-
-  simdutf_really_inline bool any() const {
-    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) != 0;
-  }
-  simdutf_really_inline bool none() const {
-    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) == 0;
-  }
-  simdutf_really_inline bool all() const {
-    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) == 0xFFFF;
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base_u8<uint8_t> {
-  static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) {
-    return __lsx_vreplgr2vr_b(_value);
-  }
-  static simdutf_really_inline simd8<uint8_t> zero() { return __lsx_vldi(0); }
-  static simdutf_really_inline simd8<uint8_t> load(const uint8_t *values) {
-    return __lsx_vld(values, 0);
-  }
-  simdutf_really_inline simd8(const __m128i _value)
-      : base_u8<uint8_t>(_value) {}
-  // Zero constructor
-  simdutf_really_inline simd8() : simd8(zero()) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Member-by-member initialization
-
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
-      : simd8((__m128i)v16u8{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                             v12, v13, v14, v15}) {}
-
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15);
-  }
-
-  // Store to array
-  simdutf_really_inline void store(uint8_t dst[16]) const {
-    return __lsx_vst(this->value, dst, 0);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return __lsx_vsadd_bu(this->value, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return __lsx_vssub_bu(this->value, other);
-  }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<uint8_t>
-  operator+(const simd8<uint8_t> other) const {
-    return __lsx_vadd_b(this->value, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  operator-(const simd8<uint8_t> other) const {
-    return __lsx_vsub_b(this->value, other);
-  }
-  simdutf_really_inline simd8<uint8_t> &operator+=(const simd8<uint8_t> other) {
-    *this = *this + other;
-    return *this;
-  }
-  simdutf_really_inline simd8<uint8_t> &operator-=(const simd8<uint8_t> other) {
-    *this = *this - other;
-    return *this;
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return __lsx_vmax_bu(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return __lsx_vmin_bu(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return __lsx_vsle_bu(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return __lsx_vsle_bu(other, *this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return __lsx_vslt_bu(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return __lsx_vslt_bu(other, *this);
-  }
-  // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true
-  // = nonzero. For ARM, returns all 1's.
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return simd8<uint8_t>(*this > other);
-  }
-  // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true
-  // = nonzero. For ARM, returns all 1's.
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return simd8<uint8_t>(*this < other);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return __lsx_vslt_bu(__lsx_vldi(0), __lsx_vand_v(this->value, bits));
-  }
-  simdutf_really_inline bool is_ascii() const {
-    return __lsx_vpickve2gr_hu(__lsx_vmskgez_b(this->value), 0) == 0xFFFF;
-  }
-
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(this->value), 0) > 0;
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return (*this & bits).any_bits_set_anywhere();
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return __lsx_vsrli_b(this->value, N);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return __lsx_vslli_b(this->value, N);
-  }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return lookup_table.apply_lookup_16_to(*this);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-
-  template <typename T>
-  simdutf_really_inline simd8<uint8_t>
-  apply_lookup_16_to(const simd8<T> original) const {
-    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
-    return __lsx_vshuf_b(__lsx_vldi(0), *this, simd8<uint8_t>(original_tmp));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> {
-  __m128i value;
-
-  static simdutf_really_inline simd8<int8_t> splat(int8_t _value) {
-    return __lsx_vreplgr2vr_b(_value);
-  }
-  static simdutf_really_inline simd8<int8_t> zero() { return __lsx_vldi(0); }
-  static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) {
-    return __lsx_vld(values, 0);
-  }
-
-  template <endianness big_endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
-    __m128i zero = __lsx_vldi(0);
-    if (match_system(big_endian)) {
-      __lsx_vst(__lsx_vilvl_b(zero, (__m128i)this->value),
-                reinterpret_cast<uint16_t *>(p), 0);
-      __lsx_vst(__lsx_vilvh_b(zero, (__m128i)this->value),
-                reinterpret_cast<uint16_t *>(p + 8), 0);
-    } else {
-      __lsx_vst(__lsx_vilvl_b((__m128i)this->value, zero),
-                reinterpret_cast<uint16_t *>(p), 0);
-      __lsx_vst(__lsx_vilvh_b((__m128i)this->value, zero),
-                reinterpret_cast<uint16_t *>(p + 8), 0);
-    }
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
-    __m128i zero = __lsx_vldi(0);
-    __m128i in16low = __lsx_vilvl_b(zero, (__m128i)this->value);
-    __m128i in16high = __lsx_vilvh_b(zero, (__m128i)this->value);
-    __m128i in32_0 = __lsx_vilvl_h(zero, in16low);
-    __m128i in32_1 = __lsx_vilvh_h(zero, in16low);
-    __m128i in32_2 = __lsx_vilvl_h(zero, in16high);
-    __m128i in32_3 = __lsx_vilvh_h(zero, in16high);
-    __lsx_vst(in32_0, reinterpret_cast<uint32_t *>(p), 0);
-    __lsx_vst(in32_1, reinterpret_cast<uint32_t *>(p + 4), 0);
-    __lsx_vst(in32_2, reinterpret_cast<uint32_t *>(p + 8), 0);
-    __lsx_vst(in32_3, reinterpret_cast<uint32_t *>(p + 12), 0);
-  }
-
-  // In places where the table can be reused, which is most uses in simdutf, it
-  // is worth it to do 4 table lookups, as there is no direct zero extension
-  // from u8 to u32.
-  simdutf_really_inline void store_ascii_as_utf32_tbl(char32_t *p) const {
-    const simd8<uint8_t> tb1{0, 255, 255, 255, 1, 255, 255, 255,
-                             2, 255, 255, 255, 3, 255, 255, 255};
-    const simd8<uint8_t> tb2{4, 255, 255, 255, 5, 255, 255, 255,
-                             6, 255, 255, 255, 7, 255, 255, 255};
-    const simd8<uint8_t> tb3{8,  255, 255, 255, 9,  255, 255, 255,
-                             10, 255, 255, 255, 11, 255, 255, 255};
-    const simd8<uint8_t> tb4{12, 255, 255, 255, 13, 255, 255, 255,
-                             14, 255, 255, 255, 15, 255, 255, 255};
-
-    // encourage store pairing and interleaving
-    const auto shuf1 = this->apply_lookup_16_to(tb1);
-    const auto shuf2 = this->apply_lookup_16_to(tb2);
-    shuf1.store(reinterpret_cast<int8_t *>(p));
-    shuf2.store(reinterpret_cast<int8_t *>(p + 4));
-
-    const auto shuf3 = this->apply_lookup_16_to(tb3);
-    const auto shuf4 = this->apply_lookup_16_to(tb4);
-    shuf3.store(reinterpret_cast<int8_t *>(p + 8));
-    shuf4.store(reinterpret_cast<int8_t *>(p + 12));
-  }
-  // Conversion from/to SIMD register
-  simdutf_really_inline simd8(const __m128i _value) : value(_value) {}
-  simdutf_really_inline operator const __m128i &() const { return this->value; }
-
-  simdutf_really_inline operator const __m128i() const { return this->value; }
-
-  simdutf_really_inline operator __m128i &() { return this->value; }
-
-  // Zero constructor
-  simdutf_really_inline simd8() : simd8(zero()) {}
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
-  // Member-by-member initialization
-
-  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
-                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
-                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
-      : simd8((__m128i)v16i8{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                             v12, v13, v14, v15}) {}
-
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15);
-  }
-
-  // Store to array
-  simdutf_really_inline void store(int8_t dst[16]) const {
-    return __lsx_vst(value, dst, 0);
-  }
-
-  simdutf_really_inline operator simd8<uint8_t>() const {
-    return ((__m128i)this->value);
-  }
-
-  simdutf_really_inline simd8<int8_t>
-  operator|(const simd8<int8_t> other) const {
-    return __lsx_vor_v((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator&(const simd8<int8_t> other) const {
-    return __lsx_vand_v((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator^(const simd8<int8_t> other) const {
-    return __lsx_vxor_v((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  bit_andnot(const simd8<int8_t> other) const {
-    return __lsx_vandn_v((__m128i)other.value, (__m128i)value);
-  }
-
-  // Math
-  simdutf_really_inline simd8<int8_t>
-  operator+(const simd8<int8_t> other) const {
-    return __lsx_vadd_b((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t>
-  operator-(const simd8<int8_t> other) const {
-    return __lsx_vsub_b((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t> &operator+=(const simd8<int8_t> other) {
-    *this = *this + other;
-    return *this;
-  }
-  simdutf_really_inline simd8<int8_t> &operator-=(const simd8<int8_t> other) {
-    *this = *this - other;
-    return *this;
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return (__lsx_vpickve2gr_hu(__lsx_vmskgez_b((__m128i)this->value), 0) ==
-            0xffff);
-  }
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return __lsx_vmax_b((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return __lsx_vmin_b((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return __lsx_vslt_b((__m128i)other.value, (__m128i)value);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return __lsx_vslt_b((__m128i)value, (__m128i)other.value);
-  }
-  simdutf_really_inline simd8<bool>
-  operator==(const simd8<int8_t> other) const {
-    return __lsx_vseq_b((__m128i)value, (__m128i)other.value);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd8<int8_t>
-  prev(const simd8<int8_t> prev_chunk) const {
-    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
-                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
-  }
-
-  // Perform a lookup assuming no value is larger than 16
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    return lookup_table.apply_lookup_16_to(*this);
-  }
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-
-  template <typename T>
-  simdutf_really_inline simd8<int8_t>
-  apply_lookup_16_to(const simd8<T> original) const {
-    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
-    return __lsx_vshuf_b(__lsx_vldi(0), (__m128i)this->value,
-                         simd8<uint8_t>(original_tmp));
-  }
-};
-
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(
-      NUM_CHUNKS == 4,
-      "LoongArch kernel should use four registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
-                                 const simd8<T> chunk2, const simd8<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
-               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    this->chunks[2] |= other.chunks[2];
-    this->chunks[3] |= other.chunks[3];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
-
-  template <endianness endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 0);
-    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 1);
-    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 2);
-    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 0);
-    this->chunks[1].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 1);
-    this->chunks[2].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 2);
-    this->chunks[3].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    __m128i mask = __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[3]), 6);
-    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[2]), 4));
-    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[1]), 2));
-    mask = __lsx_vor_v(mask, __lsx_vmsknz_b(this->chunks[0]));
-    return __lsx_vpickve2gr_du(mask, 0);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                          this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                          this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-    return simd8x64<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
-               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
-               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                          this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
-                          this->chunks[2] > mask, this->chunks[3] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
-                          this->chunks[2] >= mask, this->chunks[3] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>(simd8<uint8_t>(this->chunks[0].value) >= mask,
-                          simd8<uint8_t>(this->chunks[1].value) >= mask,
-                          simd8<uint8_t>(this->chunks[2].value) >= mask,
-                          simd8<uint8_t>(this->chunks[3].value) >= mask)
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-/* begin file src/simdutf/lsx/simd16-inl.h */
-template <typename T> struct simd16;
-
-template <typename T, typename Mask = simd16<bool>> struct base_u16 {
-  __m128i value;
-  static const int SIZE = sizeof(value);
-
-  // Conversion from/to SIMD register
-  simdutf_really_inline base_u16() = default;
-  simdutf_really_inline base_u16(const __m128i _value) : value(_value) {}
-  // Bit operations
-  simdutf_really_inline simd16<T> operator|(const simd16<T> other) const {
-    return __lsx_vor_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<T> operator&(const simd16<T> other) const {
-    return __lsx_vand_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<T> operator^(const simd16<T> other) const {
-    return __lsx_vxor_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<T> bit_andnot(const simd16<T> other) const {
-    return __lsx_vandn_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
-  simdutf_really_inline simd16<T> &operator|=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd16<T> &operator&=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline simd16<T> &operator^=(const simd16<T> other) {
-    auto this_cast = static_cast<simd16<T> *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-
-  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
-                                               const simd16<T> rhs) {
-    return __lsx_vseq_h(lhs.value, rhs.value);
-  }
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return __lsx_vor_v(__lsx_vbsll_v(*this, N * 2),
-                       __lsx_vbsrl_v(prev_chunk, 16 - N * 2));
-  }
-};
-
-template <typename T, typename Mask = simd16<bool>>
-struct base16 : base_u16<T> {
-  typedef uint16_t bitmask_t;
-  typedef uint32_t bitmask2_t;
-
-  simdutf_really_inline base16() : base_u16<T>() {}
-  simdutf_really_inline base16(const __m128i _value) : base_u16<T>(_value) {}
-  template <typename Pointer>
-  simdutf_really_inline base16(const Pointer *ptr)
-      : base16(__lsx_vld(ptr, 0)) {}
-
-  static const int SIZE = sizeof(base_u16<T>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    return __lsx_vor_v(__lsx_vbsll_v(*this, N * 2),
-                       __lsx_vbsrl_v(prev_chunk, 16 - N * 2));
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd16<bool> : base16<bool> {
-  static simdutf_really_inline simd16<bool> splat(bool _value) {
-    return __lsx_vreplgr2vr_h(uint16_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd16() : base16() {}
-  simdutf_really_inline simd16(const __m128i _value) : base16<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
-};
-
-template <typename T> struct base16_numeric : base16<T> {
-  static simdutf_really_inline simd16<T> splat(T _value) {
-    return __lsx_vreplgr2vr_h(_value);
-  }
-  static simdutf_really_inline simd16<T> zero() { return __lsx_vldi(0); }
-  static simdutf_really_inline simd16<T> load(const T values[8]) {
-    return __lsx_vld(reinterpret_cast<const uint16_t *>(values), 0);
-  }
-
-  simdutf_really_inline base16_numeric() : base16<T>() {}
-  simdutf_really_inline base16_numeric(const __m128i _value)
-      : base16<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[8]) const {
-    return __lsx_vst(this->value, dst, 0);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
-    return __lsx_vadd_b(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
-    return __lsx_vsub_b(*this, other);
-  }
-  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
-    *this = *this + other;
-    return *static_cast<simd16<T> *>(this);
-  }
-  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
-    *this = *this - other;
-    return *static_cast<simd16<T> *>(this);
-  }
-};
-
-// Signed code unitstemplate<>
-template <> struct simd16<int16_t> : base16_numeric<int16_t> {
-  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
-  simdutf_really_inline simd16(const __m128i _value)
-      : base16_numeric<int16_t>(_value) {}
-  simdutf_really_inline simd16(simd16<bool> other)
-      : base16_numeric<int16_t>(other.value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
-  simdutf_really_inline operator simd16<uint16_t>() const;
-
-  // Order-sensitive comparisons
-  simdutf_really_inline simd16<int16_t>
-  max_val(const simd16<int16_t> other) const {
-    return __lsx_vmax_h(this->value, other.value);
-  }
-  simdutf_really_inline simd16<int16_t>
-  min_val(const simd16<int16_t> other) const {
-    return __lsx_vmin_h(this->value, other.value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<int16_t> other) const {
-    return __lsx_vsle_h(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<int16_t> other) const {
-    return __lsx_vslt_h(this->value, other.value);
-  }
-};
-
-// Unsigned code unitstemplate<>
-template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
-  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
-  simdutf_really_inline simd16(const __m128i _value)
-      : base16_numeric<uint16_t>((__m128i)_value) {}
-  simdutf_really_inline simd16(simd16<bool> other)
-      : base16_numeric<uint16_t>(other.value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
-
-  // Saturated math
-  simdutf_really_inline simd16<uint16_t>
-  saturating_add(const simd16<uint16_t> other) const {
-    return __lsx_vsadd_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  saturating_sub(const simd16<uint16_t> other) const {
-    return __lsx_vssub_hu(this->value, other.value);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd16<uint16_t>
-  max_val(const simd16<uint16_t> other) const {
-    return __lsx_vmax_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  min_val(const simd16<uint16_t> other) const {
-    return __lsx_vmin_hu(this->value, other.value);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  gt_bits(const simd16<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  lt_bits(const simd16<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<=(const simd16<uint16_t> other) const {
-    return __lsx_vsle_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>=(const simd16<uint16_t> other) const {
-    return __lsx_vsle_hu(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<uint16_t> other) const {
-    return __lsx_vslt_hu(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<uint16_t> other) const {
-    return __lsx_vslt_hu(this->value, other.value);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd16<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
-    return simd16<uint16_t>(__lsx_vsrli_h(this->value, N));
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
-    return simd16<uint16_t>(__lsx_vslli_h(this->value, N));
-  }
-
-  // logical operations
-  simdutf_really_inline simd16<uint16_t>
-  operator|(const simd16<uint16_t> other) const {
-    return __lsx_vor_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  operator&(const simd16<uint16_t> other) const {
-    return __lsx_vand_v(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  operator^(const simd16<uint16_t> other) const {
-    return __lsx_vxor_v(this->value, other.value);
-  }
-
-  // Pack with the unsigned saturation of two uint16_t code units into single
-  // uint8_t vector
-  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
-                                                   const simd16<uint16_t> &v1) {
-    return __lsx_vssrlni_bu_h(v1.value, v0.value, 0);
-  }
-
-  // Change the endianness
-  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
-    return __lsx_vshuf4i_b(this->value, 0b10110001);
-  }
-};
-
-simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
-  return this->value;
-}
-
-template <typename T> struct simd16x32 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
-  static_assert(
-      NUM_CHUNKS == 4,
-      "LOONGARCH kernel should use four registers per 64-byte block.");
-  simd16<T> chunks[NUM_CHUNKS];
-
-  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
-  simd16x32<T> &
-  operator=(const simd16<T> other) = delete; // no assignment allowed
-  simd16x32() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline
-  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
-            const simd16<T> chunk2, const simd16<T> chunk3)
-      : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdutf_really_inline simd16x32(const T *ptr)
-      : chunks{simd16<T>::load(ptr),
-               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
-               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
-    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
-    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
-  }
-
-  simdutf_really_inline simd16<T> reduce_or() const {
-    return (this->chunks[0] | this->chunks[1]) |
-           (this->chunks[2] | this->chunks[3]);
-  }
-
-  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
-
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
-    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
-    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
-    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    __m128i mask = __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[3]).value), 6);
-    mask = __lsx_vor_v(
-        mask, __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[2]).value), 4));
-    mask = __lsx_vor_v(
-        mask, __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[1]).value), 2));
-    mask = __lsx_vor_v(mask, __lsx_vmsknz_b((this->chunks[0]).value));
-    return __lsx_vpickve2gr_du(mask, 0);
-  }
-
-  simdutf_really_inline void swap_bytes() {
-    this->chunks[0] = this->chunks[0].swap_bytes();
-    this->chunks[1] = this->chunks[1].swap_bytes();
-    this->chunks[2] = this->chunks[2].swap_bytes();
-    this->chunks[3] = this->chunks[3].swap_bytes();
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
-                           this->chunks[2] == mask, this->chunks[3] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
-                           this->chunks[2] <= mask, this->chunks[3] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-
-    return simd16x32<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
-               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
-               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-    return simd16x32<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
-               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
-               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
-                           this->chunks[2] < mask, this->chunks[3] < mask)
-        .to_bitmask();
-  }
-
-}; // struct simd16x32<T>
-
-template <>
-simdutf_really_inline uint64_t simd16x32<uint16_t>::not_in_range(
-    const uint16_t low, const uint16_t high) const {
-  const simd16<uint16_t> mask_low = simd16<uint16_t>::splat(low);
-  const simd16<uint16_t> mask_high = simd16<uint16_t>::splat(high);
-  simd16x32<uint16_t> x(simd16<uint16_t>((this->chunks[0] > mask_high) |
-                                         (this->chunks[0] < mask_low)),
-                        simd16<uint16_t>((this->chunks[1] > mask_high) |
-                                         (this->chunks[1] < mask_low)),
-                        simd16<uint16_t>((this->chunks[2] > mask_high) |
-                                         (this->chunks[2] < mask_low)),
-                        simd16<uint16_t>((this->chunks[3] > mask_high) |
-                                         (this->chunks[3] < mask_low)));
-  return x.to_bitmask();
-}
-/* end file src/simdutf/lsx/simd16-inl.h */
-} // namespace simd
-} // unnamed namespace
-} // namespace lsx
-} // namespace simdutf
-
-#endif // SIMDUTF_LSX_SIMD_H
-/* end file src/simdutf/lsx/simd.h */
-
-/* begin file src/simdutf/lsx/end.h */
-/* end file src/simdutf/lsx/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_LSX
-
-#endif // SIMDUTF_LSX_H
-/* end file src/simdutf/lsx.h */
-/* begin file src/simdutf/lasx.h */
-#ifndef SIMDUTF_LASX_H
-#define SIMDUTF_LASX_H
-
-#ifdef SIMDUTF_FALLBACK_H
-  #error "lasx.h must be included before fallback.h"
-#endif
-
-
-#ifndef SIMDUTF_IMPLEMENTATION_LASX
-  #define SIMDUTF_IMPLEMENTATION_LASX (SIMDUTF_IS_LASX)
-#endif
-#if SIMDUTF_IMPLEMENTATION_LASX && SIMDUTF_IS_LASX
-  #define SIMDUTF_CAN_ALWAYS_RUN_LASX 1
-#else
-  #define SIMDUTF_CAN_ALWAYS_RUN_LASX 0
-#endif
-
-#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
-
-#if SIMDUTF_IMPLEMENTATION_LASX
-
-namespace simdutf {
-/**
- * Implementation for LoongArch ASX.
- */
-namespace lasx {} // namespace lasx
-} // namespace simdutf
-
-/* begin file src/simdutf/lasx/implementation.h */
-#ifndef SIMDUTF_LASX_IMPLEMENTATION_H
-#define SIMDUTF_LASX_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace lasx {
-
-namespace {
-using namespace simdutf;
-}
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("lasx", "LOONGARCH ASX",
-                                internal::instruction_set::LSX |
-                                    internal::instruction_set::LASX) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char *input, size_t length, char *output,
-                   base64_options options) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual result
-  base64_to_binary(const char16_t *input, size_t length, char *output,
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options =
-                       last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused virtual full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-};
-
-} // namespace lasx
-} // namespace simdutf
-
-#endif // SIMDUTF_LASX_IMPLEMENTATION_H
-/* end file src/simdutf/lasx/implementation.h */
-
-/* begin file src/simdutf/lasx/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "lasx"
-// #define SIMDUTF_IMPLEMENTATION lasx
-/* end file src/simdutf/lasx/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/lasx/intrinsics.h */
-#ifndef SIMDUTF_LASX_INTRINSICS_H
-#define SIMDUTF_LASX_INTRINSICS_H
-
-
-// This should be the correct header whether
-// you use visual studio or other compilers.
-#include <lsxintrin.h>
-#include <lasxintrin.h>
-
-#if defined(__loongarch_asx)
-  #ifdef __clang__
-    #define VREGS_PREFIX "$vr"
-    #define XREGS_PREFIX "$xr"
-  #else // GCC
-    #define VREGS_PREFIX "$f"
-    #define XREGS_PREFIX "$f"
-  #endif
-  #define __ALL_REGS                                                           \
-    "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,"  \
-    "27,28,29,30,31"
-// Convert __m128i to __m256i
-static inline __m256i ____m256i(__m128i in) {
-  __m256i out = __lasx_xvldi(0);
-  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
-                   " .ifc %[out], " XREGS_PREFIX "\\i    \n\t"
-                   "  .irp j," __ALL_REGS "\n\t"
-                   "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
-                   "    xvpermi.q $xr\\i, $xr\\j, 0x0  \n\t"
-                   "   .endif                           \n\t"
-                   "  .endr                             \n\t"
-                   " .endif                             \n\t"
-                   ".endr                               \n\t"
-                   : [out] "+f"(out)
-                   : [in] "f"(in));
-  return out;
-}
-// Convert two __m128i to __m256i
-static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
-  __m256i out;
-  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
-                   " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
-                   "  .irp j," __ALL_REGS "\n\t"
-                   "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
-                   "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
-                   "   .endif                           \n\t"
-                   "  .endr                             \n\t"
-                   " .endif                             \n\t"
-                   ".endr                               \n\t"
-                   ".ifnc %[out], %[hi]                 \n\t"
-                   ".irp i," __ALL_REGS "\n\t"
-                   " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
-                   "  .irp j," __ALL_REGS "\n\t"
-                   "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
-                   "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
-                   "   .endif                           \n\t"
-                   "  .endr                             \n\t"
-                   " .endif                             \n\t"
-                   ".endr                               \n\t"
-                   ".endif                              \n\t"
-                   : [out] "=f"(out), [hi] "+f"(inhi)
-                   : [lo] "f"(inlo));
-  return out;
-}
-// Convert __m256i low part to __m128i
-static inline __m128i lasx_extracti128_lo(__m256i in) {
-  __m128i out;
-  __asm__ volatile(".ifnc %[out], %[in]                 \n\t"
-                   ".irp i," __ALL_REGS "\n\t"
-                   " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-                   "  .irp j," __ALL_REGS "\n\t"
-                   "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-                   "    vori.b $vr\\i, $vr\\j, 0        \n\t"
-                   "   .endif                           \n\t"
-                   "  .endr                             \n\t"
-                   " .endif                             \n\t"
-                   ".endr                               \n\t"
-                   ".endif                              \n\t"
-                   : [out] "=f"(out)
-                   : [in] "f"(in));
-  return out;
-}
-// Convert __m256i high part to __m128i
-static inline __m128i lasx_extracti128_hi(__m256i in) {
-  __m128i out;
-  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
-                   " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-                   "  .irp j," __ALL_REGS "\n\t"
-                   "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-                   "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
-                   "   .endif                           \n\t"
-                   "  .endr                             \n\t"
-                   " .endif                             \n\t"
-                   ".endr                               \n\t"
-                   : [out] "=f"(out)
-                   : [in] "f"(in));
-  return out;
-}
-#endif
-
-#endif //  SIMDUTF_LASX_INTRINSICS_H
-/* end file src/simdutf/lasx/intrinsics.h */
-/* begin file src/simdutf/lasx/bitmanipulation.h */
-#ifndef SIMDUTF_LASX_BITMANIPULATION_H
-#define SIMDUTF_LASX_BITMANIPULATION_H
-
-#include <limits>
-
-namespace simdutf {
-namespace lasx {
-namespace {
-
-simdutf_really_inline int count_ones(uint64_t input_num) {
-  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__lsx_vreplgr2vr_d(input_num)), 0);
-}
-
-#if SIMDUTF_NEED_TRAILING_ZEROES
-simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
-  return __builtin_ctzll(input_num);
-}
-#endif
-
-} // unnamed namespace
-} // namespace lasx
-} // namespace simdutf
-
-#endif // SIMDUTF_LASX_BITMANIPULATION_H
-/* end file src/simdutf/lasx/bitmanipulation.h */
-/* begin file src/simdutf/lasx/simd.h */
-#ifndef SIMDUTF_LASX_SIMD_H
-#define SIMDUTF_LASX_SIMD_H
-
-#include <type_traits>
-
-namespace simdutf {
-namespace lasx {
-namespace {
-namespace simd {
-
-__attribute__((aligned(32))) static const uint8_t prev_shuf_table[32][32] = {
-    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
-     0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
-    {0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
-     31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14},
-    {0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
-     30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13},
-    {0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
-     29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12},
-    {0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
-     28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
-    {0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
-     27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10},
-    {0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
-     26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9},
-    {0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8,
-     25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7,
-     24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6,
-     23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5,
-     22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4,
-     21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3,
-     20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2,
-     19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1,
-     18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1},
-    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
-     17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0},
-    {15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
-     15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,
-     14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28,
-     13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27,
-     12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26,
-     11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,
-     10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
-     9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0},
-    {8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
-     8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0},
-    {7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22,
-     7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0},
-    {6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
-     6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0},
-    {5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
-     5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0},
-    {4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
-     4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0},
-    {3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
-     3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0},
-    {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
-     2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0},
-    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
-     1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0},
-    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
-     0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
-};
-
-__attribute__((aligned(32))) static const uint8_t bitsel_mask_table[32][32] = {
-    {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0},
-    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
-     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0}};
-
-// Forward-declared so they can be used by splat and friends.
-template <typename Child> struct base {
-  __m256i value;
-
-  // Zero constructor
-  simdutf_really_inline base() : value{__m256i()} {}
-
-  // Conversion from SIMD register
-  simdutf_really_inline base(const __m256i _value) : value(_value) {}
-  // Conversion to SIMD register
-  simdutf_really_inline operator const __m256i &() const { return this->value; }
-  simdutf_really_inline operator __m256i &() { return this->value; }
-  template <endianness big_endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    if (big_endian) {
-      __m256i zero = __lasx_xvldi(0);
-      __m256i in8 = __lasx_xvpermi_d(this->value, 0b11011000);
-      __m256i inlow = __lasx_xvilvl_b(in8, zero);
-      __m256i inhigh = __lasx_xvilvh_b(in8, zero);
-      __lasx_xvst(inlow, reinterpret_cast<uint16_t *>(ptr), 0);
-      __lasx_xvst(inhigh, reinterpret_cast<uint16_t *>(ptr), 32);
-    } else {
-      __m256i inlow = __lasx_vext2xv_hu_bu(this->value);
-      __m256i inhigh = __lasx_vext2xv_hu_bu(
-          __lasx_xvpermi_q(this->value, this->value, 0b00000001));
-      __lasx_xvst(inlow, reinterpret_cast<__m256i *>(ptr), 0);
-      __lasx_xvst(inhigh, reinterpret_cast<__m256i *>(ptr), 32);
-    }
-  }
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    __m256i in32_0 = __lasx_vext2xv_wu_bu(this->value);
-    __lasx_xvst(in32_0, reinterpret_cast<uint32_t *>(ptr), 0);
-
-    __m256i in8_1 = __lasx_xvpermi_d(this->value, 0b00000001);
-    __m256i in32_1 = __lasx_vext2xv_wu_bu(in8_1);
-    __lasx_xvst(in32_1, reinterpret_cast<uint32_t *>(ptr), 32);
-
-    __m256i in8_2 = __lasx_xvpermi_d(this->value, 0b00000010);
-    __m256i in32_2 = __lasx_vext2xv_wu_bu(in8_2);
-    __lasx_xvst(in32_2, reinterpret_cast<uint32_t *>(ptr), 64);
-
-    __m256i in8_3 = __lasx_xvpermi_d(this->value, 0b00000011);
-    __m256i in32_3 = __lasx_vext2xv_wu_bu(in8_3);
-    __lasx_xvst(in32_3, reinterpret_cast<uint32_t *>(ptr), 96);
-  }
-  // Bit operations
-  simdutf_really_inline Child operator|(const Child other) const {
-    return __lasx_xvor_v(this->value, other);
-  }
-  simdutf_really_inline Child operator&(const Child other) const {
-    return __lasx_xvand_v(this->value, other);
-  }
-  simdutf_really_inline Child operator^(const Child other) const {
-    return __lasx_xvxor_v(this->value, other);
-  }
-  simdutf_really_inline Child bit_andnot(const Child other) const {
-    return __lasx_xvandn_v(this->value, other);
-  }
-  simdutf_really_inline Child &operator|=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast | other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator&=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast & other;
-    return *this_cast;
-  }
-  simdutf_really_inline Child &operator^=(const Child other) {
-    auto this_cast = static_cast<Child *>(this);
-    *this_cast = *this_cast ^ other;
-    return *this_cast;
-  }
-};
-
-template <typename T> struct simd8;
-
-template <typename T, typename Mask = simd8<bool>>
-struct base8 : base<simd8<T>> {
-  typedef uint32_t bitmask_t;
-  typedef uint64_t bitmask2_t;
-
-  simdutf_really_inline base8() : base<simd8<T>>() {}
-  simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
-  simdutf_really_inline T first() const {
-    return __lasx_xvpickve2gr_wu(this->value, 0);
-  }
-  simdutf_really_inline T last() const {
-    return __lasx_xvpickve2gr_wu(this->value, 7);
-  }
-  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
-                                               const simd8<T> rhs) {
-    return __lasx_xvseq_b(lhs, rhs);
-  }
-
-  static const int SIZE = sizeof(base<T>::value);
-
-  template <int N = 1>
-  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
-    if (!N)
-      return this->value;
-
-    __m256i zero = __lasx_xvldi(0);
-    __m256i result, shuf;
-    if (N < 16) {
-      shuf = __lasx_xvld(prev_shuf_table[N], 0);
-
-      result = __lasx_xvshuf_b(
-          __lasx_xvpermi_q(this->value, this->value, 0b00000001), this->value,
-          shuf);
-      __m256i srl_prev = __lasx_xvbsrl_v(
-          __lasx_xvpermi_q(zero, prev_chunk.value, 0b00110001), (16 - N));
-      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
-      result = __lasx_xvbitsel_v(result, srl_prev, mask);
-
-      return result;
-    } else if (N == 16) {
-      return __lasx_xvpermi_q(this->value, prev_chunk.value, 0b00100001);
-    } /*else {
-      __m256i sll_value = __lasx_xvbsll_v(
-          __lasx_xvpermi_q(zero, this->value, 0b00000011), (N - 16) % 32);
-      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
-      shuf = __lasx_xvld(prev_shuf_table[N], 0);
-      result = __lasx_xvshuf_b(
-          __lasx_xvpermi_q(prev_chunk.value, prev_chunk.value, 0b00000001),
-          prev_chunk.value, shuf);
-      result = __lasx_xvbitsel_v(sll_value, result, mask);
-      return result;
-    }*/
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd8<bool> : base8<bool> {
-  static simdutf_really_inline simd8<bool> splat(bool _value) {
-    return __lasx_xvreplgr2vr_b(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd8() : base8() {}
-  simdutf_really_inline simd8(const __m256i _value) : base8<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
-
-  simdutf_really_inline uint32_t to_bitmask() const {
-    __m256i mask = __lasx_xvmsknz_b(this->value);
-    uint32_t mask0 = __lasx_xvpickve2gr_wu(mask, 0);
-    uint32_t mask1 = __lasx_xvpickve2gr_wu(mask, 4);
-    return (mask0 | (mask1 << 16));
-  }
-  simdutf_really_inline bool any() const {
-    if (__lasx_xbz_b(this->value))
-      return false;
-    return true;
-  }
-  simdutf_really_inline bool none() const {
-    if (__lasx_xbz_b(this->value))
-      return true;
-    return false;
-  }
-  simdutf_really_inline bool all() const {
-    if (__lasx_xbnz_b(this->value))
-      return true;
-    return false;
-  }
-  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base8_numeric : base8<T> {
-  static simdutf_really_inline simd8<T> splat(T _value) {
-    return __lasx_xvreplgr2vr_b(_value);
-  }
-  static simdutf_really_inline simd8<T> zero() { return __lasx_xvldi(0); }
-  static simdutf_really_inline simd8<T> load(const T values[32]) {
-    return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
-  }
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
-                                                  T v5, T v6, T v7, T v8, T v9,
-                                                  T v10, T v11, T v12, T v13,
-                                                  T v14, T v15) {
-    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
-                    v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
-                    v12, v13, v14, v15);
-  }
-
-  simdutf_really_inline base8_numeric() : base8<T>() {}
-  simdutf_really_inline base8_numeric(const __m256i _value)
-      : base8<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[32]) const {
-    return __lasx_xvst(this->value, reinterpret_cast<__m256i *>(dst), 0);
-  }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
-    return __lasx_xvadd_b(this->value, other);
-  }
-  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
-    return __lasx_xvsub_b(this->value, other);
-  }
-  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
-    *this = *this + other;
-    return *static_cast<simd8<T> *>(this);
-  }
-  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
-    *this = *this - other;
-    return *static_cast<simd8<T> *>(this);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-
-  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
-  // for out of range values)
-  template <typename L>
-  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
-    __m256i origin = __lasx_xvand_v(this->value, __lasx_xvldi(0x1f));
-    return __lasx_xvshuf_b(__lasx_xvldi(0), lookup_table, origin);
-  }
-
-  template <typename L>
-  simdutf_really_inline simd8<L>
-  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
-            L replace5, L replace6, L replace7, L replace8, L replace9,
-            L replace10, L replace11, L replace12, L replace13, L replace14,
-            L replace15) const {
-    return lookup_16(simd8<L>::repeat_16(
-        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
-        replace7, replace8, replace9, replace10, replace11, replace12,
-        replace13, replace14, replace15));
-  }
-};
-
-// Signed bytes
-template <> struct simd8<int8_t> : base8_numeric<int8_t> {
-  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
-  simdutf_really_inline simd8(const __m256i _value)
-      : base8_numeric<int8_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
-  simdutf_really_inline operator simd8<uint8_t>() const;
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-        int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-        int8_t v12, int8_t v13, int8_t v14, int8_t v15, int8_t v16, int8_t v17,
-        int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
-        int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29,
-        int8_t v30, int8_t v31)
-      : simd8((__m256i)v32i8{v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7,
-                             v8,  v9,  v10, v11, v12, v13, v14, v15,
-                             v16, v17, v18, v19, v20, v21, v22, v23,
-                             v24, v25, v26, v27, v28, v29, v30, v31}) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<int8_t>
-  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
-            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
-            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
-    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                         v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                         v10, v11, v12, v13, v14, v15);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    __m256i ascii_mask = __lasx_xvslti_b(this->value, 0);
-    if (__lasx_xbnz_v(ascii_mask))
-      return false;
-    return true;
-  }
-  // Order-sensitive comparisons
-  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
-    return __lasx_xvmax_b(this->value, other);
-  }
-  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
-    return __lasx_xvmin_b(this->value, other);
-  }
-  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
-    return __lasx_xvslt_b(other, this->value);
-  }
-  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
-    return __lasx_xvslt_b(this->value, other);
-  }
-};
-
-// Unsigned bytes
-template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
-  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
-  simdutf_really_inline simd8(const __m256i _value)
-      : base8_numeric<uint8_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
-  // Member-by-member initialization
-  simdutf_really_inline
-  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
-        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
-        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
-        uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20,
-        uint8_t v21, uint8_t v22, uint8_t v23, uint8_t v24, uint8_t v25,
-        uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30,
-        uint8_t v31)
-      : simd8((__m256i)v32u8{v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7,
-                             v8,  v9,  v10, v11, v12, v13, v14, v15,
-                             v16, v17, v18, v19, v20, v21, v22, v23,
-                             v24, v25, v26, v27, v28, v29, v30, v31}) {}
-  // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdutf_really_inline static simd8<uint8_t>
-  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
-            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
-            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
-            uint8_t v15) {
-    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
-                          v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
-                          v10, v11, v12, v13, v14, v15);
-  }
-
-  // Saturated math
-  simdutf_really_inline simd8<uint8_t>
-  saturating_add(const simd8<uint8_t> other) const {
-    return __lasx_xvsadd_bu(this->value, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  saturating_sub(const simd8<uint8_t> other) const {
-    return __lasx_xvssub_bu(this->value, other);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd8<uint8_t>
-  max_val(const simd8<uint8_t> other) const {
-    return __lasx_xvmax_bu(*this, other);
-  }
-  simdutf_really_inline simd8<uint8_t>
-  min_val(const simd8<uint8_t> other) const {
-    return __lasx_xvmin_bu(*this, other);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  gt_bits(const simd8<uint8_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd8<uint8_t>
-  lt_bits(const simd8<uint8_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<=(const simd8<uint8_t> other) const {
-    return __lasx_xvsle_bu(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>=(const simd8<uint8_t> other) const {
-    return __lasx_xvsle_bu(other, *this);
-  }
-  simdutf_really_inline simd8<bool>
-  operator>(const simd8<uint8_t> other) const {
-    return __lasx_xvslt_bu(*this, other);
-  }
-  simdutf_really_inline simd8<bool>
-  operator<(const simd8<uint8_t> other) const {
-    return __lasx_xvslt_bu(other, *this);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd8<bool> bits_not_set() const {
-    return *this == uint8_t(0);
-  }
-  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-  simdutf_really_inline bool is_ascii() const {
-    __m256i ascii_mask = __lasx_xvslti_b(this->value, 0);
-    if (__lasx_xbnz_v(ascii_mask))
-      return false;
-    return true;
-  }
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    if (__lasx_xbnz_v(this->value))
-      return true;
-    return false;
-  }
-  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
-    return (*this & bits).any_bits_set_anywhere();
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
-    return __lasx_xvsrli_b(this->value, N);
-  }
-  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
-    return __lasx_xvslli_b(this->value, N);
-  }
-};
-simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
-  return this->value;
-}
-
-template <typename T> struct simd8x64 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
-  static_assert(NUM_CHUNKS == 2,
-                "LASX kernel should use two registers per 64-byte block.");
-  simd8<T> chunks[NUM_CHUNKS];
-
-  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
-  simd8x64<T> &
-  operator=(const simd8<T> other) = delete; // no assignment allowed
-  simd8x64() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1)
-      : chunks{chunk0, chunk1} {}
-  simdutf_really_inline simd8x64(const T *ptr)
-      : chunks{simd8<T>::load(ptr),
-               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r_hi = this->chunks[1].to_bitmask();
-    return r_lo | (r_hi << 32);
-  }
-
-  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
-    this->chunks[0] |= other.chunks[0];
-    this->chunks[1] |= other.chunks[1];
-    return *this;
-  }
-
-  simdutf_really_inline simd8<T> reduce_or() const {
-    return this->chunks[0] | this->chunks[1];
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  template <endianness endian>
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 0);
-    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
-                                                          sizeof(simd8<T>) * 1);
-  }
-
-  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
-    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
-  }
-
-  simdutf_really_inline simd8x64<T> bit_or(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<T>(this->chunks[0] | mask, this->chunks[1] | mask);
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
-    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
-                          this->chunks[1] == other.chunks[1])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-
-    return simd8x64<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd8<T> mask_low = simd8<T>::splat(low);
-    const simd8<T> mask_high = simd8<T>::splat(high);
-    return simd8x64<bool>(
-               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
-               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t gt(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq(const T m) const {
-    const simd8<T> mask = simd8<T>::splat(m);
-    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask)
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
-    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
-    return simd8x64<bool>((simd8<uint8_t>(__m256i(this->chunks[0])) >= mask),
-                          (simd8<uint8_t>(__m256i(this->chunks[1])) >= mask))
-        .to_bitmask();
-  }
-}; // struct simd8x64<T>
-
-/* begin file src/simdutf/lasx/simd16-inl.h */
-template <typename T> struct simd16;
-
-template <typename T, typename Mask = simd16<bool>>
-struct base16 : base<simd16<T>> {
-  using bitmask_type = uint32_t;
-
-  simdutf_really_inline base16() : base<simd16<T>>() {}
-  simdutf_really_inline base16(const __m256i _value)
-      : base<simd16<T>>(_value) {}
-  template <typename Pointer>
-  simdutf_really_inline base16(const Pointer *ptr)
-      : base16(__lasx_xvld(reinterpret_cast<const __m256i *>(ptr), 0)) {}
-  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
-                                               const simd16<T> rhs) {
-    return __lasx_xvseq_h(lhs.value, rhs.value);
-  }
-
-  /// the size of vector in bytes
-  static const int SIZE = sizeof(base<simd16<T>>::value);
-
-  /// the number of elements of type T a vector can hold
-  static const int ELEMENTS = SIZE / sizeof(T);
-
-  template <int N = 1>
-  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
-    if (!N)
-      return this->value;
-
-    __m256i zero = __lasx_xvldi(0);
-    __m256i result, shuf;
-    if (N < 8) {
-      shuf = __lasx_xvld(prev_shuf_table[N * 2], 0);
-
-      result = __lasx_xvshuf_b(
-          __lasx_xvpermi_q(this->value, this->value, 0b00000001), this->value,
-          shuf);
-      __m256i srl_prev = __lasx_xvbsrl_v(
-          __lasx_xvpermi_q(zero, prev_chunk, 0b00110001), (16 - N * 2));
-      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
-      result = __lasx_xvbitsel_v(result, srl_prev, mask);
-
-      return result;
-    } else if (N == 8) {
-      return __lasx_xvpermi_q(this->value, prev_chunk, 0b00100001);
-    } else {
-      __m256i sll_value = __lasx_xvbsll_v(
-          __lasx_xvpermi_q(zero, this->value, 0b00000011), (N * 2 - 16));
-      __m256i mask = __lasx_xvld(bitsel_mask_table[N * 2], 0);
-      shuf = __lasx_xvld(prev_shuf_table[N * 2], 0);
-      result =
-          __lasx_xvshuf_b(__lasx_xvpermi_q(prev_chunk, prev_chunk, 0b00000001),
-                          prev_chunk, shuf);
-      result = __lasx_xvbitsel_v(sll_value, result, mask);
-      return result;
-    }
-  }
-};
-
-// SIMD byte mask type (returned by things like eq and gt)
-template <> struct simd16<bool> : base16<bool> {
-  static simdutf_really_inline simd16<bool> splat(bool _value) {
-    return __lasx_xvreplgr2vr_h(uint8_t(-(!!_value)));
-  }
-
-  simdutf_really_inline simd16() : base16() {}
-  simdutf_really_inline simd16(const __m256i _value) : base16<bool>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
-
-  simdutf_really_inline bitmask_type to_bitmask() const {
-    __m256i mask = __lasx_xvmsknz_b(this->value);
-    bitmask_type mask0 = __lasx_xvpickve2gr_wu(mask, 0);
-    bitmask_type mask1 = __lasx_xvpickve2gr_wu(mask, 4);
-    return (mask0 | (mask1 << 16));
-  }
-  simdutf_really_inline bool any() const {
-    if (__lasx_xbz_v(this->value))
-      return false;
-    return true;
-  }
-  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
-};
-
-template <typename T> struct base16_numeric : base16<T> {
-  static simdutf_really_inline simd16<T> splat(T _value) {
-    return __lasx_xvreplgr2vr_h((uint16_t)_value);
-  }
-  static simdutf_really_inline simd16<T> zero() { return __lasx_xvldi(0); }
-  static simdutf_really_inline simd16<T> load(const T values[8]) {
-    return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
-  }
-
-  simdutf_really_inline base16_numeric() : base16<T>() {}
-  simdutf_really_inline base16_numeric(const __m256i _value)
-      : base16<T>(_value) {}
-
-  // Store to array
-  simdutf_really_inline void store(T dst[8]) const {
-    return __lasx_xvst(this->value, reinterpret_cast<__m256i *>(dst), 0);
-  }
-
-  // Override to distinguish from bool version
-  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFFFu; }
-
-  // Addition/subtraction are the same for signed and unsigned
-  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
-    return __lasx_xvadd_h(*this, other);
-  }
-  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
-    return __lasx_xvsub_h(*this, other);
-  }
-  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
-    *this = *this + other;
-    return *static_cast<simd16<T> *>(this);
-  }
-  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
-    *this = *this - other;
-    return *static_cast<simd16<T> *>(this);
-  }
-};
-
-// Signed code units
-template <> struct simd16<int16_t> : base16_numeric<int16_t> {
-  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
-  simdutf_really_inline simd16(const __m256i _value)
-      : base16_numeric<int16_t>(_value) {}
-  // Splat constructor
-  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
-  // Order-sensitive comparisons
-  simdutf_really_inline simd16<int16_t>
-  max_val(const simd16<int16_t> other) const {
-    return __lasx_xvmax_h(*this, other);
-  }
-  simdutf_really_inline simd16<int16_t>
-  min_val(const simd16<int16_t> other) const {
-    return __lasx_xvmin_h(*this, other);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<int16_t> other) const {
-    return __lasx_xvsle_h(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<int16_t> other) const {
-    return __lasx_xvslt_h(this->value, other.value);
-  }
-};
-
-// Unsigned code units
-template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
-  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
-  simdutf_really_inline simd16(const __m256i _value)
-      : base16_numeric<uint16_t>(_value) {}
-
-  // Splat constructor
-  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
-  // Array constructor
-  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
-  simdutf_really_inline simd16(const char16_t *values)
-      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
-
-  // Saturated math
-  simdutf_really_inline simd16<uint16_t>
-  saturating_add(const simd16<uint16_t> other) const {
-    return __lasx_xvsadd_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  saturating_sub(const simd16<uint16_t> other) const {
-    return __lasx_xvssub_hu(this->value, other.value);
-  }
-
-  // Order-specific operations
-  simdutf_really_inline simd16<uint16_t>
-  max_val(const simd16<uint16_t> other) const {
-    return __lasx_xvmax_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<uint16_t>
-  min_val(const simd16<uint16_t> other) const {
-    return __lasx_xvmin_hu(this->value, other.value);
-  }
-  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  gt_bits(const simd16<uint16_t> other) const {
-    return this->saturating_sub(other);
-  }
-  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdutf_really_inline simd16<uint16_t>
-  lt_bits(const simd16<uint16_t> other) const {
-    return other.saturating_sub(*this);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<=(const simd16<uint16_t> other) const {
-    return __lasx_xvsle_hu(this->value, other.value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>=(const simd16<uint16_t> other) const {
-    return __lasx_xvsle_hu(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator>(const simd16<uint16_t> other) const {
-    return __lasx_xvslt_hu(other.value, this->value);
-  }
-  simdutf_really_inline simd16<bool>
-  operator<(const simd16<uint16_t> other) const {
-    return __lasx_xvslt_hu(this->value, other.value);
-  }
-
-  // Bit-specific operations
-  simdutf_really_inline simd16<bool> bits_not_set() const {
-    return *this == uint16_t(0);
-  }
-  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
-    return (*this & bits).bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set() const {
-    return ~this->bits_not_set();
-  }
-  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
-    return ~this->bits_not_set(bits);
-  }
-
-  simdutf_really_inline bool any_bits_set_anywhere() const {
-    if (__lasx_xbnz_v(this->value))
-      return true;
-    return false;
-  }
-  simdutf_really_inline bool
-  any_bits_set_anywhere(simd16<uint16_t> bits) const {
-    return (*this & bits).any_bits_set_anywhere();
-  }
-
-  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
-    return simd16<uint16_t>(__lasx_xvsrli_h(this->value, N));
-  }
-  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
-    return simd16<uint16_t>(__lasx_xvslli_h(this->value, N));
-  }
-
-  // Change the endianness
-  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
-    return __lasx_xvshuf4i_b(this->value, 0b10110001);
-  }
-
-  // Pack with the unsigned saturation of two uint16_t code units into single
-  // uint8_t vector
-  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
-                                                   const simd16<uint16_t> &v1) {
-    return __lasx_xvpermi_d(__lasx_xvssrlni_bu_h(v1.value, v0.value, 0),
-                            0b11011000);
-  }
-};
-
-template <typename T> struct simd16x32 {
-  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
-  static_assert(NUM_CHUNKS == 2,
-                "LASX kernel should use two registers per 64-byte block.");
-  simd16<T> chunks[NUM_CHUNKS];
-
-  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
-  simd16x32<T> &
-  operator=(const simd16<T> other) = delete; // no assignment allowed
-  simd16x32() = delete;                      // no default constructor allowed
-
-  simdutf_really_inline simd16x32(const simd16<T> chunk0,
-                                  const simd16<T> chunk1)
-      : chunks{chunk0, chunk1} {}
-  simdutf_really_inline simd16x32(const T *ptr)
-      : chunks{simd16<T>::load(ptr),
-               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T))} {}
-
-  simdutf_really_inline void store(T *ptr) const {
-    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
-    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
-  }
-
-  simdutf_really_inline uint64_t to_bitmask() const {
-    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
-    uint64_t r_hi = this->chunks[1].to_bitmask();
-    return r_lo | (r_hi << 32);
-  }
-
-  simdutf_really_inline simd16<T> reduce_or() const {
-    return this->chunks[0] | this->chunks[1];
-  }
-
-  simdutf_really_inline bool is_ascii() const {
-    return this->reduce_or().is_ascii();
-  }
-
-  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
-    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
-    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>));
-  }
-
-  simdutf_really_inline simd16x32<T> bit_or(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<T>(this->chunks[0] | mask, this->chunks[1] | mask);
-  }
-
-  simdutf_really_inline void swap_bytes() {
-    this->chunks[0] = this->chunks[0].swap_bytes();
-    this->chunks[1] = this->chunks[1].swap_bytes();
-  }
-
-  simdutf_really_inline uint64_t eq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
-    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
-                           this->chunks[1] == other.chunks[1])
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t lteq(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
-        .to_bitmask();
-  }
-
-  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(low);
-    const simd16<T> mask_high = simd16<T>::splat(high);
-
-    return simd16x32<bool>(
-               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
-               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
-    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
-    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
-    return simd16x32<bool>(
-               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
-               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low))
-        .to_bitmask();
-  }
-  simdutf_really_inline uint64_t lt(const T m) const {
-    const simd16<T> mask = simd16<T>::splat(m);
-    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
-        .to_bitmask();
-  }
-}; // struct simd16x32<T>
-/* end file src/simdutf/lasx/simd16-inl.h */
-} // namespace simd
-} // unnamed namespace
-} // namespace lasx
-} // namespace simdutf
-
-#endif // SIMDUTF_LASX_SIMD_H
-/* end file src/simdutf/lasx/simd.h */
-
-/* begin file src/simdutf/lasx/end.h */
-/* end file src/simdutf/lasx/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_LASX
-
-#endif // SIMDUTF_LASX_H
-/* end file src/simdutf/lasx.h */
-/* begin file src/simdutf/fallback.h */
-#ifndef SIMDUTF_FALLBACK_H
-#define SIMDUTF_FALLBACK_H
-
-
-// Note that fallback.h is always imported last.
-
-// Default Fallback to on unless a builtin implementation has already been
-// selected.
-#ifndef SIMDUTF_IMPLEMENTATION_FALLBACK
-  #if SIMDUTF_CAN_ALWAYS_RUN_ARM64 || SIMDUTF_CAN_ALWAYS_RUN_ICELAKE ||        \
-      SIMDUTF_CAN_ALWAYS_RUN_HASWELL || SIMDUTF_CAN_ALWAYS_RUN_WESTMERE ||     \
-      SIMDUTF_CAN_ALWAYS_RUN_PPC64 || SIMDUTF_CAN_ALWAYS_RUN_RVV ||            \
-      SIMDUTF_CAN_ALWAYS_RUN_LSX || SIMDUTF_CAN_ALWAYS_RUN_LASX
-    #define SIMDUTF_IMPLEMENTATION_FALLBACK 0
-  #else
-    #define SIMDUTF_IMPLEMENTATION_FALLBACK 1
-  #endif
-#endif
-
-#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
-
-#if SIMDUTF_IMPLEMENTATION_FALLBACK
-
-namespace simdutf {
-/**
- * Fallback implementation (runs on any machine).
- */
-namespace fallback {} // namespace fallback
-} // namespace simdutf
-
-/* begin file src/simdutf/fallback/implementation.h */
-#ifndef SIMDUTF_FALLBACK_IMPLEMENTATION_H
-#define SIMDUTF_FALLBACK_IMPLEMENTATION_H
-
-
-namespace simdutf {
-namespace fallback {
-
-namespace {
-using namespace simdutf;
-}
-
-class implementation final : public simdutf::implementation {
-public:
-  simdutf_really_inline implementation()
-      : simdutf::implementation("fallback", "Generic fallback implementation",
-                                0) {}
-  simdutf_warn_unused int detect_encodings(const char *input,
-                                           size_t length) const noexcept final;
-  simdutf_warn_unused bool validate_utf8(const char *buf,
-                                         size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_ascii(const char *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result
-  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                            size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                          size_t len) const noexcept final;
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *buf, size_t len, char *utf8_output) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len, char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
-                                  char *latin1_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final;
-  simdutf_warn_unused result
-  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                      char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
-                                char *latin1_output) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
-                           char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
-                                 char16_t *utf16_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  simdutf_warn_unused size_t
-  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
-                                 char32_t *utf32_buffer) const noexcept final;
-  void change_endianness_utf16(const char16_t *buf, size_t length,
-                               char16_t *output) const noexcept final;
-  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
-                                           size_t length) const noexcept;
-  simdutf_warn_unused size_t count_utf8(const char *buf,
-                                        size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t length) const noexcept;
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept;
-  simdutf_warn_unused result base64_to_binary(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options) const noexcept;
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept;
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_options =
-          last_chunk_handling_options::loose) const noexcept;
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept;
-};
-} // namespace fallback
-} // namespace simdutf
-
-#endif // SIMDUTF_FALLBACK_IMPLEMENTATION_H
-/* end file src/simdutf/fallback/implementation.h */
-
-/* begin file src/simdutf/fallback/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "fallback"
-// #define SIMDUTF_IMPLEMENTATION fallback
-/* end file src/simdutf/fallback/begin.h */
-
-  // Declarations
-/* begin file src/simdutf/fallback/bitmanipulation.h */
-#ifndef SIMDUTF_FALLBACK_BITMANIPULATION_H
-#define SIMDUTF_FALLBACK_BITMANIPULATION_H
-
-#include <limits>
-
-namespace simdutf {
-namespace fallback {
-namespace {} // unnamed namespace
-} // namespace fallback
-} // namespace simdutf
-
-#endif // SIMDUTF_FALLBACK_BITMANIPULATION_H
-/* end file src/simdutf/fallback/bitmanipulation.h */
-
-/* begin file src/simdutf/fallback/end.h */
-/* end file src/simdutf/fallback/end.h */
-
-#endif // SIMDUTF_IMPLEMENTATION_FALLBACK
-#endif // SIMDUTF_FALLBACK_H
-/* end file src/simdutf/fallback.h */
-
-/* begin file src/scalar/utf8.h */
-#ifndef SIMDUTF_UTF8_H
-#define SIMDUTF_UTF8_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf8 {
-#if SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_RVV
-// only used by the fallback kernel.
-// credit: based on code from Google Fuchsia (Apache Licensed)
-inline simdutf_warn_unused bool validate(const char *buf, size_t len) noexcept {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  uint64_t pos = 0;
-  uint32_t code_point = 0;
-  while (pos < len) {
-    // check of the next 16 bytes are ascii.
-    uint64_t next_pos = pos + 16;
-    if (next_pos <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      std::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        pos = next_pos;
-        continue;
-      }
-    }
-    unsigned char byte = data[pos];
-
-    while (byte < 0b10000000) {
-      if (++pos == len) {
-        return true;
-      }
-      byte = data[pos];
-    }
-
-    if ((byte & 0b11100000) == 0b11000000) {
-      next_pos = pos + 2;
-      if (next_pos > len) {
-        return false;
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      // range check
-      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if ((code_point < 0x80) || (0x7ff < code_point)) {
-        return false;
-      }
-    } else if ((byte & 0b11110000) == 0b11100000) {
-      next_pos = pos + 3;
-      if (next_pos > len) {
-        return false;
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      // range check
-      code_point = (byte & 0b00001111) << 12 |
-                   (data[pos + 1] & 0b00111111) << 6 |
-                   (data[pos + 2] & 0b00111111);
-      if ((code_point < 0x800) || (0xffff < code_point) ||
-          (0xd7ff < code_point && code_point < 0xe000)) {
-        return false;
-      }
-    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
-      next_pos = pos + 4;
-      if (next_pos > len) {
-        return false;
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return false;
-      }
-      // range check
-      code_point =
-          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
-          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff || 0x10ffff < code_point) {
-        return false;
-      }
-    } else {
-      // we may have a continuation
-      return false;
-    }
-    pos = next_pos;
-  }
-  return true;
-}
-#endif
-
-inline simdutf_warn_unused result validate_with_errors(const char *buf,
-                                                       size_t len) noexcept {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  uint32_t code_point = 0;
-  while (pos < len) {
-    // check of the next 16 bytes are ascii.
-    size_t next_pos = pos + 16;
-    if (next_pos <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      std::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        pos = next_pos;
-        continue;
-      }
-    }
-    unsigned char byte = data[pos];
-
-    while (byte < 0b10000000) {
-      if (++pos == len) {
-        return result(error_code::SUCCESS, len);
-      }
-      byte = data[pos];
-    }
-
-    if ((byte & 0b11100000) == 0b11000000) {
-      next_pos = pos + 2;
-      if (next_pos > len) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if ((code_point < 0x80) || (0x7ff < code_point)) {
-        return result(error_code::OVERLONG, pos);
-      }
-    } else if ((byte & 0b11110000) == 0b11100000) {
-      next_pos = pos + 3;
-      if (next_pos > len) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      code_point = (byte & 0b00001111) << 12 |
-                   (data[pos + 1] & 0b00111111) << 6 |
-                   (data[pos + 2] & 0b00111111);
-      if ((code_point < 0x800) || (0xffff < code_point)) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0xd7ff < code_point && code_point < 0xe000) {
-        return result(error_code::SURROGATE, pos);
-      }
-    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
-      next_pos = pos + 4;
-      if (next_pos > len) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      code_point =
-          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
-          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0x10ffff < code_point) {
-        return result(error_code::TOO_LARGE, pos);
-      }
-    } else {
-      // we either have too many continuation bytes or an invalid leading byte
-      if ((byte & 0b11000000) == 0b10000000) {
-        return result(error_code::TOO_LONG, pos);
-      } else {
-        return result(error_code::HEADER_BITS, pos);
-      }
-    }
-    pos = next_pos;
-  }
-  return result(error_code::SUCCESS, len);
-}
-
-// Finds the previous leading byte starting backward from buf and validates with
-// errors from there Used to pinpoint the location of an error when an invalid
-// chunk is detected We assume that the stream starts with a leading byte, and
-// to check that it is the case, we ask that you pass a pointer to the start of
-// the stream (start).
-inline simdutf_warn_unused result rewind_and_validate_with_errors(
-    const char *start, const char *buf, size_t len) noexcept {
-  // First check that we start with a leading byte
-  if ((*start & 0b11000000) == 0b10000000) {
-    return result(error_code::TOO_LONG, 0);
-  }
-  size_t extra_len{0};
-  // A leading byte cannot be further than 4 bytes away
-  for (int i = 0; i < 5; i++) {
-    unsigned char byte = *buf;
-    if ((byte & 0b11000000) != 0b10000000) {
-      break;
-    } else {
-      buf--;
-      extra_len++;
-    }
-  }
-
-  result res = validate_with_errors(buf, len + extra_len);
-  res.count -= extra_len;
-  return res;
-}
-
-inline size_t count_code_points(const char *buf, size_t len) {
-  const int8_t *p = reinterpret_cast<const int8_t *>(buf);
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    // -65 is 0b10111111, anything larger in two-complement's should start a new
-    // code point.
-    if (p[i] > -65) {
-      counter++;
-    }
-  }
-  return counter;
-}
-
-inline size_t utf16_length_from_utf8(const char *buf, size_t len) {
-  const int8_t *p = reinterpret_cast<const int8_t *>(buf);
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    if (p[i] > -65) {
-      counter++;
-    }
-    if (uint8_t(p[i]) >= 240) {
-      counter++;
-    }
-  }
-  return counter;
-}
-
-simdutf_warn_unused inline size_t trim_partial_utf8(const char *input,
-                                                    size_t length) {
-  if (length < 3) {
-    switch (length) {
-    case 2:
-      if (uint8_t(input[length - 1]) >= 0xc0) {
-        return length - 1;
-      } // 2-, 3- and 4-byte characters with only 1 byte left
-      if (uint8_t(input[length - 2]) >= 0xe0) {
-        return length - 2;
-      } // 3- and 4-byte characters with only 2 bytes left
-      return length;
-    case 1:
-      if (uint8_t(input[length - 1]) >= 0xc0) {
-        return length - 1;
-      } // 2-, 3- and 4-byte characters with only 1 byte left
-      return length;
-    case 0:
-      return length;
-    }
-  }
-  if (uint8_t(input[length - 1]) >= 0xc0) {
-    return length - 1;
-  } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (uint8_t(input[length - 2]) >= 0xe0) {
-    return length - 2;
-  } // 3- and 4-byte characters with only 1 byte left
-  if (uint8_t(input[length - 3]) >= 0xf0) {
-    return length - 3;
-  } // 4-byte characters with only 3 bytes left
-  return length;
-}
-
-} // namespace utf8
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf8.h */
-/* begin file src/scalar/utf16.h */
-#ifndef SIMDUTF_UTF16_H
-#define SIMDUTF_UTF16_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf16 {
-
-inline simdutf_warn_unused uint16_t swap_bytes(const uint16_t word) {
-  return uint16_t((word >> 8) | (word << 8));
-}
-
-template <endianness big_endian>
-inline simdutf_warn_unused bool validate(const char16_t *data,
-                                         size_t len) noexcept {
-  uint64_t pos = 0;
-  while (pos < len) {
-    char16_t word =
-        !match_system(big_endian) ? swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xF800) == 0xD800) {
-      if (pos + 1 >= len) {
-        return false;
-      }
-      char16_t diff = char16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return false;
-      }
-      char16_t next_word =
-          !match_system(big_endian) ? swap_bytes(data[pos + 1]) : data[pos + 1];
-      char16_t diff2 = char16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return false;
-      }
-      pos += 2;
-    } else {
-      pos++;
-    }
-  }
-  return true;
-}
-
-template <endianness big_endian>
-inline simdutf_warn_unused result validate_with_errors(const char16_t *data,
-                                                       size_t len) noexcept {
-  size_t pos = 0;
-  while (pos < len) {
-    char16_t word =
-        !match_system(big_endian) ? swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xF800) == 0xD800) {
-      if (pos + 1 >= len) {
-        return result(error_code::SURROGATE, pos);
-      }
-      char16_t diff = char16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      char16_t next_word =
-          !match_system(big_endian) ? swap_bytes(data[pos + 1]) : data[pos + 1];
-      char16_t diff2 = uint16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      pos += 2;
-    } else {
-      pos++;
-    }
-  }
-  return result(error_code::SUCCESS, pos);
-}
-
-template <endianness big_endian>
-inline size_t count_code_points(const char16_t *p, size_t len) {
-  // We are not BOM aware.
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    char16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
-    counter += ((word & 0xFC00) != 0xDC00);
-  }
-  return counter;
-}
-
-template <endianness big_endian>
-inline size_t utf8_length_from_utf16(const char16_t *p, size_t len) {
-  // We are not BOM aware.
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    char16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
-    counter++; // ASCII
-    counter += static_cast<size_t>(
-        word >
-        0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes
-    counter += static_cast<size_t>((word > 0x7FF && word <= 0xD7FF) ||
-                                   (word >= 0xE000)); // three-byte
-  }
-  return counter;
-}
-
-template <endianness big_endian>
-inline size_t utf32_length_from_utf16(const char16_t *p, size_t len) {
-  // We are not BOM aware.
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    char16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
-    counter += ((word & 0xFC00) != 0xDC00);
-  }
-  return counter;
-}
-
-inline size_t latin1_length_from_utf16(size_t len) { return len; }
-
-simdutf_really_inline void
-change_endianness_utf16(const char16_t *input, size_t size, char16_t *output) {
-  for (size_t i = 0; i < size; i++) {
-    *output++ = char16_t(input[i] >> 8 | input[i] << 8);
-  }
-}
-
-template <endianness big_endian>
-simdutf_warn_unused inline size_t trim_partial_utf16(const char16_t *input,
-                                                     size_t length) {
-  if (length <= 1) {
-    return length;
-  }
-  uint16_t last_word = uint16_t(input[length - 1]);
-  last_word = !match_system(big_endian) ? swap_bytes(last_word) : last_word;
-  length -= ((last_word & 0xFC00) == 0xD800);
-  return length;
-}
-
-} // namespace utf16
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf16.h */
-/* begin file src/scalar/utf32.h */
-#ifndef SIMDUTF_UTF32_H
-#define SIMDUTF_UTF32_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf32 {
-
-inline simdutf_warn_unused bool validate(const char32_t *buf,
-                                         size_t len) noexcept {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  uint64_t pos = 0;
-  for (; pos < len; pos++) {
-    uint32_t word = data[pos];
-    if (word > 0x10FFFF || (word >= 0xD800 && word <= 0xDFFF)) {
-      return false;
-    }
-  }
-  return true;
-}
-
-inline simdutf_warn_unused result validate_with_errors(const char32_t *buf,
-                                                       size_t len) noexcept {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  for (; pos < len; pos++) {
-    uint32_t word = data[pos];
-    if (word > 0x10FFFF) {
-      return result(error_code::TOO_LARGE, pos);
-    }
-    if (word >= 0xD800 && word <= 0xDFFF) {
-      return result(error_code::SURROGATE, pos);
-    }
-  }
-  return result(error_code::SUCCESS, pos);
-}
-
-inline size_t utf8_length_from_utf32(const char32_t *buf, size_t len) {
-  // We are not BOM aware.
-  const uint32_t *p = reinterpret_cast<const uint32_t *>(buf);
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    // credit: @ttsugriy  for the vectorizable approach
-    counter++;                                     // ASCII
-    counter += static_cast<size_t>(p[i] > 0x7F);   // two-byte
-    counter += static_cast<size_t>(p[i] > 0x7FF);  // three-byte
-    counter += static_cast<size_t>(p[i] > 0xFFFF); // four-bytes
-  }
-  return counter;
-}
-
-inline size_t utf16_length_from_utf32(const char32_t *buf, size_t len) {
-  // We are not BOM aware.
-  const uint32_t *p = reinterpret_cast<const uint32_t *>(buf);
-  size_t counter{0};
-  for (size_t i = 0; i < len; i++) {
-    counter++;                                     // non-surrogate word
-    counter += static_cast<size_t>(p[i] > 0xFFFF); // surrogate pair
-  }
-  return counter;
-}
-
-inline size_t latin1_length_from_utf32(size_t len) {
-  // We are not BOM aware.
-  return len; // a utf32 codepoint will always represent 1 latin1 character
-}
-
-inline simdutf_warn_unused uint32_t swap_bytes(const uint32_t word) {
-  return ((word >> 24) & 0xff) |      // move byte 3 to byte 0
-         ((word << 8) & 0xff0000) |   // move byte 1 to byte 2
-         ((word >> 8) & 0xff00) |     // move byte 2 to byte 1
-         ((word << 24) & 0xff000000); // byte 0 to byte 3
-}
-
-} // namespace utf32
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf32.h */
-/* begin file src/scalar/base64.h */
-#ifndef SIMDUTF_BASE64_H
-#define SIMDUTF_BASE64_H
-
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <iostream>
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace base64 {
-
-// This function is not expected to be fast. Do not use in long loops.
-template <class char_type> bool is_ascii_white_space(char_type c) {
-  return c == ' ' || c == '\t' || c == '\n' || c == '\r' || c == '\f';
-}
-
-template <class char_type> bool is_ascii_white_space_or_padding(char_type c) {
-  return c == ' ' || c == '\t' || c == '\n' || c == '\r' || c == '\f' ||
-         c == '=';
-}
-
-template <class char_type> bool is_eight_byte(char_type c) {
-  if (sizeof(char_type) == 1) {
-    return true;
-  }
-  return uint8_t(c) == c;
-}
-
-// Returns true upon success. The destination buffer must be large enough.
-// This functions assumes that the padding (=) has been removed.
-template <class char_type>
-full_result
-base64_tail_decode(char *dst, const char_type *src, size_t length,
-                   size_t padded_characters, // number of padding characters
-                                             // '=', typically 0, 1, 2.
-                   base64_options options,
-                   last_chunk_handling_options last_chunk_options) {
-  // This looks like 5 branches, but we expect the compiler to resolve this to a
-  // single branch:
-  const uint8_t *to_base64 = (options & base64_url)
-                                 ? tables::base64::to_base64_url_value
-                                 : tables::base64::to_base64_value;
-  const uint32_t *d0 = (options & base64_url)
-                           ? tables::base64::base64_url::d0
-                           : tables::base64::base64_default::d0;
-  const uint32_t *d1 = (options & base64_url)
-                           ? tables::base64::base64_url::d1
-                           : tables::base64::base64_default::d1;
-  const uint32_t *d2 = (options & base64_url)
-                           ? tables::base64::base64_url::d2
-                           : tables::base64::base64_default::d2;
-  const uint32_t *d3 = (options & base64_url)
-                           ? tables::base64::base64_url::d3
-                           : tables::base64::base64_default::d3;
-
-  const char_type *srcend = src + length;
-  const char_type *srcinit = src;
-  const char *dstinit = dst;
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-
-  uint32_t x;
-  size_t idx;
-  uint8_t buffer[4];
-  while (true) {
-    while (src + 4 <= srcend && is_eight_byte(src[0]) &&
-           is_eight_byte(src[1]) && is_eight_byte(src[2]) &&
-           is_eight_byte(src[3]) &&
-           (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] |
-                d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) {
-      if (match_system(endianness::BIG)) {
-        x = scalar::utf32::swap_bytes(x);
-      }
-      std::memcpy(dst, &x, 3); // optimization opportunity: copy 4 bytes
-      dst += 3;
-      src += 4;
-    }
-    idx = 0;
-    // we need at least four characters.
-#ifdef __clang__
-    // If possible, we read four characters at a time. (It is an optimization.)
-    if (ignore_garbage && src + 4 <= srcend) {
-      char_type c0 = src[0];
-      char_type c1 = src[1];
-      char_type c2 = src[2];
-      char_type c3 = src[3];
-      uint8_t code0 = to_base64[uint8_t(c0)];
-      uint8_t code1 = to_base64[uint8_t(c1)];
-      uint8_t code2 = to_base64[uint8_t(c2)];
-      uint8_t code3 = to_base64[uint8_t(c3)];
-      buffer[idx] = code0;
-      idx += (is_eight_byte(c0) && code0 <= 63);
-      buffer[idx] = code1;
-      idx += (is_eight_byte(c1) && code1 <= 63);
-      buffer[idx] = code2;
-      idx += (is_eight_byte(c2) && code2 <= 63);
-      buffer[idx] = code3;
-      idx += (is_eight_byte(c3) && code3 <= 63);
-      src += 4;
-    }
-#endif
-    while ((idx < 4) && (src < srcend)) {
-      char_type c = *src;
-      uint8_t code = to_base64[uint8_t(c)];
-      buffer[idx] = uint8_t(code);
-      if (is_eight_byte(c) && code <= 63) {
-        idx++;
-      } else if (!ignore_garbage &&
-                 (code > 64 || !scalar::base64::is_eight_byte(c))) {
-        return {INVALID_BASE64_CHARACTER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      } else {
-        // We have a space or a newline or garbage. We ignore it.
-      }
-      src++;
-    }
-    if (idx != 4) {
-      if (!ignore_garbage &&
-          last_chunk_options == last_chunk_handling_options::strict &&
-          (idx != 1) && ((idx + padded_characters) & 3) != 0) {
-        // The partial chunk was at src - idx
-        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      } else if (!ignore_garbage &&
-                 last_chunk_options ==
-                     last_chunk_handling_options::stop_before_partial &&
-                 (idx != 1) && ((idx + padded_characters) & 3) != 0) {
-        // Rewind src to before partial chunk
-        src -= idx;
-        return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
-      } else {
-        if (idx == 2) {
-          uint32_t triple =
-              (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6);
-          if (!ignore_garbage &&
-              (last_chunk_options == last_chunk_handling_options::strict) &&
-              (triple & 0xffff)) {
-            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
-                    size_t(dst - dstinit)};
-          }
-          if (match_system(endianness::BIG)) {
-            triple <<= 8;
-            std::memcpy(dst, &triple, 1);
-          } else {
-            triple = scalar::utf32::swap_bytes(triple);
-            triple >>= 8;
-            std::memcpy(dst, &triple, 1);
-          }
-          dst += 1;
-        } else if (idx == 3) {
-          uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) +
-                            (uint32_t(buffer[1]) << 2 * 6) +
-                            (uint32_t(buffer[2]) << 1 * 6);
-          if (!ignore_garbage &&
-              (last_chunk_options == last_chunk_handling_options::strict) &&
-              (triple & 0xff)) {
-            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
-                    size_t(dst - dstinit)};
-          }
-          if (match_system(endianness::BIG)) {
-            triple <<= 8;
-            std::memcpy(dst, &triple, 2);
-          } else {
-            triple = scalar::utf32::swap_bytes(triple);
-            triple >>= 8;
-            std::memcpy(dst, &triple, 2);
-          }
-          dst += 2;
-        } else if (!ignore_garbage && idx == 1) {
-          return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
-                  size_t(dst - dstinit)};
-        }
-        return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
-      }
-    }
-
-    uint32_t triple =
-        (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) +
-        (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6);
-    if (match_system(endianness::BIG)) {
-      triple <<= 8;
-      std::memcpy(dst, &triple, 3);
-    } else {
-      triple = scalar::utf32::swap_bytes(triple);
-      triple >>= 8;
-      std::memcpy(dst, &triple, 3);
-    }
-    dst += 3;
-  }
-}
-
-// like base64_tail_decode, but it will not write past the end of the output
-// buffer. The outlen paramter is modified to reflect the number of bytes
-// written. This functions assumes that the padding (=) has been removed.
-template <class char_type>
-result base64_tail_decode_safe(
-    char *dst, size_t &outlen, const char_type *&srcr, size_t length,
-    size_t padded_characters, // number of padding characters '=', typically 0,
-                              // 1, 2.
-    base64_options options, last_chunk_handling_options last_chunk_options) {
-  const char_type *src = srcr;
-  if (length == 0) {
-    outlen = 0;
-    return {SUCCESS, 0};
-  }
-  // This looks like 5 branches, but we expect the compiler to resolve this to a
-  // single branch:
-  const uint8_t *to_base64 = (options & base64_url)
-                                 ? tables::base64::to_base64_url_value
-                                 : tables::base64::to_base64_value;
-  const uint32_t *d0 = (options & base64_url)
-                           ? tables::base64::base64_url::d0
-                           : tables::base64::base64_default::d0;
-  const uint32_t *d1 = (options & base64_url)
-                           ? tables::base64::base64_url::d1
-                           : tables::base64::base64_default::d1;
-  const uint32_t *d2 = (options & base64_url)
-                           ? tables::base64::base64_url::d2
-                           : tables::base64::base64_default::d2;
-  const uint32_t *d3 = (options & base64_url)
-                           ? tables::base64::base64_url::d3
-                           : tables::base64::base64_default::d3;
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-
-  const char_type *srcend = src + length;
-  const char_type *srcinit = src;
-  const char *dstinit = dst;
-  const char *dstend = dst + outlen;
-
-  uint32_t x;
-  size_t idx;
-  uint8_t buffer[4];
-  while (true) {
-    while (src + 4 <= srcend && is_eight_byte(src[0]) &&
-           is_eight_byte(src[1]) && is_eight_byte(src[2]) &&
-           is_eight_byte(src[3]) &&
-           (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] |
-                d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) {
-      if (dstend - dst < 3) {
-        outlen = size_t(dst - dstinit);
-        srcr = src;
-        return {OUTPUT_BUFFER_TOO_SMALL, size_t(src - srcinit)};
-      }
-      if (match_system(endianness::BIG)) {
-        x = scalar::utf32::swap_bytes(x);
-      }
-      std::memcpy(dst, &x, 3); // optimization opportunity: copy 4 bytes
-      dst += 3;
-      src += 4;
-    }
-    idx = 0;
-    const char_type *srccur = src;
-    // We need at least four characters.
-#ifdef __clang__
-    // If possible, we read four characters at a time. (It is an optimization.)
-    if (ignore_garbage && src + 4 <= srcend) {
-      char_type c0 = src[0];
-      char_type c1 = src[1];
-      char_type c2 = src[2];
-      char_type c3 = src[3];
-      uint8_t code0 = to_base64[uint8_t(c0)];
-      uint8_t code1 = to_base64[uint8_t(c1)];
-      uint8_t code2 = to_base64[uint8_t(c2)];
-      uint8_t code3 = to_base64[uint8_t(c3)];
-      buffer[idx] = code0;
-      idx += (is_eight_byte(c0) && code0 <= 63);
-      buffer[idx] = code1;
-      idx += (is_eight_byte(c1) && code1 <= 63);
-      buffer[idx] = code2;
-      idx += (is_eight_byte(c2) && code2 <= 63);
-      buffer[idx] = code3;
-      idx += (is_eight_byte(c3) && code3 <= 63);
-      src += 4;
-    }
-#endif
-    while (idx < 4 && src < srcend) {
-      char_type c = *src;
-      uint8_t code = to_base64[uint8_t(c)];
-
-      buffer[idx] = uint8_t(code);
-      if (is_eight_byte(c) && code <= 63) {
-        idx++;
-      } else if (!ignore_garbage &&
-                 (code > 64 || !scalar::base64::is_eight_byte(c))) {
-        outlen = size_t(dst - dstinit);
-        srcr = src;
-        return {INVALID_BASE64_CHARACTER, size_t(src - srcinit)};
-      } else {
-        // We have a space or a newline or garbage. We ignore it.
-      }
-      src++;
-    }
-    if (idx != 4) {
-      if (!ignore_garbage &&
-          last_chunk_options == last_chunk_handling_options::strict &&
-          ((idx + padded_characters) & 3) != 0) {
-        outlen = size_t(dst - dstinit);
-        srcr = src;
-        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit)};
-      } else if (!ignore_garbage &&
-                 last_chunk_options ==
-                     last_chunk_handling_options::stop_before_partial &&
-                 ((idx + padded_characters) & 3) != 0) {
-        // Rewind src to before partial chunk
-        srcr = srccur;
-        outlen = size_t(dst - dstinit);
-        return {SUCCESS, size_t(dst - dstinit)};
-      } else { // loose mode
-        if (idx == 0) {
-          // No data left; return success
-          outlen = size_t(dst - dstinit);
-          srcr = src;
-          return {SUCCESS, size_t(dst - dstinit)};
-        } else if (!ignore_garbage && idx == 1) {
-          // Error: Incomplete chunk of length 1 is invalid in loose mode
-          outlen = size_t(dst - dstinit);
-          srcr = src;
-          return {BASE64_INPUT_REMAINDER, size_t(src - srcinit)};
-        } else if (idx == 2 || idx == 3) {
-          // Check if there's enough space in the destination buffer
-          size_t required_space = (idx == 2) ? 1 : 2;
-          if (size_t(dstend - dst) < required_space) {
-            outlen = size_t(dst - dstinit);
-            srcr = src;
-            return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit)};
-          }
-          uint32_t triple = 0;
-          if (idx == 2) {
-            triple = (uint32_t(buffer[0]) << 18) + (uint32_t(buffer[1]) << 12);
-            if (!ignore_garbage &&
-                (last_chunk_options == last_chunk_handling_options::strict) &&
-                (triple & 0xffff)) {
-              srcr = src;
-              return {BASE64_EXTRA_BITS, size_t(src - srcinit)};
-            }
-            // Extract the first byte
-            triple >>= 16;
-            dst[0] = static_cast<char>(triple & 0xFF);
-            dst += 1;
-          } else if (idx == 3) {
-            triple = (uint32_t(buffer[0]) << 18) + (uint32_t(buffer[1]) << 12) +
-                     (uint32_t(buffer[2]) << 6);
-            if (!ignore_garbage &&
-                (last_chunk_options == last_chunk_handling_options::strict) &&
-                (triple & 0xff)) {
-              srcr = src;
-              return {BASE64_EXTRA_BITS, size_t(src - srcinit)};
-            }
-            // Extract the first two bytes
-            triple >>= 8;
-            dst[0] = static_cast<char>((triple >> 8) & 0xFF);
-            dst[1] = static_cast<char>(triple & 0xFF);
-            dst += 2;
-          }
-          outlen = size_t(dst - dstinit);
-          srcr = src;
-          return {SUCCESS, size_t(dst - dstinit)};
-        }
-      }
-    }
-
-    if (dstend - dst < 3) {
-      outlen = size_t(dst - dstinit);
-      srcr = src;
-      return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit)};
-    }
-    uint32_t triple = (uint32_t(buffer[0]) << 18) +
-                      (uint32_t(buffer[1]) << 12) + (uint32_t(buffer[2]) << 6) +
-                      (uint32_t(buffer[3]));
-    if (match_system(endianness::BIG)) {
-      triple <<= 8;
-      std::memcpy(dst, &triple, 3);
-    } else {
-      triple = scalar::utf32::swap_bytes(triple);
-      triple >>= 8;
-      std::memcpy(dst, &triple, 3);
-    }
-    dst += 3;
-  }
-}
-
-// Returns the number of bytes written. The destination buffer must be large
-// enough. It will add padding (=) if needed.
-size_t tail_encode_base64(char *dst, const char *src, size_t srclen,
-                          base64_options options) {
-  // By default, we use padding if we are not using the URL variant.
-  // This is check with ((options & base64_url) == 0) which returns true if we
-  // are not using the URL variant. However, we also allow 'inversion' of the
-  // convention with the base64_reverse_padding option. If the
-  // base64_reverse_padding option is set, we use padding if we are using the
-  // URL variant, and we omit it if we are not using the URL variant. This is
-  // checked with
-  // ((options & base64_reverse_padding) == base64_reverse_padding).
-  bool use_padding =
-      ((options & base64_url) == 0) ^
-      ((options & base64_reverse_padding) == base64_reverse_padding);
-  // This looks like 3 branches, but we expect the compiler to resolve this to
-  // a single branch:
-  const char *e0 = (options & base64_url) ? tables::base64::base64_url::e0
-                                          : tables::base64::base64_default::e0;
-  const char *e1 = (options & base64_url) ? tables::base64::base64_url::e1
-                                          : tables::base64::base64_default::e1;
-  const char *e2 = (options & base64_url) ? tables::base64::base64_url::e2
-                                          : tables::base64::base64_default::e2;
-  char *out = dst;
-  size_t i = 0;
-  uint8_t t1, t2, t3;
-  for (; i + 2 < srclen; i += 3) {
-    t1 = uint8_t(src[i]);
-    t2 = uint8_t(src[i + 1]);
-    t3 = uint8_t(src[i + 2]);
-    *out++ = e0[t1];
-    *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)];
-    *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)];
-    *out++ = e2[t3];
-  }
-  switch (srclen - i) {
-  case 0:
-    break;
-  case 1:
-    t1 = uint8_t(src[i]);
-    *out++ = e0[t1];
-    *out++ = e1[(t1 & 0x03) << 4];
-    if (use_padding) {
-      *out++ = '=';
-      *out++ = '=';
-    }
-    break;
-  default: /* case 2 */
-    t1 = uint8_t(src[i]);
-    t2 = uint8_t(src[i + 1]);
-    *out++ = e0[t1];
-    *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)];
-    *out++ = e2[(t2 & 0x0F) << 2];
-    if (use_padding) {
-      *out++ = '=';
-    }
-  }
-  return (size_t)(out - dst);
-}
-
-template <class char_type>
-simdutf_warn_unused size_t maximal_binary_length_from_base64(
-    const char_type *input, size_t length) noexcept {
-  // We follow https://infra.spec.whatwg.org/#forgiving-base64-decode
-  size_t padding = 0;
-  if (length > 0) {
-    if (input[length - 1] == '=') {
-      padding++;
-      if (length > 1 && input[length - 2] == '=') {
-        padding++;
-      }
-    }
-  }
-  size_t actual_length = length - padding;
-  if (actual_length % 4 <= 1) {
-    return actual_length / 4 * 3;
-  }
-  // if we have a valid input, then the remainder must be 2 or 3 adding one or
-  // two extra bytes.
-  return actual_length / 4 * 3 + (actual_length % 4) - 1;
-}
-
-simdutf_warn_unused size_t
-base64_length_from_binary(size_t length, base64_options options) noexcept {
-  // By default, we use padding if we are not using the URL variant.
-  // This is check with ((options & base64_url) == 0) which returns true if we
-  // are not using the URL variant. However, we also allow 'inversion' of the
-  // convention with the base64_reverse_padding option. If the
-  // base64_reverse_padding option is set, we use padding if we are using the
-  // URL variant, and we omit it if we are not using the URL variant. This is
-  // checked with
-  // ((options & base64_reverse_padding) == base64_reverse_padding).
-  bool use_padding =
-      ((options & base64_url) == 0) ^
-      ((options & base64_reverse_padding) == base64_reverse_padding);
-  if (!use_padding) {
-    return length / 3 * 4 + ((length % 3) ? (length % 3) + 1 : 0);
-  }
-  return (length + 2) / 3 *
-         4; // We use padding to make the length a multiple of 4.
-}
-
-} // namespace base64
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/base64.h */
-/* begin file src/scalar/latin1_to_utf8/latin1_to_utf8.h */
-#ifndef SIMDUTF_LATIN1_TO_UTF8_H
-#define SIMDUTF_LATIN1_TO_UTF8_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace latin1_to_utf8 {
-
-inline size_t convert(const char *buf, size_t len, char *utf8_output) {
-  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
-  size_t pos = 0;
-  size_t utf8_pos = 0;
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 |
-                 v2}; // We are only interested in these bits: 1000 1000 1000
-                      // 1000, so it makes sense to concatenate everything
-      if ((v & 0x8080808080808080) ==
-          0) { // if NONE of these are set, e.g. all of them are zero, then
-               // everything is ASCII
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          utf8_output[utf8_pos++] = char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-
-    unsigned char byte = data[pos];
-    if ((byte & 0x80) == 0) { // if ASCII
-      // will generate one UTF-8 bytes
-      utf8_output[utf8_pos++] = char(byte);
-      pos++;
-    } else {
-      // will generate two UTF-8 bytes
-      utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000);
-      utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000);
-      pos++;
-    }
-  }
-  return utf8_pos;
-}
-
-inline size_t convert_safe(const char *buf, size_t len, char *utf8_output,
-                           size_t utf8_len) {
-  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
-  size_t pos = 0;
-  size_t skip_pos = 0;
-  size_t utf8_pos = 0;
-  while (pos < len && utf8_pos < utf8_len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos >= skip_pos && pos + 16 <= len &&
-        utf8_pos + 16 <= utf8_len) { // if it is safe to read 16 more bytes,
-                                     // check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 |
-                 v2}; // We are only interested in these bits: 1000 1000 1000
-                      // 1000, so it makes sense to concatenate everything
-      if ((v & 0x8080808080808080) ==
-          0) { // if NONE of these are set, e.g. all of them are zero, then
-               // everything is ASCII
-        ::memcpy(utf8_output + utf8_pos, buf + pos, 16);
-        utf8_pos += 16;
-        pos += 16;
-      } else {
-        // At least one of the next 16 bytes are not ASCII, we will process them
-        // one by one
-        skip_pos = pos + 16;
-      }
-    } else {
-      const auto byte = data[pos];
-      if ((byte & 0x80) == 0) { // if ASCII
-        // will generate one UTF-8 bytes
-        utf8_output[utf8_pos++] = char(byte);
-        pos++;
-      } else if (utf8_pos + 2 <= utf8_len) {
-        // will generate two UTF-8 bytes
-        utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000);
-        utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000);
-        pos++;
-      } else {
-        break;
-      }
-    }
-  }
-  return utf8_pos;
-}
-
-} // namespace latin1_to_utf8
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/latin1_to_utf8/latin1_to_utf8.h */
-
-namespace simdutf {
-bool implementation::supported_by_runtime_system() const {
-  uint32_t required_instruction_sets = this->required_instruction_sets();
-  uint32_t supported_instruction_sets =
-      internal::detect_supported_architectures();
-  return ((supported_instruction_sets & required_instruction_sets) ==
-          required_instruction_sets);
-}
-
-simdutf_warn_unused encoding_type implementation::autodetect_encoding(
-    const char *input, size_t length) const noexcept {
-  // If there is a BOM, then we trust it.
-  auto bom_encoding = simdutf::BOM::check_bom(input, length);
-  if (bom_encoding != encoding_type::unspecified) {
-    return bom_encoding;
-  }
-  // UTF8 is common, it includes ASCII, and is commonly represented
-  // without a BOM, so if it fits, go with that. Note that it is still
-  // possible to get it wrong, we are only 'guessing'. If some has UTF-16
-  // data without a BOM, it could pass as UTF-8.
-  //
-  // An interesting twist might be to check for UTF-16 ASCII first (every
-  // other byte is zero).
-  if (validate_utf8(input, length)) {
-    return encoding_type::UTF8;
-  }
-  // The next most common encoding that might appear without BOM is probably
-  // UTF-16LE, so try that next.
-  if ((length % 2) == 0) {
-    // important: we need to divide by two
-    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
-                         length / 2)) {
-      return encoding_type::UTF16_LE;
-    }
-  }
-  if ((length % 4) == 0) {
-    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
-      return encoding_type::UTF32_LE;
-    }
-  }
-  return encoding_type::unspecified;
-}
-
-namespace internal {
-// When there is a single implementation, we should not pay a price
-// for dispatching to the best implementation. We should just use the
-// one we have. This is a compile-time check.
-#define SIMDUTF_SINGLE_IMPLEMENTATION                                          \
-  (SIMDUTF_IMPLEMENTATION_ICELAKE + SIMDUTF_IMPLEMENTATION_HASWELL +           \
-       SIMDUTF_IMPLEMENTATION_WESTMERE + SIMDUTF_IMPLEMENTATION_ARM64 +        \
-       SIMDUTF_IMPLEMENTATION_PPC64 + SIMDUTF_IMPLEMENTATION_LSX +             \
-       SIMDUTF_IMPLEMENTATION_LASX + SIMDUTF_IMPLEMENTATION_FALLBACK ==        \
-   1)
-
-// Static array of known implementations. We are hoping these get baked into the
-// executable without requiring a static initializer.
-
-#if SIMDUTF_IMPLEMENTATION_ICELAKE
-static const icelake::implementation *get_icelake_singleton() {
-  static const icelake::implementation icelake_singleton{};
-  return &icelake_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_HASWELL
-static const haswell::implementation *get_haswell_singleton() {
-  static const haswell::implementation haswell_singleton{};
-  return &haswell_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_WESTMERE
-static const westmere::implementation *get_westmere_singleton() {
-  static const westmere::implementation westmere_singleton{};
-  return &westmere_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_ARM64
-static const arm64::implementation *get_arm64_singleton() {
-  static const arm64::implementation arm64_singleton{};
-  return &arm64_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_PPC64
-static const ppc64::implementation *get_ppc64_singleton() {
-  static const ppc64::implementation ppc64_singleton{};
-  return &ppc64_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_RVV
-static const rvv::implementation *get_rvv_singleton() {
-  static const rvv::implementation rvv_singleton{};
-  return &rvv_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_LSX
-static const lsx::implementation *get_lsx_singleton() {
-  static const lsx::implementation lsx_singleton{};
-  return &lsx_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_LASX
-static const lasx::implementation *get_lasx_singleton() {
-  static const lasx::implementation lasx_singleton{};
-  return &lasx_singleton;
-}
-#endif
-#if SIMDUTF_IMPLEMENTATION_FALLBACK
-static const fallback::implementation *get_fallback_singleton() {
-  static const fallback::implementation fallback_singleton{};
-  return &fallback_singleton;
-}
-#endif
-
-#if SIMDUTF_SINGLE_IMPLEMENTATION
-static const implementation *get_single_implementation() {
-  return
-  #if SIMDUTF_IMPLEMENTATION_ICELAKE
-      get_icelake_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_HASWELL
-  get_haswell_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_WESTMERE
-  get_westmere_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_ARM64
-  get_arm64_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_PPC64
-  get_ppc64_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_LSX
-  get_lsx_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_LASX
-  get_lasx_singleton();
-  #endif
-  #if SIMDUTF_IMPLEMENTATION_FALLBACK
-  get_fallback_singleton();
-  #endif
-}
-#endif
-
-/**
- * @private Detects best supported implementation on first use, and sets it
- */
-class detect_best_supported_implementation_on_first_use final
-    : public implementation {
-public:
-  std::string name() const noexcept final { return set_best()->name(); }
-  std::string description() const noexcept final {
-    return set_best()->description();
-  }
-  uint32_t required_instruction_sets() const noexcept final {
-    return set_best()->required_instruction_sets();
-  }
-
-  simdutf_warn_unused int
-  detect_encodings(const char *input, size_t length) const noexcept override {
-    return set_best()->detect_encodings(input, length);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf8(const char *buf, size_t len) const noexcept final override {
-    return set_best()->validate_utf8(buf, len);
-  }
-
-  simdutf_warn_unused result validate_utf8_with_errors(
-      const char *buf, size_t len) const noexcept final override {
-    return set_best()->validate_utf8_with_errors(buf, len);
-  }
-
-  simdutf_warn_unused bool
-  validate_ascii(const char *buf, size_t len) const noexcept final override {
-    return set_best()->validate_ascii(buf, len);
-  }
-
-  simdutf_warn_unused result validate_ascii_with_errors(
-      const char *buf, size_t len) const noexcept final override {
-    return set_best()->validate_ascii_with_errors(buf, len);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf16le(const char16_t *buf,
-                   size_t len) const noexcept final override {
-    return set_best()->validate_utf16le(buf, len);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf16be(const char16_t *buf,
-                   size_t len) const noexcept final override {
-    return set_best()->validate_utf16be(buf, len);
-  }
-
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *buf, size_t len) const noexcept final override {
-    return set_best()->validate_utf16le_with_errors(buf, len);
-  }
-
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *buf, size_t len) const noexcept final override {
-    return set_best()->validate_utf16be_with_errors(buf, len);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf32(const char32_t *buf,
-                 size_t len) const noexcept final override {
-    return set_best()->validate_utf32(buf, len);
-  }
-
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *buf, size_t len) const noexcept final override {
-    return set_best()->validate_utf32_with_errors(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  convert_latin1_to_utf8(const char *buf, size_t len,
-                         char *utf8_output) const noexcept final override {
-    return set_best()->convert_latin1_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_latin1_to_utf16le(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_latin1_to_utf16be(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *buf, size_t len,
-      char32_t *latin1_output) const noexcept final override {
-    return set_best()->convert_latin1_to_utf32(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf8_to_latin1(const char *buf, size_t len,
-                         char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf8_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf8_to_latin1_with_errors(buf, len,
-                                                          latin1_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_valid_utf8_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf16le(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf16be(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf16le_with_errors(buf, len,
-                                                           utf16_output);
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf16be_with_errors(buf, len,
-                                                           utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_valid_utf8_to_utf16le(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_valid_utf8_to_utf16be(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf8_to_utf32(const char *buf, size_t len,
-                        char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf32(buf, len, utf32_output);
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf8_to_utf32_with_errors(buf, len,
-                                                         utf32_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_valid_utf8_to_utf32(buf, len, utf32_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_latin1_with_errors(buf, len,
-                                                             latin1_output);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_latin1_with_errors(buf, len,
-                                                             latin1_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
-      const char16_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_valid_utf16le_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
-      const char16_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_valid_utf16be_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf16le_to_utf8(const char16_t *buf, size_t len,
-                          char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf16be_to_utf8(const char16_t *buf, size_t len,
-                          char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *buf, size_t len,
-      char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_utf8_with_errors(buf, len,
-                                                           utf8_output);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *buf, size_t len,
-      char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_utf8_with_errors(buf, len,
-                                                           utf8_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *buf, size_t len,
-      char *utf8_output) const noexcept final override {
-    return set_best()->convert_valid_utf16le_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *buf, size_t len,
-      char *utf8_output) const noexcept final override {
-    return set_best()->convert_valid_utf16be_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                          char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf32_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused result convert_utf32_to_latin1_with_errors(
-      const char32_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf32_to_latin1_with_errors(buf, len,
-                                                           latin1_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
-      const char32_t *buf, size_t len,
-      char *latin1_output) const noexcept final override {
-    return set_best()->convert_utf32_to_latin1(buf, len, latin1_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_utf32_to_utf8(const char32_t *buf, size_t len,
-                        char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *buf, size_t len,
-      char *utf8_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t
-  convert_valid_utf32_to_utf8(const char32_t *buf, size_t len,
-                              char *utf8_output) const noexcept final override {
-    return set_best()->convert_valid_utf32_to_utf8(buf, len, utf8_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_utf16le(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf16le(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_utf16be(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf16be(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf16le_with_errors(buf, len,
-                                                            utf16_output);
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_utf32_to_utf16be_with_errors(buf, len,
-                                                            utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_valid_utf32_to_utf16le(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
-      const char32_t *buf, size_t len,
-      char16_t *utf16_output) const noexcept final override {
-    return set_best()->convert_valid_utf32_to_utf16be(buf, len, utf16_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf16le_to_utf32(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_utf32(buf, len, utf32_output);
-  }
-
-  simdutf_warn_unused size_t convert_utf16be_to_utf32(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_utf32(buf, len, utf32_output);
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf16le_to_utf32_with_errors(buf, len,
-                                                            utf32_output);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_utf16be_to_utf32_with_errors(buf, len,
-                                                            utf32_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_valid_utf16le_to_utf32(buf, len, utf32_output);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
-      const char16_t *buf, size_t len,
-      char32_t *utf32_output) const noexcept final override {
-    return set_best()->convert_valid_utf16be_to_utf32(buf, len, utf32_output);
-  }
-
-  void change_endianness_utf16(const char16_t *buf, size_t len,
-                               char16_t *output) const noexcept final override {
-    set_best()->change_endianness_utf16(buf, len, output);
-  }
-
-  simdutf_warn_unused size_t
-  count_utf16le(const char16_t *buf, size_t len) const noexcept final override {
-    return set_best()->count_utf16le(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  count_utf16be(const char16_t *buf, size_t len) const noexcept final override {
-    return set_best()->count_utf16be(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  count_utf8(const char *buf, size_t len) const noexcept final override {
-    return set_best()->count_utf8(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *buf, size_t len) const noexcept override {
-    return set_best()->latin1_length_from_utf8(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t len) const noexcept override {
-    return set_best()->latin1_length_from_utf16(len);
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t len) const noexcept override {
-    return set_best()->latin1_length_from_utf32(len);
-  }
-
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *buf, size_t len) const noexcept override {
-    return set_best()->utf8_length_from_latin1(buf, len);
-  }
-
-  simdutf_warn_unused size_t utf8_length_from_utf16le(
-      const char16_t *buf, size_t len) const noexcept override {
-    return set_best()->utf8_length_from_utf16le(buf, len);
-  }
-
-  simdutf_warn_unused size_t utf8_length_from_utf16be(
-      const char16_t *buf, size_t len) const noexcept override {
-    return set_best()->utf8_length_from_utf16be(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t len) const noexcept override {
-    return set_best()->utf16_length_from_latin1(len);
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t len) const noexcept override {
-    return set_best()->utf32_length_from_latin1(len);
-  }
-
-  simdutf_warn_unused size_t utf32_length_from_utf16le(
-      const char16_t *buf, size_t len) const noexcept override {
-    return set_best()->utf32_length_from_utf16le(buf, len);
-  }
-
-  simdutf_warn_unused size_t utf32_length_from_utf16be(
-      const char16_t *buf, size_t len) const noexcept override {
-    return set_best()->utf32_length_from_utf16be(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *buf, size_t len) const noexcept override {
-    return set_best()->utf16_length_from_utf8(buf, len);
-  }
-
-  simdutf_warn_unused size_t utf8_length_from_utf32(
-      const char32_t *buf, size_t len) const noexcept override {
-    return set_best()->utf8_length_from_utf32(buf, len);
-  }
-
-  simdutf_warn_unused size_t utf16_length_from_utf32(
-      const char32_t *buf, size_t len) const noexcept override {
-    return set_best()->utf16_length_from_utf32(buf, len);
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *buf, size_t len) const noexcept override {
-    return set_best()->utf32_length_from_utf8(buf, len);
-  }
-
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *input, size_t length) const noexcept override {
-    return set_best()->maximal_binary_length_from_base64(input, length);
-  }
-
-  simdutf_warn_unused result base64_to_binary(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_handling_options =
-          last_chunk_handling_options::loose) const noexcept override {
-    return set_best()->base64_to_binary(input, length, output, options,
-                                        last_chunk_handling_options);
-  }
-
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *input, size_t length, char *output, base64_options options,
-      last_chunk_handling_options last_chunk_handling_options =
-          last_chunk_handling_options::loose) const noexcept override {
-    return set_best()->base64_to_binary_details(input, length, output, options,
-                                                last_chunk_handling_options);
-  }
-
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *input, size_t length) const noexcept override {
-    return set_best()->maximal_binary_length_from_base64(input, length);
-  }
-
-  simdutf_warn_unused result base64_to_binary(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_handling_options =
-          last_chunk_handling_options::loose) const noexcept override {
-    return set_best()->base64_to_binary(input, length, output, options,
-                                        last_chunk_handling_options);
-  }
-
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *input, size_t length, char *output,
-      base64_options options,
-      last_chunk_handling_options last_chunk_handling_options =
-          last_chunk_handling_options::loose) const noexcept override {
-    return set_best()->base64_to_binary_details(input, length, output, options,
-                                                last_chunk_handling_options);
-  }
-
-  simdutf_warn_unused size_t base64_length_from_binary(
-      size_t length, base64_options options) const noexcept override {
-    return set_best()->base64_length_from_binary(length, options);
-  }
-
-  size_t binary_to_base64(const char *input, size_t length, char *output,
-                          base64_options options) const noexcept override {
-    return set_best()->binary_to_base64(input, length, output, options);
-  }
-
-  simdutf_really_inline
-  detect_best_supported_implementation_on_first_use() noexcept
-      : implementation("best_supported_detector",
-                       "Detects the best supported implementation and sets it",
-                       0) {}
-
-private:
-  const implementation *set_best() const noexcept;
-};
-
-static_assert(std::is_trivially_destructible<
-                  detect_best_supported_implementation_on_first_use>::value,
-              "detect_best_supported_implementation_on_first_use should be "
-              "trivially destructible");
-
-static const std::initializer_list<const implementation *> &
-get_available_implementation_pointers() {
-  static const std::initializer_list<const implementation *>
-      available_implementation_pointers{
-#if SIMDUTF_IMPLEMENTATION_ICELAKE
-          get_icelake_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_HASWELL
-          get_haswell_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_WESTMERE
-          get_westmere_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_ARM64
-          get_arm64_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_PPC64
-          get_ppc64_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_RVV
-          get_rvv_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_LSX
-          get_lsx_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_LASX
-          get_lasx_singleton(),
-#endif
-#if SIMDUTF_IMPLEMENTATION_FALLBACK
-          get_fallback_singleton(),
-#endif
-      }; // available_implementation_pointers
-  return available_implementation_pointers;
-}
-
-// So we can return UNSUPPORTED_ARCHITECTURE from the parser when there is no
-// support
-class unsupported_implementation final : public implementation {
-public:
-  simdutf_warn_unused int detect_encodings(const char *,
-                                           size_t) const noexcept override {
-    return encoding_type::unspecified;
-  }
-
-  simdutf_warn_unused bool validate_utf8(const char *,
-                                         size_t) const noexcept final override {
-    return false; // Just refuse to validate. Given that we have a fallback
-                  // implementation
-    // it seems unlikely that unsupported_implementation will ever be used. If
-    // it is used, then it will flag all strings as invalid. The alternative is
-    // to return an error_code from which the user has to figure out whether the
-    // string is valid UTF-8... which seems like a lot of work just to handle
-    // the very unlikely case that we have an unsupported implementation. And,
-    // when it does happen (that we have an unsupported implementation), what
-    // are the chances that the programmer has a fallback? Given that *we*
-    // provide the fallback, it implies that the programmer would need a
-    // fallback for our fallback.
-  }
-
-  simdutf_warn_unused result validate_utf8_with_errors(
-      const char *, size_t) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused bool
-  validate_ascii(const char *, size_t) const noexcept final override {
-    return false;
-  }
-
-  simdutf_warn_unused result validate_ascii_with_errors(
-      const char *, size_t) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf16le(const char16_t *, size_t) const noexcept final override {
-    return false;
-  }
-
-  simdutf_warn_unused bool
-  validate_utf16be(const char16_t *, size_t) const noexcept final override {
-    return false;
-  }
-
-  simdutf_warn_unused result validate_utf16le_with_errors(
-      const char16_t *, size_t) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result validate_utf16be_with_errors(
-      const char16_t *, size_t) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused bool
-  validate_utf32(const char32_t *, size_t) const noexcept final override {
-    return false;
-  }
-
-  simdutf_warn_unused result validate_utf32_with_errors(
-      const char32_t *, size_t) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf8(
-      const char *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf16le(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf16be(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_latin1_to_utf32(
-      const char *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_latin1(
-      const char *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-      const char *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-      const char *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_utf16le(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_utf16be(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-      const char *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf8_to_utf32(
-      const char *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-      const char *, size_t, char32_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-      const char *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16le_to_latin1(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16be_to_latin1(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16le_to_utf8(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16be_to_utf8(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-      const char16_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_latin1(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf32_to_latin1_with_errors(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_utf8(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-      const char32_t *, size_t, char *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_utf16le(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf32_to_utf16be(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
-      const char32_t *, size_t, char16_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16le_to_utf32(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_utf16be_to_utf32(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
-      const char16_t *, size_t, char32_t *) const noexcept final override {
-    return 0;
-  }
-
-  void change_endianness_utf16(const char16_t *, size_t,
-                               char16_t *) const noexcept final override {}
-
-  simdutf_warn_unused size_t
-  count_utf16le(const char16_t *, size_t) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  count_utf16be(const char16_t *, size_t) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t count_utf8(const char *,
-                                        size_t) const noexcept final override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf8(const char *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf16(size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  latin1_length_from_utf32(size_t) const noexcept override {
-    return 0;
-  }
-  simdutf_warn_unused size_t
-  utf8_length_from_latin1(const char *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16le(const char16_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf8_length_from_utf16be(const char16_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_utf16le(const char16_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_utf16be(const char16_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_latin1(size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf16_length_from_utf8(const char *, size_t) const noexcept override {
-    return 0;
-  }
-  simdutf_warn_unused size_t
-  utf16_length_from_latin1(size_t) const noexcept override {
-    return 0;
-  }
-  simdutf_warn_unused size_t
-  utf8_length_from_utf32(const char32_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf16_length_from_utf32(const char32_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t
-  utf32_length_from_utf8(const char *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused result
-  base64_to_binary(const char *, size_t, char *, base64_options,
-                   last_chunk_handling_options) const noexcept override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char *, size_t, char *, base64_options,
-      last_chunk_handling_options) const noexcept override {
-    return full_result(error_code::OTHER, 0, 0);
-  }
-
-  simdutf_warn_unused size_t maximal_binary_length_from_base64(
-      const char16_t *, size_t) const noexcept override {
-    return 0;
-  }
-
-  simdutf_warn_unused result
-  base64_to_binary(const char16_t *, size_t, char *, base64_options,
-                   last_chunk_handling_options) const noexcept override {
-    return result(error_code::OTHER, 0);
-  }
-
-  simdutf_warn_unused full_result base64_to_binary_details(
-      const char16_t *, size_t, char *, base64_options,
-      last_chunk_handling_options) const noexcept override {
-    return full_result(error_code::OTHER, 0, 0);
-  }
-
-  simdutf_warn_unused size_t
-  base64_length_from_binary(size_t, base64_options) const noexcept override {
-    return 0;
-  }
-
-  size_t binary_to_base64(const char *, size_t, char *,
-                          base64_options) const noexcept override {
-    return 0;
-  }
-
-  unsupported_implementation()
-      : implementation("unsupported",
-                       "Unsupported CPU (no detected SIMD instructions)", 0) {}
-};
-
-const unsupported_implementation *get_unsupported_singleton() {
-  static const unsupported_implementation unsupported_singleton{};
-  return &unsupported_singleton;
-}
-static_assert(std::is_trivially_destructible<unsupported_implementation>::value,
-              "unsupported_singleton should be trivially destructible");
-
-size_t available_implementation_list::size() const noexcept {
-  return internal::get_available_implementation_pointers().size();
-}
-const implementation *const *
-available_implementation_list::begin() const noexcept {
-  return internal::get_available_implementation_pointers().begin();
-}
-const implementation *const *
-available_implementation_list::end() const noexcept {
-  return internal::get_available_implementation_pointers().end();
-}
-const implementation *
-available_implementation_list::detect_best_supported() const noexcept {
-  // They are prelisted in priority order, so we just go down the list
-  uint32_t supported_instruction_sets =
-      internal::detect_supported_architectures();
-  for (const implementation *impl :
-       internal::get_available_implementation_pointers()) {
-    uint32_t required_instruction_sets = impl->required_instruction_sets();
-    if ((supported_instruction_sets & required_instruction_sets) ==
-        required_instruction_sets) {
-      return impl;
-    }
-  }
-  return get_unsupported_singleton(); // this should never happen?
-}
-
-const implementation *
-detect_best_supported_implementation_on_first_use::set_best() const noexcept {
-  SIMDUTF_PUSH_DISABLE_WARNINGS
-  SIMDUTF_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC:
-                                     // manually verified this is safe
-      char *force_implementation_name = getenv("SIMDUTF_FORCE_IMPLEMENTATION");
-  SIMDUTF_POP_DISABLE_WARNINGS
-
-  if (force_implementation_name) {
-    auto force_implementation =
-        get_available_implementations()[force_implementation_name];
-    if (force_implementation) {
-      return get_active_implementation() = force_implementation;
-    } else {
-      // Note: abort() and stderr usage within the library is forbidden.
-      return get_active_implementation() = get_unsupported_singleton();
-    }
-  }
-  return get_active_implementation() =
-             get_available_implementations().detect_best_supported();
-}
-
-} // namespace internal
-
-/**
- * The list of available implementations compiled into simdutf.
- */
-SIMDUTF_DLLIMPORTEXPORT const internal::available_implementation_list &
-get_available_implementations() {
-  static const internal::available_implementation_list
-      available_implementations{};
-  return available_implementations;
-}
-
-/**
- * The active implementation.
- */
-SIMDUTF_DLLIMPORTEXPORT internal::atomic_ptr<const implementation> &
-get_active_implementation() {
-#if SIMDUTF_SINGLE_IMPLEMENTATION
-  // skip runtime detection
-  static internal::atomic_ptr<const implementation> active_implementation{
-      internal::get_single_implementation()};
-  return active_implementation;
-#else
-  static const internal::detect_best_supported_implementation_on_first_use
-      detect_best_supported_implementation_on_first_use_singleton;
-  static internal::atomic_ptr<const implementation> active_implementation{
-      &detect_best_supported_implementation_on_first_use_singleton};
-  return active_implementation;
-#endif
-}
-
-#if SIMDUTF_SINGLE_IMPLEMENTATION
-const implementation *get_default_implementation() {
-  return internal::get_single_implementation();
-}
-#else
-internal::atomic_ptr<const implementation> &get_default_implementation() {
-  return get_active_implementation();
-}
-#endif
-#define SIMDUTF_GET_CURRENT_IMPLEMENTION
-
-simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept {
-  return get_default_implementation()->validate_utf8(buf, len);
-}
-simdutf_warn_unused result validate_utf8_with_errors(const char *buf,
-                                                     size_t len) noexcept {
-  return get_default_implementation()->validate_utf8_with_errors(buf, len);
-}
-simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) noexcept {
-  return get_default_implementation()->validate_ascii(buf, len);
-}
-simdutf_warn_unused result validate_ascii_with_errors(const char *buf,
-                                                      size_t len) noexcept {
-  return get_default_implementation()->validate_ascii_with_errors(buf, len);
-}
-simdutf_warn_unused size_t convert_utf8_to_utf16(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf8_to_utf16be(input, length, utf16_output);
-#else
-  return convert_utf8_to_utf16le(input, length, utf16_output);
-#endif
-}
-simdutf_warn_unused size_t convert_latin1_to_utf8(const char *buf, size_t len,
-                                                  char *utf8_output) noexcept {
-  return get_default_implementation()->convert_latin1_to_utf8(buf, len,
-                                                              utf8_output);
-}
-simdutf_warn_unused size_t convert_latin1_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_latin1_to_utf16le(buf, len,
-                                                                 utf16_output);
-}
-simdutf_warn_unused size_t convert_latin1_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_latin1_to_utf16be(buf, len,
-                                                                 utf16_output);
-}
-simdutf_warn_unused size_t convert_latin1_to_utf32(
-    const char *buf, size_t len, char32_t *latin1_output) noexcept {
-  return get_default_implementation()->convert_latin1_to_utf32(buf, len,
-                                                               latin1_output);
-}
-simdutf_warn_unused size_t convert_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_latin1(buf, len,
-                                                              latin1_output);
-}
-simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
-    const char *buf, size_t len, char *latin1_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_latin1_with_errors(
-      buf, len, latin1_output);
-}
-simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) noexcept {
-  return get_default_implementation()->convert_valid_utf8_to_latin1(
-      buf, len, latin1_output);
-}
-simdutf_warn_unused size_t convert_utf8_to_utf16le(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf16le(input, length,
-                                                               utf16_output);
-}
-simdutf_warn_unused size_t convert_utf8_to_utf16be(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf16be(input, length,
-                                                               utf16_output);
-}
-simdutf_warn_unused result convert_utf8_to_utf16_with_errors(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf8_to_utf16be_with_errors(input, length, utf16_output);
-#else
-  return convert_utf8_to_utf16le_with_errors(input, length, utf16_output);
-#endif
-}
-simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf16le_with_errors(
-      input, length, utf16_output);
-}
-simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
-    const char *input, size_t length, char16_t *utf16_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf16be_with_errors(
-      input, length, utf16_output);
-}
-simdutf_warn_unused size_t convert_utf8_to_utf32(
-    const char *input, size_t length, char32_t *utf32_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf32(input, length,
-                                                             utf32_output);
-}
-simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
-    const char *input, size_t length, char32_t *utf32_output) noexcept {
-  return get_default_implementation()->convert_utf8_to_utf32_with_errors(
-      input, length, utf32_output);
-}
-simdutf_warn_unused bool validate_utf16(const char16_t *buf,
-                                        size_t len) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return validate_utf16be(buf, len);
-#else
-  return validate_utf16le(buf, len);
-#endif
-}
-simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
-                                          size_t len) noexcept {
-  return get_default_implementation()->validate_utf16le(buf, len);
-}
-simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
-                                          size_t len) noexcept {
-  return get_default_implementation()->validate_utf16be(buf, len);
-}
-simdutf_warn_unused result validate_utf16_with_errors(const char16_t *buf,
-                                                      size_t len) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return validate_utf16be_with_errors(buf, len);
-#else
-  return validate_utf16le_with_errors(buf, len);
-#endif
-}
-simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf,
-                                                        size_t len) noexcept {
-  return get_default_implementation()->validate_utf16le_with_errors(buf, len);
-}
-simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf,
-                                                        size_t len) noexcept {
-  return get_default_implementation()->validate_utf16be_with_errors(buf, len);
-}
-simdutf_warn_unused bool validate_utf32(const char32_t *buf,
-                                        size_t len) noexcept {
-  return get_default_implementation()->validate_utf32(buf, len);
-}
-simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf,
-                                                      size_t len) noexcept {
-  return get_default_implementation()->validate_utf32_with_errors(buf, len);
-}
-simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
-    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_valid_utf8_to_utf16be(input, length, utf16_buffer);
-#else
-  return convert_valid_utf8_to_utf16le(input, length, utf16_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
-    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf8_to_utf16le(
-      input, length, utf16_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
-    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf8_to_utf16be(
-      input, length, utf16_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
-    const char *input, size_t length, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf8_to_utf32(
-      input, length, utf32_buffer);
-}
-simdutf_warn_unused size_t convert_utf16_to_utf8(const char16_t *buf,
-                                                 size_t len,
-                                                 char *utf8_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_utf8(buf, len, utf8_buffer);
-#else
-  return convert_utf16le_to_utf8(buf, len, utf8_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_utf16_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_latin1(buf, len, latin1_buffer);
-#else
-  return convert_utf16le_to_latin1(buf, len, latin1_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_latin1_to_utf16(
-    const char *buf, size_t len, char16_t *utf16_output) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_latin1_to_utf16be(buf, len, utf16_output);
-#else
-  return convert_latin1_to_utf16le(buf, len, utf16_output);
-#endif
-}
-simdutf_warn_unused size_t convert_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_latin1(buf, len,
-                                                                 latin1_buffer);
-}
-simdutf_warn_unused size_t convert_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_latin1(buf, len,
-                                                                 latin1_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16be_to_latin1(
-      buf, len, latin1_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16le_to_latin1(
-      buf, len, latin1_buffer);
-}
-simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_latin1_with_errors(
-      buf, len, latin1_buffer);
-}
-simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_latin1_with_errors(
-      buf, len, latin1_buffer);
-}
-simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t *buf,
-                                                   size_t len,
-                                                   char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_utf8(buf, len,
-                                                               utf8_buffer);
-}
-simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t *buf,
-                                                   size_t len,
-                                                   char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_utf8(buf, len,
-                                                               utf8_buffer);
-}
-simdutf_warn_unused result convert_utf16_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_utf8_with_errors(buf, len, utf8_buffer);
-#else
-  return convert_utf16le_to_utf8_with_errors(buf, len, utf8_buffer);
-#endif
-}
-simdutf_warn_unused result convert_utf16_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_latin1_with_errors(buf, len, latin1_buffer);
-#else
-  return convert_utf16le_to_latin1_with_errors(buf, len, latin1_buffer);
-#endif
-}
-simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_utf8_with_errors(
-      buf, len, utf8_buffer);
-}
-simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_utf8_with_errors(
-      buf, len, utf8_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_valid_utf16be_to_utf8(buf, len, utf8_buffer);
-#else
-  return convert_valid_utf16le_to_utf8(buf, len, utf8_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_valid_utf16be_to_latin1(buf, len, latin1_buffer);
-#else
-  return convert_valid_utf16le_to_latin1(buf, len, latin1_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16le_to_utf8(
-      buf, len, utf8_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16be_to_utf8(
-      buf, len, utf8_buffer);
-}
-simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t *buf,
-                                                 size_t len,
-                                                 char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf8(buf, len,
-                                                             utf8_buffer);
-}
-simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
-    const char32_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf8_with_errors(
-      buf, len, utf8_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf32_to_utf8(buf, len,
-                                                                   utf8_buffer);
-}
-simdutf_warn_unused size_t convert_utf32_to_utf16(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf32_to_utf16be(buf, len, utf16_buffer);
-#else
-  return convert_utf32_to_utf16le(buf, len, utf16_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_utf32_to_latin1(
-    const char32_t *input, size_t length, char *latin1_output) noexcept {
-  return get_default_implementation()->convert_utf32_to_latin1(input, length,
-                                                               latin1_output);
-}
-simdutf_warn_unused size_t convert_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf16le(buf, len,
-                                                                utf16_buffer);
-}
-simdutf_warn_unused size_t convert_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf16be(buf, len,
-                                                                utf16_buffer);
-}
-simdutf_warn_unused result convert_utf32_to_utf16_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf32_to_utf16be_with_errors(buf, len, utf16_buffer);
-#else
-  return convert_utf32_to_utf16le_with_errors(buf, len, utf16_buffer);
-#endif
-}
-simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf16le_with_errors(
-      buf, len, utf16_buffer);
-}
-simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_utf32_to_utf16be_with_errors(
-      buf, len, utf16_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf32_to_utf16(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_valid_utf32_to_utf16be(buf, len, utf16_buffer);
-#else
-  return convert_valid_utf32_to_utf16le(buf, len, utf16_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf32_to_utf16le(
-      buf, len, utf16_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf32_to_utf16be(
-      buf, len, utf16_buffer);
-}
-simdutf_warn_unused size_t convert_utf16_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_utf32(buf, len, utf32_buffer);
-#else
-  return convert_utf16le_to_utf32(buf, len, utf32_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_utf32(buf, len,
-                                                                utf32_buffer);
-}
-simdutf_warn_unused size_t convert_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_utf32(buf, len,
-                                                                utf32_buffer);
-}
-simdutf_warn_unused result convert_utf16_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_utf16be_to_utf32_with_errors(buf, len, utf32_buffer);
-#else
-  return convert_utf16le_to_utf32_with_errors(buf, len, utf32_buffer);
-#endif
-}
-simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_utf16le_to_utf32_with_errors(
-      buf, len, utf32_buffer);
-}
-simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_utf16be_to_utf32_with_errors(
-      buf, len, utf32_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return convert_valid_utf16be_to_utf32(buf, len, utf32_buffer);
-#else
-  return convert_valid_utf16le_to_utf32(buf, len, utf32_buffer);
-#endif
-}
-simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16le_to_utf32(
-      buf, len, utf32_buffer);
-}
-simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
-  return get_default_implementation()->convert_valid_utf16be_to_utf32(
-      buf, len, utf32_buffer);
-}
-void change_endianness_utf16(const char16_t *input, size_t length,
-                             char16_t *output) noexcept {
-  get_default_implementation()->change_endianness_utf16(input, length, output);
-}
-simdutf_warn_unused size_t count_utf16(const char16_t *input,
-                                       size_t length) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return count_utf16be(input, length);
-#else
-  return count_utf16le(input, length);
-#endif
-}
-simdutf_warn_unused size_t count_utf16le(const char16_t *input,
-                                         size_t length) noexcept {
-  return get_default_implementation()->count_utf16le(input, length);
-}
-simdutf_warn_unused size_t count_utf16be(const char16_t *input,
-                                         size_t length) noexcept {
-  return get_default_implementation()->count_utf16be(input, length);
-}
-simdutf_warn_unused size_t count_utf8(const char *input,
-                                      size_t length) noexcept {
-  return get_default_implementation()->count_utf8(input, length);
-}
-simdutf_warn_unused size_t latin1_length_from_utf8(const char *buf,
-                                                   size_t len) noexcept {
-  return get_default_implementation()->latin1_length_from_utf8(buf, len);
-}
-simdutf_warn_unused size_t latin1_length_from_utf16(size_t len) noexcept {
-  return get_default_implementation()->latin1_length_from_utf16(len);
-}
-simdutf_warn_unused size_t latin1_length_from_utf32(size_t len) noexcept {
-  return get_default_implementation()->latin1_length_from_utf32(len);
-}
-simdutf_warn_unused size_t utf8_length_from_latin1(const char *buf,
-                                                   size_t len) noexcept {
-  return get_default_implementation()->utf8_length_from_latin1(buf, len);
-}
-simdutf_warn_unused size_t utf8_length_from_utf16(const char16_t *input,
-                                                  size_t length) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return utf8_length_from_utf16be(input, length);
-#else
-  return utf8_length_from_utf16le(input, length);
-#endif
-}
-simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t *input,
-                                                    size_t length) noexcept {
-  return get_default_implementation()->utf8_length_from_utf16le(input, length);
-}
-simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t *input,
-                                                    size_t length) noexcept {
-  return get_default_implementation()->utf8_length_from_utf16be(input, length);
-}
-simdutf_warn_unused size_t utf32_length_from_utf16(const char16_t *input,
-                                                   size_t length) noexcept {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return utf32_length_from_utf16be(input, length);
-#else
-  return utf32_length_from_utf16le(input, length);
-#endif
-}
-simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t *input,
-                                                     size_t length) noexcept {
-  return get_default_implementation()->utf32_length_from_utf16le(input, length);
-}
-simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t *input,
-                                                     size_t length) noexcept {
-  return get_default_implementation()->utf32_length_from_utf16be(input, length);
-}
-simdutf_warn_unused size_t utf16_length_from_utf8(const char *input,
-                                                  size_t length) noexcept {
-  return get_default_implementation()->utf16_length_from_utf8(input, length);
-}
-simdutf_warn_unused size_t utf16_length_from_latin1(size_t length) noexcept {
-  return get_default_implementation()->utf16_length_from_latin1(length);
-}
-simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t *input,
-                                                  size_t length) noexcept {
-  return get_default_implementation()->utf8_length_from_utf32(input, length);
-}
-simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t *input,
-                                                   size_t length) noexcept {
-  return get_default_implementation()->utf16_length_from_utf32(input, length);
-}
-simdutf_warn_unused size_t utf32_length_from_utf8(const char *input,
-                                                  size_t length) noexcept {
-  return get_default_implementation()->utf32_length_from_utf8(input, length);
-}
-
-simdutf_warn_unused size_t
-maximal_binary_length_from_base64(const char *input, size_t length) noexcept {
-  return get_default_implementation()->maximal_binary_length_from_base64(
-      input, length);
-}
-
-simdutf_warn_unused result base64_to_binary(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_handling_options) noexcept {
-  return get_default_implementation()->base64_to_binary(
-      input, length, output, options, last_chunk_handling_options);
-}
-
-simdutf_warn_unused size_t maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) noexcept {
-  return get_default_implementation()->maximal_binary_length_from_base64(
-      input, length);
-}
-
-simdutf_warn_unused result base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_handling_options) noexcept {
-  return get_default_implementation()->base64_to_binary(
-      input, length, output, options, last_chunk_handling_options);
-}
-
-template <typename chartype>
-simdutf_warn_unused result base64_to_binary_safe_impl(
-    const chartype *input, size_t length, char *output, size_t &outlen,
-    base64_options options,
-    last_chunk_handling_options last_chunk_handling_options) noexcept {
-  static_assert(std::is_same<chartype, char>::value ||
-                    std::is_same<chartype, char16_t>::value,
-                "Only char and char16_t are supported.");
-  // The implementation could be nicer, but we expect that most times, the user
-  // will provide us with a buffer that is large enough.
-  size_t max_length = maximal_binary_length_from_base64(input, length);
-  if (outlen >= max_length) {
-    // fast path
-    full_result r = get_default_implementation()->base64_to_binary_details(
-        input, length, output, options, last_chunk_handling_options);
-    if (r.error != error_code::INVALID_BASE64_CHARACTER &&
-        r.error != error_code::BASE64_EXTRA_BITS) {
-      outlen = r.output_count;
-      if (last_chunk_handling_options == stop_before_partial) {
-        if ((r.output_count % 3) != 0) {
-          bool empty_trail = true;
-          for (size_t i = r.input_count; i < length; i++) {
-            if (!scalar::base64::is_ascii_white_space_or_padding(input[i])) {
-              empty_trail = false;
-              break;
-            }
-          }
-          if (empty_trail) {
-            r.input_count = length;
-          }
-        }
-        return {r.error, r.input_count};
-      }
-      return {r.error, length};
-    }
-    return r;
-  }
-  // The output buffer is maybe too small. We will decode a truncated version of
-  // the input.
-  size_t outlen3 = outlen / 3 * 3; // round down to multiple of 3
-  size_t safe_input = base64_length_from_binary(outlen3, options);
-  full_result r = get_default_implementation()->base64_to_binary_details(
-      input, safe_input, output, options, loose);
-  if (r.error == error_code::INVALID_BASE64_CHARACTER) {
-    return r;
-  }
-  size_t offset =
-      (r.error == error_code::BASE64_INPUT_REMAINDER)
-          ? 1
-          : ((r.output_count % 3) == 0 ? 0 : (r.output_count % 3) + 1);
-  size_t output_index = r.output_count - (r.output_count % 3);
-  size_t input_index = safe_input;
-  // offset is a value that is no larger than 3. We backtrack
-  // by up to offset characters + an undetermined number of
-  // white space characters. It is expected that the next loop
-  // runs at most 3 times + the number of white space characters
-  // in between them, so we are not worried about performance.
-  while (offset > 0 && input_index > 0) {
-    chartype c = input[--input_index];
-    if (scalar::base64::is_ascii_white_space(c)) {
-      // skipping
-    } else {
-      offset--;
-    }
-  }
-  size_t remaining_out = outlen - output_index;
-  const chartype *tail_input = input + input_index;
-  size_t tail_length = length - input_index;
-  while (tail_length > 0 &&
-         scalar::base64::is_ascii_white_space(tail_input[tail_length - 1])) {
-    tail_length--;
-  }
-  size_t padding_characts = 0;
-  if (tail_length > 0 && tail_input[tail_length - 1] == '=') {
-    tail_length--;
-    padding_characts++;
-    while (tail_length > 0 &&
-           scalar::base64::is_ascii_white_space(tail_input[tail_length - 1])) {
-      tail_length--;
-    }
-    if (tail_length > 0 && tail_input[tail_length - 1] == '=') {
-      tail_length--;
-      padding_characts++;
-    }
-  }
-  // this will advance tail_input and tail_length
-  result rr = scalar::base64::base64_tail_decode_safe(
-      output + output_index, remaining_out, tail_input, tail_length,
-      padding_characts, options, last_chunk_handling_options);
-  outlen = output_index + remaining_out;
-  if (last_chunk_handling_options != stop_before_partial &&
-      rr.error == error_code::SUCCESS && padding_characts > 0) {
-    // additional checks
-    if ((outlen % 3 == 0) || ((outlen % 3) + 1 + padding_characts != 4)) {
-      rr.error = error_code::INVALID_BASE64_CHARACTER;
-    }
-  }
-  if (rr.error == error_code::SUCCESS &&
-      last_chunk_handling_options == stop_before_partial) {
-    if (tail_input > input + input_index) {
-      rr.count = tail_input - input;
-    } else if (r.input_count > 0) {
-      rr.count = r.input_count + rr.count;
-    }
-    return rr;
-  }
-  rr.count += input_index;
-  return rr;
-}
-
-simdutf_warn_unused size_t convert_latin1_to_utf8_safe(
-    const char *buf, size_t len, char *utf8_output, size_t utf8_len) noexcept {
-  const auto start{utf8_output};
-
-  while (true) {
-    // convert_latin1_to_utf8 will never write more than input length * 2
-    auto read_len = std::min(len, utf8_len >> 1);
-    if (read_len <= 16) {
-      break;
-    }
-
-    const auto write_len =
-        simdutf::convert_latin1_to_utf8(buf, read_len, utf8_output);
-
-    utf8_output += write_len;
-    utf8_len -= write_len;
-    buf += read_len;
-    len -= read_len;
-  }
-
-  utf8_output +=
-      scalar::latin1_to_utf8::convert_safe(buf, len, utf8_output, utf8_len);
-
-  return utf8_output - start;
-}
-
-simdutf_warn_unused result base64_to_binary_safe(
-    const char *input, size_t length, char *output, size_t &outlen,
-    base64_options options,
-    last_chunk_handling_options last_chunk_handling_options) noexcept {
-  return base64_to_binary_safe_impl<char>(input, length, output, outlen,
-                                          options, last_chunk_handling_options);
-}
-simdutf_warn_unused result base64_to_binary_safe(
-    const char16_t *input, size_t length, char *output, size_t &outlen,
-    base64_options options,
-    last_chunk_handling_options last_chunk_handling_options) noexcept {
-  return base64_to_binary_safe_impl<char16_t>(
-      input, length, output, outlen, options, last_chunk_handling_options);
-}
-
-simdutf_warn_unused size_t
-base64_length_from_binary(size_t length, base64_options options) noexcept {
-  return get_default_implementation()->base64_length_from_binary(length,
-                                                                 options);
-}
-
-size_t binary_to_base64(const char *input, size_t length, char *output,
-                        base64_options options) noexcept {
-  return get_default_implementation()->binary_to_base64(input, length, output,
-                                                        options);
-}
-
-simdutf_warn_unused simdutf::encoding_type
-autodetect_encoding(const char *buf, size_t length) noexcept {
-  return get_default_implementation()->autodetect_encoding(buf, length);
-}
-simdutf_warn_unused int detect_encodings(const char *buf,
-                                         size_t length) noexcept {
-  return get_default_implementation()->detect_encodings(buf, length);
-}
-const implementation *builtin_implementation() {
-  static const implementation *builtin_impl =
-      get_available_implementations()[SIMDUTF_STRINGIFY(
-          SIMDUTF_BUILTIN_IMPLEMENTATION)];
-  return builtin_impl;
-}
-
-simdutf_warn_unused size_t trim_partial_utf8(const char *input, size_t length) {
-  return scalar::utf8::trim_partial_utf8(input, length);
-}
-
-simdutf_warn_unused size_t trim_partial_utf16be(const char16_t *input,
-                                                size_t length) {
-  return scalar::utf16::trim_partial_utf16<BIG>(input, length);
-}
-
-simdutf_warn_unused size_t trim_partial_utf16le(const char16_t *input,
-                                                size_t length) {
-  return scalar::utf16::trim_partial_utf16<LITTLE>(input, length);
-}
-
-simdutf_warn_unused size_t trim_partial_utf16(const char16_t *input,
-                                              size_t length) {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return trim_partial_utf16be(input, length);
-#else
-  return trim_partial_utf16le(input, length);
-#endif
-}
-
-} // namespace simdutf
-/* end file src/implementation.cpp */
-/* begin file src/encoding_types.cpp */
-
-namespace simdutf {
-bool match_system(endianness e) {
-#if SIMDUTF_IS_BIG_ENDIAN
-  return e == endianness::BIG;
-#else
-  return e == endianness::LITTLE;
-#endif
-}
-
-std::string to_string(encoding_type bom) {
-  switch (bom) {
-  case UTF16_LE:
-    return "UTF16 little-endian";
-  case UTF16_BE:
-    return "UTF16 big-endian";
-  case UTF32_LE:
-    return "UTF32 little-endian";
-  case UTF32_BE:
-    return "UTF32 big-endian";
-  case UTF8:
-    return "UTF8";
-  case unspecified:
-    return "unknown";
-  default:
-    return "error";
-  }
-}
-
-namespace BOM {
-// Note that BOM for UTF8 is discouraged.
-encoding_type check_bom(const uint8_t *byte, size_t length) {
-  if (length >= 2 && byte[0] == 0xff and byte[1] == 0xfe) {
-    if (length >= 4 && byte[2] == 0x00 and byte[3] == 0x0) {
-      return encoding_type::UTF32_LE;
-    } else {
-      return encoding_type::UTF16_LE;
-    }
-  } else if (length >= 2 && byte[0] == 0xfe and byte[1] == 0xff) {
-    return encoding_type::UTF16_BE;
-  } else if (length >= 4 && byte[0] == 0x00 and byte[1] == 0x00 and
-             byte[2] == 0xfe and byte[3] == 0xff) {
-    return encoding_type::UTF32_BE;
-  } else if (length >= 4 && byte[0] == 0xef and byte[1] == 0xbb and
-             byte[2] == 0xbf) {
-    return encoding_type::UTF8;
-  }
-  return encoding_type::unspecified;
-}
-
-encoding_type check_bom(const char *byte, size_t length) {
-  return check_bom(reinterpret_cast<const uint8_t *>(byte), length);
-}
-
-size_t bom_byte_size(encoding_type bom) {
-  switch (bom) {
-  case UTF16_LE:
-    return 2;
-  case UTF16_BE:
-    return 2;
-  case UTF32_LE:
-    return 4;
-  case UTF32_BE:
-    return 4;
-  case UTF8:
-    return 3;
-  case unspecified:
-    return 0;
-  default:
-    return 0;
-  }
-}
-
-} // namespace BOM
-} // namespace simdutf
-/* end file src/encoding_types.cpp */
-/* begin file src/error.cpp */
-namespace simdutf {
-// deliberately empty
-}
-/* end file src/error.cpp */
-// The large tables should be included once and they
-// should not depend on a kernel.
-/* begin file src/tables/utf8_to_utf16_tables.h */
-#ifndef SIMDUTF_UTF8_TO_UTF16_TABLES_H
-#define SIMDUTF_UTF8_TO_UTF16_TABLES_H
-#include <cstdint>
-
-namespace simdutf {
-namespace {
-namespace tables {
-namespace utf8_to_utf16 {
-/**
- * utf8bigindex uses about 8 kB
- * shufutf8 uses about 3344 B
- *
- * So we use a bit over 11 kB. It would be
- * easy to save about 4 kB by only
- * storing the index in utf8bigindex, and
- * deriving the consumed bytes otherwise.
- * However, this may come at a significant (10% to 20%)
- * performance penalty.
- */
+namespace tables {
+namespace utf8_to_utf16 {
+/**
+ * utf8bigindex uses about 8 kB
+ * shufutf8 uses about 3344 B
+ *
+ * So we use a bit over 11 kB. It would be
+ * easy to save about 4 kB by only
+ * storing the index in utf8bigindex, and
+ * deriving the consumed bytes otherwise.
+ * However, this may come at a significant (10% to 20%)
+ * performance penalty.
+ */
 
 const uint8_t shufutf8[209][16] = {
     {0, 255, 1, 255, 2, 255, 3, 255, 4, 255, 5, 255, 0, 0, 0, 0},
@@ -13422,5242 +1602,24884 @@ const uint8_t utf8bigindex[4096][2] = {
     {193, 6},  {82, 6},   {48, 8},   {8, 7},    {118, 6},  {16, 7},   {32, 7},
     {0, 6}};
 } // namespace utf8_to_utf16
-} // namespace tables
+} // namespace tables
+} // unnamed namespace
+} // namespace simdutf
+
+#endif // SIMDUTF_UTF8_TO_UTF16_TABLES_H
+/* end file src/tables/utf8_to_utf16_tables.h */
+/* begin file src/tables/utf16_to_utf8_tables.h */
+// file generated by scripts/sse_convert_utf16_to_utf8.py
+#ifndef SIMDUTF_UTF16_TO_UTF8_TABLES_H
+#define SIMDUTF_UTF16_TO_UTF8_TABLES_H
+
+namespace simdutf {
+namespace {
+namespace tables {
+namespace utf16_to_utf8 {
+
+// 1 byte for length, 16 bytes for mask
+const uint8_t pack_1_2_utf8_bytes[256][17] = {
+    {16, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14},
+    {15, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
+    {15, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80},
+    {14, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {15, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
+    {14, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {14, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80},
+    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {15, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
+    {14, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {14, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80},
+    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {15, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
+    {14, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {14, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 3, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80},
+    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
+    {13, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 3, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
+    {13, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {13, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 3, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
+    {12, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 3, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {12, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 3, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 0, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {11, 1, 0, 3, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 3, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 3, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 3, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 1, 0, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 1, 0, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 0, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80}};
+
+// 1 byte for length, 16 bytes for mask
+const uint8_t pack_1_2_3_utf8_bytes[256][17] = {
+    {12, 2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80},
+    {9, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {11, 3, 1, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 0, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 2, 3, 1, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {11, 2, 3, 1, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 3, 1, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 2, 3, 1, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 3, 1, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 2, 3, 1, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 1, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 2, 3, 1, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {11, 2, 3, 1, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 3, 1, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 2, 3, 1, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 3, 1, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 2, 3, 1, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 2, 3, 1, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 3, 1, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 0, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 2, 3, 1, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 2, 3, 1, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 2, 3, 1, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 1, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 2, 3, 1, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 2, 3, 1, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {3, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 2, 3, 1, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {2, 3, 1, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {5, 2, 3, 1, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 3, 1, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 0, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 2, 3, 1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 3, 1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 0, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 2, 3, 1, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 3, 1, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 3, 1, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 0, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 2, 3, 1, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 3, 1, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 0, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 2, 3, 1, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {3, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 3, 1, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 0, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 2, 3, 1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 3, 1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 0, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 2, 3, 1, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {3, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 2, 3, 1, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 3, 1, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 0, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {11, 2, 3, 1, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 3, 1, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 2, 3, 1, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 3, 1, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 2, 3, 1, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 2, 3, 1, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 3, 1, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 0, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 2, 3, 1, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 2, 3, 1, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {10, 2, 3, 1, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 3, 1, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 0, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 2, 3, 1, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 2, 3, 1, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 2, 3, 1, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 1, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 2, 3, 1, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 2, 3, 1, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 3, 1, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 0, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 2, 3, 1, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 2, 3, 1, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 2, 3, 1, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 3, 1, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 0, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 2, 3, 1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 3, 1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 0, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 2, 3, 1, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {3, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 2, 3, 1, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 3, 1, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 0, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {9, 2, 3, 1, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 3, 1, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 1, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 2, 3, 1, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 2, 3, 1, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 3, 1, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 2, 3, 1, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 3, 1, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 3, 1, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 3, 1, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 0, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 2, 3, 1, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 3, 1, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 0, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 2, 3, 1, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {3, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 3, 1, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80}};
+
+} // namespace utf16_to_utf8
+} // namespace tables
+} // unnamed namespace
+} // namespace simdutf
+
+#endif // SIMDUTF_UTF16_TO_UTF8_TABLES_H
+/* end file src/tables/utf16_to_utf8_tables.h */
+/* begin file src/tables/utf32_to_utf16_tables.h */
+// file generated by scripts/sse_convert_utf32_to_utf16.py
+#ifndef SIMDUTF_UTF32_TO_UTF16_TABLES_H
+#define SIMDUTF_UTF32_TO_UTF16_TABLES_H
+
+namespace simdutf {
+namespace {
+namespace tables {
+namespace utf32_to_utf16 {
+
+const uint8_t pack_utf32_to_utf16le[16][16] = {
+    {0, 1, 4, 5, 8, 9, 12, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 8, 9, 12, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 4, 5, 6, 7, 8, 9, 12, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 13, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 4, 5, 8, 9, 10, 11, 12, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 8, 9, 10, 11, 12, 13, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 0x80, 0x80},
+    {0, 1, 4, 5, 8, 9, 12, 13, 14, 15, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 8, 9, 12, 13, 14, 15, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 4, 5, 6, 7, 8, 9, 12, 13, 14, 15, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 12, 13, 14, 15, 0x80, 0x80},
+    {0, 1, 4, 5, 8, 9, 10, 11, 12, 13, 14, 15, 0x80, 0x80, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 8, 9, 10, 11, 12, 13, 14, 15, 0x80, 0x80},
+    {0, 1, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0x80, 0x80},
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+};
+
+const uint8_t pack_utf32_to_utf16be[16][16] = {
+    {1, 0, 5, 4, 9, 8, 13, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 9, 8, 13, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 5, 4, 7, 6, 9, 8, 13, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 13, 12, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 5, 4, 9, 8, 11, 10, 13, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 9, 8, 11, 10, 13, 12, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 0x80, 0x80},
+    {1, 0, 5, 4, 9, 8, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 9, 8, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 5, 4, 7, 6, 9, 8, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 13, 12, 15, 14, 0x80, 0x80},
+    {1, 0, 5, 4, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {1, 0, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
+    {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14},
+};
+
+} // namespace utf32_to_utf16
+} // namespace tables
+} // unnamed namespace
+} // namespace simdutf
+
+#endif // SIMDUTF_UTF16_TO_UTF8_TABLES_H
+/* end file src/tables/utf32_to_utf16_tables.h */
+// End of tables.
+
+// Implementations: they need to be setup before including
+// scalar/* code, as the scalar code is sometimes enabled
+// only for peculiar build targets.
+
+// The best choice should always come first!
+/* begin file src/simdutf/arm64.h */
+#ifndef SIMDUTF_ARM64_H
+#define SIMDUTF_ARM64_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "arm64.h must be included before fallback.h"
+#endif
+
+
+#ifndef SIMDUTF_IMPLEMENTATION_ARM64
+  #define SIMDUTF_IMPLEMENTATION_ARM64 (SIMDUTF_IS_ARM64)
+#endif
+#if SIMDUTF_IMPLEMENTATION_ARM64 && SIMDUTF_IS_ARM64
+  #define SIMDUTF_CAN_ALWAYS_RUN_ARM64 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_ARM64 0
+#endif
+
+
+#if SIMDUTF_IMPLEMENTATION_ARM64
+
+namespace simdutf {
+/**
+ * Implementation for NEON (ARMv8).
+ */
+namespace arm64 {} // namespace arm64
+} // namespace simdutf
+
+/* begin file src/simdutf/arm64/implementation.h */
+#ifndef SIMDUTF_ARM64_IMPLEMENTATION_H
+#define SIMDUTF_ARM64_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace arm64 {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("arm64", "ARM NEON",
+                                internal::instruction_set::NEON) {}
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace arm64
+} // namespace simdutf
+
+#endif // SIMDUTF_ARM64_IMPLEMENTATION_H
+/* end file src/simdutf/arm64/implementation.h */
+
+/* begin file src/simdutf/arm64/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "arm64"
+// #define SIMDUTF_IMPLEMENTATION arm64
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
+/* end file src/simdutf/arm64/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/arm64/intrinsics.h */
+#ifndef SIMDUTF_ARM64_INTRINSICS_H
+#define SIMDUTF_ARM64_INTRINSICS_H
+
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <arm_neon.h>
+
+#endif //  SIMDUTF_ARM64_INTRINSICS_H
+/* end file src/simdutf/arm64/intrinsics.h */
+/* begin file src/simdutf/arm64/bitmanipulation.h */
+#ifndef SIMDUTF_ARM64_BITMANIPULATION_H
+#define SIMDUTF_ARM64_BITMANIPULATION_H
+
+namespace simdutf {
+namespace arm64 {
+namespace {
+
+/* result might be undefined when input_num is zero */
+simdutf_really_inline int count_ones(uint64_t input_num) {
+  return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
+}
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+}
+#endif
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdutf
+
+#endif // SIMDUTF_ARM64_BITMANIPULATION_H
+/* end file src/simdutf/arm64/bitmanipulation.h */
+/* begin file src/simdutf/arm64/simd.h */
+#ifndef SIMDUTF_ARM64_SIMD_H
+#define SIMDUTF_ARM64_SIMD_H
+
+#include <type_traits>
+
+namespace simdutf {
+namespace arm64 {
+namespace {
+namespace simd {
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+namespace {
+  // Start of private section with Visual Studio workaround
+
+  #ifndef simdutf_make_uint8x16_t
+    #define simdutf_make_uint8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10,   \
+                                    x11, x12, x13, x14, x15, x16)              \
+      ([=]() {                                                                 \
+        uint8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,             \
+                             x9, x10, x11, x12, x13, x14, x15, x16};           \
+        return vld1q_u8(array);                                                \
+      }())
+  #endif
+  #ifndef simdutf_make_int8x16_t
+    #define simdutf_make_int8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10,    \
+                                   x11, x12, x13, x14, x15, x16)               \
+      ([=]() {                                                                 \
+        int8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,              \
+                            x9, x10, x11, x12, x13, x14, x15, x16};            \
+        return vld1q_s8(array);                                                \
+      }())
+  #endif
+
+  #ifndef simdutf_make_uint8x8_t
+    #define simdutf_make_uint8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)             \
+      ([=]() {                                                                 \
+        uint8_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                   \
+        return vld1_u8(array);                                                 \
+      }())
+  #endif
+  #ifndef simdutf_make_int8x8_t
+    #define simdutf_make_int8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)              \
+      ([=]() {                                                                 \
+        int8_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                    \
+        return vld1_s8(array);                                                 \
+      }())
+  #endif
+  #ifndef simdutf_make_uint16x8_t
+    #define simdutf_make_uint16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)            \
+      ([=]() {                                                                 \
+        uint16_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                  \
+        return vld1q_u16(array);                                               \
+      }())
+  #endif
+  #ifndef simdutf_make_int16x8_t
+    #define simdutf_make_int16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)             \
+      ([=]() {                                                                 \
+        int16_t array[8] = {x1, x2, x3, x4, x5, x6, x7, x8};                   \
+        return vld1q_s16(array);                                               \
+      }())
+  #endif
+
+// End of private section with Visual Studio workaround
+} // namespace
+#endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+
+template <typename T> struct simd8;
+
+//
+// Base class of simd8<uint8_t> and simd8<bool>, both of which use uint8x16_t
+// internally.
+//
+template <typename T, typename Mask = simd8<bool>> struct base_u8 {
+  uint8x16_t value;
+  static const int SIZE = sizeof(value);
+  void dump() const {
+    uint8_t temp[16];
+    vst1q_u8(temp, *this);
+    printf("[%04x, %04x, %04x, %04x, %04x, %04x, %04x, %04x,%04x, %04x, %04x, "
+           "%04x, %04x, %04x, %04x, %04x]\n",
+           temp[0], temp[1], temp[2], temp[3], temp[4], temp[5], temp[6],
+           temp[7], temp[8], temp[9], temp[10], temp[11], temp[12], temp[13],
+           temp[14], temp[15]);
+  }
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u8(const uint8x16_t _value) : value(_value) {}
+  simdutf_really_inline operator const uint8x16_t &() const {
+    return this->value;
+  }
+  simdutf_really_inline operator uint8x16_t &() { return this->value; }
+  simdutf_really_inline T first() const { return vgetq_lane_u8(*this, 0); }
+  simdutf_really_inline T last() const { return vgetq_lane_u8(*this, 15); }
+
+  // Bit operations
+  simdutf_really_inline simd8<T> operator|(const simd8<T> other) const {
+    return vorrq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<T> operator&(const simd8<T> other) const {
+    return vandq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<T> operator^(const simd8<T> other) const {
+    return veorq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<T> bit_andnot(const simd8<T> other) const {
+    return vbicq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+  simdutf_really_inline simd8<T> &operator|=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd8<T> &operator&=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd8<T> &operator^=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return vceqq_u8(lhs, rhs);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    return vextq_u8(prev_chunk, *this, 16 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base_u8<bool> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return vmovq_n_u8(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8(const uint8x16_t _value)
+      : base_u8<bool>(_value) {}
+  // False constructor
+  simdutf_really_inline simd8() : simd8(vdupq_n_u8(0)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return vst1q_u8(dst, *this);
+  }
+
+  // We return uint32_t instead of uint16_t because that seems to be more
+  // efficient for most purposes (cutting it down to uint16_t costs performance
+  // in some compilers).
+  simdutf_really_inline uint32_t to_bitmask() const {
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+    const uint8x16_t bit_mask =
+        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+    auto minput = *this & bit_mask;
+    uint8x16_t tmp = vpaddq_u8(minput, minput);
+    tmp = vpaddq_u8(tmp, tmp);
+    tmp = vpaddq_u8(tmp, tmp);
+    return vgetq_lane_u16(vreinterpretq_u16_u8(tmp), 0);
+  }
+
+  // Returns 4-bit out of each byte, alternating between the high 4 bits and low
+  // bits result it is 64 bit. This method is expected to be faster than none()
+  // and is equivalent when the vector register is the result of a comparison,
+  // with byte values 0xff and 0x00.
+  simdutf_really_inline uint64_t to_bitmask64() const {
+    return vget_lane_u64(
+        vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
+  }
+
+  simdutf_really_inline bool any() const {
+    return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0;
+  }
+  simdutf_really_inline bool none() const {
+    return vmaxvq_u32(vreinterpretq_u32_u8(*this)) == 0;
+  }
+  simdutf_really_inline bool all() const {
+    return vminvq_u32(vreinterpretq_u32_u8(*this)) == 0xFFFFF;
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base_u8<uint8_t> {
+  static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) {
+    return vmovq_n_u8(_value);
+  }
+  static simdutf_really_inline simd8<uint8_t> zero() { return vdupq_n_u8(0); }
+  static simdutf_really_inline simd8<uint8_t> load(const uint8_t *values) {
+    return vld1q_u8(values);
+  }
+  simdutf_really_inline simd8(const uint8x16_t _value)
+      : base_u8<uint8_t>(_value) {}
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Member-by-member initialization
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8(simdutf_make_uint8x16_t(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                                      v10, v11, v12, v13, v14, v15)) {}
+#else
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8(uint8x16_t{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15}) {}
+#endif
+
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Store to array
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return vst1q_u8(dst, *this);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return vqaddq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return vqsubq_u8(*this, other);
+  }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd8<uint8_t>
+  operator+(const simd8<uint8_t> other) const {
+    return vaddq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  operator-(const simd8<uint8_t> other) const {
+    return vsubq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t> &operator+=(const simd8<uint8_t> other) {
+    *this = *this + other;
+    return *this;
+  }
+  simdutf_really_inline simd8<uint8_t> &operator-=(const simd8<uint8_t> other) {
+    *this = *this - other;
+    return *this;
+  }
+
+  // Order-specific operations
+  simdutf_really_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
+  simdutf_really_inline uint8_t min_val() const { return vminvq_u8(*this); }
+  simdutf_really_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return vmaxq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return vminq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return vcleq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return vcgeq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return vcltq_u8(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return vcgtq_u8(*this, other);
+  }
+  // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true
+  // = nonzero. For ARM, returns all 1's.
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return simd8<uint8_t>(*this > other);
+  }
+  // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true
+  // = nonzero. For ARM, returns all 1's.
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return simd8<uint8_t>(*this < other);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return vtstq_u8(*this, bits);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return this->max_val() < 0b10000000u;
+  }
+
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return this->max_val() != 0;
+  }
+  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return (*this & bits).any_bits_set_anywhere();
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return vshrq_n_u8(*this, N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return vshlq_n_u8(*this, N);
+  }
+  simdutf_really_inline uint16_t sum_bytes() const { return vaddvq_u8(*this); }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return lookup_table.apply_lookup_16_to(*this);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<uint8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    return vqtbl1q_u8(*this, simd8<uint8_t>(original));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> {
+  int8x16_t value;
+  static const int SIZE = sizeof(value);
+
+  static simdutf_really_inline simd8<int8_t> splat(int8_t _value) {
+    return vmovq_n_s8(_value);
+  }
+  static simdutf_really_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
+  static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) {
+    return vld1q_s8(values);
+  }
+
+  // Use ST2 instead of UXTL+UXTL2 to interleave zeroes. UXTL is actually a
+  // USHLL #0, and shifting in NEON is actually quite slow.
+  //
+  // While this needs the registers to be in a specific order, bigger cores can
+  // interleave these with no overhead, and it still performs decently on little
+  // cores.
+  //    movi  v1.3d, #0
+  //      mov   v0.16b, value[0]
+  //    st2   {v0.16b, v1.16b}, [ptr], #32
+  //      mov   v0.16b, value[1]
+  //    st2   {v0.16b, v1.16b}, [ptr], #32
+  //    ...
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
+    int8x16x2_t pair = match_system(big_endian)
+                           ? int8x16x2_t{{this->value, vmovq_n_s8(0)}}
+                           : int8x16x2_t{{vmovq_n_s8(0), this->value}};
+    vst2q_s8(reinterpret_cast<int8_t *>(p), pair);
+  }
+
+  // currently unused
+  // Technically this could be done with ST4 like in store_ascii_as_utf16, but
+  // it is very much not worth it, as explicitly mentioned in the ARM Cortex-X1
+  // Core Software Optimization Guide:
+  //   4.18 Complex ASIMD instructions
+  //     The bandwidth of [ST4 with element size less than 64b] is limited by
+  //     decode constraints and it is advisable to avoid them when high
+  //     performing code is desired.
+  // Instead, it is better to use ZIP1+ZIP2 and two ST2.
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
+    const uint16x8_t low =
+        vreinterpretq_u16_s8(vzip1q_s8(this->value, vmovq_n_s8(0)));
+    const uint16x8_t high =
+        vreinterpretq_u16_s8(vzip2q_s8(this->value, vmovq_n_s8(0)));
+    const uint16x8x2_t low_pair{{low, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(p), low_pair);
+    const uint16x8x2_t high_pair{{high, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(p + 8), high_pair);
+  }
+
+  // In places where the table can be reused, which is most uses in simdutf, it
+  // is worth it to do 4 table lookups, as there is no direct zero extension
+  // from u8 to u32.
+  simdutf_really_inline void store_ascii_as_utf32_tbl(char32_t *p) const {
+    const simd8<uint8_t> tb1{0, 255, 255, 255, 1, 255, 255, 255,
+                             2, 255, 255, 255, 3, 255, 255, 255};
+    const simd8<uint8_t> tb2{4, 255, 255, 255, 5, 255, 255, 255,
+                             6, 255, 255, 255, 7, 255, 255, 255};
+    const simd8<uint8_t> tb3{8,  255, 255, 255, 9,  255, 255, 255,
+                             10, 255, 255, 255, 11, 255, 255, 255};
+    const simd8<uint8_t> tb4{12, 255, 255, 255, 13, 255, 255, 255,
+                             14, 255, 255, 255, 15, 255, 255, 255};
+
+    // encourage store pairing and interleaving
+    const auto shuf1 = this->apply_lookup_16_to(tb1);
+    const auto shuf2 = this->apply_lookup_16_to(tb2);
+    shuf1.store(reinterpret_cast<int8_t *>(p));
+    shuf2.store(reinterpret_cast<int8_t *>(p + 4));
+
+    const auto shuf3 = this->apply_lookup_16_to(tb3);
+    const auto shuf4 = this->apply_lookup_16_to(tb4);
+    shuf3.store(reinterpret_cast<int8_t *>(p + 8));
+    shuf4.store(reinterpret_cast<int8_t *>(p + 12));
+  }
+  // Conversion from/to SIMD register
+  simdutf_really_inline simd8(const int8x16_t _value) : value{_value} {}
+  simdutf_really_inline operator const int8x16_t &() const {
+    return this->value;
+  }
+#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
+  simdutf_really_inline operator const uint8x16_t() const {
+    return vreinterpretq_u8_s8(this->value);
+  }
+#endif
+  simdutf_really_inline operator int8x16_t &() { return this->value; }
+
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8(simdutf_make_int8x16_t(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                                     v10, v11, v12, v13, v14, v15)) {}
+#else
+  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8(int8x16_t{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                        v13, v14, v15}) {}
+#endif
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15);
+  }
+
+  // Store to array
+  simdutf_really_inline void store(int8_t dst[16]) const {
+    return vst1q_s8(dst, value);
+  }
+  // Explicit conversion to/from unsigned
+  //
+  // Under Visual Studio/ARM64 uint8x16_t and int8x16_t are apparently the same
+  // type. In theory, we could check this occurrence with std::same_as and
+  // std::enabled_if but it is C++14 and relatively ugly and hard to read.
+#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
+  simdutf_really_inline explicit simd8(const uint8x16_t other)
+      : simd8(vreinterpretq_s8_u8(other)) {}
+#endif
+  simdutf_really_inline operator simd8<uint8_t>() const {
+    return vreinterpretq_u8_s8(this->value);
+  }
+
+  simdutf_really_inline simd8<int8_t>
+  operator|(const simd8<int8_t> other) const {
+    return vorrq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator&(const simd8<int8_t> other) const {
+    return vandq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator^(const simd8<int8_t> other) const {
+    return veorq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  bit_andnot(const simd8<int8_t> other) const {
+    return vbicq_s8(value, other.value);
+  }
+
+  // Math
+  simdutf_really_inline simd8<int8_t>
+  operator+(const simd8<int8_t> other) const {
+    return vaddq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator-(const simd8<int8_t> other) const {
+    return vsubq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t> &operator+=(const simd8<int8_t> other) {
+    *this = *this + other;
+    return *this;
+  }
+  simdutf_really_inline simd8<int8_t> &operator-=(const simd8<int8_t> other) {
+    *this = *this - other;
+    return *this;
+  }
+
+  simdutf_really_inline int8_t max_val() const { return vmaxvq_s8(value); }
+  simdutf_really_inline int8_t min_val() const { return vminvq_s8(value); }
+  simdutf_really_inline bool is_ascii() const { return this->min_val() >= 0; }
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
+    return vmaxq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
+    return vminq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return vcgtq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return vcltq_s8(value, other.value);
+  }
+  simdutf_really_inline simd8<bool>
+  operator==(const simd8<int8_t> other) const {
+    return vceqq_s8(value, other.value);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<int8_t>
+  prev(const simd8<int8_t> prev_chunk) const {
+    return vextq_s8(prev_chunk, *this, 16 - N);
+  }
+
+  // Perform a lookup assuming no value is larger than 16
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return lookup_table.apply_lookup_16_to(*this);
+  }
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<int8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    return vqtbl1q_s8(*this, simd8<uint8_t>(original));
+  }
+};
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "ARM kernel should use four registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                 const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    this->chunks[2] |= other.chunks[2];
+    this->chunks[3] |= other.chunks[3];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 2);
+    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+    const uint8x16_t bit_mask =
+        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+    // Add each of the elements next to each other, successively, to stuff each
+    // 8 byte mask into one.
+    uint8x16_t sum0 =
+        vpaddq_u8(vandq_u8(uint8x16_t(this->chunks[0]), bit_mask),
+                  vandq_u8(uint8x16_t(this->chunks[1]), bit_mask));
+    uint8x16_t sum1 =
+        vpaddq_u8(vandq_u8(uint8x16_t(this->chunks[2]), bit_mask),
+                  vandq_u8(uint8x16_t(this->chunks[3]), bit_mask));
+    sum0 = vpaddq_u8(sum0, sum1);
+    sum0 = vpaddq_u8(sum0, sum0);
+    return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                          this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                          this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+
+    return simd8x64<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+    return simd8x64<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
+               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
+               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                          this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
+                          this->chunks[2] > mask, this->chunks[3] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
+                          this->chunks[2] >= mask, this->chunks[3] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>(simd8<uint8_t>(uint8x16_t(this->chunks[0])) >= mask,
+                          simd8<uint8_t>(uint8x16_t(this->chunks[1])) >= mask,
+                          simd8<uint8_t>(uint8x16_t(this->chunks[2])) >= mask,
+                          simd8<uint8_t>(uint8x16_t(this->chunks[3])) >= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+/* begin file src/simdutf/arm64/simd16-inl.h */
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>> struct base_u16 {
+  uint16x8_t value;
+  /// the size of vector in bytes
+  static const int SIZE = sizeof(value);
+  /// the number of elements of type T a vector can hold
+  static const int ELEMENTS = SIZE / sizeof(T);
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u16() = default;
+  simdutf_really_inline base_u16(const uint16x8_t _value) : value(_value) {}
+  simdutf_really_inline operator const uint16x8_t &() const {
+    return this->value;
+  }
+  simdutf_really_inline operator uint16x8_t &() { return this->value; }
+  // Bit operations
+  simdutf_really_inline simd16<T> operator|(const simd16<T> other) const {
+    return vorrq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator&(const simd16<T> other) const {
+    return vandq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator^(const simd16<T> other) const {
+    return veorq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> bit_andnot(const simd16<T> other) const {
+    return vbicq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
+  simdutf_really_inline simd16<T> &operator|=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd16<T> &operator&=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd16<T> &operator^=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return vceqq_u16(lhs, rhs);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return vextq_u18(prev_chunk, *this, 8 - N);
+  }
+};
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base_u16<T> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdutf_really_inline base16() : base_u16<T>() {}
+  simdutf_really_inline base16(const uint16x8_t _value) : base_u16<T>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr) : base16(vld1q_u16(ptr)) {}
+
+  static const int SIZE = sizeof(base_u16<T>::value);
+  void dump() const {
+    uint16_t temp[8];
+    vst1q_u16(temp, *this);
+    printf("[%04x, %04x, %04x, %04x, %04x, %04x, %04x, %04x]\n", temp[0],
+           temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], temp[7]);
+  }
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return vextq_u18(prev_chunk, *this, 8 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return vmovq_n_u16(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const uint16x8_t _value)
+      : base16<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return vmovq_n_u16(_value);
+  }
+  static simdutf_really_inline simd16<T> zero() { return vdupq_n_u16(0); }
+  static simdutf_really_inline simd16<T> load(const T values[8]) {
+    return vld1q_u16(reinterpret_cast<const uint16_t *>(values));
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const uint16x8_t _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return vst1q_u16(dst, *this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
+    return vaddq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
+    return vsubq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
+    *this = *this - other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Signed code units
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
+  simdutf_really_inline simd16(const uint16x8_t _value)
+      : base16_numeric<int16_t>(_value) {}
+#endif
+  simdutf_really_inline simd16(const int16x8_t _value)
+      : base16_numeric<int16_t>(vreinterpretq_u16_s16(_value)) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
+  simdutf_really_inline operator simd16<uint16_t>() const;
+  simdutf_really_inline operator const uint16x8_t &() const {
+    return this->value;
+  }
+  simdutf_really_inline operator const int16x8_t() const {
+    return vreinterpretq_s16_u16(this->value);
+  }
+
+  simdutf_really_inline int16_t max_val() const {
+    return vmaxvq_s16(vreinterpretq_s16_u16(this->value));
+  }
+  simdutf_really_inline int16_t min_val() const {
+    return vminvq_s16(vreinterpretq_s16_u16(this->value));
+  }
+  // Order-sensitive comparisons
+  simdutf_really_inline simd16<int16_t>
+  max_val(const simd16<int16_t> other) const {
+    return vmaxq_s16(vreinterpretq_s16_u16(this->value),
+                     vreinterpretq_s16_u16(other.value));
+  }
+  simdutf_really_inline simd16<int16_t>
+  min_val(const simd16<int16_t> other) const {
+    return vmaxq_s16(vreinterpretq_s16_u16(this->value),
+                     vreinterpretq_s16_u16(other.value));
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<int16_t> other) const {
+    return vcgtq_s16(vreinterpretq_s16_u16(this->value),
+                     vreinterpretq_s16_u16(other.value));
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<int16_t> other) const {
+    return vcltq_s16(vreinterpretq_s16_u16(this->value),
+                     vreinterpretq_s16_u16(other.value));
+  }
+};
+
+// Unsigned code units
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const uint16x8_t _value)
+      : base16_numeric<uint16_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+
+  simdutf_really_inline int16_t max_val() const { return vmaxvq_u16(*this); }
+  simdutf_really_inline int16_t min_val() const { return vminvq_u16(*this); }
+  // Saturated math
+  simdutf_really_inline simd16<uint16_t>
+  saturating_add(const simd16<uint16_t> other) const {
+    return vqaddq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  saturating_sub(const simd16<uint16_t> other) const {
+    return vqsubq_u16(*this, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return vmaxq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return vminq_u16(*this, other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  gt_bits(const simd16<uint16_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  lt_bits(const simd16<uint16_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return vcleq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return vcgeq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<uint16_t> other) const {
+    return vcgtq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return vcltq_u16(*this, other);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(vshrq_n_u16(*this, N));
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return simd16<uint16_t>(vshlq_n_u16(*this, N));
+  }
+
+  // logical operations
+  simdutf_really_inline simd16<uint16_t>
+  operator|(const simd16<uint16_t> other) const {
+    return vorrq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  operator&(const simd16<uint16_t> other) const {
+    return vandq_u16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  operator^(const simd16<uint16_t> other) const {
+    return veorq_u16(*this, other);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    return vqmovn_high_u16(vqmovn_u16(v0), v1);
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    return vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(*this)));
+  }
+  void dump() const {
+    uint16_t temp[8];
+    vst1q_u16(temp, *this);
+    printf("[%04x, %04x, %04x, %04x, %04x, %04x, %04x, %04x]\n", temp[0],
+           temp[1], temp[2], temp[3], temp[4], temp[5], temp[6], temp[7]);
+  }
+
+  simdutf_really_inline uint32_t sum() const { return vaddlvq_u16(value); }
+};
+simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "ARM kernel should use four registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline
+  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
+            const simd16<T> chunk2, const simd16<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
+    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
+    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+    const uint8x16_t bit_mask =
+        simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+    const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                 0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+    // Add each of the elements next to each other, successively, to stuff each
+    // 8 byte mask into one.
+    uint8x16_t sum0 = vpaddq_u8(
+        vreinterpretq_u8_u16(this->chunks[0] & vreinterpretq_u16_u8(bit_mask)),
+        vreinterpretq_u8_u16(this->chunks[1] & vreinterpretq_u16_u8(bit_mask)));
+    uint8x16_t sum1 = vpaddq_u8(
+        vreinterpretq_u8_u16(this->chunks[2] & vreinterpretq_u16_u8(bit_mask)),
+        vreinterpretq_u8_u16(this->chunks[3] & vreinterpretq_u16_u8(bit_mask)));
+    sum0 = vpaddq_u8(sum0, sum1);
+    sum0 = vpaddq_u8(sum0, sum0);
+    return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+    this->chunks[2] = this->chunks[2].swap_bytes();
+    this->chunks[3] = this->chunks[3].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                           this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                           this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+
+    return simd16x32<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+    return simd16x32<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
+               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
+               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                           this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+
+}; // struct simd16x32<T>
+template <>
+simdutf_really_inline uint64_t simd16x32<uint16_t>::not_in_range(
+    const uint16_t low, const uint16_t high) const {
+  const simd16<uint16_t> mask_low = simd16<uint16_t>::splat(low);
+  const simd16<uint16_t> mask_high = simd16<uint16_t>::splat(high);
+  simd16x32<uint16_t> x(simd16<uint16_t>((this->chunks[0] > mask_high) |
+                                         (this->chunks[0] < mask_low)),
+                        simd16<uint16_t>((this->chunks[1] > mask_high) |
+                                         (this->chunks[1] < mask_low)),
+                        simd16<uint16_t>((this->chunks[2] > mask_high) |
+                                         (this->chunks[2] < mask_low)),
+                        simd16<uint16_t>((this->chunks[3] > mask_high) |
+                                         (this->chunks[3] < mask_low)));
+  return x.to_bitmask();
+}
+
+simdutf_really_inline simd16<uint16_t> min(const simd16<uint16_t> a,
+                                           simd16<uint16_t> b) {
+  return vminq_u16(a.value, b.value);
+}
+/* end file src/simdutf/arm64/simd16-inl.h */
+/* begin file src/simdutf/arm64/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  static const size_t SIZE = sizeof(uint32x4_t);
+  static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  uint32x4_t value;
+
+  simdutf_really_inline simd32(const uint32x4_t v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd32(const Pointer *ptr)
+      : value(vld1q_u32(reinterpret_cast<const uint32_t *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const { return vaddvq_u32(value); }
+
+  simdutf_really_inline simd32<uint32_t> swap_bytes() const {
+    return vreinterpretq_u32_u8(vrev32q_u8(vreinterpretq_u8_u32(value)));
+  }
+
+  template <int N> simdutf_really_inline simd32<uint32_t> shr() const {
+    return vshrq_n_u32(value, N);
+  }
+
+  template <int N> simdutf_really_inline simd32<uint32_t> shl() const {
+    return vshlq_n_u32(value, N);
+  }
+
+  void dump() const {
+    uint32_t temp[4];
+    vst1q_u32(temp, value);
+    printf("[%08x, %08x, %08x, %08x]\n", temp[0], temp[1], temp[2], temp[3]);
+  }
+
+  // operators
+  simdutf_really_inline simd32 &operator+=(const simd32 other) {
+    value = vaddq_u32(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd32<uint32_t> zero() {
+    return vdupq_n_u32(0);
+  }
+
+  simdutf_really_inline static simd32<uint32_t> splat(uint32_t v) {
+    return vdupq_n_u32(v);
+  }
+};
+
+//----------------------------------------------------------------------
+
+template <> struct simd32<bool> {
+  uint32x4_t value;
+
+  simdutf_really_inline simd32(const uint32x4_t v) : value(v) {}
+
+  simdutf_really_inline bool any() const { return vmaxvq_u32(value) != 0; }
+};
+
+//----------------------------------------------------------------------
+
+template <typename T>
+simdutf_really_inline simd32<T> operator|(const simd32<T> a,
+                                          const simd32<T> b) {
+  return vorrq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> min(const simd32<uint32_t> a,
+                                           const simd32<uint32_t> b) {
+  return vminq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> max(const simd32<uint32_t> a,
+                                           const simd32<uint32_t> b) {
+  return vmaxq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator==(const simd32<uint32_t> a,
+                                              uint32_t b) {
+  return vceqq_u32(a.value, vdupq_n_u32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return vandq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return vandq_u32(a.value, vdupq_n_u32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator|(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return vorrq_u32(a.value, vdupq_n_u32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator+(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return vaddq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator-(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return vsubq_u32(a.value, vdupq_n_u32(b));
+}
+
+simdutf_really_inline simd32<bool> operator>=(const simd32<uint32_t> a,
+                                              const simd32<uint32_t> b) {
+  return vcgeq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator!(const simd32<bool> v) {
+  return vmvnq_u32(v.value);
+}
+
+simdutf_really_inline simd32<bool> operator>(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return vcgtq_u32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> select(const simd32<bool> cond,
+                                              const simd32<uint32_t> v_true,
+                                              const simd32<uint32_t> v_false) {
+  return vbslq_u32(cond.value, v_true.value, v_false.value);
+}
+/* end file src/simdutf/arm64/simd32-inl.h */
+/* begin file src/simdutf/arm64/simd64-inl.h */
+template <typename T> struct simd64;
+
+template <> struct simd64<uint64_t> {
+  uint64x2_t value;
+
+  simdutf_really_inline simd64(const uint64x2_t v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd64(const Pointer *ptr)
+      : value(vld1q_u64(reinterpret_cast<const uint64_t *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const { return vaddvq_u64(value); }
+
+  // operators
+  simdutf_really_inline simd64 &operator+=(const simd64 other) {
+    value = vaddq_u64(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd64<uint64_t> zero() {
+    return vdupq_n_u64(0);
+  }
+
+  simdutf_really_inline static simd64<uint64_t> splat(uint64_t v) {
+    return vdupq_n_u64(v);
+  }
+};
+/* end file src/simdutf/arm64/simd64-inl.h */
+
+simdutf_really_inline simd64<uint64_t> sum_8bytes(const simd8<uint8_t> v) {
+  // We do it as 3 instructions. There might be a faster way.
+  // We hope that these 3 instructions are cheap.
+  uint16x8_t first_sum = vpaddlq_u8(v);
+  uint32x4_t second_sum = vpaddlq_u16(first_sum);
+  return vpaddlq_u32(second_sum);
+}
+
+} // namespace simd
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdutf
+
+#endif // SIMDUTF_ARM64_SIMD_H
+/* end file src/simdutf/arm64/simd.h */
+
+/* begin file src/simdutf/arm64/end.h */
+/* end file src/simdutf/arm64/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_ARM64
+
+#endif // SIMDUTF_ARM64_H
+/* end file src/simdutf/arm64.h */
+/* begin file src/simdutf/icelake.h */
+#ifndef SIMDUTF_ICELAKE_H
+#define SIMDUTF_ICELAKE_H
+
+
+#ifdef __has_include
+  // How do we detect that a compiler supports vbmi2?
+  // For sure if the following header is found, we are ok?
+  #if __has_include(<avx512vbmi2intrin.h>)
+    #define SIMDUTF_COMPILER_SUPPORTS_VBMI2 1
+  #endif
+#endif
+
+#ifdef _MSC_VER
+  #if _MSC_VER >= 1930
+    // Visual Studio 2022 and up support VBMI2 under x64 even if the header
+    // avx512vbmi2intrin.h is not found.
+    // Visual Studio 2019 technically supports VBMI2, but the implementation
+    // might be unreliable. Search for visualstudio2019icelakeissue in our
+    // tests.
+    #define SIMDUTF_COMPILER_SUPPORTS_VBMI2 1
+  #endif
+#endif
+
+// We allow icelake on x64 as long as the compiler is known to support VBMI2.
+#ifndef SIMDUTF_IMPLEMENTATION_ICELAKE
+  #define SIMDUTF_IMPLEMENTATION_ICELAKE                                       \
+    ((SIMDUTF_IS_X86_64) && (SIMDUTF_COMPILER_SUPPORTS_VBMI2))
+#endif
+
+// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdutf/simdutf/issues/1247
+#if ((SIMDUTF_IMPLEMENTATION_ICELAKE) && (SIMDUTF_IS_X86_64) && (__AVX2__) &&  \
+     (SIMDUTF_HAS_AVX512F && SIMDUTF_HAS_AVX512DQ && SIMDUTF_HAS_AVX512VL &&   \
+      SIMDUTF_HAS_AVX512VBMI2) &&                                              \
+     (!SIMDUTF_IS_32BITS))
+  #define SIMDUTF_CAN_ALWAYS_RUN_ICELAKE 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_ICELAKE 0
+#endif
+
+#if SIMDUTF_IMPLEMENTATION_ICELAKE
+  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+    #define SIMDUTF_TARGET_ICELAKE
+  #else
+    #define SIMDUTF_TARGET_ICELAKE                                             \
+      SIMDUTF_TARGET_REGION(                                                   \
+          "avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,"         \
+          "avx512vl,avx2,bmi,bmi2,pclmul,lzcnt,popcnt,avx512vpopcntdq")
+  #endif
+
+namespace simdutf {
+namespace icelake {} // namespace icelake
+} // namespace simdutf
+
+  //
+  // These two need to be included outside SIMDUTF_TARGET_REGION
+  //
+/* begin file src/simdutf/icelake/intrinsics.h */
+#ifndef SIMDUTF_ICELAKE_INTRINSICS_H
+#define SIMDUTF_ICELAKE_INTRINSICS_H
+
+
+#ifdef SIMDUTF_VISUAL_STUDIO
+  // under clang within visual studio, this will include <x86intrin.h>
+  #include <intrin.h> // visual studio or clang
+  #include <immintrin.h>
+#else
+
+  #if SIMDUTF_GCC11ORMORE
+// We should not get warnings while including <x86intrin.h> yet we do
+// under some versions of GCC.
+// If the x86intrin.h header has uninitialized values that are problematic,
+// it is a GCC issue, we want to ignore these warnings.
+SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
+  #endif
+
+  #include <x86intrin.h> // elsewhere
+
+  #if SIMDUTF_GCC11ORMORE
+// cancels the suppression of the -Wuninitialized
+SIMDUTF_POP_DISABLE_WARNINGS
+  #endif
+
+  #ifndef _tzcnt_u64
+    #define _tzcnt_u64(x) __tzcnt_u64(x)
+  #endif // _tzcnt_u64
+#endif   // SIMDUTF_VISUAL_STUDIO
+
+#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
+  /**
+   * You are not supposed, normally, to include these
+   * headers directly. Instead you should either include intrin.h
+   * or x86intrin.h. However, when compiling with clang
+   * under Windows (i.e., when _MSC_VER is set), these headers
+   * only get included *if* the corresponding features are detected
+   * from macros:
+   * e.g., if __AVX2__ is set... in turn,  we normally set these
+   * macros by compiling against the corresponding architecture
+   * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+   * software with these advanced instructions. In simdutf, we
+   * want to compile the whole program for a generic target,
+   * and only target our specific kernels. As a workaround,
+   * we directly include the needed headers. These headers would
+   * normally guard against such usage, but we carefully included
+   * <x86intrin.h>  (or <intrin.h>) before, so the headers
+   * are fooled.
+   */
+  #include <bmiintrin.h>   // for _blsr_u64
+  #include <bmi2intrin.h>  // for _pext_u64, _pdep_u64
+  #include <lzcntintrin.h> // for  __lzcnt64
+  #include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+  #include <smmintrin.h>
+  #include <tmmintrin.h>
+  #include <avxintrin.h>
+  #include <avx2intrin.h>
+  // Important: we need the AVX-512 headers:
+  #include <avx512fintrin.h>
+  #include <avx512dqintrin.h>
+  #include <avx512cdintrin.h>
+  #include <avx512bwintrin.h>
+  #include <avx512vlintrin.h>
+  #include <avx512vlbwintrin.h>
+  #include <avx512vbmiintrin.h>
+  #include <avx512vbmi2intrin.h>
+  #include <avx512vpopcntdqintrin.h>
+  #include <avx512vpopcntdqvlintrin.h>
+  // unfortunately, we may not get _blsr_u64, but, thankfully, clang
+  // has it as a macro.
+  #ifndef _blsr_u64
+    // we roll our own
+    #define _blsr_u64(n) ((n - 1) & n)
+  #endif //  _blsr_u64
+#endif   // SIMDUTF_CLANG_VISUAL_STUDIO
+
+#if defined(__GNUC__) && !defined(__clang__)
+
+  #if __GNUC__ == 8
+    #define SIMDUTF_GCC8 1
+  #elif __GNUC__ == 9
+    #define SIMDUTF_GCC9 1
+  #endif //  __GNUC__ == 8 || __GNUC__ == 9
+
+#endif // defined(__GNUC__) && !defined(__clang__)
+
+#if SIMDUTF_GCC8
+  #pragma GCC push_options
+  #pragma GCC target("avx512f")
+/**
+ * GCC 8 fails to provide _mm512_set_epi8. We roll our own.
+ */
+inline __m512i
+_mm512_set_epi8(uint8_t a0, uint8_t a1, uint8_t a2, uint8_t a3, uint8_t a4,
+                uint8_t a5, uint8_t a6, uint8_t a7, uint8_t a8, uint8_t a9,
+                uint8_t a10, uint8_t a11, uint8_t a12, uint8_t a13, uint8_t a14,
+                uint8_t a15, uint8_t a16, uint8_t a17, uint8_t a18, uint8_t a19,
+                uint8_t a20, uint8_t a21, uint8_t a22, uint8_t a23, uint8_t a24,
+                uint8_t a25, uint8_t a26, uint8_t a27, uint8_t a28, uint8_t a29,
+                uint8_t a30, uint8_t a31, uint8_t a32, uint8_t a33, uint8_t a34,
+                uint8_t a35, uint8_t a36, uint8_t a37, uint8_t a38, uint8_t a39,
+                uint8_t a40, uint8_t a41, uint8_t a42, uint8_t a43, uint8_t a44,
+                uint8_t a45, uint8_t a46, uint8_t a47, uint8_t a48, uint8_t a49,
+                uint8_t a50, uint8_t a51, uint8_t a52, uint8_t a53, uint8_t a54,
+                uint8_t a55, uint8_t a56, uint8_t a57, uint8_t a58, uint8_t a59,
+                uint8_t a60, uint8_t a61, uint8_t a62, uint8_t a63) {
+  return _mm512_set_epi64(
+      uint64_t(a7) + (uint64_t(a6) << 8) + (uint64_t(a5) << 16) +
+          (uint64_t(a4) << 24) + (uint64_t(a3) << 32) + (uint64_t(a2) << 40) +
+          (uint64_t(a1) << 48) + (uint64_t(a0) << 56),
+      uint64_t(a15) + (uint64_t(a14) << 8) + (uint64_t(a13) << 16) +
+          (uint64_t(a12) << 24) + (uint64_t(a11) << 32) +
+          (uint64_t(a10) << 40) + (uint64_t(a9) << 48) + (uint64_t(a8) << 56),
+      uint64_t(a23) + (uint64_t(a22) << 8) + (uint64_t(a21) << 16) +
+          (uint64_t(a20) << 24) + (uint64_t(a19) << 32) +
+          (uint64_t(a18) << 40) + (uint64_t(a17) << 48) + (uint64_t(a16) << 56),
+      uint64_t(a31) + (uint64_t(a30) << 8) + (uint64_t(a29) << 16) +
+          (uint64_t(a28) << 24) + (uint64_t(a27) << 32) +
+          (uint64_t(a26) << 40) + (uint64_t(a25) << 48) + (uint64_t(a24) << 56),
+      uint64_t(a39) + (uint64_t(a38) << 8) + (uint64_t(a37) << 16) +
+          (uint64_t(a36) << 24) + (uint64_t(a35) << 32) +
+          (uint64_t(a34) << 40) + (uint64_t(a33) << 48) + (uint64_t(a32) << 56),
+      uint64_t(a47) + (uint64_t(a46) << 8) + (uint64_t(a45) << 16) +
+          (uint64_t(a44) << 24) + (uint64_t(a43) << 32) +
+          (uint64_t(a42) << 40) + (uint64_t(a41) << 48) + (uint64_t(a40) << 56),
+      uint64_t(a55) + (uint64_t(a54) << 8) + (uint64_t(a53) << 16) +
+          (uint64_t(a52) << 24) + (uint64_t(a51) << 32) +
+          (uint64_t(a50) << 40) + (uint64_t(a49) << 48) + (uint64_t(a48) << 56),
+      uint64_t(a63) + (uint64_t(a62) << 8) + (uint64_t(a61) << 16) +
+          (uint64_t(a60) << 24) + (uint64_t(a59) << 32) +
+          (uint64_t(a58) << 40) + (uint64_t(a57) << 48) +
+          (uint64_t(a56) << 56));
+}
+  #pragma GCC pop_options
+#endif // SIMDUTF_GCC8
+
+#endif // SIMDUTF_HASWELL_INTRINSICS_H
+/* end file src/simdutf/icelake/intrinsics.h */
+/* begin file src/simdutf/icelake/implementation.h */
+#ifndef SIMDUTF_ICELAKE_IMPLEMENTATION_H
+#define SIMDUTF_ICELAKE_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace icelake {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation(
+            "icelake",
+            "Intel AVX512 (AVX-512BW, AVX-512CD, AVX-512VL, AVX-512VBMI2 "
+            "extensions)",
+            internal::instruction_set::AVX2 | internal::instruction_set::BMI1 |
+                internal::instruction_set::BMI2 |
+                internal::instruction_set::AVX512BW |
+                internal::instruction_set::AVX512CD |
+                internal::instruction_set::AVX512VL |
+                internal::instruction_set::AVX512VBMI2 |
+                internal::instruction_set::AVX512VPOPCNTDQ) {}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace icelake
+} // namespace simdutf
+
+#endif // SIMDUTF_ICELAKE_IMPLEMENTATION_H
+/* end file src/simdutf/icelake/implementation.h */
+
+  //
+  // The rest need to be inside the region
+  //
+/* begin file src/simdutf/icelake/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "icelake"
+// #define SIMDUTF_IMPLEMENTATION icelake
+
+#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+// nothing needed.
+#else
+SIMDUTF_TARGET_ICELAKE
+#endif
+
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+// clang-format off
+SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
+// clang-format on
+#endif // end of workaround
+/* end file src/simdutf/icelake/begin.h */
+  // Declarations
+/* begin file src/simdutf/icelake/bitmanipulation.h */
+#ifndef SIMDUTF_ICELAKE_BITMANIPULATION_H
+#define SIMDUTF_ICELAKE_BITMANIPULATION_H
+
+namespace simdutf {
+namespace icelake {
+namespace {
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdutf_really_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+// simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+//   #if SIMDUTF_REGULAR_VISUAL_STUDIO
+//   return (int)_tzcnt_u64(input_num);
+//   #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
+//   return __builtin_ctzll(input_num);
+//   #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+// }
+#endif
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdutf
+
+#endif // SIMDUTF_ICELAKE_BITMANIPULATION_H
+/* end file src/simdutf/icelake/bitmanipulation.h */
+/* begin file src/simdutf/icelake/simd.h */
+#ifndef SIMDUTF_ICELAKE_SIMD_H
+#define SIMDUTF_ICELAKE_SIMD_H
+
+namespace simdutf {
+namespace icelake {
+namespace {
+namespace simd {
+
+/* begin file src/simdutf/icelake/simd16-inl.h */
+template <typename T> struct simd16;
+
+template <> struct simd16<uint16_t> {
+  static const size_t SIZE = sizeof(__m512i);
+  static const size_t ELEMENTS = SIZE / sizeof(uint16_t);
+
+  template <typename Pointer>
+  static simdutf_really_inline simd16<uint16_t> load(const Pointer *ptr) {
+    return simd16<uint16_t>(ptr);
+  }
+
+  __m512i value;
+
+  simdutf_really_inline simd16(const __m512i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd16(const Pointer *ptr)
+      : value(_mm512_loadu_si512(reinterpret_cast<const __m512i *>(ptr))) {}
+
+  // operators
+  simdutf_really_inline simd16 &operator+=(const simd16 other) {
+    value = _mm512_add_epi32(value, other.value);
+    return *this;
+  }
+
+  simdutf_really_inline simd16 &operator-=(const simd16 other) {
+    value = _mm512_sub_epi32(value, other.value);
+    return *this;
+  }
+
+  // methods
+  simdutf_really_inline simd16 swap_bytes() const {
+    const __m512i byteflip = _mm512_setr_epi64(
+        0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+        0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+        0x0607040502030001, 0x0e0f0c0d0a0b0809);
+
+    return _mm512_shuffle_epi8(value, byteflip);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    const auto lo = _mm512_and_si512(value, _mm512_set1_epi32(0xffff));
+    const auto hi = _mm512_srli_epi32(value, 16);
+    const auto sum32 = _mm512_add_epi32(lo, hi);
+
+    return _mm512_reduce_add_epi32(sum32);
+  }
+
+  // static members
+  simdutf_really_inline static simd16<uint16_t> zero() {
+    return _mm512_setzero_si512();
+  }
+
+  simdutf_really_inline static simd16<uint16_t> splat(uint16_t v) {
+    return _mm512_set1_epi16(v);
+  }
+};
+
+template <> struct simd16<bool> {
+  __mmask32 value;
+
+  simdutf_really_inline simd16(const __mmask32 v) : value(v) {}
+};
+
+// ------------------------------------------------------------
+
+simdutf_really_inline simd16<uint16_t> min(const simd16<uint16_t> b,
+                                           const simd16<uint16_t> a) {
+  return _mm512_min_epu16(a.value, b.value);
+}
+
+simdutf_really_inline simd16<uint16_t> operator&(const simd16<uint16_t> a,
+                                                 uint16_t b) {
+  return _mm512_and_si512(a.value, _mm512_set1_epi16(b));
+}
+
+simdutf_really_inline simd16<uint16_t> operator^(const simd16<uint16_t> a,
+                                                 uint16_t b) {
+  return _mm512_xor_si512(a.value, _mm512_set1_epi16(b));
+}
+
+simdutf_really_inline simd16<uint16_t> operator^(const simd16<uint16_t> a,
+                                                 const simd16<uint16_t> b) {
+  return _mm512_xor_si512(a.value, b.value);
+}
+
+simdutf_really_inline simd16<bool> operator==(const simd16<uint16_t> a,
+                                              uint16_t b) {
+  return _mm512_cmpeq_epi16_mask(a.value, _mm512_set1_epi16(b));
+}
+/* end file src/simdutf/icelake/simd16-inl.h */
+/* begin file src/simdutf/icelake/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  static const size_t SIZE = sizeof(__m512i);
+  static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  __m512i value;
+
+  simdutf_really_inline simd32(const __m512i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd32(const Pointer *ptr)
+      : value(_mm512_loadu_si512(reinterpret_cast<const __m512i *>(ptr))) {}
+
+  uint64_t sum() const {
+    const __m512i mask = _mm512_set1_epi64(0xffffffff);
+    const __m512i t0 = _mm512_and_si512(value, mask);
+    const __m512i t1 = _mm512_srli_epi64(value, 32);
+    const __m512i t2 = _mm512_add_epi64(t0, t1);
+    return _mm512_reduce_add_epi64(t2);
+  }
+
+  // operators
+  simdutf_really_inline simd32 &operator+=(const simd32 other) {
+    value = _mm512_add_epi32(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd32<uint32_t> zero() {
+    return _mm512_setzero_si512();
+  }
+
+  simdutf_really_inline static simd32<uint32_t> splat(uint32_t v) {
+    return _mm512_set1_epi32(v);
+  }
+};
+
+simdutf_really_inline simd32<uint32_t> min(const simd32<uint32_t> b,
+                                           const simd32<uint32_t> a) {
+  return _mm512_min_epu32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> b,
+                                                 const simd32<uint32_t> a) {
+  return _mm512_and_si512(a.value, b.value);
+}
+/* end file src/simdutf/icelake/simd32-inl.h */
+
+} // namespace simd
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdutf
+
+#endif // SIMDUTF_ICELAKE_SIMD_H
+/* end file src/simdutf/icelake/simd.h */
+
+/* begin file src/simdutf/icelake/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
+
+
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+SIMDUTF_POP_DISABLE_WARNINGS
+#endif // end of workaround
+/* end file src/simdutf/icelake/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_ICELAKE
+#endif // SIMDUTF_ICELAKE_H
+/* end file src/simdutf/icelake.h */
+/* begin file src/simdutf/haswell.h */
+#ifndef SIMDUTF_HASWELL_H
+#define SIMDUTF_HASWELL_H
+
+#ifdef SIMDUTF_WESTMERE_H
+  #error "haswell.h must be included before westmere.h"
+#endif
+#ifdef SIMDUTF_FALLBACK_H
+  #error "haswell.h must be included before fallback.h"
+#endif
+
+
+// Default Haswell to on if this is x86-64. Even if we are not compiled for it,
+// it could be selected at runtime.
+#ifndef SIMDUTF_IMPLEMENTATION_HASWELL
+  //
+  // You do not want to restrict it like so: SIMDUTF_IS_X86_64 && __AVX2__
+  // because we want to rely on *runtime dispatch*.
+  //
+  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+    #define SIMDUTF_IMPLEMENTATION_HASWELL 0
+  #else
+    #define SIMDUTF_IMPLEMENTATION_HASWELL (SIMDUTF_IS_X86_64)
+  #endif
+
+#endif
+// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdutf/simdutf/issues/1247
+#if ((SIMDUTF_IMPLEMENTATION_HASWELL) && (SIMDUTF_IS_X86_64) && (__AVX2__))
+  #define SIMDUTF_CAN_ALWAYS_RUN_HASWELL 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_HASWELL 0
+#endif
+
+#if SIMDUTF_IMPLEMENTATION_HASWELL
+
+  #define SIMDUTF_TARGET_HASWELL SIMDUTF_TARGET_REGION("avx2,bmi,lzcnt,popcnt")
+
+namespace simdutf {
+/**
+ * Implementation for Haswell (Intel AVX2).
+ */
+namespace haswell {} // namespace haswell
+} // namespace simdutf
+
+  //
+  // These two need to be included outside SIMDUTF_TARGET_REGION
+  //
+/* begin file src/simdutf/haswell/implementation.h */
+#ifndef SIMDUTF_HASWELL_IMPLEMENTATION_H
+#define SIMDUTF_HASWELL_IMPLEMENTATION_H
+
+
+// The constructor may be executed on any host, so we take care not to use
+// SIMDUTF_TARGET_REGION
+namespace simdutf {
+namespace haswell {
+
+using namespace simdutf;
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("haswell", "Intel/AMD AVX2",
+                                internal::instruction_set::AVX2 |
+                                    internal::instruction_set::BMI1 |
+                                    internal::instruction_set::BMI2) {}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace haswell
+} // namespace simdutf
+
+#endif // SIMDUTF_HASWELL_IMPLEMENTATION_H
+/* end file src/simdutf/haswell/implementation.h */
+/* begin file src/simdutf/haswell/intrinsics.h */
+#ifndef SIMDUTF_HASWELL_INTRINSICS_H
+#define SIMDUTF_HASWELL_INTRINSICS_H
+
+
+#ifdef SIMDUTF_VISUAL_STUDIO
+  // under clang within visual studio, this will include <x86intrin.h>
+  #include <intrin.h> // visual studio or clang
+#else
+
+  #if SIMDUTF_GCC11ORMORE
+// We should not get warnings while including <x86intrin.h> yet we do
+// under some versions of GCC.
+// If the x86intrin.h header has uninitialized values that are problematic,
+// it is a GCC issue, we want to ignore these warnings.
+SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
+  #endif
+
+  #include <x86intrin.h> // elsewhere
+
+  #if SIMDUTF_GCC11ORMORE
+// cancels the suppression of the -Wuninitialized
+SIMDUTF_POP_DISABLE_WARNINGS
+  #endif
+
+#endif // SIMDUTF_VISUAL_STUDIO
+
+#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
+  /**
+   * You are not supposed, normally, to include these
+   * headers directly. Instead you should either include intrin.h
+   * or x86intrin.h. However, when compiling with clang
+   * under Windows (i.e., when _MSC_VER is set), these headers
+   * only get included *if* the corresponding features are detected
+   * from macros:
+   * e.g., if __AVX2__ is set... in turn,  we normally set these
+   * macros by compiling against the corresponding architecture
+   * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+   * software with these advanced instructions. In simdutf, we
+   * want to compile the whole program for a generic target,
+   * and only target our specific kernels. As a workaround,
+   * we directly include the needed headers. These headers would
+   * normally guard against such usage, but we carefully included
+   * <x86intrin.h>  (or <intrin.h>) before, so the headers
+   * are fooled.
+   */
+  #include <bmiintrin.h>   // for _blsr_u64
+  #include <lzcntintrin.h> // for  __lzcnt64
+  #include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+  #include <smmintrin.h>
+  #include <tmmintrin.h>
+  #include <avxintrin.h>
+  #include <avx2intrin.h>
+  // unfortunately, we may not get _blsr_u64, but, thankfully, clang
+  // has it as a macro.
+  #ifndef _blsr_u64
+    // we roll our own
+    #define _blsr_u64(n) (((n) - 1) & (n))
+  #endif //  _blsr_u64
+  // Same issue with _blsmsk_u32:
+  #ifndef _blsmsk_u32
+    // we roll our own
+    #define _blsmsk_u32(n) (((n) - 1) ^ (n))
+  #endif //  _blsmsk_u32
+#endif   // SIMDUTF_CLANG_VISUAL_STUDIO
+
+#endif // SIMDUTF_HASWELL_INTRINSICS_H
+/* end file src/simdutf/haswell/intrinsics.h */
+
+  //
+  // The rest need to be inside the region
+  //
+/* begin file src/simdutf/haswell/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "haswell"
+// #define SIMDUTF_IMPLEMENTATION haswell
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
+
+#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
+// nothing needed.
+#else
+SIMDUTF_TARGET_HASWELL
+#endif
+
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+// clang-format off
+SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
+// clang-format on
+#endif // end of workaround
+/* end file src/simdutf/haswell/begin.h */
+  // Declarations
+/* begin file src/simdutf/haswell/bitmanipulation.h */
+#ifndef SIMDUTF_HASWELL_BITMANIPULATION_H
+#define SIMDUTF_HASWELL_BITMANIPULATION_H
+
+namespace simdutf {
+namespace haswell {
+namespace {
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdutf_really_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+  #if SIMDUTF_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+}
+#endif
+
+template <typename T> bool is_power_of_two(T x) { return (x & (x - 1)) == 0; }
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+
+#endif // SIMDUTF_HASWELL_BITMANIPULATION_H
+/* end file src/simdutf/haswell/bitmanipulation.h */
+/* begin file src/simdutf/haswell/simd.h */
+#ifndef SIMDUTF_HASWELL_SIMD_H
+#define SIMDUTF_HASWELL_SIMD_H
+
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace simd {
+
+// Forward-declared so they can be used by splat and friends.
+template <typename Child> struct base {
+  __m256i value;
+
+  // Zero constructor
+  simdutf_really_inline base() : value{__m256i()} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base(const __m256i _value) : value(_value) {}
+  // Conversion to SIMD register
+  simdutf_really_inline operator const __m256i &() const { return this->value; }
+  simdutf_really_inline operator __m256i &() { return this->value; }
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this));
+    __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this, 1));
+    if (big_endian) {
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      first = _mm256_shuffle_epi8(first, swap);
+      second = _mm256_shuffle_epi8(second, swap);
+    }
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first);
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second);
+  }
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr),
+                        _mm256_cvtepu8_epi32(_mm256_castsi256_si128(*this)));
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 8),
+                        _mm256_cvtepu8_epi32(_mm256_castsi256_si128(
+                            _mm256_srli_si256(*this, 8))));
+    _mm256_storeu_si256(
+        reinterpret_cast<__m256i *>(ptr + 16),
+        _mm256_cvtepu8_epi32(_mm256_extractf128_si256(*this, 1)));
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 24),
+                        _mm256_cvtepu8_epi32(_mm_srli_si128(
+                            _mm256_extractf128_si256(*this, 1), 8)));
+  }
+  // Bit operations
+  simdutf_really_inline Child operator|(const Child other) const {
+    return _mm256_or_si256(*this, other);
+  }
+  simdutf_really_inline Child operator&(const Child other) const {
+    return _mm256_and_si256(*this, other);
+  }
+  simdutf_really_inline Child operator^(const Child other) const {
+    return _mm256_xor_si256(*this, other);
+  }
+  simdutf_really_inline Child bit_andnot(const Child other) const {
+    return _mm256_andnot_si256(other, *this);
+  }
+  simdutf_really_inline Child &operator|=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator&=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator^=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+};
+
+// Forward-declared so they can be used by splat and friends.
+template <typename T> struct simd8;
+
+template <typename T, typename Mask = simd8<bool>>
+struct base8 : base<simd8<T>> {
+  typedef uint32_t bitmask_t;
+  typedef uint64_t bitmask2_t;
+
+  simdutf_really_inline base8() : base<simd8<T>>() {}
+  simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+  simdutf_really_inline T first() const {
+    return _mm256_extract_epi8(*this, 0);
+  }
+  simdutf_really_inline T last() const {
+    return _mm256_extract_epi8(*this, 31);
+  }
+  friend simdutf_always_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return _mm256_cmpeq_epi8(lhs, rhs);
+  }
+
+  static const int SIZE = sizeof(base<T>::value);
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    return _mm256_alignr_epi8(
+        *this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return _mm256_set1_epi8(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8() : base8() {}
+  simdutf_really_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+  simdutf_really_inline uint32_t to_bitmask() const {
+    return uint32_t(_mm256_movemask_epi8(*this));
+  }
+  simdutf_really_inline bool any() const {
+    return !_mm256_testz_si256(*this, *this);
+  }
+  simdutf_really_inline bool none() const {
+    return _mm256_testz_si256(*this, *this);
+  }
+  simdutf_really_inline bool all() const {
+    return static_cast<uint32_t>(_mm256_movemask_epi8(*this)) == 0xFFFFFFFF;
+  }
+  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  static simdutf_really_inline simd8<T> splat(T _value) {
+    return _mm256_set1_epi8(_value);
+  }
+  static simdutf_really_inline simd8<T> zero() {
+    return _mm256_setzero_si256();
+  }
+  static simdutf_really_inline simd8<T> load(const T values[32]) {
+    return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
+  }
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                  T v5, T v6, T v7, T v8, T v9,
+                                                  T v10, T v11, T v12, T v13,
+                                                  T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                    v12, v13, v14, v15);
+  }
+
+  simdutf_really_inline base8_numeric() : base8<T>() {}
+  simdutf_really_inline base8_numeric(const __m256i _value)
+      : base8<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[32]) const {
+    return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this);
+  }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
+    return _mm256_add_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
+    return _mm256_sub_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
+    *this = *this + other;
+    return *static_cast<simd8<T> *>(this);
+  }
+  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
+    *this = *this - other;
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return _mm256_shuffle_epi8(lookup_table, *this);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
+  simdutf_really_inline simd8(const __m256i _value)
+      : base8_numeric<int8_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+  simdutf_really_inline operator simd8<uint8_t>() const;
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+        int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+        int8_t v12, int8_t v13, int8_t v14, int8_t v15, int8_t v16, int8_t v17,
+        int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+        int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29,
+        int8_t v30, int8_t v31)
+      : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                               v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+                               v22, v23, v24, v25, v26, v27, v28, v29, v30,
+                               v31)) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                         v10, v11, v12, v13, v14, v15);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return _mm256_movemask_epi8(*this) == 0;
+  }
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
+    return _mm256_max_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
+    return _mm256_min_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return _mm256_cmpgt_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return _mm256_cmpgt_epi8(other, *this);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
+  simdutf_really_inline simd8(const __m256i _value)
+      : base8_numeric<uint8_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+        uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20,
+        uint8_t v21, uint8_t v22, uint8_t v23, uint8_t v24, uint8_t v25,
+        uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30,
+        uint8_t v31)
+      : simd8(_mm256_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                               v12, v13, v14, v15, v16, v17, v18, v19, v20, v21,
+                               v22, v23, v24, v25, v26, v27, v28, v29, v30,
+                               v31)) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                          v10, v11, v12, v13, v14, v15);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return _mm256_adds_epu8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return _mm256_subs_epu8(*this, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return _mm256_max_epu8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return _mm256_min_epu8(other, *this);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdutf_really_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return this->lt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> bits_not_set() const {
+    return *this == uint8_t(0);
+  }
+  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return _mm256_movemask_epi8(*this) == 0;
+  }
+  simdutf_really_inline bool bits_not_set_anywhere() const {
+    return _mm256_testz_si256(*this, *this);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+    return _mm256_testz_si256(*this, bits);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N);
+  }
+  // Get one of the bits and make a bitmask out of it.
+  // e.g. value.get_bit<7>() gets the high bit
+  template <int N> simdutf_really_inline int get_bit() const {
+    return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7 - N));
+  }
+
+  simdutf_really_inline uint64_t sum_bytes() const {
+    const auto tmp = _mm256_sad_epu8(value, _mm256_setzero_si256());
+
+    return _mm256_extract_epi64(tmp, 0) + _mm256_extract_epi64(tmp, 1) +
+           _mm256_extract_epi64(tmp, 2) + _mm256_extract_epi64(tmp, 3);
+  }
+};
+simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 2,
+                "Haswell kernel should use two registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1)
+      : chunks{chunk0, chunk1} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r_hi = this->chunks[1].to_bitmask();
+    return r_lo | (r_hi << 32);
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return this->chunks[0] | this->chunks[1];
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
+  }
+
+  simdutf_really_inline simd8x64<T> bit_or(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<T>(this->chunks[0] | mask, this->chunks[1] | mask);
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
+                          this->chunks[1] == other.chunks[1])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+
+    return simd8x64<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+    return simd8x64<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>((simd8<uint8_t>(__m256i(this->chunks[0])) >= mask),
+                          (simd8<uint8_t>(__m256i(this->chunks[1])) >= mask))
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+/* begin file src/simdutf/haswell/simd16-inl.h */
+#ifdef __GNUC__
+  #if __GNUC__ < 8
+    #define _mm256_set_m128i(xmm1, xmm2)                                       \
+      _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1),                  \
+                                _mm256_castsi128_si256(xmm2), 2)
+    #define _mm256_setr_m128i(xmm2, xmm1)                                      \
+      _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1),                  \
+                                _mm256_castsi128_si256(xmm2), 2)
+  #endif
+#endif
+
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base<simd16<T>> {
+  using bitmask_type = uint32_t;
+
+  simdutf_really_inline base16() : base<simd16<T>>() {}
+  simdutf_really_inline base16(const __m256i _value)
+      : base<simd16<T>>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr)
+      : base16(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(ptr))) {}
+  friend simdutf_always_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return _mm256_cmpeq_epi16(lhs, rhs);
+  }
+
+  /// the size of vector in bytes
+  static const int SIZE = sizeof(base<simd16<T>>::value);
+
+  /// the number of elements of type T a vector can hold
+  static const int ELEMENTS = SIZE / sizeof(T);
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return _mm256_alignr_epi8(*this, prev_chunk, 16 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return _mm256_set1_epi16(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const __m256i _value) : base16<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+
+  simdutf_really_inline bitmask_type to_bitmask() const {
+    return _mm256_movemask_epi8(*this);
+  }
+  simdutf_really_inline bool any() const {
+    return !_mm256_testz_si256(*this, *this);
+  }
+  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return _mm256_set1_epi16(_value);
+  }
+  static simdutf_really_inline simd16<T> zero() {
+    return _mm256_setzero_si256();
+  }
+  static simdutf_really_inline simd16<T> load(const T values[8]) {
+    return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const __m256i _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
+    return _mm256_add_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
+    return _mm256_sub_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
+    *this = *this - other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Signed code units
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+  simdutf_really_inline simd16(const __m256i _value)
+      : base16_numeric<int16_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
+  // Order-sensitive comparisons
+  simdutf_really_inline simd16<int16_t>
+  max_val(const simd16<int16_t> other) const {
+    return _mm256_max_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<int16_t>
+  min_val(const simd16<int16_t> other) const {
+    return _mm256_min_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<int16_t> other) const {
+    return _mm256_cmpgt_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<int16_t> other) const {
+    return _mm256_cmpgt_epi16(other, *this);
+  }
+};
+
+// Unsigned code units
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const __m256i _value)
+      : base16_numeric<uint16_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+  simdutf_really_inline simd16(const simd16<bool> bm) : simd16(bm.value) {}
+
+  // Saturated math
+  simdutf_really_inline simd16<uint16_t>
+  saturating_add(const simd16<uint16_t> other) const {
+    return _mm256_adds_epu16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  saturating_sub(const simd16<uint16_t> other) const {
+    return _mm256_subs_epu16(*this, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return _mm256_max_epu16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return _mm256_min_epu16(*this, other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  gt_bits(const simd16<uint16_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  lt_bits(const simd16<uint16_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<uint16_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+
+  simdutf_really_inline bool bits_not_set_anywhere() const {
+    return _mm256_testz_si256(*this, *this);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdutf_really_inline bool
+  bits_not_set_anywhere(simd16<uint16_t> bits) const {
+    return _mm256_testz_si256(*this, bits);
+  }
+  simdutf_really_inline bool
+  any_bits_set_anywhere(simd16<uint16_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(_mm256_srli_epi16(*this, N));
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return simd16<uint16_t>(_mm256_slli_epi16(*this, N));
+  }
+  // Get one of the bits and make a bitmask out of it.
+  // e.g. value.get_bit<7>() gets the high bit
+  template <int N> simdutf_really_inline int get_bit() const {
+    return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 15 - N));
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    const __m256i swap = _mm256_setr_epi8(
+        1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+        21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+    return _mm256_shuffle_epi8(*this, swap);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    // Note: the AVX2 variant of pack operates on 128-bit lanes, thus
+    //       we have to shuffle lanes in order to produce bytes in the
+    //       correct order.
+
+    // get the 0th lanes
+    const __m128i lo_0 = _mm256_extracti128_si256(v0, 0);
+    const __m128i lo_1 = _mm256_extracti128_si256(v1, 0);
+
+    // get the 1st lanes
+    const __m128i hi_0 = _mm256_extracti128_si256(v0, 1);
+    const __m128i hi_1 = _mm256_extracti128_si256(v1, 1);
+
+    // build new vectors (shuffle lanes)
+    const __m256i t0 = _mm256_set_m128i(lo_1, lo_0);
+    const __m256i t1 = _mm256_set_m128i(hi_1, hi_0);
+
+    // pack code units in linear order from v0 and v1
+    return _mm256_packus_epi16(t0, t1);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    const auto lo_u16 = _mm256_and_si256(value, _mm256_set1_epi32(0x0000ffff));
+    const auto hi_u16 = _mm256_srli_epi32(value, 16);
+    const auto sum_u32 = _mm256_add_epi32(lo_u16, hi_u16);
+
+    const auto lo_u32 =
+        _mm256_and_si256(sum_u32, _mm256_set1_epi64x(0xffffffff));
+    const auto hi_u32 = _mm256_srli_epi64(sum_u32, 32);
+    const auto sum_u64 = _mm256_add_epi64(lo_u32, hi_u32);
+
+    return uint64_t(_mm256_extract_epi64(sum_u64, 0)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 1)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 2)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 3));
+  }
+};
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(NUM_CHUNKS == 2,
+                "Haswell kernel should use two registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd16x32(const simd16<T> chunk0,
+                                  const simd16<T> chunk1)
+      : chunks{chunk0, chunk1} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r_hi = this->chunks[1].to_bitmask();
+    return r_lo | (r_hi << 32);
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return this->chunks[0] | this->chunks[1];
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>));
+  }
+
+  simdutf_really_inline simd16x32<T> bit_or(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<T>(this->chunks[0] | mask, this->chunks[1] | mask);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
+    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
+                           this->chunks[1] == other.chunks[1])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+
+    return simd16x32<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
+    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
+    return simd16x32<bool>(
+               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
+               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
+        .to_bitmask();
+  }
+}; // struct simd16x32<T>
+
+simd16<uint16_t> min(const simd16<uint16_t> a, simd16<uint16_t> b) {
+  return _mm256_min_epu16(a.value, b.value);
+}
+/* end file src/simdutf/haswell/simd16-inl.h */
+/* begin file src/simdutf/haswell/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  static const size_t SIZE = sizeof(__m256i);
+  static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  __m256i value;
+
+  simdutf_really_inline simd32(const __m256i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd32(const Pointer *ptr)
+      : value(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const {
+    const __m256i mask = _mm256_set1_epi64x(0xffffffff);
+    const __m256i t0 = _mm256_and_si256(value, mask);
+    const __m256i t1 = _mm256_srli_epi64(value, 32);
+    const __m256i t2 = _mm256_add_epi64(t0, t1);
+
+    return uint64_t(_mm256_extract_epi64(t2, 0)) +
+           uint64_t(_mm256_extract_epi64(t2, 1)) +
+           uint64_t(_mm256_extract_epi64(t2, 2)) +
+           uint64_t(_mm256_extract_epi64(t2, 3));
+  }
+
+  simdutf_really_inline simd32<uint32_t> swap_bytes() const {
+    const __m256i shuffle =
+        _mm256_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12,
+                         3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12);
+
+    return _mm256_shuffle_epi8(value, shuffle);
+  }
+
+  // operators
+  simdutf_really_inline simd32 &operator+=(const simd32 other) {
+    value = _mm256_add_epi32(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd32<uint32_t> zero() {
+    return _mm256_setzero_si256();
+  }
+
+  simdutf_really_inline static simd32<uint32_t> splat(uint32_t v) {
+    return _mm256_set1_epi32(v);
+  }
+};
+
+//----------------------------------------------------------------------
+
+template <> struct simd32<bool> {
+  // static const size_t SIZE = sizeof(__m128i);
+  // static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  __m256i value;
+
+  simdutf_really_inline simd32(const __m256i v) : value(v) {}
+
+  simdutf_really_inline bool any() const {
+    return _mm256_movemask_epi8(value) != 0;
+  }
+};
+
+//----------------------------------------------------------------------
+
+template <typename T>
+simdutf_really_inline simd32<T> operator|(const simd32<T> a,
+                                          const simd32<T> b) {
+  return _mm256_or_si256(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> min(const simd32<uint32_t> b,
+                                           const simd32<uint32_t> a) {
+  return _mm256_min_epu32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> max(const simd32<uint32_t> a,
+                                           const simd32<uint32_t> b) {
+  return _mm256_max_epu32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> b,
+                                                 const simd32<uint32_t> a) {
+  return _mm256_and_si256(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator+(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return _mm256_add_epi32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator>=(const simd32<uint32_t> a,
+                                              const simd32<uint32_t> b) {
+  return _mm256_cmpeq_epi32(_mm256_max_epu32(a.value, b.value), a.value);
+}
+
+simdutf_really_inline simd32<bool> operator!(const simd32<bool> v) {
+  return _mm256_xor_si256(v.value, _mm256_set1_epi8(-1));
+}
+
+simdutf_really_inline simd32<bool> operator>(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return !(b >= a);
+}
+/* end file src/simdutf/haswell/simd32-inl.h */
+/* begin file src/simdutf/haswell/simd64-inl.h */
+template <typename T> struct simd64;
+
+template <> struct simd64<uint64_t> {
+  // static const size_t SIZE = sizeof(__m256i);
+  // static const size_t ELEMENTS = SIZE / sizeof(uint64_t);
+
+  __m256i value;
+
+  simdutf_really_inline simd64(const __m256i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd64(const Pointer *ptr)
+      : value(_mm256_loadu_si256(reinterpret_cast<const __m256i *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const {
+    return _mm256_extract_epi64(value, 0) + _mm256_extract_epi64(value, 1) +
+           _mm256_extract_epi64(value, 2) + _mm256_extract_epi64(value, 3);
+  }
+
+  // operators
+  simdutf_really_inline simd64 &operator+=(const simd64 other) {
+    value = _mm256_add_epi64(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd64<uint64_t> zero() {
+    return _mm256_setzero_si256();
+  }
+
+  simdutf_really_inline static simd64<uint64_t> splat(uint64_t v) {
+    return _mm256_set1_epi64x(v);
+  }
+};
+/* end file src/simdutf/haswell/simd64-inl.h */
+
+simdutf_really_inline simd64<uint64_t> sum_8bytes(const simd8<uint8_t> v) {
+  return _mm256_sad_epu8(v.value, simd8<uint8_t>::zero());
+}
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+
+#endif // SIMDUTF_HASWELL_SIMD_H
+/* end file src/simdutf/haswell/simd.h */
+
+/* begin file src/simdutf/haswell/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
+
+#undef SIMDUTF_SIMD_HAS_BYTEMASK
+
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+SIMDUTF_POP_DISABLE_WARNINGS
+#endif // end of workaround
+/* end file src/simdutf/haswell/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_HASWELL
+#endif // SIMDUTF_HASWELL_COMMON_H
+/* end file src/simdutf/haswell.h */
+/* begin file src/simdutf/westmere.h */
+#ifndef SIMDUTF_WESTMERE_H
+#define SIMDUTF_WESTMERE_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "westmere.h must be included before fallback.h"
+#endif
+
+
+// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
+#ifndef SIMDUTF_IMPLEMENTATION_WESTMERE
+  //
+  // You do not want to set it to (SIMDUTF_IS_X86_64 &&
+  // !SIMDUTF_REQUIRES_HASWELL) because you want to rely on runtime dispatch!
+  //
+  #if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE || SIMDUTF_CAN_ALWAYS_RUN_HASWELL
+    #define SIMDUTF_IMPLEMENTATION_WESTMERE 0
+  #else
+    #define SIMDUTF_IMPLEMENTATION_WESTMERE (SIMDUTF_IS_X86_64)
+  #endif
+
+#endif
+
+#if (SIMDUTF_IMPLEMENTATION_WESTMERE && SIMDUTF_IS_X86_64 && __SSE4_2__)
+  #define SIMDUTF_CAN_ALWAYS_RUN_WESTMERE 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_WESTMERE 0
+#endif
+
+#if SIMDUTF_IMPLEMENTATION_WESTMERE
+
+  #define SIMDUTF_TARGET_WESTMERE SIMDUTF_TARGET_REGION("sse4.2,popcnt")
+
+namespace simdutf {
+/**
+ * Implementation for Westmere (Intel SSE4.2).
+ */
+namespace westmere {} // namespace westmere
+} // namespace simdutf
+
+  //
+  // These two need to be included outside SIMDUTF_TARGET_REGION
+  //
+/* begin file src/simdutf/westmere/implementation.h */
+#ifndef SIMDUTF_WESTMERE_IMPLEMENTATION_H
+#define SIMDUTF_WESTMERE_IMPLEMENTATION_H
+
+
+// The constructor may be executed on any host, so we take care not to use
+// SIMDUTF_TARGET_REGION
+namespace simdutf {
+namespace westmere {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("westmere", "Intel/AMD SSE4.2",
+                                internal::instruction_set::SSE42) {}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace westmere
+} // namespace simdutf
+
+#endif // SIMDUTF_WESTMERE_IMPLEMENTATION_H
+/* end file src/simdutf/westmere/implementation.h */
+/* begin file src/simdutf/westmere/intrinsics.h */
+#ifndef SIMDUTF_WESTMERE_INTRINSICS_H
+#define SIMDUTF_WESTMERE_INTRINSICS_H
+
+#ifdef SIMDUTF_VISUAL_STUDIO
+  // under clang within visual studio, this will include <x86intrin.h>
+  #include <intrin.h> // visual studio or clang
+#else
+
+  #if SIMDUTF_GCC11ORMORE
+// We should not get warnings while including <x86intrin.h> yet we do
+// under some versions of GCC.
+// If the x86intrin.h header has uninitialized values that are problematic,
+// it is a GCC issue, we want to ignore these warnings.
+SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
+  #endif
+
+  #include <x86intrin.h> // elsewhere
+
+  #if SIMDUTF_GCC11ORMORE
+// cancels the suppression of the -Wuninitialized
+SIMDUTF_POP_DISABLE_WARNINGS
+  #endif
+
+#endif // SIMDUTF_VISUAL_STUDIO
+
+#ifdef SIMDUTF_CLANG_VISUAL_STUDIO
+  /**
+   * You are not supposed, normally, to include these
+   * headers directly. Instead you should either include intrin.h
+   * or x86intrin.h. However, when compiling with clang
+   * under Windows (i.e., when _MSC_VER is set), these headers
+   * only get included *if* the corresponding features are detected
+   * from macros:
+   */
+  #include <smmintrin.h> // for _mm_alignr_epi8
+#endif
+
+#endif // SIMDUTF_WESTMERE_INTRINSICS_H
+/* end file src/simdutf/westmere/intrinsics.h */
+
+  //
+  // The rest need to be inside the region
+  //
+/* begin file src/simdutf/westmere/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "westmere"
+// #define SIMDUTF_IMPLEMENTATION westmere
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
+
+#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
+// nothing needed.
+#else
+SIMDUTF_TARGET_WESTMERE
+#endif
+/* end file src/simdutf/westmere/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/westmere/bitmanipulation.h */
+#ifndef SIMDUTF_WESTMERE_BITMANIPULATION_H
+#define SIMDUTF_WESTMERE_BITMANIPULATION_H
+
+namespace simdutf {
+namespace westmere {
+namespace {
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+simdutf_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdutf_really_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+  #if SIMDUTF_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+  #else  // SIMDUTF_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+  #endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+}
+#endif
+
+template <typename T> bool is_power_of_two(T x) { return (x & (x - 1)) == 0; }
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+
+#endif // SIMDUTF_WESTMERE_BITMANIPULATION_H
+/* end file src/simdutf/westmere/bitmanipulation.h */
+/* begin file src/simdutf/westmere/simd.h */
+#ifndef SIMDUTF_WESTMERE_SIMD_H
+#define SIMDUTF_WESTMERE_SIMD_H
+
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace simd {
+
+template <typename Child> struct base {
+  __m128i value;
+
+  // Zero constructor
+  simdutf_really_inline base() : value{__m128i()} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base(const __m128i _value) : value(_value) {}
+  // Conversion to SIMD register
+  simdutf_really_inline operator const __m128i &() const { return this->value; }
+  simdutf_really_inline operator __m128i &() { return this->value; }
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
+    __m128i first = _mm_cvtepu8_epi16(*this);
+    __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this, 8));
+    if (big_endian) {
+      const __m128i swap =
+          _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+      first = _mm_shuffle_epi8(first, swap);
+      second = _mm_shuffle_epi8(second, swap);
+    }
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first);
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8), second);
+  }
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p), _mm_cvtepu8_epi32(*this));
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 4),
+                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 4)));
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 8),
+                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 8)));
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(p + 12),
+                     _mm_cvtepu8_epi32(_mm_srli_si128(*this, 12)));
+  }
+  // Bit operations
+  simdutf_really_inline Child operator|(const Child other) const {
+    return _mm_or_si128(*this, other);
+  }
+  simdutf_really_inline Child operator&(const Child other) const {
+    return _mm_and_si128(*this, other);
+  }
+  simdutf_really_inline Child operator^(const Child other) const {
+    return _mm_xor_si128(*this, other);
+  }
+  simdutf_really_inline Child bit_andnot(const Child other) const {
+    return _mm_andnot_si128(other, *this);
+  }
+  simdutf_really_inline Child &operator|=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator&=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator^=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+};
+
+// Forward-declared so they can be used by splat and friends.
+template <typename T> struct simd8;
+
+template <typename T, typename Mask = simd8<bool>>
+struct base8 : base<simd8<T>> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdutf_really_inline T first() const { return _mm_extract_epi8(*this, 0); }
+  simdutf_really_inline T last() const { return _mm_extract_epi8(*this, 15); }
+  simdutf_really_inline base8() : base<simd8<T>>() {}
+  simdutf_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return _mm_cmpeq_epi8(lhs, rhs);
+  }
+
+  static const int SIZE = sizeof(base<simd8<T>>::value);
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return _mm_set1_epi8(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8() : base8() {}
+  simdutf_really_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+  simdutf_really_inline int to_bitmask() const {
+    return _mm_movemask_epi8(*this);
+  }
+  simdutf_really_inline bool any() const {
+    return !_mm_testz_si128(*this, *this);
+  }
+  simdutf_really_inline bool none() const {
+    return _mm_testz_si128(*this, *this);
+  }
+  simdutf_really_inline bool all() const {
+    return _mm_movemask_epi8(*this) == 0xFFFF;
+  }
+  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  static simdutf_really_inline simd8<T> splat(T _value) {
+    return _mm_set1_epi8(_value);
+  }
+  static simdutf_really_inline simd8<T> zero() { return _mm_setzero_si128(); }
+  static simdutf_really_inline simd8<T> load(const T values[16]) {
+    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+  }
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                  T v5, T v6, T v7, T v8, T v9,
+                                                  T v10, T v11, T v12, T v13,
+                                                  T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15);
+  }
+
+  simdutf_really_inline base8_numeric() : base8<T>() {}
+  simdutf_really_inline base8_numeric(const __m128i _value)
+      : base8<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[16]) const {
+    return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
+    return _mm_add_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
+    return _mm_sub_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
+    *this = *this + other;
+    return *static_cast<simd8<T> *>(this);
+  }
+  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
+    *this = *this - other;
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return _mm_shuffle_epi8(lookup_table, *this);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
+  simdutf_really_inline simd8(const __m128i _value)
+      : base8_numeric<int8_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                            v12, v13, v14, v15)) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15);
+  }
+  simdutf_really_inline operator simd8<uint8_t>() const;
+  simdutf_really_inline bool is_ascii() const {
+    return _mm_movemask_epi8(*this) == 0;
+  }
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
+    return _mm_max_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
+    return _mm_min_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return _mm_cmpgt_epi8(*this, other);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return _mm_cmpgt_epi8(other, *this);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
+  simdutf_really_inline simd8(const __m128i _value)
+      : base8_numeric<uint8_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8(_mm_setr_epi8(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                            v12, v13, v14, v15)) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return _mm_adds_epu8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return _mm_subs_epu8(*this, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return _mm_max_epu8(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return _mm_min_epu8(*this, other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdutf_really_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> bits_not_set() const {
+    return *this == uint8_t(0);
+  }
+  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return _mm_movemask_epi8(*this) == 0;
+  }
+
+  simdutf_really_inline bool bits_not_set_anywhere() const {
+    return _mm_testz_si128(*this, *this);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+    return _mm_testz_si128(*this, bits);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N);
+  }
+  // Get one of the bits and make a bitmask out of it.
+  // e.g. value.get_bit<7>() gets the high bit
+  template <int N> simdutf_really_inline int get_bit() const {
+    return _mm_movemask_epi8(_mm_slli_epi16(*this, 7 - N));
+  }
+
+  simdutf_really_inline uint64_t sum_bytes() const {
+    const auto tmp = _mm_sad_epu8(value, _mm_setzero_si128());
+    return _mm_extract_epi64(tmp, 0) + _mm_extract_epi64(tmp, 1);
+  }
+};
+
+simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "Westmere kernel should use four registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                 const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    this->chunks[2] |= other.chunks[2];
+    this->chunks[3] |= other.chunks[3];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 2);
+    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                          this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
+                          this->chunks[1] == other.chunks[1],
+                          this->chunks[2] == other.chunks[2],
+                          this->chunks[3] == other.chunks[3])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                          this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+
+    return simd8x64<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low - 1);
+    const simd8<T> mask_high = simd8<T>::splat(high + 1);
+    return simd8x64<bool>(
+               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
+               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
+               (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
+               (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                          this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
+                          this->chunks[2] > mask, this->chunks[3] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
+                          this->chunks[2] >= mask, this->chunks[3] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>(simd8<uint8_t>(__m128i(this->chunks[0])) >= mask,
+                          simd8<uint8_t>(__m128i(this->chunks[1])) >= mask,
+                          simd8<uint8_t>(__m128i(this->chunks[2])) >= mask,
+                          simd8<uint8_t>(__m128i(this->chunks[3])) >= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+/* begin file src/simdutf/westmere/simd16-inl.h */
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base<simd16<T>> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdutf_really_inline base16() : base<simd16<T>>() {}
+  simdutf_really_inline base16(const __m128i _value)
+      : base<simd16<T>>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr)
+      : base16(_mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr))) {}
+
+  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return _mm_cmpeq_epi16(lhs, rhs);
+  }
+
+  /// the size of vector in bytes
+  static const int SIZE = sizeof(base<simd16<T>>::value);
+
+  /// the number of elements of type T a vector can hold
+  static const int ELEMENTS = SIZE / sizeof(T);
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return _mm_set1_epi16(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const __m128i _value) : base16<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+
+  simdutf_really_inline int to_bitmask() const {
+    return _mm_movemask_epi8(*this);
+  }
+  simdutf_really_inline bool any() const {
+    return !_mm_testz_si128(*this, *this);
+  }
+  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return _mm_set1_epi16(_value);
+  }
+  static simdutf_really_inline simd16<T> zero() { return _mm_setzero_si128(); }
+  static simdutf_really_inline simd16<T> load(const T values[8]) {
+    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const __m128i _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
+    return _mm_add_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
+    return _mm_sub_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
+    *this = *this - other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Signed code units
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+  simdutf_really_inline simd16(const __m128i _value)
+      : base16_numeric<int16_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
+  // Member-by-member initialization
+  simdutf_really_inline simd16(int16_t v0, int16_t v1, int16_t v2, int16_t v3,
+                               int16_t v4, int16_t v5, int16_t v6, int16_t v7)
+      : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
+  simdutf_really_inline operator simd16<uint16_t>() const;
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd16<int16_t>
+  max_val(const simd16<int16_t> other) const {
+    return _mm_max_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<int16_t>
+  min_val(const simd16<int16_t> other) const {
+    return _mm_min_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<int16_t> other) const {
+    return _mm_cmpgt_epi16(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<int16_t> other) const {
+    return _mm_cmpgt_epi16(other, *this);
+  }
+};
+
+// Unsigned code units
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const __m128i _value)
+      : base16_numeric<uint16_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+  simdutf_really_inline simd16(const simd16<bool> bm) : simd16(bm.value) {}
+  // Member-by-member initialization
+  simdutf_really_inline simd16(uint16_t v0, uint16_t v1, uint16_t v2,
+                               uint16_t v3, uint16_t v4, uint16_t v5,
+                               uint16_t v6, uint16_t v7)
+      : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd16<uint16_t>
+  repeat_16(uint16_t v0, uint16_t v1, uint16_t v2, uint16_t v3, uint16_t v4,
+            uint16_t v5, uint16_t v6, uint16_t v7) {
+    return simd16<uint16_t>(v0, v1, v2, v3, v4, v5, v6, v7);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd16<uint16_t>
+  saturating_add(const simd16<uint16_t> other) const {
+    return _mm_adds_epu16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  saturating_sub(const simd16<uint16_t> other) const {
+    return _mm_subs_epu16(*this, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return _mm_max_epu16(*this, other);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return _mm_min_epu16(*this, other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  gt_bits(const simd16<uint16_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  lt_bits(const simd16<uint16_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<uint16_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+
+  simdutf_really_inline bool bits_not_set_anywhere() const {
+    return _mm_testz_si128(*this, *this);
+  }
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdutf_really_inline bool
+  bits_not_set_anywhere(simd16<uint16_t> bits) const {
+    return _mm_testz_si128(*this, bits);
+  }
+  simdutf_really_inline bool
+  any_bits_set_anywhere(simd16<uint16_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(_mm_srli_epi16(*this, N));
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return simd16<uint16_t>(_mm_slli_epi16(*this, N));
+  }
+  // Get one of the bits and make a bitmask out of it.
+  // e.g. value.get_bit<7>() gets the high bit
+  template <int N> simdutf_really_inline int get_bit() const {
+    return _mm_movemask_epi8(_mm_slli_epi16(*this, 7 - N));
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    const __m128i swap =
+        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+    return _mm_shuffle_epi8(*this, swap);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    return _mm_packus_epi16(v0, v1);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    const auto lo_u16 = _mm_and_si128(value, _mm_set1_epi32(0x0000ffff));
+    const auto hi_u16 = _mm_srli_epi32(value, 16);
+    const auto sum_u32 = _mm_add_epi32(lo_u16, hi_u16);
+
+    const auto lo_u32 = _mm_and_si128(sum_u32, _mm_set1_epi64x(0xffffffff));
+    const auto hi_u32 = _mm_srli_epi64(sum_u32, 32);
+    const auto sum_u64 = _mm_add_epi64(lo_u32, hi_u32);
+
+    return uint64_t(_mm_extract_epi64(sum_u64, 0)) +
+           uint64_t(_mm_extract_epi64(sum_u64, 1));
+  }
+};
+
+simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "Westmere kernel should use four registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline
+  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
+            const simd16<T> chunk2, const simd16<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
+    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
+    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+    this->chunks[2] = this->chunks[2].swap_bytes();
+    this->chunks[3] = this->chunks[3].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                           this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
+    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
+                           this->chunks[1] == other.chunks[1],
+                           this->chunks[2] == other.chunks[2],
+                           this->chunks[3] == other.chunks[3])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                           this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+
+    return simd16x32<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
+    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
+    return simd16x32<bool>(
+               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
+               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
+               (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
+               (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                           this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+}; // struct simd16x32<T>
+
+simd16<uint16_t> min(const simd16<uint16_t> a, simd16<uint16_t> b) {
+  return _mm_min_epu16(a.value, b.value);
+}
+/* end file src/simdutf/westmere/simd16-inl.h */
+/* begin file src/simdutf/westmere/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  static const size_t SIZE = sizeof(__m128i);
+  static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  __m128i value;
+
+  simdutf_really_inline simd32(const __m128i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd32(const Pointer *ptr)
+      : value(_mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const {
+    return uint64_t(_mm_extract_epi32(value, 0)) +
+           uint64_t(_mm_extract_epi32(value, 1)) +
+           uint64_t(_mm_extract_epi32(value, 2)) +
+           uint64_t(_mm_extract_epi32(value, 3));
+  }
+
+  simdutf_really_inline simd32<uint32_t> swap_bytes() const {
+    const __m128i shuffle =
+        _mm_setr_epi8(3, 2, 1, 0, 7, 6, 5, 4, 8, 9, 10, 11, 15, 14, 13, 12);
+
+    return _mm_shuffle_epi8(value, shuffle);
+  }
+
+  template <int N> simdutf_really_inline simd32<uint32_t> shr() const {
+    return _mm_srli_epi32(value, N);
+  }
+
+  template <int N> simdutf_really_inline simd32<uint32_t> shl() const {
+    return _mm_slli_epi32(value, N);
+  }
+
+  void dump() const {
+    printf("[%08x, %08x, %08x, %08x]\n", uint32_t(_mm_extract_epi32(value, 0)),
+           uint32_t(_mm_extract_epi32(value, 1)),
+           uint32_t(_mm_extract_epi32(value, 2)),
+           uint32_t(_mm_extract_epi32(value, 3)));
+  }
+
+  // operators
+  simdutf_really_inline simd32 &operator+=(const simd32 other) {
+    value = _mm_add_epi32(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd32<uint32_t> zero() {
+    return _mm_setzero_si128();
+  }
+
+  simdutf_really_inline static simd32<uint32_t> splat(uint32_t v) {
+    return _mm_set1_epi32(v);
+  }
+};
+
+//----------------------------------------------------------------------
+
+template <> struct simd32<bool> {
+  // static const size_t SIZE = sizeof(__m128i);
+  // static const size_t ELEMENTS = SIZE / sizeof(uint32_t);
+
+  __m128i value;
+
+  simdutf_really_inline simd32(const __m128i v) : value(v) {}
+
+  simdutf_really_inline bool any() const {
+    return _mm_movemask_epi8(value) != 0;
+  }
+
+  simdutf_really_inline uint8_t to_4bit_bitmask() const {
+    return uint8_t(_mm_movemask_ps(_mm_castsi128_ps(value)));
+  }
+};
+
+//----------------------------------------------------------------------
+
+template <typename T>
+simdutf_really_inline simd32<T> operator|(const simd32<T> a,
+                                          const simd32<T> b) {
+  return _mm_or_si128(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> min(const simd32<uint32_t> a,
+                                           const simd32<uint32_t> b) {
+  return _mm_min_epu32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> max(const simd32<uint32_t> a,
+                                           const simd32<uint32_t> b) {
+  return _mm_max_epu32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator==(const simd32<uint32_t> a,
+                                              uint32_t b) {
+  return _mm_cmpeq_epi32(a.value, _mm_set1_epi32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return _mm_and_si128(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return _mm_and_si128(a.value, _mm_set1_epi32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator|(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return _mm_or_si128(a.value, _mm_set1_epi32(b));
+}
+
+simdutf_really_inline simd32<uint32_t> operator+(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return _mm_add_epi32(a.value, b.value);
+}
+
+simdutf_really_inline simd32<uint32_t> operator-(const simd32<uint32_t> a,
+                                                 uint32_t b) {
+  return _mm_sub_epi32(a.value, _mm_set1_epi32(b));
+}
+
+simdutf_really_inline simd32<bool> operator>=(const simd32<uint32_t> a,
+                                              const simd32<uint32_t> b) {
+  return _mm_cmpeq_epi32(_mm_max_epu32(a.value, b.value), a.value);
+}
+
+simdutf_really_inline simd32<bool> operator!(const simd32<bool> v) {
+  return _mm_xor_si128(v.value, _mm_set1_epi8(-1));
+}
+
+simdutf_really_inline simd32<bool> operator>(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return !(b >= a);
+}
+
+simdutf_really_inline simd32<uint32_t> select(const simd32<bool> cond,
+                                              const simd32<uint32_t> v_true,
+                                              const simd32<uint32_t> v_false) {
+  return _mm_blendv_epi8(v_false.value, v_true.value, cond.value);
+}
+/* end file src/simdutf/westmere/simd32-inl.h */
+/* begin file src/simdutf/westmere/simd64-inl.h */
+template <typename T> struct simd64;
+
+template <> struct simd64<uint64_t> {
+  // static const size_t SIZE = sizeof(__m128i);
+  // static const size_t ELEMENTS = SIZE / sizeof(uint64_t);
+
+  __m128i value;
+
+  simdutf_really_inline simd64(const __m128i v) : value(v) {}
+
+  template <typename Pointer>
+  simdutf_really_inline simd64(const Pointer *ptr)
+      : value(_mm_loadu_si128(reinterpret_cast<const __m128i *>(ptr))) {}
+
+  simdutf_really_inline uint64_t sum() const {
+    return _mm_extract_epi64(value, 0) + _mm_extract_epi64(value, 1);
+  }
+
+  // operators
+  simdutf_really_inline simd64 &operator+=(const simd64 other) {
+    value = _mm_add_epi64(value, other.value);
+    return *this;
+  }
+
+  // static members
+  simdutf_really_inline static simd64<uint64_t> zero() {
+    return _mm_setzero_si128();
+  }
+
+  simdutf_really_inline static simd64<uint64_t> splat(uint64_t v) {
+    return _mm_set1_epi64x(v);
+  }
+};
+/* end file src/simdutf/westmere/simd64-inl.h */
+
+simdutf_really_inline simd64<uint64_t> sum_8bytes(const simd8<uint8_t> v) {
+  return _mm_sad_epu8(v.value, simd8<uint8_t>::zero());
+}
+
+simdutf_really_inline simd8<uint8_t> as_vector_u8(const simd32<uint32_t> v) {
+  return simd8<uint8_t>(v.value);
+}
+
+} // namespace simd
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+
+#endif // SIMDUTF_WESTMERE_SIMD_INPUT_H
+/* end file src/simdutf/westmere/simd.h */
+
+/* begin file src/simdutf/westmere/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
+
+#undef SIMDUTF_SIMD_HAS_BYTEMASK
+/* end file src/simdutf/westmere/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_WESTMERE
+#endif // SIMDUTF_WESTMERE_COMMON_H
+/* end file src/simdutf/westmere.h */
+/* begin file src/simdutf/ppc64.h */
+#ifndef SIMDUTF_PPC64_H
+#define SIMDUTF_PPC64_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "ppc64.h must be included before fallback.h"
+#endif
+
+
+#ifndef SIMDUTF_IMPLEMENTATION_PPC64
+  #define SIMDUTF_IMPLEMENTATION_PPC64 (SIMDUTF_IS_PPC64)
+#endif
+#define SIMDUTF_CAN_ALWAYS_RUN_PPC64                                           \
+  SIMDUTF_IMPLEMENTATION_PPC64 &&SIMDUTF_IS_PPC64
+
+
+#if SIMDUTF_IMPLEMENTATION_PPC64
+
+namespace simdutf {
+/**
+ * Implementation for ALTIVEC (PPC64).
+ */
+namespace ppc64 {} // namespace ppc64
+} // namespace simdutf
+
+/* begin file src/simdutf/ppc64/implementation.h */
+#ifndef SIMDUTF_PPC64_IMPLEMENTATION_H
+#define SIMDUTF_PPC64_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace ppc64 {
+
+namespace {
+using namespace simdutf;
+
+template <size_t N> simdutf_really_inline size_t align_down(size_t size) {
+  return N * (size / N);
+}
+} // namespace
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("ppc64", "PPC64 ALTIVEC",
+                                internal::instruction_set::ALTIVEC) {}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused size_t maximal_binary_length_from_base64(
+      const char *input, size_t length) const noexcept;
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+
+#ifdef SIMDUTF_INTERNAL_TESTS
+  virtual std::vector<TestProcedure> internal_tests() const override;
+#endif
+};
+
+} // namespace ppc64
+} // namespace simdutf
+
+#endif // SIMDUTF_PPC64_IMPLEMENTATION_H
+/* end file src/simdutf/ppc64/implementation.h */
+
+/* begin file src/simdutf/ppc64/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
+// #define SIMDUTF_IMPLEMENTATION ppc64
+/* end file src/simdutf/ppc64/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/ppc64/intrinsics.h */
+#ifndef SIMDUTF_PPC64_INTRINSICS_H
+#define SIMDUTF_PPC64_INTRINSICS_H
+
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <altivec.h>
+
+// These are defined by altivec.h in GCC toolchain, it is safe to undef them.
+#ifdef bool
+  #undef bool
+#endif
+
+#ifdef vector
+  #undef vector
+#endif
+
+#endif //  SIMDUTF_PPC64_INTRINSICS_H
+/* end file src/simdutf/ppc64/intrinsics.h */
+/* begin file src/simdutf/ppc64/bitmanipulation.h */
+#ifndef SIMDUTF_PPC64_BITMANIPULATION_H
+#define SIMDUTF_PPC64_BITMANIPULATION_H
+
+namespace simdutf {
+namespace ppc64 {
+namespace {
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+simdutf_really_inline int count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdutf_really_inline int count_ones(uint64_t input_num) {
+  return __builtin_popcountll(input_num);
+}
+#endif
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+#endif
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+
+#endif // SIMDUTF_PPC64_BITMANIPULATION_H
+/* end file src/simdutf/ppc64/bitmanipulation.h */
+/* begin file src/simdutf/ppc64/simd.h */
+#ifndef SIMDUTF_PPC64_SIMD_H
+#define SIMDUTF_PPC64_SIMD_H
+
+#include <type_traits>
+
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace simd {
+
+using vec_bool_t = __vector __bool char;
+using vec_bool16_t = __vector __bool short;
+using vec_bool32_t = __vector __bool int;
+using vec_u8_t = __vector unsigned char;
+using vec_i8_t = __vector signed char;
+using vec_u16_t = __vector unsigned short;
+using vec_i16_t = __vector signed short;
+using vec_u32_t = __vector unsigned int;
+using vec_i32_t = __vector signed int;
+using vec_u64_t = __vector unsigned long long;
+using vec_i64_t = __vector signed long long;
+
+// clang-format off
+template <typename T> struct vector_u8_type_for_element_aux {
+  using type = typename std::conditional<std::is_same<T, bool>::value,    vec_bool_t,
+               typename std::conditional<std::is_same<T, uint8_t>::value, vec_u8_t,
+               typename std::conditional<std::is_same<T, int8_t>::value,  vec_i8_t, void>::type>::type>::type;
+
+  static_assert(not std::is_same<type, void>::value,
+                "accepted element types are 8 bit integers or bool");
+};
+
+template <typename T> struct vector_u16_type_for_element_aux {
+  using type = typename std::conditional<std::is_same<T, bool>::value,     vec_bool16_t,
+               typename std::conditional<std::is_same<T, uint16_t>::value, vec_u16_t,
+               typename std::conditional<std::is_same<T, int16_t>::value,  vec_i16_t, void>::type>::type>::type;
+
+  static_assert(not std::is_same<type, void>::value,
+                "accepted element types are 16 bit integers or bool");
+};
+
+template <typename T> struct vector_u32_type_for_element_aux {
+  using type = typename std::conditional<std::is_same<T, bool>::value,     vec_bool32_t,
+               typename std::conditional<std::is_same<T, uint32_t>::value, vec_u32_t,
+               typename std::conditional<std::is_same<T, int32_t>::value,  vec_i32_t, void>::type>::type>::type;
+
+  static_assert(not std::is_same<type, void>::value,
+                "accepted element types are 32 bit integers or bool");
+};
+// clang-format on
+
+template <typename T>
+using vector_u8_type_for_element =
+    typename vector_u8_type_for_element_aux<T>::type;
+
+template <typename T>
+using vector_u16_type_for_element =
+    typename vector_u16_type_for_element_aux<T>::type;
+
+template <typename T>
+using vector_u32_type_for_element =
+    typename vector_u32_type_for_element_aux<T>::type;
+
+template <typename T> uint16_t move_mask_u8(T vec) {
+  const vec_u8_t perm_mask = {15 * 8, 14 * 8, 13 * 8, 12 * 8, 11 * 8, 10 * 8,
+                              9 * 8,  8 * 8,  7 * 8,  6 * 8,  5 * 8,  4 * 8,
+                              3 * 8,  2 * 8,  1 * 8,  0 * 8};
+
+  const auto result = (vec_u64_t)vec_vbpermq((vec_u8_t)vec, perm_mask);
+#if SIMDUTF_IS_BIG_ENDIAN
+  return static_cast<uint16_t>(result[0]);
+#else
+  return static_cast<uint16_t>(result[1]);
+#endif
+}
+
+/* begin file src/simdutf/ppc64/simd8-inl.h */
+// file included directly
+
+template <typename T> struct base8 {
+  using vector_type = vector_u8_type_for_element<T>;
+  vector_type value;
+  static const int SIZE = sizeof(vector_type);
+  static const int ELEMENTS = sizeof(vector_type) / sizeof(T);
+
+  // Zero constructor
+  simdutf_really_inline base8() : value{vec_splats(T(0))} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base8(const vector_type _value) : value{_value} {}
+
+  // Splat scalar
+  simdutf_really_inline base8(T v) : value{vec_splats(v)} {}
+
+  // Conversion to SIMD register
+  simdutf_really_inline operator const vector_type &() const {
+    return this->value;
+  }
+
+  template <typename U> simdutf_really_inline void store(U *ptr) const {
+    vec_xst(value, 0, reinterpret_cast<T *>(ptr));
+  }
+
+  template <typename SIMD8> void operator|=(const SIMD8 other) {
+    this->value = vec_or(this->value, other.value);
+  }
+
+  template <int N = 1> vector_type prev_aux(vector_type prev_chunk) const {
+    vector_type chunk = this->value;
+#if !SIMDUTF_IS_BIG_ENDIAN
+    chunk = (vector_type)vec_reve(this->value);
+    prev_chunk = (vector_type)vec_reve((vector_type)prev_chunk);
+#endif
+    chunk = (vector_type)vec_sld((vector_type)prev_chunk, (vector_type)chunk,
+                                 16 - N);
+#if !SIMDUTF_IS_BIG_ENDIAN
+    chunk = (vector_type)vec_reve((vector_type)chunk);
+#endif
+    return chunk;
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return move_mask_u8(this->value) == 0;
+  }
+
+  simdutf_really_inline uint16_t to_bitmask() const {
+    return move_mask_u8(value);
+  }
+
+  template <endianness big_endian>
+  simdutf_really_inline void store_bytes_as_utf16(char16_t *p) const {
+    const vector_type zero = vec_splats(T(0));
+
+    if (big_endian) {
+      const vec_u8_t perm_lo = {16, 0, 16, 1, 16, 2, 16, 3,
+                                16, 4, 16, 5, 16, 6, 16, 7};
+      const vec_u8_t perm_hi = {16, 8,  16, 9,  16, 10, 16, 11,
+                                16, 12, 16, 13, 16, 14, 16, 15};
+
+      const vector_type v0 = vec_perm(value, zero, perm_lo);
+      const vector_type v1 = vec_perm(value, zero, perm_hi);
+
+#if defined(__clang__)
+      vec_xst(v0, 0, reinterpret_cast<T *>(p));
+      vec_xst(v1, 16, reinterpret_cast<T *>(p));
+#else
+      vec_xst(v0, 0, reinterpret_cast<vector_type *>(p));
+      vec_xst(v1, 16, reinterpret_cast<vector_type *>(p));
+#endif // defined(__clang__)
+    } else {
+      const vec_u8_t perm_lo = {0, 16, 1, 16, 2, 16, 3, 16,
+                                4, 16, 5, 16, 6, 16, 7, 16};
+      const vec_u8_t perm_hi = {8,  16, 9,  16, 10, 16, 11, 16,
+                                12, 16, 13, 16, 14, 16, 15, 16};
+
+      const vector_type v0 = vec_perm(value, zero, perm_lo);
+      const vector_type v1 = vec_perm(value, zero, perm_hi);
+
+#if defined(__clang__)
+      vec_xst(v0, 0, reinterpret_cast<T *>(p));
+      vec_xst(v1, 16, reinterpret_cast<T *>(p));
+#else
+      vec_xst(v0, 0, reinterpret_cast<vector_type *>(p));
+      vec_xst(v1, 16, reinterpret_cast<vector_type *>(p));
+#endif // defined(__clang__)
+    }
+  }
+
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
+    store_bytes_as_utf16<big_endian>(p);
+  }
+
+  simdutf_really_inline void store_bytes_as_utf32(char32_t *p) const {
+    const vector_type zero = vec_splats(T(0));
+
+#if SIMDUTF_IS_BIG_ENDIAN
+    const vec_u8_t perm0 = {16, 16, 16, 0, 16, 16, 16, 1,
+                            16, 16, 16, 2, 16, 16, 16, 3};
+
+    const vec_u8_t perm1 = {16, 16, 16, 4, 16, 16, 16, 5,
+                            16, 16, 16, 6, 16, 16, 16, 7};
+
+    const vec_u8_t perm2 = {16, 16, 16, 8,  16, 16, 16, 9,
+                            16, 16, 16, 10, 16, 16, 16, 11};
+
+    const vec_u8_t perm3 = {16, 16, 16, 12, 16, 16, 16, 13,
+                            16, 16, 16, 14, 16, 16, 16, 15};
+#else
+    const vec_u8_t perm0 = {0, 16, 16, 16, 1, 16, 16, 16,
+                            2, 16, 16, 16, 3, 16, 16, 16};
+
+    const vec_u8_t perm1 = {4, 16, 16, 16, 5, 16, 16, 16,
+                            6, 16, 16, 16, 7, 16, 16, 16};
+
+    const vec_u8_t perm2 = {8,  16, 16, 16, 9,  16, 16, 16,
+                            10, 16, 16, 16, 11, 16, 16, 16};
+
+    const vec_u8_t perm3 = {12, 16, 16, 16, 13, 16, 16, 16,
+                            14, 16, 16, 16, 15, 16, 16, 16};
+#endif // SIMDUTF_IS_BIG_ENDIAN
+
+    const vector_type v0 = vec_perm(value, zero, perm0);
+    const vector_type v1 = vec_perm(value, zero, perm1);
+    const vector_type v2 = vec_perm(value, zero, perm2);
+    const vector_type v3 = vec_perm(value, zero, perm3);
+
+    constexpr size_t n = base8<T>::SIZE;
+
+#if defined(__clang__)
+    vec_xst(v0, 0 * n, reinterpret_cast<T *>(p));
+    vec_xst(v1, 1 * n, reinterpret_cast<T *>(p));
+    vec_xst(v2, 2 * n, reinterpret_cast<T *>(p));
+    vec_xst(v3, 3 * n, reinterpret_cast<T *>(p));
+#else
+    vec_xst(v0, 0 * n, reinterpret_cast<vector_type *>(p));
+    vec_xst(v1, 1 * n, reinterpret_cast<vector_type *>(p));
+    vec_xst(v2, 2 * n, reinterpret_cast<vector_type *>(p));
+    vec_xst(v3, 3 * n, reinterpret_cast<vector_type *>(p));
+#endif // defined(__clang__)
+  }
+
+  simdutf_really_inline void store_words_as_utf32(char32_t *p) const {
+    const vector_type zero = vec_splats(T(0));
+
+#if SIMDUTF_IS_BIG_ENDIAN
+    const vec_u8_t perm0 = {16, 16, 0, 1, 16, 16, 2, 3,
+                            16, 16, 4, 5, 16, 16, 6, 7};
+    const vec_u8_t perm1 = {16, 16, 8,  9,  16, 16, 10, 11,
+                            16, 16, 12, 13, 16, 16, 14, 15};
+#else
+    const vec_u8_t perm0 = {0, 1, 16, 16, 2, 3, 16, 16,
+                            4, 5, 16, 16, 6, 7, 16, 16};
+    const vec_u8_t perm1 = {8,  9,  16, 16, 10, 11, 16, 16,
+                            12, 13, 16, 16, 14, 15, 16, 16};
+#endif // SIMDUTF_IS_BIG_ENDIAN
+
+    const vector_type v0 = vec_perm(value, zero, perm0);
+    const vector_type v1 = vec_perm(value, zero, perm1);
+
+    constexpr size_t n = base8<T>::SIZE;
+
+#if defined(__clang__)
+    vec_xst(v0, 0 * n, reinterpret_cast<T *>(p));
+    vec_xst(v1, 1 * n, reinterpret_cast<T *>(p));
+#else
+    vec_xst(v0, 0 * n, reinterpret_cast<vector_type *>(p));
+    vec_xst(v1, 1 * n, reinterpret_cast<vector_type *>(p));
+#endif // defined(__clang__)
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
+    store_bytes_as_utf32(p);
+  }
+};
+
+// Forward declaration
+template <typename T> struct simd8;
+
+template <typename T>
+simd8<bool> operator==(const simd8<T> a, const simd8<T> b);
+
+template <typename T>
+simd8<bool> operator!=(const simd8<T> a, const simd8<T> b);
+
+template <typename T> simd8<T> operator&(const simd8<T> a, const simd8<T> b);
+
+template <typename T> simd8<T> operator|(const simd8<T> a, const simd8<T> b);
+
+template <typename T> simd8<T> operator^(const simd8<T> a, const simd8<T> b);
+
+template <typename T> simd8<T> operator+(const simd8<T> a, const simd8<T> b);
+
+template <typename T> simd8<bool> operator<(const simd8<T> a, const simd8<T> b);
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  using super = base8<bool>;
+
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return (vector_type)vec_splats((unsigned char)(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8() : super(vector_type()) {}
+  simdutf_really_inline simd8(const vector_type _value) : super(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+  template <typename T>
+  simdutf_really_inline simd8(simd8<T> other)
+      : simd8(vector_type(other.value)) {}
+
+  simdutf_really_inline uint16_t to_bitmask() const {
+    return move_mask_u8(value);
+  }
+
+  simdutf_really_inline bool any() const {
+    return !vec_all_eq(this->value, (vector_type)vec_splats(0));
+  }
+
+  simdutf_really_inline bool all() const { return to_bitmask() == 0xffff; }
+
+  simdutf_really_inline simd8<bool> operator~() const {
+    return this->value ^ (vector_type)splat(true);
+  }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  using super = base8<T>;
+  using vector_type = typename super::vector_type;
+
+  static simdutf_really_inline simd8<T> splat(T value) {
+    return (vector_type)vec_splats(value);
+  }
+
+  static simdutf_really_inline simd8<T> zero() { return splat(0); }
+
+  template <typename U>
+  static simdutf_really_inline simd8<T> load(const U *values) {
+    return vec_xl(0, reinterpret_cast<const T *>(values));
+  }
+
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                  T v5, T v6, T v7, T v8, T v9,
+                                                  T v10, T v11, T v12, T v13,
+                                                  T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15);
+  }
+
+  simdutf_really_inline base8_numeric() : base8<T>() {}
+  simdutf_really_inline base8_numeric(const vector_type _value)
+      : base8<T>(_value) {}
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
+    this->value = vec_sub(this->value, other.value);
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return (vector_type)vec_perm((vector_type)lookup_table,
+                                 (vector_type)lookup_table, this->value);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_32(const simd8<L> lookup_table_lo,
+            const simd8<L> lookup_table_hi) const {
+    return (vector_type)vec_perm(lookup_table_lo.value, lookup_table_hi.value,
+                                 this->value);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  using Self = simd8<uint8_t>;
+
+  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
+  simdutf_really_inline simd8(const vector_type _value)
+      : base8_numeric<uint8_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8((vector_type){v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                            v12, v13, v14, v15}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return move_mask_u8(this->value) == 0;
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8(simd8<T> other)
+      : simd8(vector_type(other.value)) {}
+
+  template <int N>
+  simdutf_really_inline Self prev(const Self prev_chunk) const {
+    return prev_aux<N>(prev_chunk.value);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return (vector_type)vec_subs(this->value, (vector_type)other);
+  }
+
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline bool bits_not_set_anywhere() const {
+    return vec_all_eq(this->value, (vector_type)vec_splats(0));
+  }
+
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(
+        (vector_type)vec_sr(this->value, (vector_type)vec_splat_u8(N)));
+  }
+
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return simd8<uint8_t>(
+        (vector_type)vec_sl(this->value, (vector_type)vec_splat_u8(N)));
+  }
+
+  void dump() const {
+    uint8_t tmp[16];
+    store(tmp);
+    for (int i = 0; i < 16; i++) {
+      if (i == 0) {
+        printf("[%02x", tmp[i]);
+      } else if (i == 15) {
+        printf(" %02x]", tmp[i]);
+      } else {
+        printf(" %02x", tmp[i]);
+      }
+    }
+    putchar('\n');
+  }
+
+  void dump_ascii() const {
+    uint8_t tmp[16];
+    store(tmp);
+    for (int i = 0; i < 16; i++) {
+      if (i == 0) {
+        printf("[%c", tmp[i]);
+      } else if (i == 15) {
+        printf("%c]", tmp[i]);
+      } else {
+        printf("%c", tmp[i]);
+      }
+    }
+    putchar('\n');
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
+  simdutf_really_inline simd8(const vector_type _value)
+      : base8_numeric<int8_t>(_value) {}
+
+  template <typename T>
+  simdutf_really_inline simd8(simd8<T> other)
+      : simd8(vector_type(other.value)) {}
+
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+
+  simdutf_really_inline operator simd8<uint8_t>() const;
+
+  // Saturated math
+  simdutf_really_inline simd8<int8_t>
+  saturating_add(const simd8<int8_t> other) const {
+    return (vector_type)vec_adds(this->value, other.value);
+  }
+
+  void dump() const {
+    int8_t tmp[16];
+    store(tmp);
+    for (int i = 0; i < 16; i++) {
+      if (i == 0) {
+        printf("[%02x", tmp[i]);
+      } else if (i == 15) {
+        printf("%02x]", tmp[i]);
+      } else {
+        printf("%02x", tmp[i]);
+      }
+    }
+    putchar('\n');
+  }
+};
+
+template <typename T>
+simd8<bool> operator==(const simd8<T> a, const simd8<T> b) {
+  return vec_cmpeq(a.value, b.value);
+}
+
+template <typename T>
+simd8<bool> operator!=(const simd8<T> a, const simd8<T> b) {
+  return vec_cmpne(a.value, b.value);
+}
+
+template <typename T> simd8<T> operator&(const simd8<T> a, const simd8<T> b) {
+  return vec_and(a.value, b.value);
+}
+
+template <typename T, typename U> simd8<T> operator&(const simd8<T> a, U b) {
+  return vec_and(a.value, vec_splats(T(b)));
+}
+
+template <typename T> simd8<T> operator|(const simd8<T> a, const simd8<T> b) {
+  return vec_or(a.value, b.value);
+}
+
+template <typename T> simd8<T> operator^(const simd8<T> a, const simd8<T> b) {
+  return vec_xor(a.value, b.value);
+}
+
+template <typename T, typename U> simd8<T> operator^(const simd8<T> a, U b) {
+  return vec_xor(a.value, vec_splats(T(b)));
+}
+
+template <typename T> simd8<T> operator+(const simd8<T> a, const simd8<T> b) {
+  return vec_add(a.value, b.value);
+}
+
+template <typename T, typename U> simd8<T> operator+(const simd8<T> a, U b) {
+  return vec_add(a.value, vec_splats(T(b)));
+}
+
+simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
+  return (simd8<uint8_t>::vector_type)value;
+}
+
+template <typename T>
+simd8<bool> operator<(const simd8<T> a, const simd8<T> b) {
+  return vec_cmplt(a.value, b.value);
+}
+
+template <typename T>
+simd8<bool> operator>(const simd8<T> a, const simd8<T> b) {
+  return vec_cmpgt(a.value, b.value);
+}
+
+template <typename T>
+simd8<bool> operator>=(const simd8<T> a, const simd8<T> b) {
+  return vec_cmpge(a.value, b.value);
+}
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static constexpr size_t ELEMENTS = simd8<T>::ELEMENTS;
+
+  static_assert(NUM_CHUNKS == 4,
+                "PPC64 kernel should use four registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+  simd8x64(simd8x64<T> &&) = default;
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                 const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + ELEMENTS * 0);
+    this->chunks[1].store(ptr + ELEMENTS * 1);
+    this->chunks[2].store(ptr + ELEMENTS * 2);
+    this->chunks[3].store(ptr + ELEMENTS * 3);
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    this->chunks[2] |= other.chunks[2];
+    this->chunks[3] |= other.chunks[3];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 2);
+    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                          this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
+                          this->chunks[2] > mask, this->chunks[3] > mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>(simd8<uint8_t>(this->chunks[0]) >= mask,
+                          simd8<uint8_t>(this->chunks[1]) >= mask,
+                          simd8<uint8_t>(this->chunks[2]) >= mask,
+                          simd8<uint8_t>(this->chunks[3]) >= mask)
+        .to_bitmask();
+  }
+
+  void dump() const {
+    puts("");
+    for (int i = 0; i < 4; i++) {
+      printf("chunk[%d] = ", i);
+      this->chunks[i].dump();
+    }
+  }
+}; // struct simd8x64<T>
+
+simdutf_really_inline simd8<uint8_t> avg(const simd8<uint8_t> a,
+                                         const simd8<uint8_t> b) {
+  return vec_avg(a.value, b.value);
+}
+/* end file src/simdutf/ppc64/simd8-inl.h */
+/* begin file src/simdutf/ppc64/simd16-inl.h */
+// file included directly
+
+template <typename T> struct simd16;
+
+template <typename T> struct base16 {
+  using vector_type = vector_u16_type_for_element<T>;
+  static const int SIZE = sizeof(vector_type);
+  static const int ELEMENTS = sizeof(vector_type) / sizeof(T);
+
+  vector_type value;
+
+  // Zero constructor
+  simdutf_really_inline base16() : value{vector_type()} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base16(const vector_type _value) : value{_value} {}
+
+  void dump() const {
+    uint16_t tmp[8];
+    vec_xst(value, 0, reinterpret_cast<vector_type *>(tmp));
+    for (int i = 0; i < 8; i++) {
+      if (i == 0) {
+        printf("[%04x", tmp[i]);
+      } else if (i == 8 - 1) {
+        printf(" %04x]", tmp[i]);
+      } else {
+        printf(" %04x", tmp[i]);
+      }
+    }
+    putchar('\n');
+  }
+};
+
+// Forward declaration
+template <typename> struct simd16;
+
+template <typename T>
+simd16<bool> operator==(const simd16<T> a, const simd16<T> b);
+
+template <typename T, typename U>
+simd16<bool> operator==(const simd16<T> a, U b);
+
+template <typename T> simd16<T> operator&(const simd16<T> a, const simd16<T> b);
+
+template <typename T> simd16<T> operator|(const simd16<T> a, const simd16<T> b);
+
+template <typename T, typename U> simd16<T> operator|(const simd16<T> a, U b);
+
+template <typename T, typename U> simd16<T> operator^(const simd16<T> a, U b);
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return (vector_type)vec_splats(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+
+  simdutf_really_inline simd16(const vector_type _value)
+      : base16<bool>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+
+  simdutf_really_inline uint16_t to_bitmask() const {
+    return move_mask_u8(value);
+  }
+
+  simdutf_really_inline bool any() const {
+    const auto tmp = vec_u64_t(value);
+
+    return tmp[0] || tmp[1]; // Note: logical or, not binary one
+  }
+
+  simdutf_really_inline bool is_zero() const {
+    const auto tmp = vec_u64_t(value);
+
+    return (tmp[0] | tmp[1]) == 0;
+  }
+
+  simdutf_really_inline simd16<bool> &operator|=(const simd16<bool> rhs) {
+    value = vec_or(this->value, rhs.value);
+    return *this;
+  }
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  using vector_type = typename base16<T>::vector_type;
+
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return vec_splats(_value);
+  }
+
+  static simdutf_really_inline simd16<T> zero() { return splat(0); }
+
+  template <typename U>
+  static simdutf_really_inline simd16<T> load(const U *ptr) {
+    return vec_xl(0, reinterpret_cast<const T *>(ptr));
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const vector_type _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  template <typename U> simdutf_really_inline void store(U *dst) const {
+#if defined(__clang__)
+    return vec_xst(this->value, 0, reinterpret_cast<T *>(dst));
+#else
+    return vec_xst(this->value, 0, reinterpret_cast<vector_type *>(dst));
+#endif // defined(__clang__)
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const {
+    return vec_xor(this->value, vec_splats(T(0xffff)));
+  }
+};
+
+// Signed code units
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+  simdutf_really_inline simd16(const vector_type _value)
+      : base16_numeric<int16_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline operator simd16<uint16_t>() const;
+};
+
+// Unsigned code units
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const vector_type _value)
+      : base16_numeric<uint16_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+
+  // Array constructor
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+
+  simdutf_really_inline bool is_ascii() const {
+    return vec_all_lt(value, vec_splats(uint16_t(128)));
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return vec_max(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return vec_min(this->value, other.value);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return other.max_val(*this) == other;
+  }
+
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return other.min_val(*this) == other;
+  }
+
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return vec_cmplt(value, other.value);
+  }
+
+  // Bit-specific operations
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return vec_sr(value, vec_splats(uint16_t(N)));
+  }
+
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return vec_sl(value, vec_splats(uint16_t(N)));
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    return vec_revb(value);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    return vec_packs(v0.value, v1.value);
+  }
+};
+
+template <typename T>
+simd16<bool> operator==(const simd16<T> a, const simd16<T> b) {
+  return vec_cmpeq(a.value, b.value);
+}
+
+template <typename T, typename U>
+simd16<bool> operator==(const simd16<T> a, U b) {
+  return vec_cmpeq(a.value, vec_splats(T(b)));
+}
+
+template <typename T>
+simd16<T> operator&(const simd16<T> a, const simd16<T> b) {
+  return vec_and(a.value, b.value);
+}
+
+template <typename T, typename U> simd16<T> operator&(const simd16<T> a, U b) {
+  return vec_and(a.value, vec_splats(T(b)));
+}
+
+template <typename T>
+simd16<T> operator|(const simd16<T> a, const simd16<T> b) {
+  return vec_or(a.value, b.value);
+}
+
+template <typename T, typename U> simd16<T> operator|(const simd16<T> a, U b) {
+  return vec_or(a.value, vec_splats(T(b)));
+}
+
+template <typename T>
+simd16<T> operator^(const simd16<T> a, const simd16<T> b) {
+  return vec_xor(a.value, b.value);
+}
+
+template <typename T, typename U> simd16<T> operator^(const simd16<T> a, U b) {
+  return vec_xor(a.value, vec_splats(T(b)));
+}
+
+simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
+  return (vec_u16_t)(value);
+}
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "AltiVec kernel should use four registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline
+  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
+            const simd16<T> chunk2, const simd16<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
+    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
+    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+    this->chunks[2] = this->chunks[2].swap_bytes();
+    this->chunks[3] = this->chunks[3].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                           this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
+    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
+    return simd16x32<bool>(
+               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
+               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
+               (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
+               (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low))
+        .to_bitmask();
+  }
+}; // struct simd16x32<T>
+/* end file src/simdutf/ppc64/simd16-inl.h */
+/* begin file src/simdutf/ppc64/simd32-inl.h */
+// file included directly
+
+template <typename T> struct simd32;
+
+template <typename T> struct base32 {
+  using vector_type = vector_u32_type_for_element<T>;
+  static const int SIZE = sizeof(vector_type);
+  static const int ELEMENTS = sizeof(vector_type) / sizeof(T);
+
+  vector_type value;
+
+  // Zero constructor
+  simdutf_really_inline base32() : value{vector_type()} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base32(const vector_type _value) : value{_value} {}
+
+  // Splat for scalar
+  simdutf_really_inline base32(T scalar) : value{vec_splats(scalar)} {}
+
+  template <typename Pointer>
+  simdutf_really_inline base32(const Pointer *ptr)
+      : base32(vec_xl(0, reinterpret_cast<const T *>(ptr))) {}
+
+  // Store to array
+  template <typename U> simdutf_really_inline void store(U *dst) const {
+#if defined(__clang__)
+    return vec_xst(this->value, 0, reinterpret_cast<T *>(dst));
+#else
+    return vec_xst(this->value, 0, reinterpret_cast<vector_type *>(dst));
+#endif // defined(__clang__)
+  }
+
+  void dump(const char *name = nullptr) const {
+    if (name != nullptr) {
+      printf("%-10s = ", name);
+    }
+
+    uint32_t tmp[4];
+    vec_xst(value, 0, reinterpret_cast<vector_type *>(tmp));
+    for (int i = 0; i < 4; i++) {
+      if (i == 0) {
+        printf("[%08x", tmp[i]);
+      } else if (i == 4 - 1) {
+        printf(" %08x]", tmp[i]);
+      } else {
+        printf(" %08x", tmp[i]);
+      }
+    }
+    putchar('\n');
+  }
+};
+
+template <typename T> struct base32_numeric : base32<T> {
+  using super = base32<T>;
+  using vector_type = typename super::vector_type;
+
+  static simdutf_really_inline simd32<T> splat(T _value) {
+    return vec_splats(_value);
+  }
+
+  static simdutf_really_inline simd32<T> zero() { return splat(0); }
+
+  template <typename U>
+  static simdutf_really_inline simd32<T> load(const U *values) {
+    return vec_xl(0, reinterpret_cast<const T *>(values));
+  }
+
+  simdutf_really_inline base32_numeric() : base32<T>() {}
+
+  simdutf_really_inline base32_numeric(const vector_type _value)
+      : base32<T>(_value) {}
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd32<T> operator+(const simd32<T> other) const {
+    return vec_add(this->value, other.value);
+  }
+
+  simdutf_really_inline simd32<T> operator-(const simd32<T> other) const {
+    return vec_sub(this->value, other.value);
+  }
+
+  simdutf_really_inline simd32<T> &operator+=(const simd32<T> other) {
+    *this = *this + other;
+    return *static_cast<simd32<T> *>(this);
+  }
+
+  simdutf_really_inline simd32<T> &operator-=(const simd32<T> other) {
+    *this = *this - other;
+    return *static_cast<simd32<T> *>(this);
+  }
+};
+
+// Forward declaration
+template <typename> struct simd32;
+
+template <typename T>
+simd32<bool> operator==(const simd32<T> a, const simd32<T> b);
+
+template <typename T>
+simd32<bool> operator!=(const simd32<T> a, const simd32<T> b);
+
+template <typename T>
+simd32<bool> operator>(const simd32<T> a, const simd32<T> b);
+
+template <typename T> simd32<bool> operator==(const simd32<T> a, T b);
+
+template <typename T> simd32<bool> operator!=(const simd32<T> a, T b);
+
+template <typename T> simd32<T> operator&(const simd32<T> a, const simd32<T> b);
+
+template <typename T> simd32<T> operator|(const simd32<T> a, const simd32<T> b);
+
+template <typename T> simd32<T> operator^(const simd32<T> a, const simd32<T> b);
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd32<bool> : base32<bool> {
+  static simdutf_really_inline simd32<bool> splat(bool _value) {
+    return (vector_type)vec_splats(uint32_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd32(const vector_type _value)
+      : base32<bool>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd32(bool _value) : base32<bool>(splat(_value)) {}
+
+  simdutf_really_inline uint16_t to_bitmask() const {
+    return move_mask_u8(value);
+  }
+
+  simdutf_really_inline bool any() const {
+    const vec_u64_t tmp = (vec_u64_t)value;
+
+    return tmp[0] || tmp[1]; // Note: logical or, not binary one
+  }
+
+  simdutf_really_inline bool is_zero() const {
+    const vec_u64_t tmp = (vec_u64_t)value;
+
+    return (tmp[0] | tmp[1]) == 0;
+  }
+
+  simdutf_really_inline simd32<bool> operator~() const {
+    return (vec_bool32_t)vec_xor(this->value, vec_splats(uint32_t(0xffffffff)));
+  }
+};
+
+// Unsigned code units
+template <> struct simd32<uint32_t> : base32_numeric<uint32_t> {
+  simdutf_really_inline simd32() : base32_numeric<uint32_t>() {}
+
+  simdutf_really_inline simd32(const vector_type _value)
+      : base32_numeric<uint32_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd32(uint32_t _value) : simd32(splat(_value)) {}
+
+  // Array constructor
+  simdutf_really_inline simd32(const char32_t *values)
+      : simd32(load(reinterpret_cast<const uint32_t *>(values))) {}
+
+  // Bit-specific operations
+  template <int N> simdutf_really_inline simd32<uint32_t> shr() const {
+    return vec_sr(value, vec_splats(uint32_t(N)));
+  }
+
+  template <int N> simdutf_really_inline simd32<uint32_t> shl() const {
+    return vec_sl(value, vec_splats(uint32_t(N)));
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd32<uint32_t> swap_bytes() const {
+    return vec_revb(value);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    return uint64_t(value[0]) + uint64_t(value[1]) + uint64_t(value[2]) +
+           uint64_t(value[3]);
+  }
+
+  static simdutf_really_inline simd16<uint16_t>
+  pack(const simd32<uint32_t> &v0, const simd32<uint32_t> &v1) {
+    return vec_packs(v0.value, v1.value);
+  }
+};
+
+template <typename T>
+simd32<bool> operator==(const simd32<T> a, const simd32<T> b) {
+  return vec_cmpeq(a.value, b.value);
+}
+
+template <typename T>
+simd32<bool> operator!=(const simd32<T> a, const simd32<T> b) {
+  return vec_cmpne(a.value, b.value);
+}
+
+template <typename T> simd32<bool> operator==(const simd32<T> a, T b) {
+  return vec_cmpeq(a.value, vec_splats(b));
+}
+
+template <typename T> simd32<bool> operator!=(const simd32<T> a, T b) {
+  return vec_cmpne(a.value, vec_splats(b));
+}
+
+template <typename T>
+simd32<bool> operator>(const simd32<T> a, const simd32<T> b) {
+  return vec_cmpgt(a.value, b.value);
+}
+
+template <typename T>
+simd32<T> operator&(const simd32<T> a, const simd32<T> b) {
+  return vec_and(a.value, b.value);
+}
+
+template <typename T, typename U> simd32<T> operator&(const simd32<T> a, U b) {
+  return vec_and(a.value, vec_splats(T(b)));
+}
+
+template <typename T>
+simd32<T> operator|(const simd32<T> a, const simd32<T> b) {
+  return vec_or(a.value, b.value);
+}
+
+template <typename T>
+simd32<T> operator^(const simd32<T> a, const simd32<T> b) {
+  return vec_xor(a.value, b.value);
+}
+
+template <typename T, typename U> simd32<T> operator^(const simd32<T> a, U b) {
+  return vec_xor(a.value, vec_splats(T(b)));
+}
+
+template <typename T> simd32<T> max_val(const simd32<T> a, const simd32<T> b) {
+  return vec_max(a.value, b.value);
+}
+
+template <typename T>
+simdutf_really_inline simd32<T> min(const simd32<T> b, const simd32<T> a) {
+  return vec_min(a.value, b.value);
+}
+/* end file src/simdutf/ppc64/simd32-inl.h */
+
+template <typename T>
+simd8<T> select(const simd8<T> cond, const simd8<T> val_true,
+                const simd8<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, cond.value);
+}
+
+template <typename T>
+simd8<T> select(const T cond, const simd8<T> val_true,
+                const simd8<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, vec_splats(cond));
+}
+
+template <typename T>
+simd16<T> select(const simd16<T> cond, const simd16<T> val_true,
+                 const simd16<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, cond.value);
+}
+
+template <typename T>
+simd16<T> select(const T cond, const simd16<T> val_true,
+                 const simd16<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, vec_splats(cond));
+}
+
+template <typename T>
+simd32<T> select(const simd32<T> cond, const simd32<T> val_true,
+                 const simd32<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, cond.value);
+}
+
+template <typename T>
+simd32<T> select(const T cond, const simd32<T> val_true,
+                 const simd32<T> val_false) {
+  return vec_sel(val_false.value, val_true.value, vec_splats(cond));
+}
+
+using vector_u8 = simd8<uint8_t>;
+using vector_u16 = simd16<uint16_t>;
+using vector_u32 = simd32<uint32_t>;
+using vector_i8 = simd8<int8_t>;
+
+simdutf_really_inline vector_u8 as_vector_u8(const vector_u16 v) {
+  return vector_u8::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u8 as_vector_u8(const vector_u32 v) {
+  return vector_u8::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u8 as_vector_u8(const vector_i8 v) {
+  return vector_u8::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u8 as_vector_u8(const simd16<bool> v) {
+  return vector_u8::vector_type(v.value);
+}
+
+simdutf_really_inline vector_i8 as_vector_i8(const vector_u8 v) {
+  return vector_i8::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u16 as_vector_u16(const vector_u8 v) {
+  return vector_u16::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u16 as_vector_u16(const simd16<bool> v) {
+  return vector_u16::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u32 as_vector_u32(const vector_u8 v) {
+  return vector_u32::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u32 as_vector_u32(const vector_u16 v) {
+  return vector_u32::vector_type(v.value);
+}
+
+simdutf_really_inline vector_u32 max(vector_u32 a, vector_u32 b) {
+  return vec_max(a.value, b.value);
+}
+
+simdutf_really_inline vector_u32 max(vector_u32 a, vector_u32 b, vector_u32 c) {
+  return max(max(a, b), c);
+}
+
+simdutf_really_inline vector_u32 sum4bytes(vector_u8 bytes, vector_u32 acc) {
+  return vec_sum4s(bytes.value, acc.value);
+}
+
+} // namespace simd
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+
+#endif // SIMDUTF_PPC64_SIMD_INPUT_H
+/* end file src/simdutf/ppc64/simd.h */
+
+/* begin file src/simdutf/ppc64/end.h */
+/* end file src/simdutf/ppc64/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_PPC64
+
+#endif // SIMDUTF_PPC64_H
+/* end file src/simdutf/ppc64.h */
+/* begin file src/simdutf/rvv.h */
+#ifndef SIMDUTF_RVV_H
+#define SIMDUTF_RVV_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "rvv.h must be included before fallback.h"
+#endif
+
+
+#define SIMDUTF_CAN_ALWAYS_RUN_RVV SIMDUTF_IS_RVV
+
+#ifndef SIMDUTF_IMPLEMENTATION_RVV
+  #define SIMDUTF_IMPLEMENTATION_RVV                                           \
+    (SIMDUTF_CAN_ALWAYS_RUN_RVV ||                                             \
+     (SIMDUTF_IS_RISCV64 && SIMDUTF_HAS_RVV_INTRINSICS &&                      \
+      SIMDUTF_HAS_RVV_TARGET_REGION))
+#endif
+
+#if SIMDUTF_IMPLEMENTATION_RVV
+
+  #if SIMDUTF_CAN_ALWAYS_RUN_RVV
+    #define SIMDUTF_TARGET_RVV
+  #else
+    #define SIMDUTF_TARGET_RVV SIMDUTF_TARGET_REGION("arch=+v")
+  #endif
+  #if !SIMDUTF_IS_ZVBB && SIMDUTF_HAS_ZVBB_INTRINSICS
+    #define SIMDUTF_TARGET_ZVBB SIMDUTF_TARGET_REGION("arch=+v,+zvbb")
+  #endif
+
+namespace simdutf {
+namespace rvv {} // namespace rvv
+} // namespace simdutf
+
+/* begin file src/simdutf/rvv/implementation.h */
+#ifndef SIMDUTF_RVV_IMPLEMENTATION_H
+#define SIMDUTF_RVV_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace rvv {
+
+namespace {
+using namespace simdutf;
+} // namespace
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("rvv", "RISC-V Vector Extension",
+                                internal::instruction_set::RVV),
+        _supports_zvbb(internal::detect_supported_architectures() &
+                       internal::instruction_set::ZVBB) {}
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+private:
+  const bool _supports_zvbb;
+
+#if SIMDUTF_IS_ZVBB
+  bool supports_zvbb() const { return true; }
+#elif SIMDUTF_HAS_ZVBB_INTRINSICS
+  bool supports_zvbb() const { return _supports_zvbb; }
+#else
+  bool supports_zvbb() const { return false; }
+#endif
+};
+
+} // namespace rvv
+} // namespace simdutf
+
+#endif // SIMDUTF_RVV_IMPLEMENTATION_H
+/* end file src/simdutf/rvv/implementation.h */
+/* begin file src/simdutf/rvv/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "rvv"
+// #define SIMDUTF_IMPLEMENTATION rvv
+
+#if SIMDUTF_CAN_ALWAYS_RUN_RVV
+// nothing needed.
+#else
+SIMDUTF_TARGET_RVV
+#endif
+/* end file src/simdutf/rvv/begin.h */
+/* begin file src/simdutf/rvv/intrinsics.h */
+#ifndef SIMDUTF_RVV_INTRINSICS_H
+#define SIMDUTF_RVV_INTRINSICS_H
+
+
+#include <riscv_vector.h>
+
+#if __riscv_v_intrinsic >= 1000000 || __GCC__ >= 14
+  #define simdutf_vrgather_u8m1x2(tbl, idx)                                    \
+    __riscv_vcreate_v_u8m1_u8m2(                                               \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 0),        \
+                                 __riscv_vsetvlmax_e8m1()),                    \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),        \
+                                 __riscv_vsetvlmax_e8m1()));
+
+  #define simdutf_vrgather_u8m1x4(tbl, idx)                                    \
+    __riscv_vcreate_v_u8m1_u8m4(                                               \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),        \
+                                 __riscv_vsetvlmax_e8m1()),                    \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 1),        \
+                                 __riscv_vsetvlmax_e8m1()),                    \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),        \
+                                 __riscv_vsetvlmax_e8m1()),                    \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),        \
+                                 __riscv_vsetvlmax_e8m1()));
+#else
+  // This has worse codegen on gcc
+  #define simdutf_vrgather_u8m1x2(tbl, idx)                                    \
+    __riscv_vset_v_u8m1_u8m2(                                                  \
+        __riscv_vlmul_ext_v_u8m1_u8m2(__riscv_vrgather_vv_u8m1(                \
+            tbl, __riscv_vget_v_u8m2_u8m1(idx, 0), __riscv_vsetvlmax_e8m1())), \
+        1,                                                                     \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),        \
+                                 __riscv_vsetvlmax_e8m1()))
+
+  #define simdutf_vrgather_u8m1x4(tbl, idx)                                    \
+    __riscv_vset_v_u8m1_u8m4(                                                  \
+        __riscv_vset_v_u8m1_u8m4(                                              \
+            __riscv_vset_v_u8m1_u8m4(                                          \
+                __riscv_vlmul_ext_v_u8m1_u8m4(__riscv_vrgather_vv_u8m1(        \
+                    tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),                     \
+                    __riscv_vsetvlmax_e8m1())),                                \
+                1,                                                             \
+                __riscv_vrgather_vv_u8m1(tbl,                                  \
+                                         __riscv_vget_v_u8m4_u8m1(idx, 1),     \
+                                         __riscv_vsetvlmax_e8m1())),           \
+            2,                                                                 \
+            __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),    \
+                                     __riscv_vsetvlmax_e8m1())),               \
+        3,                                                                     \
+        __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),        \
+                                 __riscv_vsetvlmax_e8m1()))
+#endif
+
+/* Zvbb adds dedicated support for endianness swaps with vrev8, but if we can't
+ * use that, we have to emulate it with the standard V extension.
+ * Using LMUL=1 vrgathers could be faster than the srl+macc variant, but that
+ * would increase register pressure, and vrgather implementations performance
+ * varies a lot. */
+enum class simdutf_ByteFlip { NONE, V, ZVBB };
+
+template <simdutf_ByteFlip method>
+simdutf_really_inline static uint16_t simdutf_byteflip(uint16_t v) {
+  if (method != simdutf_ByteFlip::NONE)
+    return (uint16_t)((v * 1u) << 8 | (v * 1u) >> 8);
+  return v;
+}
+
+#ifdef SIMDUTF_TARGET_ZVBB
+SIMDUTF_UNTARGET_REGION
+SIMDUTF_TARGET_ZVBB
+#endif
+
+template <simdutf_ByteFlip method>
+simdutf_really_inline static vuint16m1_t simdutf_byteflip(vuint16m1_t v,
+                                                          size_t vl) {
+#if SIMDUTF_HAS_ZVBB_INTRINSICS
+  if (method == simdutf_ByteFlip::ZVBB)
+    return __riscv_vrev8_v_u16m1(v, vl);
+#endif
+  if (method == simdutf_ByteFlip::V)
+    return __riscv_vmacc_vx_u16m1(__riscv_vsrl_vx_u16m1(v, 8, vl), 0x100, v,
+                                  vl);
+  return v;
+}
+
+template <simdutf_ByteFlip method>
+simdutf_really_inline static vuint16m2_t simdutf_byteflip(vuint16m2_t v,
+                                                          size_t vl) {
+#if SIMDUTF_HAS_ZVBB_INTRINSICS
+  if (method == simdutf_ByteFlip::ZVBB)
+    return __riscv_vrev8_v_u16m2(v, vl);
+#endif
+  if (method == simdutf_ByteFlip::V)
+    return __riscv_vmacc_vx_u16m2(__riscv_vsrl_vx_u16m2(v, 8, vl), 0x100, v,
+                                  vl);
+  return v;
+}
+
+template <simdutf_ByteFlip method>
+simdutf_really_inline static vuint16m4_t simdutf_byteflip(vuint16m4_t v,
+                                                          size_t vl) {
+#if SIMDUTF_HAS_ZVBB_INTRINSICS
+  if (method == simdutf_ByteFlip::ZVBB)
+    return __riscv_vrev8_v_u16m4(v, vl);
+#endif
+  if (method == simdutf_ByteFlip::V)
+    return __riscv_vmacc_vx_u16m4(__riscv_vsrl_vx_u16m4(v, 8, vl), 0x100, v,
+                                  vl);
+  return v;
+}
+
+template <simdutf_ByteFlip method>
+simdutf_really_inline static vuint16m8_t simdutf_byteflip(vuint16m8_t v,
+                                                          size_t vl) {
+#if SIMDUTF_HAS_ZVBB_INTRINSICS
+  if (method == simdutf_ByteFlip::ZVBB)
+    return __riscv_vrev8_v_u16m8(v, vl);
+#endif
+  if (method == simdutf_ByteFlip::V)
+    return __riscv_vmacc_vx_u16m8(__riscv_vsrl_vx_u16m8(v, 8, vl), 0x100, v,
+                                  vl);
+  return v;
+}
+
+#ifdef SIMDUTF_TARGET_ZVBB
+SIMDUTF_UNTARGET_REGION
+SIMDUTF_TARGET_RVV
+#endif
+
+#endif //  SIMDUTF_RVV_INTRINSICS_H
+/* end file src/simdutf/rvv/intrinsics.h */
+/* begin file src/simdutf/rvv/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_RVV
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
+
+/* end file src/simdutf/rvv/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_RVV
+
+#endif // SIMDUTF_RVV_H
+/* end file src/simdutf/rvv.h */
+/* begin file src/simdutf/lsx.h */
+#ifndef SIMDUTF_LSX_H
+#define SIMDUTF_LSX_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "lsx.h must be included before fallback.h"
+#endif
+
+
+#ifndef SIMDUTF_IMPLEMENTATION_LSX
+  #define SIMDUTF_IMPLEMENTATION_LSX (SIMDUTF_IS_LSX)
+#endif
+#if SIMDUTF_IMPLEMENTATION_LSX && SIMDUTF_IS_LSX
+  #define SIMDUTF_CAN_ALWAYS_RUN_LSX 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_LSX 0
+#endif
+
+#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
+
+#if SIMDUTF_IMPLEMENTATION_LSX
+
+namespace simdutf {
+/**
+ * Implementation for LoongArch SX.
+ */
+namespace lsx {} // namespace lsx
+} // namespace simdutf
+
+/* begin file src/simdutf/lsx/implementation.h */
+#ifndef SIMDUTF_LSX_IMPLEMENTATION_H
+#define SIMDUTF_LSX_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace lsx {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("lsx", "LOONGARCH SX",
+                                internal::instruction_set::LSX) {}
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace lsx
+} // namespace simdutf
+
+#endif // SIMDUTF_LSX_IMPLEMENTATION_H
+/* end file src/simdutf/lsx/implementation.h */
+
+/* begin file src/simdutf/lsx/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "lsx"
+// #define SIMDUTF_IMPLEMENTATION lsx
+#define SIMDUTF_SIMD_HAS_UNSIGNED_CMP 1
+/* end file src/simdutf/lsx/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/lsx/intrinsics.h */
+#ifndef SIMDUTF_LSX_INTRINSICS_H
+#define SIMDUTF_LSX_INTRINSICS_H
+
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <lsxintrin.h>
+
+/*
+Encoding of argument for LoongArch64 xvldi instruction.  See:
+https://jia.je/unofficial-loongarch-intrinsics-guide/lasx/misc/#__m256i-__lasx_xvldi-imm_n1024_1023-imm
+
+1: imm[12:8]=0b10000: broadcast imm[7:0] as 32-bit elements to all lanes
+
+2: imm[12:8]=0b10001: broadcast imm[7:0] << 8 as 32-bit elements to all lanes
+
+3: imm[12:8]=0b10010: broadcast imm[7:0] << 16 as 32-bit elements to all lanes
+
+4: imm[12:8]=0b10011: broadcast imm[7:0] << 24 as 32-bit elements to all lanes
+
+5: imm[12:8]=0b10100: broadcast imm[7:0] as 16-bit elements to all lanes
+
+6: imm[12:8]=0b10101: broadcast imm[7:0] << 8 as 16-bit elements to all lanes
+
+7: imm[12:8]=0b10110: broadcast (imm[7:0] << 8) | 0xFF as 32-bit elements to all
+lanes
+
+8: imm[12:8]=0b10111: broadcast (imm[7:0] << 16) | 0xFFFF as 32-bit elements to
+all lanes
+
+9: imm[12:8]=0b11000: broadcast imm[7:0] as 8-bit elements to all lanes
+
+10: imm[12:8]=0b11001: repeat each bit of imm[7:0] eight times, and broadcast
+the result as 64-bit elements to all lanes
+*/
+
+namespace vldi {
+
+template <uint16_t v> class const_u16 {
+  constexpr static const uint8_t b0 = ((v >> 0 * 8) & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 1 * 8) & 0xff);
+
+  constexpr static bool is_case5 = uint16_t(b0) == v;
+  constexpr static bool is_case6 = (uint16_t(b1) << 8) == v;
+  constexpr static bool is_case9 = (b0 == b1);
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00));
+
+public:
+  constexpr static uint16_t operation = is_case5    ? 0b10100
+                                        : is_case6  ? 0b10101
+                                        : is_case9  ? 0b11000
+                                        : is_case10 ? 0x11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_case5    ? b0
+      : is_case6  ? b1
+      : is_case9  ? b0
+      : is_case10 ? ((b0 ? 0x55 : 0x00) | (b1 ? 0xaa : 0x00))
+                  : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+
+template <uint32_t v> class const_u32 {
+  constexpr static const uint8_t b0 = (v & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 8) & 0xff);
+  constexpr static const uint8_t b2 = ((v >> 16) & 0xff);
+  constexpr static const uint8_t b3 = ((v >> 24) & 0xff);
+
+  constexpr static bool is_case1 = (uint32_t(b0) == v);
+  constexpr static bool is_case2 = ((uint32_t(b1) << 8) == v);
+  constexpr static bool is_case3 = ((uint32_t(b2) << 16) == v);
+  constexpr static bool is_case4 = ((uint32_t(b3) << 24) == v);
+  constexpr static bool is_case5 = (b0 == b2) && (b1 == 0) && (b3 == 0);
+  constexpr static bool is_case6 = (b1 == b3) && (b0 == 0) && (b2 == 0);
+  constexpr static bool is_case7 = (b3 == 0) && (b2 == 0) && (b0 == 0xff);
+  constexpr static bool is_case8 = (b3 == 0) && (b1 == 0xff) && (b0 == 0xff);
+  constexpr static bool is_case9 = (b0 == b1) && (b0 == b2) && (b0 == b3);
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00)) &&
+      ((b2 == 0xff) || (b2 == 0x00)) && ((b3 == 0xff) || (b3 == 0x00));
+
+public:
+  constexpr static uint16_t operation = is_case1    ? 0b10000
+                                        : is_case2  ? 0b10001
+                                        : is_case3  ? 0b10010
+                                        : is_case4  ? 0b10011
+                                        : is_case5  ? 0b10100
+                                        : is_case6  ? 0b10101
+                                        : is_case7  ? 0b10110
+                                        : is_case8  ? 0b10111
+                                        : is_case9  ? 0b11000
+                                        : is_case10 ? 0b11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_case1    ? b0
+      : is_case2  ? b1
+      : is_case3  ? b2
+      : is_case4  ? b3
+      : is_case5  ? b0
+      : is_case6  ? b1
+      : is_case7  ? b1
+      : is_case8  ? b2
+      : is_case9  ? b0
+      : is_case10 ? ((b0 ? 0x11 : 0x00) | (b1 ? 0x22 : 0x00) |
+                     (b2 ? 0x44 : 0x00) | (b3 ? 0x88 : 0x00))
+                  : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+
+template <uint64_t v> class const_u64 {
+  constexpr static const uint8_t b0 = ((v >> 0 * 8) & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 1 * 8) & 0xff);
+  constexpr static const uint8_t b2 = ((v >> 2 * 8) & 0xff);
+  constexpr static const uint8_t b3 = ((v >> 3 * 8) & 0xff);
+  constexpr static const uint8_t b4 = ((v >> 4 * 8) & 0xff);
+  constexpr static const uint8_t b5 = ((v >> 5 * 8) & 0xff);
+  constexpr static const uint8_t b6 = ((v >> 6 * 8) & 0xff);
+  constexpr static const uint8_t b7 = ((v >> 7 * 8) & 0xff);
+
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00)) &&
+      ((b2 == 0xff) || (b2 == 0x00)) && ((b3 == 0xff) || (b3 == 0x00)) &&
+      ((b4 == 0xff) || (b4 == 0x00)) && ((b5 == 0xff) || (b5 == 0x00)) &&
+      ((b6 == 0xff) || (b6 == 0x00)) && ((b7 == 0xff) || (b7 == 0x00));
+
+public:
+  constexpr static bool is_32bit =
+      ((v & 0xffffffff) == (v >> 32)) && const_u32<(v >> 32)>::value;
+  constexpr static uint8_t op_32bit = const_u32<(v >> 32)>::operation;
+  constexpr static uint8_t byte_32bit = const_u32<(v >> 32)>::byte;
+
+  constexpr static uint16_t operation = is_32bit    ? op_32bit
+                                        : is_case10 ? 0x11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_32bit ? byte_32bit
+      : is_case10
+          ? ((b0 ? 0x01 : 0x00) | (b1 ? 0x02 : 0x00) | (b2 ? 0x04 : 0x00) |
+             (b3 ? 0x08 : 0x00) | (b4 ? 0x10 : 0x00) | (b5 ? 0x20 : 0x00) |
+             (b6 ? 0x40 : 0x00) | (b7 ? 0x80 : 0x00))
+          : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+} // namespace vldi
+
+// Uncomment when running under QEMU affected
+// by bug https://gitlab.com/qemu-project/qemu/-/issues/2865
+// Versions <= 9.2.2 are affected, likely anything newer is correct.
+#ifndef QEMU_VLDI_BUG
+// #define QEMU_VLDI_BUG 1
+#endif
+
+#ifdef QEMU_VLDI_BUG
+  #define lsx_splat_u16(v) __lsx_vreplgr2vr_h(v)
+  #define lsx_splat_u32(v) __lsx_vreplgr2vr_w(v)
+#else
+template <uint16_t x> constexpr __m128i lsx_splat_u16_aux() {
+  constexpr bool is_imm10 = (int16_t(x) < 512) && (int16_t(x) > -512);
+  constexpr uint16_t imm10 = is_imm10 ? x : 0;
+  constexpr bool is_vldi = vldi::const_u16<x>::valid;
+  constexpr int vldi_imm = is_vldi ? vldi::const_u16<x>::value : 0;
+
+  return is_imm10  ? __lsx_vrepli_h(int16_t(imm10))
+         : is_vldi ? __lsx_vldi(vldi_imm)
+                   : __lsx_vreplgr2vr_h(x);
+}
+
+template <uint32_t x> constexpr __m128i lsx_splat_u32_aux() {
+  constexpr bool is_imm10 = (int32_t(x) < 512) && (int32_t(x) > -512);
+  constexpr uint32_t imm10 = is_imm10 ? x : 0;
+  constexpr bool is_vldi = vldi::const_u32<x>::valid;
+  constexpr int vldi_imm = is_vldi ? vldi::const_u32<x>::value : 0;
+
+  return is_imm10  ? __lsx_vrepli_w(int32_t(imm10))
+         : is_vldi ? __lsx_vldi(vldi_imm)
+                   : __lsx_vreplgr2vr_w(x);
+}
+
+  #define lsx_splat_u16(v) lsx_splat_u16_aux<(v)>()
+  #define lsx_splat_u32(v) lsx_splat_u32_aux<(v)>()
+#endif // QEMU_VLDI_BUG
+
+#endif //  SIMDUTF_LSX_INTRINSICS_H
+/* end file src/simdutf/lsx/intrinsics.h */
+/* begin file src/simdutf/lsx/bitmanipulation.h */
+#ifndef SIMDUTF_LSX_BITMANIPULATION_H
+#define SIMDUTF_LSX_BITMANIPULATION_H
+
+#include <limits>
+
+namespace simdutf {
+namespace lsx {
+namespace {
+
+simdutf_really_inline int count_ones(uint64_t input_num) {
+  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__lsx_vreplgr2vr_d(input_num)), 0);
+}
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+// simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+//   return __builtin_ctzll(input_num);
+// }
+#endif
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdutf
+
+#endif // SIMDUTF_LSX_BITMANIPULATION_H
+/* end file src/simdutf/lsx/bitmanipulation.h */
+/* begin file src/simdutf/lsx/simd.h */
+#ifndef SIMDUTF_LSX_SIMD_H
+#define SIMDUTF_LSX_SIMD_H
+
+#include <type_traits>
+
+namespace simdutf {
+namespace lsx {
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+
+//
+// Base class of simd8<uint8_t> and simd8<bool>, both of which use __m128i
+// internally.
+//
+template <typename T, typename Mask = simd8<bool>> struct base_u8 {
+  __m128i value;
+  static const int SIZE = sizeof(value);
+
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u8(const __m128i _value) : value(_value) {}
+  simdutf_really_inline operator const __m128i &() const { return this->value; }
+  simdutf_really_inline operator __m128i &() { return this->value; }
+  simdutf_really_inline T first() const {
+    return __lsx_vpickve2gr_bu(this->value, 0);
+  }
+  simdutf_really_inline T last() const {
+    return __lsx_vpickve2gr_bu(this->value, 15);
+  }
+
+  // Bit operations
+  simdutf_really_inline simd8<T> operator|(const simd8<T> other) const {
+    return __lsx_vor_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator&(const simd8<T> other) const {
+    return __lsx_vand_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator^(const simd8<T> other) const {
+    return __lsx_vxor_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> bit_andnot(const simd8<T> other) const {
+    return __lsx_vandn_v(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+  simdutf_really_inline simd8<T> &operator|=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd8<T> &operator&=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd8<T> &operator^=(const simd8<T> other) {
+    auto this_cast = static_cast<simd8<T> *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return __lsx_vseq_b(lhs, rhs);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
+                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base_u8<bool> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return __lsx_vreplgr2vr_b(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8(const __m128i _value) : base_u8<bool>(_value) {}
+  // False constructor
+  simdutf_really_inline simd8() : simd8(__lsx_vldi(0)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  simdutf_really_inline uint32_t to_bitmask() const {
+    return __lsx_vpickve2gr_wu(__lsx_vmsknz_b(*this), 0);
+  }
+
+  simdutf_really_inline bool any() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) != 0;
+  }
+  simdutf_really_inline bool none() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) == 0;
+  }
+  simdutf_really_inline bool all() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0) == 0xFFFF;
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base_u8<uint8_t> {
+  static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) {
+    return __lsx_vreplgr2vr_b(_value);
+  }
+  static simdutf_really_inline simd8<uint8_t> zero() { return __lsx_vldi(0); }
+  static simdutf_really_inline simd8<uint8_t> load(const uint8_t *values) {
+    return __lsx_vld(values, 0);
+  }
+  simdutf_really_inline simd8(const __m128i _value)
+      : base_u8<uint8_t>(_value) {}
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Member-by-member initialization
+
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8((__m128i)v16u8{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                             v12, v13, v14, v15}) {}
+
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Store to array
+  simdutf_really_inline void store(uint8_t dst[16]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return __lsx_vsadd_bu(this->value, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return __lsx_vssub_bu(this->value, other);
+  }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd8<uint8_t>
+  operator+(const simd8<uint8_t> other) const {
+    return __lsx_vadd_b(this->value, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  operator-(const simd8<uint8_t> other) const {
+    return __lsx_vsub_b(this->value, other);
+  }
+  simdutf_really_inline simd8<uint8_t> &operator+=(const simd8<uint8_t> other) {
+    *this = *this + other;
+    return *this;
+  }
+  simdutf_really_inline simd8<uint8_t> &operator-=(const simd8<uint8_t> other) {
+    *this = *this - other;
+    return *this;
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return __lsx_vmax_bu(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return __lsx_vmin_bu(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return __lsx_vsle_bu(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return __lsx_vsle_bu(other, *this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return __lsx_vslt_bu(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return __lsx_vslt_bu(other, *this);
+  }
+  // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true
+  // = nonzero. For ARM, returns all 1's.
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return simd8<uint8_t>(*this > other);
+  }
+  // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true
+  // = nonzero. For ARM, returns all 1's.
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return simd8<uint8_t>(*this < other);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return __lsx_vslt_bu(__lsx_vldi(0), __lsx_vand_v(this->value, bits));
+  }
+  simdutf_really_inline bool is_ascii() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmskgez_b(this->value), 0) == 0xFFFF;
+  }
+
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    return __lsx_vpickve2gr_hu(__lsx_vmsknz_b(this->value), 0) > 0;
+  }
+  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return (*this & bits).any_bits_set_anywhere();
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return __lsx_vsrli_b(this->value, N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return __lsx_vslli_b(this->value, N);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return lookup_table.apply_lookup_16_to(*this);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<uint8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
+    return __lsx_vshuf_b(__lsx_vldi(0), *this, simd8<uint8_t>(original_tmp));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> {
+  __m128i value;
+
+  static simdutf_really_inline simd8<int8_t> splat(int8_t _value) {
+    return __lsx_vreplgr2vr_b(_value);
+  }
+  static simdutf_really_inline simd8<int8_t> zero() { return __lsx_vldi(0); }
+  static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) {
+    return __lsx_vld(values, 0);
+  }
+
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *p) const {
+    __m128i zero = __lsx_vldi(0);
+    if (match_system(big_endian)) {
+      __lsx_vst(__lsx_vilvl_b(zero, (__m128i)this->value),
+                reinterpret_cast<uint16_t *>(p), 0);
+      __lsx_vst(__lsx_vilvh_b(zero, (__m128i)this->value),
+                reinterpret_cast<uint16_t *>(p + 8), 0);
+    } else {
+      __lsx_vst(__lsx_vilvl_b((__m128i)this->value, zero),
+                reinterpret_cast<uint16_t *>(p), 0);
+      __lsx_vst(__lsx_vilvh_b((__m128i)this->value, zero),
+                reinterpret_cast<uint16_t *>(p + 8), 0);
+    }
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *p) const {
+    __m128i zero = __lsx_vldi(0);
+    __m128i in16low = __lsx_vilvl_b(zero, (__m128i)this->value);
+    __m128i in16high = __lsx_vilvh_b(zero, (__m128i)this->value);
+    __m128i in32_0 = __lsx_vilvl_h(zero, in16low);
+    __m128i in32_1 = __lsx_vilvh_h(zero, in16low);
+    __m128i in32_2 = __lsx_vilvl_h(zero, in16high);
+    __m128i in32_3 = __lsx_vilvh_h(zero, in16high);
+    __lsx_vst(in32_0, reinterpret_cast<uint32_t *>(p), 0);
+    __lsx_vst(in32_1, reinterpret_cast<uint32_t *>(p + 4), 0);
+    __lsx_vst(in32_2, reinterpret_cast<uint32_t *>(p + 8), 0);
+    __lsx_vst(in32_3, reinterpret_cast<uint32_t *>(p + 12), 0);
+  }
+
+  // In places where the table can be reused, which is most uses in simdutf, it
+  // is worth it to do 4 table lookups, as there is no direct zero extension
+  // from u8 to u32.
+  simdutf_really_inline void store_ascii_as_utf32_tbl(char32_t *p) const {
+    const simd8<uint8_t> tb1{0, 255, 255, 255, 1, 255, 255, 255,
+                             2, 255, 255, 255, 3, 255, 255, 255};
+    const simd8<uint8_t> tb2{4, 255, 255, 255, 5, 255, 255, 255,
+                             6, 255, 255, 255, 7, 255, 255, 255};
+    const simd8<uint8_t> tb3{8,  255, 255, 255, 9,  255, 255, 255,
+                             10, 255, 255, 255, 11, 255, 255, 255};
+    const simd8<uint8_t> tb4{12, 255, 255, 255, 13, 255, 255, 255,
+                             14, 255, 255, 255, 15, 255, 255, 255};
+
+    // encourage store pairing and interleaving
+    const auto shuf1 = this->apply_lookup_16_to(tb1);
+    const auto shuf2 = this->apply_lookup_16_to(tb2);
+    shuf1.store(reinterpret_cast<int8_t *>(p));
+    shuf2.store(reinterpret_cast<int8_t *>(p + 4));
+
+    const auto shuf3 = this->apply_lookup_16_to(tb3);
+    const auto shuf4 = this->apply_lookup_16_to(tb4);
+    shuf3.store(reinterpret_cast<int8_t *>(p + 8));
+    shuf4.store(reinterpret_cast<int8_t *>(p + 12));
+  }
+  // Conversion from/to SIMD register
+  simdutf_really_inline simd8(const __m128i _value) : value(_value) {}
+  simdutf_really_inline operator const __m128i &() const { return this->value; }
+
+  simdutf_really_inline operator const __m128i() const { return this->value; }
+
+  simdutf_really_inline operator __m128i &() { return this->value; }
+
+  // Zero constructor
+  simdutf_really_inline simd8() : simd8(zero()) {}
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+
+  simdutf_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                              int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                              int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                              int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8((__m128i)v16i8{v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                             v12, v13, v14, v15}) {}
+
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15);
+  }
+
+  // Store to array
+  simdutf_really_inline void store(int8_t dst[16]) const {
+    return __lsx_vst(value, dst, 0);
+  }
+
+  simdutf_really_inline operator simd8<uint8_t>() const {
+    return ((__m128i)this->value);
+  }
+
+  simdutf_really_inline simd8<int8_t>
+  operator|(const simd8<int8_t> other) const {
+    return __lsx_vor_v((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator&(const simd8<int8_t> other) const {
+    return __lsx_vand_v((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator^(const simd8<int8_t> other) const {
+    return __lsx_vxor_v((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  bit_andnot(const simd8<int8_t> other) const {
+    return __lsx_vandn_v((__m128i)other.value, (__m128i)value);
+  }
+
+  // Math
+  simdutf_really_inline simd8<int8_t>
+  operator+(const simd8<int8_t> other) const {
+    return __lsx_vadd_b((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t>
+  operator-(const simd8<int8_t> other) const {
+    return __lsx_vsub_b((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t> &operator+=(const simd8<int8_t> other) {
+    *this = *this + other;
+    return *this;
+  }
+  simdutf_really_inline simd8<int8_t> &operator-=(const simd8<int8_t> other) {
+    *this = *this - other;
+    return *this;
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return (__lsx_vpickve2gr_hu(__lsx_vmskgez_b((__m128i)this->value), 0) ==
+            0xffff);
+  }
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
+    return __lsx_vmax_b((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
+    return __lsx_vmin_b((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return __lsx_vslt_b((__m128i)other.value, (__m128i)value);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return __lsx_vslt_b((__m128i)value, (__m128i)other.value);
+  }
+  simdutf_really_inline simd8<bool>
+  operator==(const simd8<int8_t> other) const {
+    return __lsx_vseq_b((__m128i)value, (__m128i)other.value);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd8<int8_t>
+  prev(const simd8<int8_t> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(this->value, N),
+                       __lsx_vbsrl_v(prev_chunk.value, 16 - N));
+  }
+
+  // Perform a lookup assuming no value is larger than 16
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return lookup_table.apply_lookup_16_to(*this);
+  }
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+
+  template <typename T>
+  simdutf_really_inline simd8<int8_t>
+  apply_lookup_16_to(const simd8<T> original) const {
+    __m128i original_tmp = __lsx_vand_v(original, __lsx_vldi(0x1f));
+    return __lsx_vshuf_b(__lsx_vldi(0), (__m128i)this->value,
+                         simd8<uint8_t>(original_tmp));
+  }
+};
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(
+      NUM_CHUNKS == 4,
+      "LoongArch kernel should use four registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                 const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 2 * sizeof(simd8<T>) / sizeof(T)),
+               simd8<T>::load(ptr + 3 * sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    this->chunks[2] |= other.chunks[2];
+    this->chunks[3] |= other.chunks[3];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+    this->chunks[2].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 2);
+    this->chunks[3].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store_ascii_as_utf32_tbl(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    __m128i mask = __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[3]), 6);
+    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[2]), 4));
+    mask = __lsx_vor_v(mask, __lsx_vbsll_v(__lsx_vmsknz_b(this->chunks[1]), 2));
+    mask = __lsx_vor_v(mask, __lsx_vmsknz_b(this->chunks[0]));
+    return __lsx_vpickve2gr_du(mask, 0);
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                          this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                          this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+
+    return simd8x64<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+    return simd8x64<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
+               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
+               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                          this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask,
+                          this->chunks[2] > mask, this->chunks[3] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask,
+                          this->chunks[2] >= mask, this->chunks[3] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>(simd8<uint8_t>(this->chunks[0].value) >= mask,
+                          simd8<uint8_t>(this->chunks[1].value) >= mask,
+                          simd8<uint8_t>(this->chunks[2].value) >= mask,
+                          simd8<uint8_t>(this->chunks[3].value) >= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+/* begin file src/simdutf/lsx/simd16-inl.h */
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>> struct base_u16 {
+  __m128i value;
+  static const int SIZE = sizeof(value);
+
+  // Conversion from/to SIMD register
+  simdutf_really_inline base_u16() = default;
+  simdutf_really_inline base_u16(const __m128i _value) : value(_value) {}
+  // Bit operations
+  simdutf_really_inline simd16<T> operator|(const simd16<T> other) const {
+    return __lsx_vor_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> operator&(const simd16<T> other) const {
+    return __lsx_vand_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> operator^(const simd16<T> other) const {
+    return __lsx_vxor_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> bit_andnot(const simd16<T> other) const {
+    return __lsx_vandn_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
+  simdutf_really_inline simd16<T> &operator|=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd16<T> &operator&=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline simd16<T> &operator^=(const simd16<T> other) {
+    auto this_cast = static_cast<simd16<T> *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+
+  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return __lsx_vseq_h(lhs.value, rhs.value);
+  }
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(*this, N * 2),
+                       __lsx_vbsrl_v(prev_chunk, 16 - N * 2));
+  }
+};
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base_u16<T> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdutf_really_inline base16() : base_u16<T>() {}
+  simdutf_really_inline base16(const __m128i _value) : base_u16<T>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr)
+      : base16(__lsx_vld(ptr, 0)) {}
+
+  static const int SIZE = sizeof(base_u16<T>::value);
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    return __lsx_vor_v(__lsx_vbsll_v(*this, N * 2),
+                       __lsx_vbsrl_v(prev_chunk, 16 - N * 2));
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return __lsx_vreplgr2vr_h(uint16_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const __m128i _value) : base16<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return __lsx_vreplgr2vr_h(_value);
+  }
+  static simdutf_really_inline simd16<T> zero() { return __lsx_vldi(0); }
+  static simdutf_really_inline simd16<T> load(const T values[8]) {
+    return __lsx_vld(reinterpret_cast<const uint16_t *>(values), 0);
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const __m128i _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return __lsx_vst(this->value, dst, 0);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
+    return __lsx_vadd_b(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
+    return __lsx_vsub_b(*this, other);
+  }
+  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
+    *this = *this - other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Signed code unitstemplate<>
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+  simdutf_really_inline simd16(const __m128i _value)
+      : base16_numeric<int16_t>(_value) {}
+  simdutf_really_inline simd16(simd16<bool> other)
+      : base16_numeric<int16_t>(other.value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
+  simdutf_really_inline operator simd16<uint16_t>() const;
+
+  // Order-sensitive comparisons
+  simdutf_really_inline simd16<int16_t>
+  max_val(const simd16<int16_t> other) const {
+    return __lsx_vmax_h(this->value, other.value);
+  }
+  simdutf_really_inline simd16<int16_t>
+  min_val(const simd16<int16_t> other) const {
+    return __lsx_vmin_h(this->value, other.value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<int16_t> other) const {
+    return __lsx_vsle_h(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<int16_t> other) const {
+    return __lsx_vslt_h(this->value, other.value);
+  }
+};
+
+// Unsigned code unitstemplate<>
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const __m128i _value)
+      : base16_numeric<uint16_t>((__m128i)_value) {}
+  simdutf_really_inline simd16(simd16<bool> other)
+      : base16_numeric<uint16_t>(other.value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+
+  // Saturated math
+  simdutf_really_inline simd16<uint16_t>
+  saturating_add(const simd16<uint16_t> other) const {
+    return __lsx_vsadd_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  saturating_sub(const simd16<uint16_t> other) const {
+    return __lsx_vssub_hu(this->value, other.value);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return __lsx_vmax_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return __lsx_vmin_hu(this->value, other.value);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  gt_bits(const simd16<uint16_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  lt_bits(const simd16<uint16_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return __lsx_vsle_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return __lsx_vsle_hu(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<uint16_t> other) const {
+    return __lsx_vslt_hu(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return __lsx_vslt_hu(this->value, other.value);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(__lsx_vsrli_h(this->value, N));
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return simd16<uint16_t>(__lsx_vslli_h(this->value, N));
+  }
+
+  // logical operations
+  simdutf_really_inline simd16<uint16_t>
+  operator|(const simd16<uint16_t> other) const {
+    return __lsx_vor_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  operator&(const simd16<uint16_t> other) const {
+    return __lsx_vand_v(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  operator^(const simd16<uint16_t> other) const {
+    return __lsx_vxor_v(this->value, other.value);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+    return __lsx_vssrlni_bu_h(v1.value, v0.value, 0);
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    return __lsx_vshuf4i_b(this->value, 0b10110001);
+  }
+};
+
+simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(
+      NUM_CHUNKS == 4,
+      "LOONGARCH kernel should use four registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline
+  simd16x32(const simd16<T> chunk0, const simd16<T> chunk1,
+            const simd16<T> chunk2, const simd16<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 2 * sizeof(simd16<T>) / sizeof(T)),
+               simd16<T>::load(ptr + 3 * sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+    this->chunks[2].store(ptr + sizeof(simd16<T>) * 2 / sizeof(T));
+    this->chunks[3].store(ptr + sizeof(simd16<T>) * 3 / sizeof(T));
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdutf_really_inline bool is_ascii() const { return reduce_or().is_ascii(); }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 1);
+    this->chunks[2].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 2);
+    this->chunks[3].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 3);
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    __m128i mask = __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[3]).value), 6);
+    mask = __lsx_vor_v(
+        mask, __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[2]).value), 4));
+    mask = __lsx_vor_v(
+        mask, __lsx_vbsll_v(__lsx_vmsknz_b((this->chunks[1]).value), 2));
+    mask = __lsx_vor_v(mask, __lsx_vmsknz_b((this->chunks[0]).value));
+    return __lsx_vpickve2gr_du(mask, 0);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+    this->chunks[2] = this->chunks[2].swap_bytes();
+    this->chunks[3] = this->chunks[3].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                           this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                           this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+
+    return simd16x32<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
+               (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
+               (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+    return simd16x32<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
+               (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
+               (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask,
+                           this->chunks[2] < mask, this->chunks[3] < mask)
+        .to_bitmask();
+  }
+
+}; // struct simd16x32<T>
+
+template <>
+simdutf_really_inline uint64_t simd16x32<uint16_t>::not_in_range(
+    const uint16_t low, const uint16_t high) const {
+  const simd16<uint16_t> mask_low = simd16<uint16_t>::splat(low);
+  const simd16<uint16_t> mask_high = simd16<uint16_t>::splat(high);
+  simd16x32<uint16_t> x(simd16<uint16_t>((this->chunks[0] > mask_high) |
+                                         (this->chunks[0] < mask_low)),
+                        simd16<uint16_t>((this->chunks[1] > mask_high) |
+                                         (this->chunks[1] < mask_low)),
+                        simd16<uint16_t>((this->chunks[2] > mask_high) |
+                                         (this->chunks[2] < mask_low)),
+                        simd16<uint16_t>((this->chunks[3] > mask_high) |
+                                         (this->chunks[3] < mask_low)));
+  return x.to_bitmask();
+}
+/* end file src/simdutf/lsx/simd16-inl.h */
+/* begin file src/simdutf/lsx/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  __m128i value;
+  static const int SIZE = sizeof(value);
+  static const int ELEMENTS = SIZE / sizeof(uint32_t);
+
+  // constructors
+  simdutf_really_inline simd32(__m128i v) : value(v) {}
+
+  template <typename Ptr>
+  simdutf_really_inline simd32(Ptr *ptr) : value(__lsx_vld(ptr, 0)) {}
+
+  // in-place operators
+  simdutf_really_inline simd32 &operator-=(const simd32 other) {
+    value = __lsx_vsub_w(value, other.value);
+    return *this;
+  }
+
+  // members
+  simdutf_really_inline uint64_t sum() const {
+    return uint64_t(__lsx_vpickve2gr_wu(value, 0)) +
+           uint64_t(__lsx_vpickve2gr_wu(value, 1)) +
+           uint64_t(__lsx_vpickve2gr_wu(value, 2)) +
+           uint64_t(__lsx_vpickve2gr_wu(value, 3));
+  }
+
+  // static members
+  static simdutf_really_inline simd32<uint32_t> splat(uint32_t x) {
+    return __lsx_vreplgr2vr_w(x);
+  }
+
+  static simdutf_really_inline simd32<uint32_t> zero() {
+    return __lsx_vrepli_w(0);
+  }
+};
+
+// ------------------------------------------------------------
+
+template <> struct simd32<bool> {
+  __m128i value;
+  static const int SIZE = sizeof(value);
+
+  // constructors
+  simdutf_really_inline simd32(__m128i v) : value(v) {}
+};
+
+// ------------------------------------------------------------
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return __lsx_vor_v(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator<(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return __lsx_vslt_wu(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator>(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return __lsx_vslt_wu(b.value, a.value);
+}
+
+// ------------------------------------------------------------
+
+simdutf_really_inline simd32<uint32_t> as_vector_u32(const simd32<bool> v) {
+  return v.value;
+}
+/* end file src/simdutf/lsx/simd32-inl.h */
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdutf
+
+#endif // SIMDUTF_LSX_SIMD_H
+/* end file src/simdutf/lsx/simd.h */
+
+/* begin file src/simdutf/lsx/end.h */
+#undef SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+/* end file src/simdutf/lsx/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_LSX
+
+#endif // SIMDUTF_LSX_H
+/* end file src/simdutf/lsx.h */
+/* begin file src/simdutf/lasx.h */
+#ifndef SIMDUTF_LASX_H
+#define SIMDUTF_LASX_H
+
+#ifdef SIMDUTF_FALLBACK_H
+  #error "lasx.h must be included before fallback.h"
+#endif
+
+
+#ifndef SIMDUTF_IMPLEMENTATION_LASX
+  #define SIMDUTF_IMPLEMENTATION_LASX (SIMDUTF_IS_LASX)
+#endif
+#if SIMDUTF_IMPLEMENTATION_LASX && SIMDUTF_IS_LASX
+  #define SIMDUTF_CAN_ALWAYS_RUN_LASX 1
+#else
+  #define SIMDUTF_CAN_ALWAYS_RUN_LASX 0
+#endif
+
+#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
+
+#if SIMDUTF_IMPLEMENTATION_LASX
+
+namespace simdutf {
+/**
+ * Implementation for LoongArch ASX.
+ */
+namespace lasx {} // namespace lasx
+} // namespace simdutf
+
+/* begin file src/simdutf/lasx/implementation.h */
+#ifndef SIMDUTF_LASX_IMPLEMENTATION_H
+#define SIMDUTF_LASX_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace lasx {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("lasx", "LOONGARCH ASX",
+                                internal::instruction_set::LSX |
+                                    internal::instruction_set::LASX) {}
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+
+} // namespace lasx
+} // namespace simdutf
+
+#endif // SIMDUTF_LASX_IMPLEMENTATION_H
+/* end file src/simdutf/lasx/implementation.h */
+
+/* begin file src/simdutf/lasx/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "lasx"
+// #define SIMDUTF_IMPLEMENTATION lasx
+#define SIMDUTF_SIMD_HAS_UNSIGNED_CMP 1
+/* end file src/simdutf/lasx/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/lasx/intrinsics.h */
+#ifndef SIMDUTF_LASX_INTRINSICS_H
+#define SIMDUTF_LASX_INTRINSICS_H
+
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <lsxintrin.h>
+#include <lasxintrin.h>
+
+#if defined(__loongarch_asx)
+  #ifdef __clang__
+    #define VREGS_PREFIX "$vr"
+    #define XREGS_PREFIX "$xr"
+  #else // GCC
+    #define VREGS_PREFIX "$f"
+    #define XREGS_PREFIX "$f"
+  #endif
+  #define __ALL_REGS                                                           \
+    "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,"  \
+    "27,28,29,30,31"
+// Convert __m128i to __m256i
+static inline __m256i ____m256i(__m128i in) {
+  __m256i out = __lasx_xvldi(0);
+  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
+                   " .ifc %[out], " XREGS_PREFIX "\\i    \n\t"
+                   "  .irp j," __ALL_REGS "\n\t"
+                   "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
+                   "    xvpermi.q $xr\\i, $xr\\j, 0x0  \n\t"
+                   "   .endif                           \n\t"
+                   "  .endr                             \n\t"
+                   " .endif                             \n\t"
+                   ".endr                               \n\t"
+                   : [out] "+f"(out)
+                   : [in] "f"(in));
+  return out;
+}
+// Convert two __m128i to __m256i
+static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
+  __m256i out;
+  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
+                   " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
+                   "  .irp j," __ALL_REGS "\n\t"
+                   "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
+                   "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+                   "   .endif                           \n\t"
+                   "  .endr                             \n\t"
+                   " .endif                             \n\t"
+                   ".endr                               \n\t"
+                   ".ifnc %[out], %[hi]                 \n\t"
+                   ".irp i," __ALL_REGS "\n\t"
+                   " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
+                   "  .irp j," __ALL_REGS "\n\t"
+                   "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
+                   "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
+                   "   .endif                           \n\t"
+                   "  .endr                             \n\t"
+                   " .endif                             \n\t"
+                   ".endr                               \n\t"
+                   ".endif                              \n\t"
+                   : [out] "=f"(out), [hi] "+f"(inhi)
+                   : [lo] "f"(inlo));
+  return out;
+}
+// Convert __m256i low part to __m128i
+static inline __m128i lasx_extracti128_lo(__m256i in) {
+  __m128i out;
+  __asm__ volatile(".ifnc %[out], %[in]                 \n\t"
+                   ".irp i," __ALL_REGS "\n\t"
+                   " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+                   "  .irp j," __ALL_REGS "\n\t"
+                   "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+                   "    vori.b $vr\\i, $vr\\j, 0        \n\t"
+                   "   .endif                           \n\t"
+                   "  .endr                             \n\t"
+                   " .endif                             \n\t"
+                   ".endr                               \n\t"
+                   ".endif                              \n\t"
+                   : [out] "=f"(out)
+                   : [in] "f"(in));
+  return out;
+}
+// Convert __m256i high part to __m128i
+static inline __m128i lasx_extracti128_hi(__m256i in) {
+  __m128i out;
+  __asm__ volatile(".irp i," __ALL_REGS "\n\t"
+                   " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+                   "  .irp j," __ALL_REGS "\n\t"
+                   "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+                   "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
+                   "   .endif                           \n\t"
+                   "  .endr                             \n\t"
+                   " .endif                             \n\t"
+                   ".endr                               \n\t"
+                   : [out] "=f"(out)
+                   : [in] "f"(in));
+  return out;
+}
+#endif
+
+/*
+Encoding of argument for LoongArch64 xvldi instruction.  See:
+https://jia.je/unofficial-loongarch-intrinsics-guide/lasx/misc/#__m256i-__lasx_xvldi-imm_n1024_1023-imm
+
+1: imm[12:8]=0b10000: broadcast imm[7:0] as 32-bit elements to all lanes
+
+2: imm[12:8]=0b10001: broadcast imm[7:0] << 8 as 32-bit elements to all lanes
+
+3: imm[12:8]=0b10010: broadcast imm[7:0] << 16 as 32-bit elements to all lanes
+
+4: imm[12:8]=0b10011: broadcast imm[7:0] << 24 as 32-bit elements to all lanes
+
+5: imm[12:8]=0b10100: broadcast imm[7:0] as 16-bit elements to all lanes
+
+6: imm[12:8]=0b10101: broadcast imm[7:0] << 8 as 16-bit elements to all lanes
+
+7: imm[12:8]=0b10110: broadcast (imm[7:0] << 8) | 0xFF as 32-bit elements to all
+lanes
+
+8: imm[12:8]=0b10111: broadcast (imm[7:0] << 16) | 0xFFFF as 32-bit elements to
+all lanes
+
+9: imm[12:8]=0b11000: broadcast imm[7:0] as 8-bit elements to all lanes
+
+10: imm[12:8]=0b11001: repeat each bit of imm[7:0] eight times, and broadcast
+the result as 64-bit elements to all lanes
+*/
+
+namespace lasx_vldi {
+
+template <uint16_t v> class const_u16 {
+  constexpr static const uint8_t b0 = ((v >> 0 * 8) & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 1 * 8) & 0xff);
+
+  constexpr static bool is_case5 = uint16_t(b0) == v;
+  constexpr static bool is_case6 = (uint16_t(b1) << 8) == v;
+  constexpr static bool is_case9 = (b0 == b1);
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00));
+
+public:
+  constexpr static uint16_t operation = is_case5    ? 0b10100
+                                        : is_case6  ? 0b10101
+                                        : is_case9  ? 0b11000
+                                        : is_case10 ? 0x11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_case5    ? b0
+      : is_case6  ? b1
+      : is_case9  ? b0
+      : is_case10 ? ((b0 ? 0x55 : 0x00) | (b1 ? 0xaa : 0x00))
+                  : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+
+template <uint32_t v> class const_u32 {
+  constexpr static const uint8_t b0 = (v & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 8) & 0xff);
+  constexpr static const uint8_t b2 = ((v >> 16) & 0xff);
+  constexpr static const uint8_t b3 = ((v >> 24) & 0xff);
+
+  constexpr static bool is_case1 = (uint32_t(b0) == v);
+  constexpr static bool is_case2 = ((uint32_t(b1) << 8) == v);
+  constexpr static bool is_case3 = ((uint32_t(b2) << 16) == v);
+  constexpr static bool is_case4 = ((uint32_t(b3) << 24) == v);
+  constexpr static bool is_case5 = (b0 == b2) && (b1 == 0) && (b3 == 0);
+  constexpr static bool is_case6 = (b1 == b3) && (b0 == 0) && (b2 == 0);
+  constexpr static bool is_case7 = (b3 == 0) && (b2 == 0) && (b0 == 0xff);
+  constexpr static bool is_case8 = (b3 == 0) && (b1 == 0xff) && (b0 == 0xff);
+  constexpr static bool is_case9 = (b0 == b1) && (b0 == b2) && (b0 == b3);
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00)) &&
+      ((b2 == 0xff) || (b2 == 0x00)) && ((b3 == 0xff) || (b3 == 0x00));
+
+public:
+  constexpr static uint16_t operation = is_case1    ? 0b10000
+                                        : is_case2  ? 0b10001
+                                        : is_case3  ? 0b10010
+                                        : is_case4  ? 0b10011
+                                        : is_case5  ? 0b10100
+                                        : is_case6  ? 0b10101
+                                        : is_case7  ? 0b10110
+                                        : is_case8  ? 0b10111
+                                        : is_case9  ? 0b11000
+                                        : is_case10 ? 0b11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_case1    ? b0
+      : is_case2  ? b1
+      : is_case3  ? b2
+      : is_case4  ? b3
+      : is_case5  ? b0
+      : is_case6  ? b1
+      : is_case7  ? b1
+      : is_case8  ? b2
+      : is_case9  ? b0
+      : is_case10 ? ((b0 ? 0x11 : 0x00) | (b1 ? 0x22 : 0x00) |
+                     (b2 ? 0x44 : 0x00) | (b3 ? 0x88 : 0x00))
+                  : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+
+template <uint64_t v> class const_u64 {
+  constexpr static const uint8_t b0 = ((v >> 0 * 8) & 0xff);
+  constexpr static const uint8_t b1 = ((v >> 1 * 8) & 0xff);
+  constexpr static const uint8_t b2 = ((v >> 2 * 8) & 0xff);
+  constexpr static const uint8_t b3 = ((v >> 3 * 8) & 0xff);
+  constexpr static const uint8_t b4 = ((v >> 4 * 8) & 0xff);
+  constexpr static const uint8_t b5 = ((v >> 5 * 8) & 0xff);
+  constexpr static const uint8_t b6 = ((v >> 6 * 8) & 0xff);
+  constexpr static const uint8_t b7 = ((v >> 7 * 8) & 0xff);
+
+  constexpr static bool is_case10 =
+      ((b0 == 0xff) || (b0 == 0x00)) && ((b1 == 0xff) || (b1 == 0x00)) &&
+      ((b2 == 0xff) || (b2 == 0x00)) && ((b3 == 0xff) || (b3 == 0x00)) &&
+      ((b4 == 0xff) || (b4 == 0x00)) && ((b5 == 0xff) || (b5 == 0x00)) &&
+      ((b6 == 0xff) || (b6 == 0x00)) && ((b7 == 0xff) || (b7 == 0x00));
+
+public:
+  constexpr static bool is_32bit =
+      ((v & 0xffffffff) == (v >> 32)) && const_u32<(v >> 32)>::value;
+  constexpr static uint8_t op_32bit = const_u32<(v >> 32)>::operation;
+  constexpr static uint8_t byte_32bit = const_u32<(v >> 32)>::byte;
+
+  constexpr static uint16_t operation = is_32bit    ? op_32bit
+                                        : is_case10 ? 0x11001
+                                                    : 0xffff;
+
+  constexpr static uint16_t byte =
+      is_32bit ? byte_32bit
+      : is_case10
+          ? ((b0 ? 0x01 : 0x00) | (b1 ? 0x02 : 0x00) | (b2 ? 0x04 : 0x00) |
+             (b3 ? 0x08 : 0x00) | (b4 ? 0x10 : 0x00) | (b5 ? 0x20 : 0x00) |
+             (b6 ? 0x40 : 0x00) | (b7 ? 0x80 : 0x00))
+          : 0xffff;
+
+  constexpr static int value = int((operation << 8) | byte) - 8192;
+  constexpr static bool valid = operation != 0xffff;
+};
+
+} // namespace lasx_vldi
+
+// Uncomment when running under QEMU affected
+// by bug https://gitlab.com/qemu-project/qemu/-/issues/2865
+// Versions <= 9.2.2 are affected, likely anything newer is correct.
+#ifndef QEMU_VLDI_BUG
+// #define QEMU_VLDI_BUG 1
+#endif
+
+#ifdef QEMU_VLDI_BUG
+  #define lasx_splat_u16(v) __lasx_xvreplgr2vr_h(v)
+  #define lasx_splat_u32(v) __lasx_xvreplgr2vr_w(v)
+#else
+template <uint16_t x> constexpr __m256i lasx_splat_u16_aux() {
+  constexpr bool is_imm10 = (int16_t(x) < 512) && (int16_t(x) > -512);
+  constexpr uint16_t imm10 = is_imm10 ? x : 0;
+  constexpr bool is_vldi = lasx_vldi::const_u16<x>::valid;
+  constexpr int vldi_imm = is_vldi ? lasx_vldi::const_u16<x>::value : 0;
+
+  return is_imm10  ? __lasx_xvrepli_h(int16_t(imm10))
+         : is_vldi ? __lasx_xvldi(vldi_imm)
+                   : __lasx_xvreplgr2vr_h(x);
+}
+
+template <uint32_t x> constexpr __m256i lasx_splat_u32_aux() {
+  constexpr bool is_imm10 = (int32_t(x) < 512) && (int32_t(x) > -512);
+  constexpr uint32_t imm10 = is_imm10 ? x : 0;
+  constexpr bool is_vldi = lasx_vldi::const_u32<x>::valid;
+  constexpr int vldi_imm = is_vldi ? lasx_vldi::const_u32<x>::value : 0;
+
+  return is_imm10  ? __lasx_xvrepli_w(int32_t(imm10))
+         : is_vldi ? __lasx_xvldi(vldi_imm)
+                   : __lasx_xvreplgr2vr_w(x);
+}
+
+  #define lasx_splat_u16(v) lasx_splat_u16_aux<(v)>()
+  #define lasx_splat_u32(v) lasx_splat_u32_aux<(v)>()
+#endif // QEMU_VLDI_BUG
+
+#endif //  SIMDUTF_LASX_INTRINSICS_H
+/* end file src/simdutf/lasx/intrinsics.h */
+/* begin file src/simdutf/lasx/bitmanipulation.h */
+#ifndef SIMDUTF_LASX_BITMANIPULATION_H
+#define SIMDUTF_LASX_BITMANIPULATION_H
+
+#include <limits>
+
+namespace simdutf {
+namespace lasx {
+namespace {
+
+simdutf_really_inline int count_ones(uint64_t input_num) {
+  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__lsx_vreplgr2vr_d(input_num)), 0);
+}
+
+#if SIMDUTF_NEED_TRAILING_ZEROES
+// simdutf_really_inline int trailing_zeroes(uint64_t input_num) {
+//   return __builtin_ctzll(input_num);
+// }
+#endif
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdutf
+
+#endif // SIMDUTF_LASX_BITMANIPULATION_H
+/* end file src/simdutf/lasx/bitmanipulation.h */
+/* begin file src/simdutf/lasx/simd.h */
+#ifndef SIMDUTF_LASX_SIMD_H
+#define SIMDUTF_LASX_SIMD_H
+
+#include <type_traits>
+
+namespace simdutf {
+namespace lasx {
+namespace {
+namespace simd {
+
+__attribute__((aligned(32))) static const uint8_t prev_shuf_table[32][32] = {
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+     0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+    {0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14,
+     31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14},
+    {0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13,
+     30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13},
+    {0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
+     29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12},
+    {0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
+     28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
+    {0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,
+     27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10},
+    {0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
+     26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9},
+    {0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7, 8,
+     25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6, 7,
+     24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5, 6,
+     23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4, 5,
+     22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4, 5},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3, 4,
+     21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3, 4},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2, 3,
+     20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2, 3},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1, 2,
+     19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1, 2},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0, 1,
+     18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0, 1},
+    {0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+     17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 0},
+    {15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30,
+     15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29,
+     14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28,
+     13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27,
+     12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26,
+     11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25,
+     10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
+     9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0,  0},
+    {8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23,
+     8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0,  0},
+    {7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22,
+     7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0,  0},
+    {6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21,
+     6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0,  0},
+    {5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
+     5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0,  0},
+    {4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
+     4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0,  0},
+    {3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
+     3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0,  0},
+    {2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
+     2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0,  0},
+    {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
+     1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0},
+    {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15,
+     0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15},
+};
+
+__attribute__((aligned(32))) static const uint8_t bitsel_mask_table[32][32] = {
+    {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0,  0x0},
+    {0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF,
+     0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x0}};
+
+// Forward-declared so they can be used by splat and friends.
+template <typename Child> struct base {
+  __m256i value;
+
+  // Zero constructor
+  simdutf_really_inline base() : value{__m256i()} {}
+
+  // Conversion from SIMD register
+  simdutf_really_inline base(const __m256i _value) : value(_value) {}
+  // Conversion to SIMD register
+  simdutf_really_inline operator const __m256i &() const { return this->value; }
+  simdutf_really_inline operator __m256i &() { return this->value; }
+  template <endianness big_endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    if (big_endian) {
+      __m256i zero = __lasx_xvldi(0);
+      __m256i in8 = __lasx_xvpermi_d(this->value, 0b11011000);
+      __m256i inlow = __lasx_xvilvl_b(in8, zero);
+      __m256i inhigh = __lasx_xvilvh_b(in8, zero);
+      __lasx_xvst(inlow, reinterpret_cast<uint16_t *>(ptr), 0);
+      __lasx_xvst(inhigh, reinterpret_cast<uint16_t *>(ptr), 32);
+    } else {
+      __m256i inlow = __lasx_vext2xv_hu_bu(this->value);
+      __m256i inhigh = __lasx_vext2xv_hu_bu(
+          __lasx_xvpermi_q(this->value, this->value, 0b00000001));
+      __lasx_xvst(inlow, reinterpret_cast<__m256i *>(ptr), 0);
+      __lasx_xvst(inhigh, reinterpret_cast<__m256i *>(ptr), 32);
+    }
+  }
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    __m256i in32_0 = __lasx_vext2xv_wu_bu(this->value);
+    __lasx_xvst(in32_0, reinterpret_cast<uint32_t *>(ptr), 0);
+
+    __m256i in8_1 = __lasx_xvpermi_d(this->value, 0b00000001);
+    __m256i in32_1 = __lasx_vext2xv_wu_bu(in8_1);
+    __lasx_xvst(in32_1, reinterpret_cast<uint32_t *>(ptr), 32);
+
+    __m256i in8_2 = __lasx_xvpermi_d(this->value, 0b00000010);
+    __m256i in32_2 = __lasx_vext2xv_wu_bu(in8_2);
+    __lasx_xvst(in32_2, reinterpret_cast<uint32_t *>(ptr), 64);
+
+    __m256i in8_3 = __lasx_xvpermi_d(this->value, 0b00000011);
+    __m256i in32_3 = __lasx_vext2xv_wu_bu(in8_3);
+    __lasx_xvst(in32_3, reinterpret_cast<uint32_t *>(ptr), 96);
+  }
+  // Bit operations
+  simdutf_really_inline Child operator|(const Child other) const {
+    return __lasx_xvor_v(this->value, other);
+  }
+  simdutf_really_inline Child operator&(const Child other) const {
+    return __lasx_xvand_v(this->value, other);
+  }
+  simdutf_really_inline Child operator^(const Child other) const {
+    return __lasx_xvxor_v(this->value, other);
+  }
+  simdutf_really_inline Child bit_andnot(const Child other) const {
+    return __lasx_xvandn_v(this->value, other);
+  }
+  simdutf_really_inline Child &operator|=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator&=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdutf_really_inline Child &operator^=(const Child other) {
+    auto this_cast = static_cast<Child *>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+};
+
+template <typename T> struct simd8;
+
+template <typename T, typename Mask = simd8<bool>>
+struct base8 : base<simd8<T>> {
+  typedef uint32_t bitmask_t;
+  typedef uint64_t bitmask2_t;
+
+  simdutf_really_inline base8() : base<simd8<T>>() {}
+  simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+  simdutf_really_inline T first() const {
+    return __lasx_xvpickve2gr_wu(this->value, 0);
+  }
+  simdutf_really_inline T last() const {
+    return __lasx_xvpickve2gr_wu(this->value, 7);
+  }
+  friend simdutf_really_inline Mask operator==(const simd8<T> lhs,
+                                               const simd8<T> rhs) {
+    return __lasx_xvseq_b(lhs, rhs);
+  }
+
+  static const int SIZE = sizeof(base<T>::value);
+
+  template <int N = 1>
+  simdutf_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    if (!N)
+      return this->value;
+
+    __m256i zero = __lasx_xvldi(0);
+    __m256i result, shuf;
+    if (N < 16) {
+      shuf = __lasx_xvld(prev_shuf_table[N], 0);
+
+      result = __lasx_xvshuf_b(
+          __lasx_xvpermi_q(this->value, this->value, 0b00000001), this->value,
+          shuf);
+      __m256i srl_prev = __lasx_xvbsrl_v(
+          __lasx_xvpermi_q(zero, prev_chunk.value, 0b00110001), (16 - N));
+      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
+      result = __lasx_xvbitsel_v(result, srl_prev, mask);
+
+      return result;
+    } else if (N == 16) {
+      return __lasx_xvpermi_q(this->value, prev_chunk.value, 0b00100001);
+    } /*else {
+      __m256i sll_value = __lasx_xvbsll_v(
+          __lasx_xvpermi_q(zero, this->value, 0b00000011), (N - 16) % 32);
+      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
+      shuf = __lasx_xvld(prev_shuf_table[N], 0);
+      result = __lasx_xvshuf_b(
+          __lasx_xvpermi_q(prev_chunk.value, prev_chunk.value, 0b00000001),
+          prev_chunk.value, shuf);
+      result = __lasx_xvbitsel_v(sll_value, result, mask);
+      return result;
+    }*/
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  static simdutf_really_inline simd8<bool> splat(bool _value) {
+    return __lasx_xvreplgr2vr_b(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd8() : base8() {}
+  simdutf_really_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+  simdutf_really_inline uint32_t to_bitmask() const {
+    __m256i mask = __lasx_xvmsknz_b(this->value);
+    uint32_t mask0 = __lasx_xvpickve2gr_wu(mask, 0);
+    uint32_t mask1 = __lasx_xvpickve2gr_wu(mask, 4);
+    return (mask0 | (mask1 << 16));
+  }
+  simdutf_really_inline bool any() const {
+    if (__lasx_xbz_b(this->value))
+      return false;
+    return true;
+  }
+  simdutf_really_inline bool none() const {
+    if (__lasx_xbz_b(this->value))
+      return true;
+    return false;
+  }
+  simdutf_really_inline bool all() const {
+    if (__lasx_xbnz_b(this->value))
+      return true;
+    return false;
+  }
+  simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  static simdutf_really_inline simd8<T> splat(T _value) {
+    return __lasx_xvreplgr2vr_b(_value);
+  }
+  static simdutf_really_inline simd8<T> zero() { return __lasx_xvldi(0); }
+  static simdutf_really_inline simd8<T> load(const T values[32]) {
+    return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
+  }
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdutf_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                  T v5, T v6, T v7, T v8, T v9,
+                                                  T v10, T v11, T v12, T v13,
+                                                  T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11,
+                    v12, v13, v14, v15);
+  }
+
+  simdutf_really_inline base8_numeric() : base8<T>() {}
+  simdutf_really_inline base8_numeric(const __m256i _value)
+      : base8<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[32]) const {
+    return __lasx_xvst(this->value, reinterpret_cast<__m256i *>(dst), 0);
+  }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
+    return __lasx_xvadd_b(this->value, other);
+  }
+  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
+    return __lasx_xvsub_b(this->value, other);
+  }
+  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
+    *this = *this + other;
+    return *static_cast<simd8<T> *>(this);
+  }
+  simdutf_really_inline simd8<T> &operator-=(const simd8<T> other) {
+    *this = *this - other;
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    __m256i origin = __lasx_xvand_v(this->value, __lasx_xvldi(0x1f));
+    return __lasx_xvshuf_b(__lasx_xvldi(0), lookup_table, origin);
+  }
+
+  template <typename L>
+  simdutf_really_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
+  simdutf_really_inline simd8(const __m256i _value)
+      : base8_numeric<int8_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+  simdutf_really_inline operator simd8<uint8_t>() const;
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+        int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+        int8_t v12, int8_t v13, int8_t v14, int8_t v15, int8_t v16, int8_t v17,
+        int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+        int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29,
+        int8_t v30, int8_t v31)
+      : simd8((__m256i)v32i8{v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7,
+                             v8,  v9,  v10, v11, v12, v13, v14, v15,
+                             v16, v17, v18, v19, v20, v21, v22, v23,
+                             v24, v25, v26, v27, v28, v29, v30, v31}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                         v10, v11, v12, v13, v14, v15);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    __m256i ascii_mask = __lasx_xvslti_b(this->value, 0);
+    if (__lasx_xbnz_v(ascii_mask))
+      return false;
+    return true;
+  }
+  // Order-sensitive comparisons
+  simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const {
+    return __lasx_xvmax_b(this->value, other);
+  }
+  simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const {
+    return __lasx_xvmin_b(this->value, other);
+  }
+  simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const {
+    return __lasx_xvslt_b(other, this->value);
+  }
+  simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const {
+    return __lasx_xvslt_b(this->value, other);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
+  simdutf_really_inline simd8(const __m256i _value)
+      : base8_numeric<uint8_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdutf_really_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+        uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20,
+        uint8_t v21, uint8_t v22, uint8_t v23, uint8_t v24, uint8_t v25,
+        uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30,
+        uint8_t v31)
+      : simd8((__m256i)v32u8{v0,  v1,  v2,  v3,  v4,  v5,  v6,  v7,
+                             v8,  v9,  v10, v11, v12, v13, v14, v15,
+                             v16, v17, v18, v19, v20, v21, v22, v23,
+                             v24, v25, v26, v27, v28, v29, v30, v31}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdutf_really_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15, v0, v1, v2, v3, v4, v5, v6, v7, v8, v9,
+                          v10, v11, v12, v13, v14, v15);
+  }
+
+  // Saturated math
+  simdutf_really_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return __lasx_xvsadd_bu(this->value, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return __lasx_xvssub_bu(this->value, other);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return __lasx_xvmax_bu(*this, other);
+  }
+  simdutf_really_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return __lasx_xvmin_bu(*this, other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return __lasx_xvsle_bu(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return __lasx_xvsle_bu(other, *this);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return __lasx_xvslt_bu(*this, other);
+  }
+  simdutf_really_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return __lasx_xvslt_bu(other, *this);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd8<bool> bits_not_set() const {
+    return *this == uint8_t(0);
+  }
+  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+  simdutf_really_inline bool is_ascii() const {
+    __m256i ascii_mask = __lasx_xvslti_b(this->value, 0);
+    if (__lasx_xbnz_v(ascii_mask))
+      return false;
+    return true;
+  }
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    if (__lasx_xbnz_v(this->value))
+      return true;
+    return false;
+  }
+  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return (*this & bits).any_bits_set_anywhere();
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return __lasx_xvsrli_b(this->value, N);
+  }
+  template <int N> simdutf_really_inline simd8<uint8_t> shl() const {
+    return __lasx_xvslli_b(this->value, N);
+  }
+};
+simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const {
+  return this->value;
+}
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 2,
+                "LASX kernel should use two registers per 64-byte block.");
+  simd8<T> chunks[NUM_CHUNKS];
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T> other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1)
+      : chunks{chunk0, chunk1} {}
+  simdutf_really_inline simd8x64(const T *ptr)
+      : chunks{simd8<T>::load(ptr),
+               simd8<T>::load(ptr + sizeof(simd8<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1 / sizeof(T));
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r_hi = this->chunks[1].to_bitmask();
+    return r_lo | (r_hi << 32);
+  }
+
+  simdutf_really_inline simd8x64<T> &operator|=(const simd8x64<T> &other) {
+    this->chunks[0] |= other.chunks[0];
+    this->chunks[1] |= other.chunks[1];
+    return *this;
+  }
+
+  simdutf_really_inline simd8<T> reduce_or() const {
+    return this->chunks[0] | this->chunks[1];
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  template <endianness endian>
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 0);
+    this->chunks[1].template store_ascii_as_utf16<endian>(ptr +
+                                                          sizeof(simd8<T>) * 1);
+  }
+
+  simdutf_really_inline void store_ascii_as_utf32(char32_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store_ascii_as_utf32(ptr + sizeof(simd8<T>) * 1);
+  }
+
+  simdutf_really_inline simd8x64<T> bit_or(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<T>(this->chunks[0] | mask, this->chunks[1] | mask);
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
+                          this->chunks[1] == other.chunks[1])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+
+    return simd8x64<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd8<T> mask_low = simd8<T>::splat(low);
+    const simd8<T> mask_high = simd8<T>::splat(high);
+    return simd8x64<bool>(
+               (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
+               (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t gt(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] > mask, this->chunks[1] > mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] >= mask, this->chunks[1] >= mask)
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
+    const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
+    return simd8x64<bool>((simd8<uint8_t>(__m256i(this->chunks[0])) >= mask),
+                          (simd8<uint8_t>(__m256i(this->chunks[1])) >= mask))
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+/* begin file src/simdutf/lasx/simd16-inl.h */
+template <typename T> struct simd16;
+
+template <typename T, typename Mask = simd16<bool>>
+struct base16 : base<simd16<T>> {
+  using bitmask_type = uint32_t;
+
+  simdutf_really_inline base16() : base<simd16<T>>() {}
+  simdutf_really_inline base16(const __m256i _value)
+      : base<simd16<T>>(_value) {}
+  template <typename Pointer>
+  simdutf_really_inline base16(const Pointer *ptr)
+      : base16(__lasx_xvld(reinterpret_cast<const __m256i *>(ptr), 0)) {}
+  friend simdutf_really_inline Mask operator==(const simd16<T> lhs,
+                                               const simd16<T> rhs) {
+    return __lasx_xvseq_h(lhs.value, rhs.value);
+  }
+
+  /// the size of vector in bytes
+  static const int SIZE = sizeof(base<simd16<T>>::value);
+
+  /// the number of elements of type T a vector can hold
+  static const int ELEMENTS = SIZE / sizeof(T);
+
+  template <int N = 1>
+  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
+    if (!N)
+      return this->value;
+
+    __m256i zero = __lasx_xvldi(0);
+    __m256i result, shuf;
+    if (N < 8) {
+      shuf = __lasx_xvld(prev_shuf_table[N * 2], 0);
+
+      result = __lasx_xvshuf_b(
+          __lasx_xvpermi_q(this->value, this->value, 0b00000001), this->value,
+          shuf);
+      __m256i srl_prev = __lasx_xvbsrl_v(
+          __lasx_xvpermi_q(zero, prev_chunk, 0b00110001), (16 - N * 2));
+      __m256i mask = __lasx_xvld(bitsel_mask_table[N], 0);
+      result = __lasx_xvbitsel_v(result, srl_prev, mask);
+
+      return result;
+    } else if (N == 8) {
+      return __lasx_xvpermi_q(this->value, prev_chunk, 0b00100001);
+    } else {
+      __m256i sll_value = __lasx_xvbsll_v(
+          __lasx_xvpermi_q(zero, this->value, 0b00000011), (N * 2 - 16));
+      __m256i mask = __lasx_xvld(bitsel_mask_table[N * 2], 0);
+      shuf = __lasx_xvld(prev_shuf_table[N * 2], 0);
+      result =
+          __lasx_xvshuf_b(__lasx_xvpermi_q(prev_chunk, prev_chunk, 0b00000001),
+                          prev_chunk, shuf);
+      result = __lasx_xvbitsel_v(sll_value, result, mask);
+      return result;
+    }
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd16<bool> : base16<bool> {
+  static simdutf_really_inline simd16<bool> splat(bool _value) {
+    return __lasx_xvreplgr2vr_h(uint8_t(-(!!_value)));
+  }
+
+  simdutf_really_inline simd16() : base16() {}
+  simdutf_really_inline simd16(const __m256i _value) : base16<bool>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+
+  simdutf_really_inline bitmask_type to_bitmask() const {
+    __m256i mask = __lasx_xvmsknz_b(this->value);
+    bitmask_type mask0 = __lasx_xvpickve2gr_wu(mask, 0);
+    bitmask_type mask1 = __lasx_xvpickve2gr_wu(mask, 4);
+    return (mask0 | (mask1 << 16));
+  }
+  simdutf_really_inline bool any() const {
+    if (__lasx_xbz_v(this->value))
+      return false;
+    return true;
+  }
+  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
+};
+
+template <typename T> struct base16_numeric : base16<T> {
+  static simdutf_really_inline simd16<T> splat(T _value) {
+    return __lasx_xvreplgr2vr_h((uint16_t)_value);
+  }
+  static simdutf_really_inline simd16<T> zero() { return __lasx_xvldi(0); }
+  static simdutf_really_inline simd16<T> load(const T values[8]) {
+    return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
+  }
+
+  simdutf_really_inline base16_numeric() : base16<T>() {}
+  simdutf_really_inline base16_numeric(const __m256i _value)
+      : base16<T>(_value) {}
+
+  // Store to array
+  simdutf_really_inline void store(T dst[8]) const {
+    return __lasx_xvst(this->value, reinterpret_cast<__m256i *>(dst), 0);
+  }
+
+  // Override to distinguish from bool version
+  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const {
+    return __lasx_xvadd_h(*this, other);
+  }
+  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const {
+    return __lasx_xvsub_h(*this, other);
+  }
+  simdutf_really_inline simd16<T> &operator+=(const simd16<T> other) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+  simdutf_really_inline simd16<T> &operator-=(const simd16<T> other) {
+    *this = *this - other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Signed code units
+template <> struct simd16<int16_t> : base16_numeric<int16_t> {
+  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
+  simdutf_really_inline simd16(const __m256i _value)
+      : base16_numeric<int16_t>(_value) {}
+  // Splat constructor
+  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const int16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const int16_t *>(values))) {}
+  // Order-sensitive comparisons
+  simdutf_really_inline simd16<int16_t>
+  max_val(const simd16<int16_t> other) const {
+    return __lasx_xvmax_h(*this, other);
+  }
+  simdutf_really_inline simd16<int16_t>
+  min_val(const simd16<int16_t> other) const {
+    return __lasx_xvmin_h(*this, other);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<int16_t> other) const {
+    return __lasx_xvsle_h(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<int16_t> other) const {
+    return __lasx_xvslt_h(this->value, other.value);
+  }
+};
+
+// Unsigned code units
+template <> struct simd16<uint16_t> : base16_numeric<uint16_t> {
+  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
+  simdutf_really_inline simd16(const __m256i _value)
+      : base16_numeric<uint16_t>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
+  // Array constructor
+  simdutf_really_inline simd16(const uint16_t *values) : simd16(load(values)) {}
+  simdutf_really_inline simd16(const char16_t *values)
+      : simd16(load(reinterpret_cast<const uint16_t *>(values))) {}
+
+  // Saturated math
+  simdutf_really_inline simd16<uint16_t>
+  saturating_add(const simd16<uint16_t> other) const {
+    return __lasx_xvsadd_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  saturating_sub(const simd16<uint16_t> other) const {
+    return __lasx_xvssub_hu(this->value, other.value);
+  }
+
+  // Order-specific operations
+  simdutf_really_inline simd16<uint16_t>
+  max_val(const simd16<uint16_t> other) const {
+    return __lasx_xvmax_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<uint16_t>
+  min_val(const simd16<uint16_t> other) const {
+    return __lasx_xvmin_hu(this->value, other.value);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  gt_bits(const simd16<uint16_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdutf_really_inline simd16<uint16_t>
+  lt_bits(const simd16<uint16_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<=(const simd16<uint16_t> other) const {
+    return __lasx_xvsle_hu(this->value, other.value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>=(const simd16<uint16_t> other) const {
+    return __lasx_xvsle_hu(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator>(const simd16<uint16_t> other) const {
+    return __lasx_xvslt_hu(other.value, this->value);
+  }
+  simdutf_really_inline simd16<bool>
+  operator<(const simd16<uint16_t> other) const {
+    return __lasx_xvslt_hu(this->value, other.value);
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+
+  simdutf_really_inline bool any_bits_set_anywhere() const {
+    if (__lasx_xbnz_v(this->value))
+      return true;
+    return false;
+  }
+  simdutf_really_inline bool
+  any_bits_set_anywhere(simd16<uint16_t> bits) const {
+    return (*this & bits).any_bits_set_anywhere();
+  }
+
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(__lasx_xvsrli_h(this->value, N));
+  }
+  template <int N> simdutf_really_inline simd16<uint16_t> shl() const {
+    return simd16<uint16_t>(__lasx_xvslli_h(this->value, N));
+  }
+
+  // Change the endianness
+  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
+    return __lasx_xvshuf4i_b(this->value, 0b10110001);
+  }
+
+  template <unsigned N>
+  static simdutf_really_inline simd8<uint8_t>
+  pack_shifted_right(const simd16<uint16_t> &v0, const simd16<uint16_t> &v1) {
+    return __lasx_xvpermi_d(__lasx_xvssrlni_bu_h(v1.value, v0.value, N),
+                            0b11011000);
+  }
+
+  // Pack with the unsigned saturation of two uint16_t code units into single
+  // uint8_t vector
+  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t> &v0,
+                                                   const simd16<uint16_t> &v1) {
+
+    return pack_shifted_right<0>(v0, v1);
+  }
+};
+
+template <typename T> struct simd16x32 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
+  static_assert(NUM_CHUNKS == 2,
+                "LASX kernel should use two registers per 64-byte block.");
+  simd16<T> chunks[NUM_CHUNKS];
+
+  simd16x32(const simd16x32<T> &o) = delete; // no copy allowed
+  simd16x32<T> &
+  operator=(const simd16<T> other) = delete; // no assignment allowed
+  simd16x32() = delete;                      // no default constructor allowed
+
+  simdutf_really_inline simd16x32(const simd16<T> chunk0,
+                                  const simd16<T> chunk1)
+      : chunks{chunk0, chunk1} {}
+  simdutf_really_inline simd16x32(const T *ptr)
+      : chunks{simd16<T>::load(ptr),
+               simd16<T>::load(ptr + sizeof(simd16<T>) / sizeof(T))} {}
+
+  simdutf_really_inline void store(T *ptr) const {
+    this->chunks[0].store(ptr + sizeof(simd16<T>) * 0 / sizeof(T));
+    this->chunks[1].store(ptr + sizeof(simd16<T>) * 1 / sizeof(T));
+  }
+
+  simdutf_really_inline uint64_t to_bitmask() const {
+    uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r_hi = this->chunks[1].to_bitmask();
+    return r_lo | (r_hi << 32);
+  }
+
+  simdutf_really_inline simd16<T> reduce_or() const {
+    return this->chunks[0] | this->chunks[1];
+  }
+
+  simdutf_really_inline bool is_ascii() const {
+    return this->reduce_or().is_ascii();
+  }
+
+  simdutf_really_inline void store_ascii_as_utf16(char16_t *ptr) const {
+    this->chunks[0].store_ascii_as_utf16(ptr + sizeof(simd16<T>) * 0);
+    this->chunks[1].store_ascii_as_utf16(ptr + sizeof(simd16<T>));
+  }
+
+  simdutf_really_inline simd16x32<T> bit_or(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<T>(this->chunks[0] | mask, this->chunks[1] | mask);
+  }
+
+  simdutf_really_inline void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+  }
+
+  simdutf_really_inline uint64_t eq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] == mask, this->chunks[1] == mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
+    return simd16x32<bool>(this->chunks[0] == other.chunks[0],
+                           this->chunks[1] == other.chunks[1])
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t lteq(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask)
+        .to_bitmask();
+  }
+
+  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(low);
+    const simd16<T> mask_high = simd16<T>::splat(high);
+
+    return simd16x32<bool>(
+               (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
+               (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
+    const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low - 1));
+    const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high + 1));
+    return simd16x32<bool>(
+               (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
+               (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low))
+        .to_bitmask();
+  }
+  simdutf_really_inline uint64_t lt(const T m) const {
+    const simd16<T> mask = simd16<T>::splat(m);
+    return simd16x32<bool>(this->chunks[0] < mask, this->chunks[1] < mask)
+        .to_bitmask();
+  }
+}; // struct simd16x32<T>
+/* end file src/simdutf/lasx/simd16-inl.h */
+/* begin file src/simdutf/lasx/simd32-inl.h */
+template <typename T> struct simd32;
+
+template <> struct simd32<uint32_t> {
+  __m256i value;
+  static const int SIZE = sizeof(value);
+  static const int ELEMENTS = SIZE / sizeof(uint32_t);
+
+  // constructors
+  simdutf_really_inline simd32(__m256i v) : value(v) {}
+
+  template <typename Ptr>
+  simdutf_really_inline simd32(Ptr *ptr) : value(__lasx_xvld(ptr, 0)) {}
+
+  // in-place operators
+  simdutf_really_inline simd32 &operator-=(const simd32 other) {
+    value = __lasx_xvsub_w(value, other.value);
+    return *this;
+  }
+
+  // members
+  simdutf_really_inline uint64_t sum() const {
+    const auto odd = __lasx_xvsrli_d(value, 32);
+    const auto even = __lasx_xvand_v(value, __lasx_xvreplgr2vr_d(0xffffffff));
+
+    const auto sum64 = __lasx_xvadd_d(odd, even);
+
+    return uint64_t(__lasx_xvpickve2gr_du(sum64, 0)) +
+           uint64_t(__lasx_xvpickve2gr_du(sum64, 1)) +
+           uint64_t(__lasx_xvpickve2gr_du(sum64, 2)) +
+           uint64_t(__lasx_xvpickve2gr_du(sum64, 3));
+  }
+
+  // static members
+  static simdutf_really_inline simd32<uint32_t> splat(uint32_t x) {
+    return __lasx_xvreplgr2vr_w(x);
+  }
+
+  static simdutf_really_inline simd32<uint32_t> zero() {
+    return __lasx_xvrepli_w(0);
+  }
+};
+
+// ------------------------------------------------------------
+
+template <> struct simd32<bool> {
+  __m256i value;
+  static const int SIZE = sizeof(value);
+
+  // constructors
+  simdutf_really_inline simd32(__m256i v) : value(v) {}
+};
+
+// ------------------------------------------------------------
+
+simdutf_really_inline simd32<uint32_t> operator&(const simd32<uint32_t> a,
+                                                 const simd32<uint32_t> b) {
+  return __lasx_xvor_v(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator<(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return __lasx_xvslt_wu(a.value, b.value);
+}
+
+simdutf_really_inline simd32<bool> operator>(const simd32<uint32_t> a,
+                                             const simd32<uint32_t> b) {
+  return __lasx_xvslt_wu(b.value, a.value);
+}
+
+// ------------------------------------------------------------
+
+simdutf_really_inline simd32<uint32_t> as_vector_u32(const simd32<bool> v) {
+  return v.value;
+}
+/* end file src/simdutf/lasx/simd32-inl.h */
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdutf
+
+#endif // SIMDUTF_LASX_SIMD_H
+/* end file src/simdutf/lasx/simd.h */
+
+/* begin file src/simdutf/lasx/end.h */
+#undef SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+/* end file src/simdutf/lasx/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_LASX
+
+#endif // SIMDUTF_LASX_H
+/* end file src/simdutf/lasx.h */
+/* begin file src/simdutf/fallback.h */
+#ifndef SIMDUTF_FALLBACK_H
+#define SIMDUTF_FALLBACK_H
+
+
+// Note that fallback.h is always imported last.
+
+// Default Fallback to on unless a builtin implementation has already been
+// selected.
+#ifndef SIMDUTF_IMPLEMENTATION_FALLBACK
+  #if SIMDUTF_CAN_ALWAYS_RUN_ARM64 || SIMDUTF_CAN_ALWAYS_RUN_ICELAKE ||        \
+      SIMDUTF_CAN_ALWAYS_RUN_HASWELL || SIMDUTF_CAN_ALWAYS_RUN_WESTMERE ||     \
+      SIMDUTF_CAN_ALWAYS_RUN_PPC64 || SIMDUTF_CAN_ALWAYS_RUN_RVV ||            \
+      SIMDUTF_CAN_ALWAYS_RUN_LSX || SIMDUTF_CAN_ALWAYS_RUN_LASX
+    #define SIMDUTF_IMPLEMENTATION_FALLBACK 0
+  #else
+    #define SIMDUTF_IMPLEMENTATION_FALLBACK 1
+  #endif
+#endif
+
+#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)
+
+#if SIMDUTF_IMPLEMENTATION_FALLBACK
+
+namespace simdutf {
+/**
+ * Fallback implementation (runs on any machine).
+ */
+namespace fallback {} // namespace fallback
+} // namespace simdutf
+
+/* begin file src/simdutf/fallback/implementation.h */
+#ifndef SIMDUTF_FALLBACK_IMPLEMENTATION_H
+#define SIMDUTF_FALLBACK_IMPLEMENTATION_H
+
+
+namespace simdutf {
+namespace fallback {
+
+namespace {
+using namespace simdutf;
+}
+
+class implementation final : public simdutf::implementation {
+public:
+  simdutf_really_inline implementation()
+      : simdutf::implementation("fallback", "Generic fallback implementation",
+                                0) {}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *input,
+                                           size_t length) const noexcept final;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *buf,
+                                         size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result
+  validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool validate_ascii(const char *buf,
+                                          size_t len) const noexcept final;
+  simdutf_warn_unused result
+  validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                            size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                            size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                          size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *buf, size_t len, char *utf8_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len, char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len, char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len, char16_t *utf16_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len, char16_t *utf16_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len, char32_t *utf32_output) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len, char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_latin1(const char16_t *buf, size_t len,
+                                  char *latin1_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *buf, size_t len, char *utf8_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final;
+  simdutf_warn_unused result
+  convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                      char *latin1_output) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_latin1(const char32_t *buf, size_t len,
+                                char *latin1_output) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16le(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf16be(const char32_t *buf, size_t len,
+                           char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16le(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf16be(const char32_t *buf, size_t len,
+                                 char16_t *utf16_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16le_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+  simdutf_warn_unused size_t
+  convert_valid_utf16be_to_utf32(const char16_t *buf, size_t len,
+                                 char32_t *utf32_buffer) const noexcept final;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t length,
+                               char16_t *output) const noexcept final;
+  simdutf_warn_unused size_t count_utf16le(const char16_t *buf,
+                                           size_t length) const noexcept;
+  simdutf_warn_unused size_t count_utf16be(const char16_t *buf,
+                                           size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *buf,
+                                        size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *input, size_t length) const noexcept;
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *input, size_t length) const noexcept;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *input, size_t length, char *output,
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options =
+                       last_chunk_handling_options::loose) const noexcept;
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_options =
+          last_chunk_handling_options::loose) const noexcept;
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept;
+#endif // SIMDUTF_FEATURE_BASE64
+};
+} // namespace fallback
+} // namespace simdutf
+
+#endif // SIMDUTF_FALLBACK_IMPLEMENTATION_H
+/* end file src/simdutf/fallback/implementation.h */
+
+/* begin file src/simdutf/fallback/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "fallback"
+// #define SIMDUTF_IMPLEMENTATION fallback
+/* end file src/simdutf/fallback/begin.h */
+
+  // Declarations
+/* begin file src/simdutf/fallback/bitmanipulation.h */
+#ifndef SIMDUTF_FALLBACK_BITMANIPULATION_H
+#define SIMDUTF_FALLBACK_BITMANIPULATION_H
+
+#include <limits>
+
+namespace simdutf {
+namespace fallback {
+namespace {} // unnamed namespace
+} // namespace fallback
+} // namespace simdutf
+
+#endif // SIMDUTF_FALLBACK_BITMANIPULATION_H
+/* end file src/simdutf/fallback/bitmanipulation.h */
+
+/* begin file src/simdutf/fallback/end.h */
+/* end file src/simdutf/fallback/end.h */
+
+#endif // SIMDUTF_IMPLEMENTATION_FALLBACK
+#endif // SIMDUTF_FALLBACK_H
+/* end file src/simdutf/fallback.h */
+
+// The scalar routines should be included once.
+/* begin file src/scalar/swap_bytes.h */
+#ifndef SIMDUTF_SWAP_BYTES_H
+#define SIMDUTF_SWAP_BYTES_H
+
+namespace simdutf {
+namespace scalar {
+
+inline simdutf_warn_unused uint16_t u16_swap_bytes(const uint16_t word) {
+  return uint16_t((word >> 8) | (word << 8));
+}
+
+inline simdutf_warn_unused uint32_t u32_swap_bytes(const uint32_t word) {
+  return ((word >> 24) & 0xff) |      // move byte 3 to byte 0
+         ((word << 8) & 0xff0000) |   // move byte 1 to byte 2
+         ((word >> 8) & 0xff00) |     // move byte 2 to byte 1
+         ((word << 24) & 0xff000000); // byte 0 to byte 3
+}
+
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/swap_bytes.h */
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/scalar/ascii.h */
+#ifndef SIMDUTF_ASCII_H
+#define SIMDUTF_ASCII_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace ascii {
+#if SIMDUTF_IMPLEMENTATION_FALLBACK
+// Only used by the fallback kernel.
+inline simdutf_warn_unused bool validate(const char *buf, size_t len) noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  uint64_t pos = 0;
+  // process in blocks of 16 bytes when possible
+  for (; pos + 16 <= len; pos += 16) {
+    uint64_t v1;
+    std::memcpy(&v1, data + pos, sizeof(uint64_t));
+    uint64_t v2;
+    std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+    uint64_t v{v1 | v2};
+    if ((v & 0x8080808080808080) != 0) {
+      return false;
+    }
+  }
+  // process the tail byte-by-byte
+  for (; pos < len; pos++) {
+    if (data[pos] >= 0b10000000) {
+      return false;
+    }
+  }
+  return true;
+}
+#endif
+
+inline simdutf_warn_unused result validate_with_errors(const char *buf,
+                                                       size_t len) noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  // process in blocks of 16 bytes when possible
+  for (; pos + 16 <= len; pos += 16) {
+    uint64_t v1;
+    std::memcpy(&v1, data + pos, sizeof(uint64_t));
+    uint64_t v2;
+    std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+    uint64_t v{v1 | v2};
+    if ((v & 0x8080808080808080) != 0) {
+      for (; pos < len; pos++) {
+        if (data[pos] >= 0b10000000) {
+          return result(error_code::TOO_LARGE, pos);
+        }
+      }
+    }
+  }
+  // process the tail byte-by-byte
+  for (; pos < len; pos++) {
+    if (data[pos] >= 0b10000000) {
+      return result(error_code::TOO_LARGE, pos);
+    }
+  }
+  return result(error_code::SUCCESS, pos);
+}
+
+} // namespace ascii
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/ascii.h */
+#endif // SIMDUTF_FEATURE_ASCII
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/scalar/utf8.h */
+#ifndef SIMDUTF_UTF8_H
+#define SIMDUTF_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8 {
+#if SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_RVV
+// only used by the fallback kernel.
+// credit: based on code from Google Fuchsia (Apache Licensed)
+inline simdutf_warn_unused bool validate(const char *buf, size_t len) noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  uint64_t pos = 0;
+  uint32_t code_point = 0;
+  while (pos < len) {
+    // check of the next 16 bytes are ascii.
+    uint64_t next_pos = pos + 16;
+    if (next_pos <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      std::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        pos = next_pos;
+        continue;
+      }
+    }
+    unsigned char byte = data[pos];
+
+    while (byte < 0b10000000) {
+      if (++pos == len) {
+        return true;
+      }
+      byte = data[pos];
+    }
+
+    if ((byte & 0b11100000) == 0b11000000) {
+      next_pos = pos + 2;
+      if (next_pos > len) {
+        return false;
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      // range check
+      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if ((code_point < 0x80) || (0x7ff < code_point)) {
+        return false;
+      }
+    } else if ((byte & 0b11110000) == 0b11100000) {
+      next_pos = pos + 3;
+      if (next_pos > len) {
+        return false;
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      // range check
+      code_point = (byte & 0b00001111) << 12 |
+                   (data[pos + 1] & 0b00111111) << 6 |
+                   (data[pos + 2] & 0b00111111);
+      if ((code_point < 0x800) || (0xffff < code_point) ||
+          (0xd7ff < code_point && code_point < 0xe000)) {
+        return false;
+      }
+    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
+      next_pos = pos + 4;
+      if (next_pos > len) {
+        return false;
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return false;
+      }
+      // range check
+      code_point =
+          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
+          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff || 0x10ffff < code_point) {
+        return false;
+      }
+    } else {
+      // we may have a continuation
+      return false;
+    }
+    pos = next_pos;
+  }
+  return true;
+}
+#endif
+
+inline simdutf_warn_unused result validate_with_errors(const char *buf,
+                                                       size_t len) noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  uint32_t code_point = 0;
+  while (pos < len) {
+    // check of the next 16 bytes are ascii.
+    size_t next_pos = pos + 16;
+    if (next_pos <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      std::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        pos = next_pos;
+        continue;
+      }
+    }
+    unsigned char byte = data[pos];
+
+    while (byte < 0b10000000) {
+      if (++pos == len) {
+        return result(error_code::SUCCESS, len);
+      }
+      byte = data[pos];
+    }
+
+    if ((byte & 0b11100000) == 0b11000000) {
+      next_pos = pos + 2;
+      if (next_pos > len) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if ((code_point < 0x80) || (0x7ff < code_point)) {
+        return result(error_code::OVERLONG, pos);
+      }
+    } else if ((byte & 0b11110000) == 0b11100000) {
+      next_pos = pos + 3;
+      if (next_pos > len) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      code_point = (byte & 0b00001111) << 12 |
+                   (data[pos + 1] & 0b00111111) << 6 |
+                   (data[pos + 2] & 0b00111111);
+      if ((code_point < 0x800) || (0xffff < code_point)) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0xd7ff < code_point && code_point < 0xe000) {
+        return result(error_code::SURROGATE, pos);
+      }
+    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
+      next_pos = pos + 4;
+      if (next_pos > len) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      code_point =
+          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
+          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0x10ffff < code_point) {
+        return result(error_code::TOO_LARGE, pos);
+      }
+    } else {
+      // we either have too many continuation bytes or an invalid leading byte
+      if ((byte & 0b11000000) == 0b10000000) {
+        return result(error_code::TOO_LONG, pos);
+      } else {
+        return result(error_code::HEADER_BITS, pos);
+      }
+    }
+    pos = next_pos;
+  }
+  return result(error_code::SUCCESS, len);
+}
+
+// Finds the previous leading byte starting backward from buf and validates with
+// errors from there Used to pinpoint the location of an error when an invalid
+// chunk is detected We assume that the stream starts with a leading byte, and
+// to check that it is the case, we ask that you pass a pointer to the start of
+// the stream (start).
+inline simdutf_warn_unused result rewind_and_validate_with_errors(
+    const char *start, const char *buf, size_t len) noexcept {
+  // First check that we start with a leading byte
+  if ((*start & 0b11000000) == 0b10000000) {
+    return result(error_code::TOO_LONG, 0);
+  }
+  size_t extra_len{0};
+  // A leading byte cannot be further than 4 bytes away
+  for (int i = 0; i < 5; i++) {
+    unsigned char byte = *buf;
+    if ((byte & 0b11000000) != 0b10000000) {
+      break;
+    } else {
+      buf--;
+      extra_len++;
+    }
+  }
+
+  result res = validate_with_errors(buf, len + extra_len);
+  res.count -= extra_len;
+  return res;
+}
+
+inline size_t count_code_points(const char *buf, size_t len) {
+  const int8_t *p = reinterpret_cast<const int8_t *>(buf);
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    // -65 is 0b10111111, anything larger in two-complement's should start a new
+    // code point.
+    if (p[i] > -65) {
+      counter++;
+    }
+  }
+  return counter;
+}
+
+inline size_t utf16_length_from_utf8(const char *buf, size_t len) {
+  const int8_t *p = reinterpret_cast<const int8_t *>(buf);
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    if (p[i] > -65) {
+      counter++;
+    }
+    if (uint8_t(p[i]) >= 240) {
+      counter++;
+    }
+  }
+  return counter;
+}
+
+simdutf_warn_unused inline size_t trim_partial_utf8(const char *input,
+                                                    size_t length) {
+  if (length < 3) {
+    switch (length) {
+    case 2:
+      if (uint8_t(input[length - 1]) >= 0xc0) {
+        return length - 1;
+      } // 2-, 3- and 4-byte characters with only 1 byte left
+      if (uint8_t(input[length - 2]) >= 0xe0) {
+        return length - 2;
+      } // 3- and 4-byte characters with only 2 bytes left
+      return length;
+    case 1:
+      if (uint8_t(input[length - 1]) >= 0xc0) {
+        return length - 1;
+      } // 2-, 3- and 4-byte characters with only 1 byte left
+      return length;
+    case 0:
+      return length;
+    }
+  }
+  if (uint8_t(input[length - 1]) >= 0xc0) {
+    return length - 1;
+  } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (uint8_t(input[length - 2]) >= 0xe0) {
+    return length - 2;
+  } // 3- and 4-byte characters with only 1 byte left
+  if (uint8_t(input[length - 3]) >= 0xf0) {
+    return length - 3;
+  } // 4-byte characters with only 3 bytes left
+  return length;
+}
+
+} // namespace utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    (SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1)
+/* begin file src/scalar/utf16.h */
+#ifndef SIMDUTF_UTF16_H
+#define SIMDUTF_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16 {
+
+template <endianness big_endian>
+inline simdutf_warn_unused bool validate(const char16_t *data,
+                                         size_t len) noexcept {
+  uint64_t pos = 0;
+  while (pos < len) {
+    char16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xF800) == 0xD800) {
+      if (pos + 1 >= len) {
+        return false;
+      }
+      char16_t diff = char16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return false;
+      }
+      char16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      char16_t diff2 = char16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return false;
+      }
+      pos += 2;
+    } else {
+      pos++;
+    }
+  }
+  return true;
+}
+
+template <endianness big_endian>
+inline simdutf_warn_unused result validate_with_errors(const char16_t *data,
+                                                       size_t len) noexcept {
+  size_t pos = 0;
+  while (pos < len) {
+    char16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xF800) == 0xD800) {
+      if (pos + 1 >= len) {
+        return result(error_code::SURROGATE, pos);
+      }
+      char16_t diff = char16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      char16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      char16_t diff2 = uint16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      pos += 2;
+    } else {
+      pos++;
+    }
+  }
+  return result(error_code::SUCCESS, pos);
+}
+
+template <endianness big_endian>
+inline size_t count_code_points(const char16_t *p, size_t len) {
+  // We are not BOM aware.
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    char16_t word = !match_system(big_endian) ? u16_swap_bytes(p[i]) : p[i];
+    counter += ((word & 0xFC00) != 0xDC00);
+  }
+  return counter;
+}
+
+template <endianness big_endian>
+inline size_t utf8_length_from_utf16(const char16_t *p, size_t len) {
+  // We are not BOM aware.
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    char16_t word = !match_system(big_endian) ? u16_swap_bytes(p[i]) : p[i];
+    counter++; // ASCII
+    counter += static_cast<size_t>(
+        word >
+        0x7F); // non-ASCII is at least 2 bytes, surrogates are 2*2 == 4 bytes
+    counter += static_cast<size_t>((word > 0x7FF && word <= 0xD7FF) ||
+                                   (word >= 0xE000)); // three-byte
+  }
+  return counter;
+}
+
+template <endianness big_endian>
+inline size_t utf32_length_from_utf16(const char16_t *p, size_t len) {
+  // We are not BOM aware.
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    char16_t word = !match_system(big_endian) ? u16_swap_bytes(p[i]) : p[i];
+    counter += ((word & 0xFC00) != 0xDC00);
+  }
+  return counter;
+}
+
+simdutf_really_inline void
+change_endianness_utf16(const char16_t *input, size_t size, char16_t *output) {
+  for (size_t i = 0; i < size; i++) {
+    *output++ = char16_t(input[i] >> 8 | input[i] << 8);
+  }
+}
+
+template <endianness big_endian>
+simdutf_warn_unused inline size_t trim_partial_utf16(const char16_t *input,
+                                                     size_t length) {
+  if (length <= 1) {
+    return length;
+  }
+  uint16_t last_word = uint16_t(input[length - 1]);
+  last_word = !match_system(big_endian) ? u16_swap_bytes(last_word) : last_word;
+  length -= ((last_word & 0xFC00) == 0xD800);
+  return length;
+}
+
+} // namespace utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // (SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1)
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/scalar/utf32.h */
+#ifndef SIMDUTF_UTF32_H
+#define SIMDUTF_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32 {
+
+inline simdutf_warn_unused bool validate(const char32_t *buf,
+                                         size_t len) noexcept {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  uint64_t pos = 0;
+  for (; pos < len; pos++) {
+    uint32_t word = data[pos];
+    if (word > 0x10FFFF || (word >= 0xD800 && word <= 0xDFFF)) {
+      return false;
+    }
+  }
+  return true;
+}
+
+inline simdutf_warn_unused result validate_with_errors(const char32_t *buf,
+                                                       size_t len) noexcept {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  for (; pos < len; pos++) {
+    uint32_t word = data[pos];
+    if (word > 0x10FFFF) {
+      return result(error_code::TOO_LARGE, pos);
+    }
+    if (word >= 0xD800 && word <= 0xDFFF) {
+      return result(error_code::SURROGATE, pos);
+    }
+  }
+  return result(error_code::SUCCESS, pos);
+}
+
+inline size_t utf8_length_from_utf32(const char32_t *buf, size_t len) {
+  // We are not BOM aware.
+  const uint32_t *p = reinterpret_cast<const uint32_t *>(buf);
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    // credit: @ttsugriy  for the vectorizable approach
+    counter++;                                     // ASCII
+    counter += static_cast<size_t>(p[i] > 0x7F);   // two-byte
+    counter += static_cast<size_t>(p[i] > 0x7FF);  // three-byte
+    counter += static_cast<size_t>(p[i] > 0xFFFF); // four-bytes
+  }
+  return counter;
+}
+
+inline size_t utf16_length_from_utf32(const char32_t *buf, size_t len) {
+  // We are not BOM aware.
+  const uint32_t *p = reinterpret_cast<const uint32_t *>(buf);
+  size_t counter{0};
+  for (size_t i = 0; i < len; i++) {
+    counter++;                                     // non-surrogate word
+    counter += static_cast<size_t>(p[i] > 0xFFFF); // surrogate pair
+  }
+  return counter;
+}
+
+} // namespace utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/latin1.h */
+#ifndef SIMDUTF_LATIN1_H
+#define SIMDUTF_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace latin1 {
+
+simdutf_really_inline size_t utf8_length_from_latin1(const char *buf,
+                                                     size_t len) {
+  const uint8_t *c = reinterpret_cast<const uint8_t *>(buf);
+  size_t answer = 0;
+  for (size_t i = 0; i < len; i++) {
+    if ((c[i] >> 7)) {
+      answer++;
+    }
+  }
+  return answer + len;
+}
+
+} // namespace latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/latin1.h */
+#endif // SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/scalar/base64.h */
+#ifndef SIMDUTF_BASE64_H
+#define SIMDUTF_BASE64_H
+
+#include <cstddef>
+#include <cstdint>
+#include <cstring>
+#include <iostream>
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace base64 {
+
+// This function is not expected to be fast. Do not use in long loops.
+template <class char_type> bool is_ascii_white_space(char_type c) {
+  return c == ' ' || c == '\t' || c == '\n' || c == '\r' || c == '\f';
+}
+
+template <class char_type> bool is_ascii_white_space_or_padding(char_type c) {
+  return c == ' ' || c == '\t' || c == '\n' || c == '\r' || c == '\f' ||
+         c == '=';
+}
+
+template <class char_type> bool is_eight_byte(char_type c) {
+  if (sizeof(char_type) == 1) {
+    return true;
+  }
+  return uint8_t(c) == c;
+}
+
+// Returns true upon success. The destination buffer must be large enough.
+// This functions assumes that the padding (=) has been removed.
+template <class char_type>
+full_result
+base64_tail_decode(char *dst, const char_type *src, size_t length,
+                   size_t padded_characters, // number of padding characters
+                                             // '=', typically 0, 1, 2.
+                   base64_options options,
+                   last_chunk_handling_options last_chunk_options) {
+  // This looks like 5 branches, but we expect the compiler to resolve this to a
+  // single branch:
+  const uint8_t *to_base64 = (options & base64_url)
+                                 ? tables::base64::to_base64_url_value
+                                 : tables::base64::to_base64_value;
+  const uint32_t *d0 = (options & base64_url)
+                           ? tables::base64::base64_url::d0
+                           : tables::base64::base64_default::d0;
+  const uint32_t *d1 = (options & base64_url)
+                           ? tables::base64::base64_url::d1
+                           : tables::base64::base64_default::d1;
+  const uint32_t *d2 = (options & base64_url)
+                           ? tables::base64::base64_url::d2
+                           : tables::base64::base64_default::d2;
+  const uint32_t *d3 = (options & base64_url)
+                           ? tables::base64::base64_url::d3
+                           : tables::base64::base64_default::d3;
+
+  const char_type *srcend = src + length;
+  const char_type *srcinit = src;
+  const char *dstinit = dst;
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+
+  uint32_t x;
+  size_t idx;
+  uint8_t buffer[4];
+  while (true) {
+    while (src + 4 <= srcend && is_eight_byte(src[0]) &&
+           is_eight_byte(src[1]) && is_eight_byte(src[2]) &&
+           is_eight_byte(src[3]) &&
+           (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] |
+                d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) {
+      if (match_system(endianness::BIG)) {
+        x = scalar::u32_swap_bytes(x);
+      }
+      std::memcpy(dst, &x, 3); // optimization opportunity: copy 4 bytes
+      dst += 3;
+      src += 4;
+    }
+    idx = 0;
+    // we need at least four characters.
+#ifdef __clang__
+    // If possible, we read four characters at a time. (It is an optimization.)
+    if (ignore_garbage && src + 4 <= srcend) {
+      char_type c0 = src[0];
+      char_type c1 = src[1];
+      char_type c2 = src[2];
+      char_type c3 = src[3];
+      uint8_t code0 = to_base64[uint8_t(c0)];
+      uint8_t code1 = to_base64[uint8_t(c1)];
+      uint8_t code2 = to_base64[uint8_t(c2)];
+      uint8_t code3 = to_base64[uint8_t(c3)];
+      buffer[idx] = code0;
+      idx += (is_eight_byte(c0) && code0 <= 63);
+      buffer[idx] = code1;
+      idx += (is_eight_byte(c1) && code1 <= 63);
+      buffer[idx] = code2;
+      idx += (is_eight_byte(c2) && code2 <= 63);
+      buffer[idx] = code3;
+      idx += (is_eight_byte(c3) && code3 <= 63);
+      src += 4;
+    }
+#endif
+    while ((idx < 4) && (src < srcend)) {
+      char_type c = *src;
+      uint8_t code = to_base64[uint8_t(c)];
+      buffer[idx] = uint8_t(code);
+      if (is_eight_byte(c) && code <= 63) {
+        idx++;
+      } else if (!ignore_garbage &&
+                 (code > 64 || !scalar::base64::is_eight_byte(c))) {
+        return {INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      } else {
+        // We have a space or a newline or garbage. We ignore it.
+      }
+      src++;
+    }
+    if (idx != 4) {
+      if (!ignore_garbage &&
+          last_chunk_options == last_chunk_handling_options::strict &&
+          (idx != 1) && ((idx + padded_characters) & 3) != 0) {
+        // The partial chunk was at src - idx
+        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      } else if (!ignore_garbage &&
+                 last_chunk_options ==
+                     last_chunk_handling_options::stop_before_partial &&
+                 (idx != 1) && ((idx + padded_characters) & 3) != 0) {
+        // Rewind src to before partial chunk
+        src -= idx;
+        return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
+      } else {
+        if (idx == 2) {
+          uint32_t triple =
+              (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6);
+          if (!ignore_garbage &&
+              (last_chunk_options == last_chunk_handling_options::strict) &&
+              (triple & 0xffff)) {
+            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                    size_t(dst - dstinit)};
+          }
+          if (match_system(endianness::BIG)) {
+            triple <<= 8;
+            std::memcpy(dst, &triple, 1);
+          } else {
+            triple = scalar::u32_swap_bytes(triple);
+            triple >>= 8;
+            std::memcpy(dst, &triple, 1);
+          }
+          dst += 1;
+        } else if (idx == 3) {
+          uint32_t triple = (uint32_t(buffer[0]) << 3 * 6) +
+                            (uint32_t(buffer[1]) << 2 * 6) +
+                            (uint32_t(buffer[2]) << 1 * 6);
+          if (!ignore_garbage &&
+              (last_chunk_options == last_chunk_handling_options::strict) &&
+              (triple & 0xff)) {
+            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                    size_t(dst - dstinit)};
+          }
+          if (match_system(endianness::BIG)) {
+            triple <<= 8;
+            std::memcpy(dst, &triple, 2);
+          } else {
+            triple = scalar::u32_swap_bytes(triple);
+            triple >>= 8;
+            std::memcpy(dst, &triple, 2);
+          }
+          dst += 2;
+        } else if (!ignore_garbage && idx == 1) {
+          return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+                  size_t(dst - dstinit)};
+        }
+        return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
+      }
+    }
+
+    uint32_t triple =
+        (uint32_t(buffer[0]) << 3 * 6) + (uint32_t(buffer[1]) << 2 * 6) +
+        (uint32_t(buffer[2]) << 1 * 6) + (uint32_t(buffer[3]) << 0 * 6);
+    if (match_system(endianness::BIG)) {
+      triple <<= 8;
+      std::memcpy(dst, &triple, 3);
+    } else {
+      triple = scalar::u32_swap_bytes(triple);
+      triple >>= 8;
+      std::memcpy(dst, &triple, 3);
+    }
+    dst += 3;
+  }
+}
+
+// like base64_tail_decode, but it will not write past the end of the output
+// buffer. The outlen paramter is modified to reflect the number of bytes
+// written. This functions assumes that the padding (=) has been removed.
+template <class char_type>
+result base64_tail_decode_safe(
+    char *dst, size_t &outlen, const char_type *&srcr, size_t length,
+    size_t padded_characters, // number of padding characters '=', typically 0,
+                              // 1, 2.
+    base64_options options, last_chunk_handling_options last_chunk_options) {
+  const char_type *src = srcr;
+  if (length == 0) {
+    outlen = 0;
+    return {SUCCESS, 0};
+  }
+  // This looks like 5 branches, but we expect the compiler to resolve this to a
+  // single branch:
+  const uint8_t *to_base64 = (options & base64_url)
+                                 ? tables::base64::to_base64_url_value
+                                 : tables::base64::to_base64_value;
+  const uint32_t *d0 = (options & base64_url)
+                           ? tables::base64::base64_url::d0
+                           : tables::base64::base64_default::d0;
+  const uint32_t *d1 = (options & base64_url)
+                           ? tables::base64::base64_url::d1
+                           : tables::base64::base64_default::d1;
+  const uint32_t *d2 = (options & base64_url)
+                           ? tables::base64::base64_url::d2
+                           : tables::base64::base64_default::d2;
+  const uint32_t *d3 = (options & base64_url)
+                           ? tables::base64::base64_url::d3
+                           : tables::base64::base64_default::d3;
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+
+  const char_type *srcend = src + length;
+  const char_type *srcinit = src;
+  const char *dstinit = dst;
+  const char *dstend = dst + outlen;
+
+  uint32_t x;
+  size_t idx;
+  uint8_t buffer[4];
+  while (true) {
+    while (src + 4 <= srcend && is_eight_byte(src[0]) &&
+           is_eight_byte(src[1]) && is_eight_byte(src[2]) &&
+           is_eight_byte(src[3]) &&
+           (x = d0[uint8_t(src[0])] | d1[uint8_t(src[1])] |
+                d2[uint8_t(src[2])] | d3[uint8_t(src[3])]) < 0x01FFFFFF) {
+      if (dstend - dst < 3) {
+        outlen = size_t(dst - dstinit);
+        srcr = src;
+        return {OUTPUT_BUFFER_TOO_SMALL, size_t(src - srcinit)};
+      }
+      if (match_system(endianness::BIG)) {
+        x = scalar::u32_swap_bytes(x);
+      }
+      std::memcpy(dst, &x, 3); // optimization opportunity: copy 4 bytes
+      dst += 3;
+      src += 4;
+    }
+    idx = 0;
+    const char_type *srccur = src;
+    // We need at least four characters.
+#ifdef __clang__
+    // If possible, we read four characters at a time. (It is an optimization.)
+    if (ignore_garbage && src + 4 <= srcend) {
+      char_type c0 = src[0];
+      char_type c1 = src[1];
+      char_type c2 = src[2];
+      char_type c3 = src[3];
+      uint8_t code0 = to_base64[uint8_t(c0)];
+      uint8_t code1 = to_base64[uint8_t(c1)];
+      uint8_t code2 = to_base64[uint8_t(c2)];
+      uint8_t code3 = to_base64[uint8_t(c3)];
+      buffer[idx] = code0;
+      idx += (is_eight_byte(c0) && code0 <= 63);
+      buffer[idx] = code1;
+      idx += (is_eight_byte(c1) && code1 <= 63);
+      buffer[idx] = code2;
+      idx += (is_eight_byte(c2) && code2 <= 63);
+      buffer[idx] = code3;
+      idx += (is_eight_byte(c3) && code3 <= 63);
+      src += 4;
+    }
+#endif
+    while (idx < 4 && src < srcend) {
+      char_type c = *src;
+      uint8_t code = to_base64[uint8_t(c)];
+
+      buffer[idx] = uint8_t(code);
+      if (is_eight_byte(c) && code <= 63) {
+        idx++;
+      } else if (!ignore_garbage &&
+                 (code > 64 || !scalar::base64::is_eight_byte(c))) {
+        outlen = size_t(dst - dstinit);
+        srcr = src;
+        return {INVALID_BASE64_CHARACTER, size_t(src - srcinit)};
+      } else {
+        // We have a space or a newline or garbage. We ignore it.
+      }
+      src++;
+    }
+    if (idx != 4) {
+      if (!ignore_garbage &&
+          last_chunk_options == last_chunk_handling_options::strict &&
+          ((idx + padded_characters) & 3) != 0) {
+        outlen = size_t(dst - dstinit);
+        srcr = src;
+        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit)};
+      } else if (!ignore_garbage &&
+                 last_chunk_options ==
+                     last_chunk_handling_options::stop_before_partial &&
+                 ((idx + padded_characters) & 3) != 0) {
+        // Rewind src to before partial chunk
+        srcr = srccur;
+        outlen = size_t(dst - dstinit);
+        return {SUCCESS, size_t(dst - dstinit)};
+      } else { // loose mode
+        if (idx == 0) {
+          // No data left; return success
+          outlen = size_t(dst - dstinit);
+          srcr = src;
+          return {SUCCESS, size_t(dst - dstinit)};
+        } else if (!ignore_garbage && idx == 1) {
+          // Error: Incomplete chunk of length 1 is invalid in loose mode
+          outlen = size_t(dst - dstinit);
+          srcr = src;
+          return {BASE64_INPUT_REMAINDER, size_t(src - srcinit)};
+        } else if (idx == 2 || idx == 3) {
+          // Check if there's enough space in the destination buffer
+          size_t required_space = (idx == 2) ? 1 : 2;
+          if (size_t(dstend - dst) < required_space) {
+            outlen = size_t(dst - dstinit);
+            srcr = src;
+            return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit)};
+          }
+          uint32_t triple = 0;
+          if (idx == 2) {
+            triple = (uint32_t(buffer[0]) << 18) + (uint32_t(buffer[1]) << 12);
+            if (!ignore_garbage &&
+                (last_chunk_options == last_chunk_handling_options::strict) &&
+                (triple & 0xffff)) {
+              srcr = src;
+              return {BASE64_EXTRA_BITS, size_t(src - srcinit)};
+            }
+            // Extract the first byte
+            triple >>= 16;
+            dst[0] = static_cast<char>(triple & 0xFF);
+            dst += 1;
+          } else if (idx == 3) {
+            triple = (uint32_t(buffer[0]) << 18) + (uint32_t(buffer[1]) << 12) +
+                     (uint32_t(buffer[2]) << 6);
+            if (!ignore_garbage &&
+                (last_chunk_options == last_chunk_handling_options::strict) &&
+                (triple & 0xff)) {
+              srcr = src;
+              return {BASE64_EXTRA_BITS, size_t(src - srcinit)};
+            }
+            // Extract the first two bytes
+            triple >>= 8;
+            dst[0] = static_cast<char>((triple >> 8) & 0xFF);
+            dst[1] = static_cast<char>(triple & 0xFF);
+            dst += 2;
+          }
+          outlen = size_t(dst - dstinit);
+          srcr = src;
+          return {SUCCESS, size_t(dst - dstinit)};
+        }
+      }
+    }
+
+    if (dstend - dst < 3) {
+      outlen = size_t(dst - dstinit);
+      srcr = src;
+      return {OUTPUT_BUFFER_TOO_SMALL, size_t(srccur - srcinit)};
+    }
+    uint32_t triple = (uint32_t(buffer[0]) << 18) +
+                      (uint32_t(buffer[1]) << 12) + (uint32_t(buffer[2]) << 6) +
+                      (uint32_t(buffer[3]));
+    if (match_system(endianness::BIG)) {
+      triple <<= 8;
+      std::memcpy(dst, &triple, 3);
+    } else {
+      triple = scalar::u32_swap_bytes(triple);
+      triple >>= 8;
+      std::memcpy(dst, &triple, 3);
+    }
+    dst += 3;
+  }
+}
+
+// Returns the number of bytes written. The destination buffer must be large
+// enough. It will add padding (=) if needed.
+size_t tail_encode_base64(char *dst, const char *src, size_t srclen,
+                          base64_options options) {
+  // By default, we use padding if we are not using the URL variant.
+  // This is check with ((options & base64_url) == 0) which returns true if we
+  // are not using the URL variant. However, we also allow 'inversion' of the
+  // convention with the base64_reverse_padding option. If the
+  // base64_reverse_padding option is set, we use padding if we are using the
+  // URL variant, and we omit it if we are not using the URL variant. This is
+  // checked with
+  // ((options & base64_reverse_padding) == base64_reverse_padding).
+  bool use_padding =
+      ((options & base64_url) == 0) ^
+      ((options & base64_reverse_padding) == base64_reverse_padding);
+  // This looks like 3 branches, but we expect the compiler to resolve this to
+  // a single branch:
+  const char *e0 = (options & base64_url) ? tables::base64::base64_url::e0
+                                          : tables::base64::base64_default::e0;
+  const char *e1 = (options & base64_url) ? tables::base64::base64_url::e1
+                                          : tables::base64::base64_default::e1;
+  const char *e2 = (options & base64_url) ? tables::base64::base64_url::e2
+                                          : tables::base64::base64_default::e2;
+  char *out = dst;
+  size_t i = 0;
+  uint8_t t1, t2, t3;
+  for (; i + 2 < srclen; i += 3) {
+    t1 = uint8_t(src[i]);
+    t2 = uint8_t(src[i + 1]);
+    t3 = uint8_t(src[i + 2]);
+    *out++ = e0[t1];
+    *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)];
+    *out++ = e1[((t2 & 0x0F) << 2) | ((t3 >> 6) & 0x03)];
+    *out++ = e2[t3];
+  }
+  switch (srclen - i) {
+  case 0:
+    break;
+  case 1:
+    t1 = uint8_t(src[i]);
+    *out++ = e0[t1];
+    *out++ = e1[(t1 & 0x03) << 4];
+    if (use_padding) {
+      *out++ = '=';
+      *out++ = '=';
+    }
+    break;
+  default: /* case 2 */
+    t1 = uint8_t(src[i]);
+    t2 = uint8_t(src[i + 1]);
+    *out++ = e0[t1];
+    *out++ = e1[((t1 & 0x03) << 4) | ((t2 >> 4) & 0x0F)];
+    *out++ = e2[(t2 & 0x0F) << 2];
+    if (use_padding) {
+      *out++ = '=';
+    }
+  }
+  return (size_t)(out - dst);
+}
+
+template <class char_type>
+simdutf_warn_unused size_t maximal_binary_length_from_base64(
+    const char_type *input, size_t length) noexcept {
+  // We follow https://infra.spec.whatwg.org/#forgiving-base64-decode
+  size_t padding = 0;
+  if (length > 0) {
+    if (input[length - 1] == '=') {
+      padding++;
+      if (length > 1 && input[length - 2] == '=') {
+        padding++;
+      }
+    }
+  }
+  size_t actual_length = length - padding;
+  if (actual_length % 4 <= 1) {
+    return actual_length / 4 * 3;
+  }
+  // if we have a valid input, then the remainder must be 2 or 3 adding one or
+  // two extra bytes.
+  return actual_length / 4 * 3 + (actual_length % 4) - 1;
+}
+
+simdutf_warn_unused size_t
+base64_length_from_binary(size_t length, base64_options options) noexcept {
+  // By default, we use padding if we are not using the URL variant.
+  // This is check with ((options & base64_url) == 0) which returns true if we
+  // are not using the URL variant. However, we also allow 'inversion' of the
+  // convention with the base64_reverse_padding option. If the
+  // base64_reverse_padding option is set, we use padding if we are using the
+  // URL variant, and we omit it if we are not using the URL variant. This is
+  // checked with
+  // ((options & base64_reverse_padding) == base64_reverse_padding).
+  bool use_padding =
+      ((options & base64_url) == 0) ^
+      ((options & base64_reverse_padding) == base64_reverse_padding);
+  if (!use_padding) {
+    return length / 3 * 4 + ((length % 3) ? (length % 3) + 1 : 0);
+  }
+  return (length + 2) / 3 *
+         4; // We use padding to make the length a multiple of 4.
+}
+
+} // namespace base64
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/base64.h */
+#endif // SIMDUTF_FEATURE_BASE64
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
+#ifndef SIMDUTF_VALID_UTF32_TO_UTF8_H
+#define SIMDUTF_VALID_UTF32_TO_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_utf8 {
+
+#if SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64
+// only used by the fallback and POWER kernel
+inline size_t convert_valid(const char32_t *buf, size_t len,
+                            char *utf8_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 2 ASCII characters
+    if (pos + 2 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
+        *utf8_output++ = char(buf[pos]);
+        *utf8_output++ = char(buf[pos + 1]);
+        pos += 2;
+        continue;
+      }
+    }
+    uint32_t word = data[pos];
+    if ((word & 0xFFFFFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xFFFFF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xFFFF0000) == 0) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 18) | 0b11110000);
+      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    }
+  }
+  return utf8_output - start;
+}
+#endif // SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64
+
+} // namespace utf32_to_utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
+/* begin file src/scalar/utf32_to_utf8/utf32_to_utf8.h */
+#ifndef SIMDUTF_UTF32_TO_UTF8_H
+#define SIMDUTF_UTF32_TO_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_utf8 {
+
+inline size_t convert(const char32_t *buf, size_t len, char *utf8_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 2 ASCII characters
+    if (pos + 2 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
+        *utf8_output++ = char(buf[pos]);
+        *utf8_output++ = char(buf[pos + 1]);
+        pos += 2;
+        continue;
+      }
+    }
+    uint32_t word = data[pos];
+    if ((word & 0xFFFFFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xFFFFF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xFFFF0000) == 0) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      if (word >= 0xD800 && word <= 0xDFFF) {
+        return 0;
+      }
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      if (word > 0x10FFFF) {
+        return 0;
+      }
+      *utf8_output++ = char((word >> 18) | 0b11110000);
+      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    }
+  }
+  return utf8_output - start;
+}
+
+inline result convert_with_errors(const char32_t *buf, size_t len,
+                                  char *utf8_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 2 ASCII characters
+    if (pos + 2 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
+        *utf8_output++ = char(buf[pos]);
+        *utf8_output++ = char(buf[pos + 1]);
+        pos += 2;
+        continue;
+      }
+    }
+    uint32_t word = data[pos];
+    if ((word & 0xFFFFFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xFFFFF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xFFFF0000) == 0) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      if (word >= 0xD800 && word <= 0xDFFF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      if (word > 0x10FFFF) {
+        return result(error_code::TOO_LARGE, pos);
+      }
+      *utf8_output++ = char((word >> 18) | 0b11110000);
+      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    }
+  }
+  return result(error_code::SUCCESS, utf8_output - start);
+}
+
+} // namespace utf32_to_utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_utf8/utf32_to_utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
+#ifndef SIMDUTF_VALID_UTF32_TO_UTF16_H
+#define SIMDUTF_VALID_UTF32_TO_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_utf16 {
+
+template <endianness big_endian>
+inline size_t convert_valid(const char32_t *buf, size_t len,
+                            char16_t *utf16_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    uint32_t word = data[pos];
+    if ((word & 0xFFFF0000) == 0) {
+      // will not generate a surrogate pair
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(uint16_t(word)))
+                            : char16_t(word);
+      pos++;
+    } else {
+      // will generate a surrogate pair
+      word -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+      pos++;
+    }
+  }
+  return utf16_output - start;
+}
+
+} // namespace utf32_to_utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
+/* begin file src/scalar/utf32_to_utf16/utf32_to_utf16.h */
+#ifndef SIMDUTF_UTF32_TO_UTF16_H
+#define SIMDUTF_UTF32_TO_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_utf16 {
+
+template <endianness big_endian>
+inline size_t convert(const char32_t *buf, size_t len, char16_t *utf16_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    uint32_t word = data[pos];
+    if ((word & 0xFFFF0000) == 0) {
+      if (word >= 0xD800 && word <= 0xDFFF) {
+        return 0;
+      }
+      // will not generate a surrogate pair
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(uint16_t(word)))
+                            : char16_t(word);
+    } else {
+      // will generate a surrogate pair
+      if (word > 0x10FFFF) {
+        return 0;
+      }
+      word -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+    }
+    pos++;
+  }
+  return utf16_output - start;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char32_t *buf, size_t len,
+                                  char16_t *utf16_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    uint32_t word = data[pos];
+    if ((word & 0xFFFF0000) == 0) {
+      if (word >= 0xD800 && word <= 0xDFFF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      // will not generate a surrogate pair
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(uint16_t(word)))
+                            : char16_t(word);
+    } else {
+      // will generate a surrogate pair
+      if (word > 0x10FFFF) {
+        return result(error_code::TOO_LARGE, pos);
+      }
+      word -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+    }
+    pos++;
+  }
+  return result(error_code::SUCCESS, utf16_output - start);
+}
+
+} // namespace utf32_to_utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_utf16/utf32_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
+#ifndef SIMDUTF_VALID_UTF16_TO_UTF8_H
+#define SIMDUTF_VALID_UTF16_TO_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_utf8 {
+
+template <endianness big_endian>
+inline size_t convert_valid(const char16_t *buf, size_t len,
+                            char *utf8_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 4 ASCII characters
+    if (pos + 4 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if (!match_system(big_endian)) {
+        v = (v >> 8) | (v << (64 - 8));
+      }
+      if ((v & 0xFF80FF80FF80FF80) == 0) {
+        size_t final_pos = pos + 4;
+        while (pos < final_pos) {
+          *utf8_output++ = !match_system(big_endian)
+                               ? char(u16_swap_bytes(buf[pos]))
+                               : char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xF800) != 0xD800) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((value >> 18) | 0b11110000);
+      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((value & 0b111111) | 0b10000000);
+      pos += 2;
+    }
+  }
+  return utf8_output - start;
+}
+
+} // namespace utf16_to_utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
+/* begin file src/scalar/utf16_to_utf8/utf16_to_utf8.h */
+#ifndef SIMDUTF_UTF16_TO_UTF8_H
+#define SIMDUTF_UTF16_TO_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_utf8 {
+
+template <endianness big_endian>
+inline size_t convert(const char16_t *buf, size_t len, char *utf8_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 8 bytes
+    if (pos + 4 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if (!match_system(big_endian)) {
+        v = (v >> 8) | (v << (64 - 8));
+      }
+      if ((v & 0xFF80FF80FF80FF80) == 0) {
+        size_t final_pos = pos + 4;
+        while (pos < final_pos) {
+          *utf8_output++ = !match_system(big_endian)
+                               ? char(u16_swap_bytes(buf[pos]))
+                               : char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xF800) != 0xD800) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      if (pos + 1 >= len) {
+        return 0;
+      }
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return 0;
+      }
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return 0;
+      }
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((value >> 18) | 0b11110000);
+      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((value & 0b111111) | 0b10000000);
+      pos += 2;
+    }
+  }
+  return utf8_output - start;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char16_t *buf, size_t len,
+                                  char *utf8_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *start{utf8_output};
+  while (pos < len) {
+    // try to convert the next block of 8 bytes
+    if (pos + 4 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if (!match_system(big_endian))
+        v = (v >> 8) | (v << (64 - 8));
+      if ((v & 0xFF80FF80FF80FF80) == 0) {
+        size_t final_pos = pos + 4;
+        while (pos < final_pos) {
+          *utf8_output++ = !match_system(big_endian)
+                               ? char(u16_swap_bytes(buf[pos]))
+                               : char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xFF80) == 0) {
+      // will generate one UTF-8 bytes
+      *utf8_output++ = char(word);
+      pos++;
+    } else if ((word & 0xF800) == 0) {
+      // will generate two UTF-8 bytes
+      // we have 0b110XXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 6) | 0b11000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else if ((word & 0xF800) != 0xD800) {
+      // will generate three UTF-8 bytes
+      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((word >> 12) | 0b11100000);
+      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((word & 0b111111) | 0b10000000);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      if (pos + 1 >= len) {
+        return result(error_code::SURROGATE, pos);
+      }
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      // will generate four UTF-8 bytes
+      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
+      *utf8_output++ = char((value >> 18) | 0b11110000);
+      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+      *utf8_output++ = char((value & 0b111111) | 0b10000000);
+      pos += 2;
+    }
+  }
+  return result(error_code::SUCCESS, utf8_output - start);
+}
+
+} // namespace utf16_to_utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_utf8/utf16_to_utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
+#ifndef SIMDUTF_VALID_UTF16_TO_UTF32_H
+#define SIMDUTF_VALID_UTF16_TO_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_utf32 {
+
+template <endianness big_endian>
+inline size_t convert_valid(const char16_t *buf, size_t len,
+                            char32_t *utf32_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xF800) != 0xD800) {
+      // No surrogate pair, extend 16-bit word to 32-bit word
+      *utf32_output++ = char32_t(word);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      *utf32_output++ = char32_t(value);
+      pos += 2;
+    }
+  }
+  return utf32_output - start;
+}
+
+} // namespace utf16_to_utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
+/* begin file src/scalar/utf16_to_utf32/utf16_to_utf32.h */
+#ifndef SIMDUTF_UTF16_TO_UTF32_H
+#define SIMDUTF_UTF16_TO_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_utf32 {
+
+template <endianness big_endian>
+inline size_t convert(const char16_t *buf, size_t len, char32_t *utf32_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xF800) != 0xD800) {
+      // No surrogate pair, extend 16-bit word to 32-bit word
+      *utf32_output++ = char32_t(word);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return 0;
+      }
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return 0;
+      }
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      *utf32_output++ = char32_t(value);
+      pos += 2;
+    }
+  }
+  return utf32_output - start;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char16_t *buf, size_t len,
+                                  char32_t *utf32_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    uint16_t word =
+        !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xF800) != 0xD800) {
+      // No surrogate pair, extend 16-bit word to 32-bit word
+      *utf32_output++ = char32_t(word);
+      pos++;
+    } else {
+      // must be a surrogate pair
+      uint16_t diff = uint16_t(word - 0xD800);
+      if (diff > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      if (pos + 1 >= len) {
+        return result(error_code::SURROGATE, pos);
+      } // minimal bound checking
+      uint16_t next_word = !match_system(big_endian)
+                               ? u16_swap_bytes(data[pos + 1])
+                               : data[pos + 1];
+      uint16_t diff2 = uint16_t(next_word - 0xDC00);
+      if (diff2 > 0x3FF) {
+        return result(error_code::SURROGATE, pos);
+      }
+      uint32_t value = (diff << 10) + diff2 + 0x10000;
+      *utf32_output++ = char32_t(value);
+      pos += 2;
+    }
+  }
+  return result(error_code::SUCCESS, utf32_output - start);
+}
+
+} // namespace utf16_to_utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_utf32/utf16_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 &&                                                    \
+    (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_LATIN1)
+/* begin file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
+#ifndef SIMDUTF_VALID_UTF8_TO_UTF16_H
+#define SIMDUTF_VALID_UTF8_TO_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_utf16 {
+
+template <endianness big_endian>
+inline size_t convert_valid(const char *buf, size_t len,
+                            char16_t *utf16_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    // try to convert the next block of 8 ASCII bytes
+    if (pos + 8 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 8;
+        while (pos < final_pos) {
+          *utf16_output++ = !match_system(big_endian)
+                                ? char16_t(u16_swap_bytes(buf[pos]))
+                                : char16_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(leading_byte))
+                            : char16_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 1 >= len) {
+        break;
+      } // minimal bound checking
+      uint16_t code_point = uint16_t(((leading_byte & 0b00011111) << 6) |
+                                     (data[pos + 1] & 0b00111111));
+      if (!match_system(big_endian)) {
+        code_point = u16_swap_bytes(uint16_t(code_point));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 2 >= len) {
+        break;
+      } // minimal bound checking
+      uint16_t code_point = uint16_t(((leading_byte & 0b00001111) << 12) |
+                                     ((data[pos + 1] & 0b00111111) << 6) |
+                                     (data[pos + 2] & 0b00111111));
+      if (!match_system(big_endian)) {
+        code_point = u16_swap_bytes(uint16_t(code_point));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        break;
+      } // minimal bound checking
+      uint32_t code_point = ((leading_byte & 0b00000111) << 18) |
+                            ((data[pos + 1] & 0b00111111) << 12) |
+                            ((data[pos + 2] & 0b00111111) << 6) |
+                            (data[pos + 3] & 0b00111111);
+      code_point -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+      pos += 4;
+    } else {
+      // we may have a continuation but we do not do error checking
+      return 0;
+    }
+  }
+  return utf16_output - start;
+}
+
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
+/* begin file src/scalar/utf8_to_utf16/utf8_to_utf16.h */
+#ifndef SIMDUTF_UTF8_TO_UTF16_H
+#define SIMDUTF_UTF8_TO_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_utf16 {
+
+template <endianness big_endian>
+inline size_t convert(const char *buf, size_t len, char16_t *utf16_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *utf16_output++ = !match_system(big_endian)
+                                ? char16_t(u16_swap_bytes(buf[pos]))
+                                : char16_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(leading_byte))
+                            : char16_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      // range check
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if (code_point < 0x80 || 0x7ff < code_point) {
+        return 0;
+      }
+      if (!match_system(big_endian)) {
+        code_point = uint32_t(u16_swap_bytes(uint16_t(code_point)));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 2 >= len) {
+        return 0;
+      } // minimal bound checking
+
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      // range check
+      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
+                            (data[pos + 1] & 0b00111111) << 6 |
+                            (data[pos + 2] & 0b00111111);
+      if (code_point < 0x800 || 0xffff < code_point ||
+          (0xd7ff < code_point && code_point < 0xe000)) {
+        return 0;
+      }
+      if (!match_system(big_endian)) {
+        code_point = uint32_t(u16_swap_bytes(uint16_t(code_point)));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        return 0;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+
+      // range check
+      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
+                            (data[pos + 1] & 0b00111111) << 12 |
+                            (data[pos + 2] & 0b00111111) << 6 |
+                            (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff || 0x10ffff < code_point) {
+        return 0;
+      }
+      code_point -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+      pos += 4;
+    } else {
+      return 0;
+    }
+  }
+  return utf16_output - start;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char *buf, size_t len,
+                                  char16_t *utf16_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *utf16_output++ = !match_system(big_endian)
+                                ? char16_t(u16_swap_bytes(buf[pos]))
+                                : char16_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf16_output++ = !match_system(big_endian)
+                            ? char16_t(u16_swap_bytes(leading_byte))
+                            : char16_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 1 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if (code_point < 0x80 || 0x7ff < code_point) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (!match_system(big_endian)) {
+        code_point = uint32_t(u16_swap_bytes(uint16_t(code_point)));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8, it should become
+      // a single UTF-16 word.
+      if (pos + 2 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
+                            (data[pos + 1] & 0b00111111) << 6 |
+                            (data[pos + 2] & 0b00111111);
+      if ((code_point < 0x800) || (0xffff < code_point)) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0xd7ff < code_point && code_point < 0xe000) {
+        return result(error_code::SURROGATE, pos);
+      }
+      if (!match_system(big_endian)) {
+        code_point = uint32_t(u16_swap_bytes(uint16_t(code_point)));
+      }
+      *utf16_output++ = char16_t(code_point);
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+
+      // range check
+      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
+                            (data[pos + 1] & 0b00111111) << 12 |
+                            (data[pos + 2] & 0b00111111) << 6 |
+                            (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0x10ffff < code_point) {
+        return result(error_code::TOO_LARGE, pos);
+      }
+      code_point -= 0x10000;
+      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
+      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
+      if (!match_system(big_endian)) {
+        high_surrogate = u16_swap_bytes(high_surrogate);
+        low_surrogate = u16_swap_bytes(low_surrogate);
+      }
+      *utf16_output++ = char16_t(high_surrogate);
+      *utf16_output++ = char16_t(low_surrogate);
+      pos += 4;
+    } else {
+      // we either have too many continuation bytes or an invalid leading byte
+      if ((leading_byte & 0b11000000) == 0b10000000) {
+        return result(error_code::TOO_LONG, pos);
+      } else {
+        return result(error_code::HEADER_BITS, pos);
+      }
+    }
+  }
+  return result(error_code::SUCCESS, utf16_output - start);
+}
+
+/**
+ * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and
+ * we have up to len input bytes left, and we encountered some error. It is
+ * possible that the error is at 'buf' exactly, but it could also be in the
+ * previous bytes  (up to 3 bytes back).
+ *
+ * prior_bytes indicates how many bytes, prior to 'buf' may belong to the
+ * current memory section and can be safely accessed. We prior_bytes to access
+ * safely up to three bytes before 'buf'.
+ *
+ * The caller is responsible to ensure that len > 0.
+ *
+ * If the error is believed to have occurred prior to 'buf', the count value
+ * contain in the result will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
+ */
+template <endianness endian>
+inline result rewind_and_convert_with_errors(size_t prior_bytes,
+                                             const char *buf, size_t len,
+                                             char16_t *utf16_output) {
+  size_t extra_len{0};
+  // We potentially need to go back in time and find a leading byte.
+  // In theory '3' would be sufficient, but sometimes the error can go back
+  // quite far.
+  size_t how_far_back = prior_bytes;
+  // size_t how_far_back = 3; // 3 bytes in the past + current position
+  // if(how_far_back >= prior_bytes) { how_far_back = prior_bytes; }
+  bool found_leading_bytes{false};
+  // important: it is i <= how_far_back and not 'i < how_far_back'.
+  for (size_t i = 0; i <= how_far_back; i++) {
+    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
+    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
+    if (found_leading_bytes) {
+      if (i > 0 && byte < 128) {
+        // If we had to go back and the leading byte is ascii
+        // then we can stop right away.
+        return result(error_code::TOO_LONG, 0 - i + 1);
+      }
+      buf -= i;
+      extra_len = i;
+      break;
+    }
+  }
+  //
+  // It is possible for this function to return a negative count in its result.
+  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
+  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
+  // unsigned integral type of the result of the sizeof operator
+  //
+  // An unsigned type will simply wrap round arithmetically (well defined).
+  //
+  if (!found_leading_bytes) {
+    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
+    // [....] [continuation] [continuation] [continuation] | [buf is
+    // continuation] Or we possibly have a stream that does not start with a
+    // leading byte.
+    return result(error_code::TOO_LONG, 0 - how_far_back);
+  }
+  result res = convert_with_errors<endian>(buf, len + extra_len, utf16_output);
+  if (res.error) {
+    res.count -= extra_len;
+  }
+  return res;
+}
+
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_utf16/utf8_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_LATIN1)
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
+/* begin file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
+#ifndef SIMDUTF_VALID_UTF8_TO_UTF32_H
+#define SIMDUTF_VALID_UTF8_TO_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_utf32 {
+
+inline size_t convert_valid(const char *buf, size_t len,
+                            char32_t *utf32_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    // try to convert the next block of 8 ASCII bytes
+    if (pos + 8 <=
+        len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 8;
+        while (pos < final_pos) {
+          *utf32_output++ = char32_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf32_output++ = char32_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        break;
+      } // minimal bound checking
+      *utf32_output++ = char32_t(((leading_byte & 0b00011111) << 6) |
+                                 (data[pos + 1] & 0b00111111));
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8
+      if (pos + 2 >= len) {
+        break;
+      } // minimal bound checking
+      *utf32_output++ = char32_t(((leading_byte & 0b00001111) << 12) |
+                                 ((data[pos + 1] & 0b00111111) << 6) |
+                                 (data[pos + 2] & 0b00111111));
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        break;
+      } // minimal bound checking
+      uint32_t code_word = ((leading_byte & 0b00000111) << 18) |
+                           ((data[pos + 1] & 0b00111111) << 12) |
+                           ((data[pos + 2] & 0b00111111) << 6) |
+                           (data[pos + 3] & 0b00111111);
+      *utf32_output++ = char32_t(code_word);
+      pos += 4;
+    } else {
+      // we may have a continuation but we do not do error checking
+      return 0;
+    }
+  }
+  return utf32_output - start;
+}
+
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
+/* begin file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
+#ifndef SIMDUTF_UTF8_TO_UTF32_H
+#define SIMDUTF_UTF8_TO_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_utf32 {
+
+inline size_t convert(const char *buf, size_t len, char32_t *utf32_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *utf32_output++ = char32_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf32_output++ = char32_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      // range check
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if (code_point < 0x80 || 0x7ff < code_point) {
+        return 0;
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8
+      if (pos + 2 >= len) {
+        return 0;
+      } // minimal bound checking
+
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      // range check
+      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
+                            (data[pos + 1] & 0b00111111) << 6 |
+                            (data[pos + 2] & 0b00111111);
+      if (code_point < 0x800 || 0xffff < code_point ||
+          (0xd7ff < code_point && code_point < 0xe000)) {
+        return 0;
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        return 0;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return 0;
+      }
+
+      // range check
+      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
+                            (data[pos + 1] & 0b00111111) << 12 |
+                            (data[pos + 2] & 0b00111111) << 6 |
+                            (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff || 0x10ffff < code_point) {
+        return 0;
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 4;
+    } else {
+      return 0;
+    }
+  }
+  return utf32_output - start;
+}
+
+inline result convert_with_errors(const char *buf, size_t len,
+                                  char32_t *utf32_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *utf32_output++ = char32_t(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *utf32_output++ = char32_t(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) == 0b11000000) {
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      if (code_point < 0x80 || 0x7ff < code_point) {
+        return result(error_code::OVERLONG, pos);
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8
+      if (pos + 2 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      // range check
+      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
+                            (data[pos + 1] & 0b00111111) << 6 |
+                            (data[pos + 2] & 0b00111111);
+      if (code_point < 0x800 || 0xffff < code_point) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0xd7ff < code_point && code_point < 0xe000) {
+        return result(error_code::SURROGATE, pos);
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 3;
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      if (pos + 3 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      }
+
+      // range check
+      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
+                            (data[pos + 1] & 0b00111111) << 12 |
+                            (data[pos + 2] & 0b00111111) << 6 |
+                            (data[pos + 3] & 0b00111111);
+      if (code_point <= 0xffff) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0x10ffff < code_point) {
+        return result(error_code::TOO_LARGE, pos);
+      }
+      *utf32_output++ = char32_t(code_point);
+      pos += 4;
+    } else {
+      // we either have too many continuation bytes or an invalid leading byte
+      if ((leading_byte & 0b11000000) == 0b10000000) {
+        return result(error_code::TOO_LONG, pos);
+      } else {
+        return result(error_code::HEADER_BITS, pos);
+      }
+    }
+  }
+  return result(error_code::SUCCESS, utf32_output - start);
+}
+
+/**
+ * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and
+ * we have up to len input bytes left, and we encountered some error. It is
+ * possible that the error is at 'buf' exactly, but it could also be in the
+ * previous bytes location (up to 3 bytes back).
+ *
+ * prior_bytes indicates how many bytes, prior to 'buf' may belong to the
+ * current memory section and can be safely accessed. We prior_bytes to access
+ * safely up to three bytes before 'buf'.
+ *
+ * The caller is responsible to ensure that len > 0.
+ *
+ * If the error is believed to have occurred prior to 'buf', the count value
+ * contain in the result will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
+ */
+inline result rewind_and_convert_with_errors(size_t prior_bytes,
+                                             const char *buf, size_t len,
+                                             char32_t *utf32_output) {
+  size_t extra_len{0};
+  // We potentially need to go back in time and find a leading byte.
+  size_t how_far_back = 3; // 3 bytes in the past + current position
+  if (how_far_back > prior_bytes) {
+    how_far_back = prior_bytes;
+  }
+  bool found_leading_bytes{false};
+  // important: it is i <= how_far_back and not 'i < how_far_back'.
+  for (size_t i = 0; i <= how_far_back; i++) {
+    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
+    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
+    if (found_leading_bytes) {
+      if (i > 0 && byte < 128) {
+        // If we had to go back and the leading byte is ascii
+        // then we can stop right away.
+        return result(error_code::TOO_LONG, 0 - i + 1);
+      }
+      buf -= i;
+      extra_len = i;
+      break;
+    }
+  }
+  //
+  // It is possible for this function to return a negative count in its result.
+  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
+  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
+  // unsigned integral type of the result of the sizeof operator
+  //
+  // An unsigned type will simply wrap round arithmetically (well defined).
+  //
+  if (!found_leading_bytes) {
+    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
+    // [....] [continuation] [continuation] [continuation] | [buf is
+    // continuation] Or we possibly have a stream that does not start with a
+    // leading byte.
+    return result(error_code::TOO_LONG, 0 - how_far_back);
+  }
+
+  result res = convert_with_errors(buf, len + extra_len, utf32_output);
+  if (res.error) {
+    res.count -= extra_len;
+  }
+  return res;
+}
+
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/latin1_to_utf8/latin1_to_utf8.h */
+#ifndef SIMDUTF_LATIN1_TO_UTF8_H
+#define SIMDUTF_LATIN1_TO_UTF8_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace latin1_to_utf8 {
+
+inline size_t convert(const char *buf, size_t len, char *utf8_output) {
+  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
+  size_t pos = 0;
+  size_t utf8_pos = 0;
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 |
+                 v2}; // We are only interested in these bits: 1000 1000 1000
+                      // 1000, so it makes sense to concatenate everything
+      if ((v & 0x8080808080808080) ==
+          0) { // if NONE of these are set, e.g. all of them are zero, then
+               // everything is ASCII
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          utf8_output[utf8_pos++] = char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+
+    unsigned char byte = data[pos];
+    if ((byte & 0x80) == 0) { // if ASCII
+      // will generate one UTF-8 bytes
+      utf8_output[utf8_pos++] = char(byte);
+      pos++;
+    } else {
+      // will generate two UTF-8 bytes
+      utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000);
+      utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000);
+      pos++;
+    }
+  }
+  return utf8_pos;
+}
+
+inline size_t convert_safe(const char *buf, size_t len, char *utf8_output,
+                           size_t utf8_len) {
+  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
+  size_t pos = 0;
+  size_t skip_pos = 0;
+  size_t utf8_pos = 0;
+  while (pos < len && utf8_pos < utf8_len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos >= skip_pos && pos + 16 <= len &&
+        utf8_pos + 16 <= utf8_len) { // if it is safe to read 16 more bytes,
+                                     // check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 |
+                 v2}; // We are only interested in these bits: 1000 1000 1000
+                      // 1000, so it makes sense to concatenate everything
+      if ((v & 0x8080808080808080) ==
+          0) { // if NONE of these are set, e.g. all of them are zero, then
+               // everything is ASCII
+        ::memcpy(utf8_output + utf8_pos, buf + pos, 16);
+        utf8_pos += 16;
+        pos += 16;
+      } else {
+        // At least one of the next 16 bytes are not ASCII, we will process them
+        // one by one
+        skip_pos = pos + 16;
+      }
+    } else {
+      const auto byte = data[pos];
+      if ((byte & 0x80) == 0) { // if ASCII
+        // will generate one UTF-8 bytes
+        utf8_output[utf8_pos++] = char(byte);
+        pos++;
+      } else if (utf8_pos + 2 <= utf8_len) {
+        // will generate two UTF-8 bytes
+        utf8_output[utf8_pos++] = char((byte >> 6) | 0b11000000);
+        utf8_output[utf8_pos++] = char((byte & 0b111111) | 0b10000000);
+        pos++;
+      } else {
+        break;
+      }
+    }
+  }
+  return utf8_pos;
+}
+
+} // namespace latin1_to_utf8
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/latin1_to_utf8/latin1_to_utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/latin1_to_utf16/latin1_to_utf16.h */
+#ifndef SIMDUTF_LATIN1_TO_UTF16_H
+#define SIMDUTF_LATIN1_TO_UTF16_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace latin1_to_utf16 {
+
+template <endianness big_endian>
+inline size_t convert(const char *buf, size_t len, char16_t *utf16_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+
+  while (pos < len) {
+    uint16_t word =
+        uint16_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point
+    *utf16_output++ =
+        char16_t(match_system(big_endian) ? word : u16_swap_bytes(word));
+    pos++;
+  }
+
+  return utf16_output - start;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char *buf, size_t len,
+                                  char16_t *utf16_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+
+  while (pos < len) {
+    uint16_t word =
+        uint16_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point
+    *utf16_output++ =
+        char16_t(match_system(big_endian) ? word : u16_swap_bytes(word));
+    pos++;
+  }
+
+  return result(error_code::SUCCESS, utf16_output - start);
+}
+
+} // namespace latin1_to_utf16
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/latin1_to_utf16/latin1_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/latin1_to_utf32/latin1_to_utf32.h */
+#ifndef SIMDUTF_LATIN1_TO_UTF32_H
+#define SIMDUTF_LATIN1_TO_UTF32_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace latin1_to_utf32 {
+
+inline size_t convert(const char *buf, size_t len, char32_t *utf32_output) {
+  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
+  char32_t *start{utf32_output};
+  for (size_t i = 0; i < len; i++) {
+    *utf32_output++ = (char32_t)data[i];
+  }
+  return utf32_output - start;
+}
+
+} // namespace latin1_to_utf32
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/latin1_to_utf32/latin1_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf8_to_latin1/utf8_to_latin1.h */
+#ifndef SIMDUTF_UTF8_TO_LATIN1_H
+#define SIMDUTF_UTF8_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_latin1 {
+
+inline size_t convert(const char *buf, size_t len, char *latin_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char *start{latin_output};
+
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000
+                           // 1000 1000 .... etc
+      if ((v & 0x8080808080808080) ==
+          0) { // if NONE of these are set, e.g. all of them are zero, then
+               // everything is ASCII
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *latin_output++ = char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+
+    // suppose it is not an all ASCII byte sequence
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *latin_output++ = char(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) ==
+               0b11000000) { // the first three bits indicate:
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        return 0;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      } // checks if the next byte is a valid continuation byte in UTF-8. A
+        // valid continuation byte starts with 10.
+      // range check -
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 |
+          (data[pos + 1] &
+           0b00111111); // assembles the Unicode code point from the two bytes.
+                        // It does this by discarding the leading 110 and 10
+                        // bits from the two bytes, shifting the remaining bits
+                        // of the first byte, and then combining the results
+                        // with a bitwise OR operation.
+      if (code_point < 0x80 || 0xFF < code_point) {
+        return 0; // We only care about the range 129-255 which is Non-ASCII
+                  // latin1 characters. A code_point beneath 0x80 is invalid as
+                  // it is already covered by bytes whose leading bit is zero.
+      }
+      *latin_output++ = char(code_point);
+      pos += 2;
+    } else {
+      return 0;
+    }
+  }
+  return latin_output - start;
+}
+
+inline result convert_with_errors(const char *buf, size_t len,
+                                  char *latin_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  size_t pos = 0;
+  char *start{latin_output};
+
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000
+                           // 1000 1000...etc
+      if ((v & 0x8080808080808080) ==
+          0) { // if NONE of these are set, e.g. all of them are zero, then
+               // everything is ASCII
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *latin_output++ = char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    // suppose it is not an all ASCII byte sequence
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *latin_output++ = char(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) ==
+               0b11000000) { // the first three bits indicate:
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        return result(error_code::TOO_SHORT, pos);
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return result(error_code::TOO_SHORT, pos);
+      } // checks if the next byte is a valid continuation byte in UTF-8. A
+        // valid continuation byte starts with 10.
+      // range check -
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 |
+          (data[pos + 1] &
+           0b00111111); // assembles the Unicode code point from the two bytes.
+                        // It does this by discarding the leading 110 and 10
+                        // bits from the two bytes, shifting the remaining bits
+                        // of the first byte, and then combining the results
+                        // with a bitwise OR operation.
+      if (code_point < 0x80) {
+        return result(error_code::OVERLONG, pos);
+      }
+      if (0xFF < code_point) {
+        return result(error_code::TOO_LARGE, pos);
+      } // We only care about the range 129-255 which is Non-ASCII latin1
+        // characters
+      *latin_output++ = char(code_point);
+      pos += 2;
+    } else if ((leading_byte & 0b11110000) == 0b11100000) {
+      // We have a three-byte UTF-8
+      return result(error_code::TOO_LARGE, pos);
+    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
+      // we have a 4-byte UTF-8 word.
+      return result(error_code::TOO_LARGE, pos);
+    } else {
+      // we either have too many continuation bytes or an invalid leading byte
+      if ((leading_byte & 0b11000000) == 0b10000000) {
+        return result(error_code::TOO_LONG, pos);
+      }
+
+      return result(error_code::HEADER_BITS, pos);
+    }
+  }
+  return result(error_code::SUCCESS, latin_output - start);
+}
+
+inline result rewind_and_convert_with_errors(size_t prior_bytes,
+                                             const char *buf, size_t len,
+                                             char *latin1_output) {
+  size_t extra_len{0};
+  // We potentially need to go back in time and find a leading byte.
+  // In theory '3' would be sufficient, but sometimes the error can go back
+  // quite far.
+  size_t how_far_back = prior_bytes;
+  // size_t how_far_back = 3; // 3 bytes in the past + current position
+  // if(how_far_back >= prior_bytes) { how_far_back = prior_bytes; }
+  bool found_leading_bytes{false};
+  // important: it is i <= how_far_back and not 'i < how_far_back'.
+  for (size_t i = 0; i <= how_far_back; i++) {
+    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
+    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
+    if (found_leading_bytes) {
+      if (i > 0 && byte < 128) {
+        // If we had to go back and the leading byte is ascii
+        // then we can stop right away.
+        return result(error_code::TOO_LONG, 0 - i + 1);
+      }
+      buf -= i;
+      extra_len = i;
+      break;
+    }
+  }
+  //
+  // It is possible for this function to return a negative count in its result.
+  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
+  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
+  // unsigned integral type of the result of the sizeof operator
+  //
+  // An unsigned type will simply wrap round arithmetically (well defined).
+  //
+  if (!found_leading_bytes) {
+    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
+    // [....] [continuation] [continuation] [continuation] | [buf is
+    // continuation] Or we possibly have a stream that does not start with a
+    // leading byte.
+    return result(error_code::TOO_LONG, 0 - how_far_back);
+  }
+  result res = convert_with_errors(buf, len + extra_len, latin1_output);
+  if (res.error) {
+    res.count -= extra_len;
+  }
+  return res;
+}
+
+} // namespace utf8_to_latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_latin1/utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf16_to_latin1/utf16_to_latin1.h */
+#ifndef SIMDUTF_UTF16_TO_LATIN1_H
+#define SIMDUTF_UTF16_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_latin1 {
+
+#include <cstring> // for std::memcpy
+
+template <endianness big_endian>
+inline size_t convert(const char16_t *buf, size_t len, char *latin_output) {
+  if (len == 0) {
+    return 0;
+  }
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *current_write = latin_output;
+  uint16_t word = 0;
+  uint16_t too_large = 0;
+
+  while (pos < len) {
+    word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    too_large |= word;
+    *current_write++ = char(word & 0xFF);
+    pos++;
+  }
+  if ((too_large & 0xFF00) != 0) {
+    return 0;
+  }
+
+  return current_write - latin_output;
+}
+
+template <endianness big_endian>
+inline result convert_with_errors(const char16_t *buf, size_t len,
+                                  char *latin_output) {
+  if (len == 0) {
+    return result(error_code::SUCCESS, 0);
+  }
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *start{latin_output};
+  uint16_t word;
+
+  while (pos < len) {
+    if (pos + 16 <= len) { // if it is safe to read 32 more bytes, check that
+                           // they are Latin1
+      uint64_t v1, v2, v3, v4;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      ::memcpy(&v2, data + pos + 4, sizeof(uint64_t));
+      ::memcpy(&v3, data + pos + 8, sizeof(uint64_t));
+      ::memcpy(&v4, data + pos + 12, sizeof(uint64_t));
+
+      if (!match_system(big_endian)) {
+        v1 = (v1 >> 8) | (v1 << (64 - 8));
+      }
+      if (!match_system(big_endian)) {
+        v2 = (v2 >> 8) | (v2 << (64 - 8));
+      }
+      if (!match_system(big_endian)) {
+        v3 = (v3 >> 8) | (v3 << (64 - 8));
+      }
+      if (!match_system(big_endian)) {
+        v4 = (v4 >> 8) | (v4 << (64 - 8));
+      }
+
+      if (((v1 | v2 | v3 | v4) & 0xFF00FF00FF00FF00) == 0) {
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *latin_output++ = !match_system(big_endian)
+                                ? char(u16_swap_bytes(data[pos]))
+                                : char(data[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+    word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    if ((word & 0xFF00) == 0) {
+      *latin_output++ = char(word & 0xFF);
+      pos++;
+    } else {
+      return result(error_code::TOO_LARGE, pos);
+    }
+  }
+  return result(error_code::SUCCESS, latin_output - start);
+}
+
+} // namespace utf16_to_latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_latin1/utf16_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf32_to_latin1/utf32_to_latin1.h */
+#ifndef SIMDUTF_UTF32_TO_LATIN1_H
+#define SIMDUTF_UTF32_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_latin1 {
+
+inline size_t convert(const char32_t *buf, size_t len, char *latin1_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  char *start = latin1_output;
+  uint32_t utf32_char;
+  size_t pos = 0;
+  uint32_t too_large = 0;
+
+  while (pos < len) {
+    utf32_char = (uint32_t)data[pos];
+    too_large |= utf32_char;
+    *latin1_output++ = (char)(utf32_char & 0xFF);
+    pos++;
+  }
+  if ((too_large & 0xFFFFFF00) != 0) {
+    return 0;
+  }
+  return latin1_output - start;
+}
+
+inline result convert_with_errors(const char32_t *buf, size_t len,
+                                  char *latin1_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  char *start{latin1_output};
+  size_t pos = 0;
+  while (pos < len) {
+    if (pos + 2 <=
+        len) { // if it is safe to read 8 more bytes, check that they are Latin1
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0xFFFFFF00FFFFFF00) == 0) {
+        *latin1_output++ = char(buf[pos]);
+        *latin1_output++ = char(buf[pos + 1]);
+        pos += 2;
+        continue;
+      }
+    }
+    uint32_t utf32_char = data[pos];
+    if ((utf32_char & 0xFFFFFF00) ==
+        0) { // Check if the character can be represented in Latin-1
+      *latin1_output++ = (char)(utf32_char & 0xFF);
+      pos++;
+    } else {
+      return result(error_code::TOO_LARGE, pos);
+    };
+  }
+  return result(error_code::SUCCESS, latin1_output - start);
+}
+
+} // namespace utf32_to_latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_latin1/utf32_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf8_to_latin1/valid_utf8_to_latin1.h */
+#ifndef SIMDUTF_VALID_UTF8_TO_LATIN1_H
+#define SIMDUTF_VALID_UTF8_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf8_to_latin1 {
+
+inline size_t convert_valid(const char *buf, size_t len, char *latin_output) {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+
+  size_t pos = 0;
+  char *start{latin_output};
+
+  while (pos < len) {
+    // try to convert the next block of 16 ASCII bytes
+    if (pos + 16 <=
+        len) { // if it is safe to read 16 more bytes, check that they are ascii
+      uint64_t v1;
+      ::memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 |
+                 v2}; // We are only interested in these bits: 1000 1000 1000
+                      // 1000, so it makes sense to concatenate everything
+      if ((v & 0x8080808080808080) ==
+          0) { // if NONE of these are set, e.g. all of them are zero, then
+               // everything is ASCII
+        size_t final_pos = pos + 16;
+        while (pos < final_pos) {
+          *latin_output++ = char(buf[pos]);
+          pos++;
+        }
+        continue;
+      }
+    }
+
+    // suppose it is not an all ASCII byte sequence
+    uint8_t leading_byte = data[pos]; // leading byte
+    if (leading_byte < 0b10000000) {
+      // converting one ASCII byte !!!
+      *latin_output++ = char(leading_byte);
+      pos++;
+    } else if ((leading_byte & 0b11100000) ==
+               0b11000000) { // the first three bits indicate:
+      // We have a two-byte UTF-8
+      if (pos + 1 >= len) {
+        break;
+      } // minimal bound checking
+      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
+        return 0;
+      } // checks if the next byte is a valid continuation byte in UTF-8. A
+        // valid continuation byte starts with 10.
+      // range check -
+      uint32_t code_point =
+          (leading_byte & 0b00011111) << 6 |
+          (data[pos + 1] &
+           0b00111111); // assembles the Unicode code point from the two bytes.
+                        // It does this by discarding the leading 110 and 10
+                        // bits from the two bytes, shifting the remaining bits
+                        // of the first byte, and then combining the results
+                        // with a bitwise OR operation.
+      *latin_output++ = char(code_point);
+      pos += 2;
+    } else {
+      // we may have a continuation but we do not do error checking
+      return 0;
+    }
+  }
+  return latin_output - start;
+}
+
+} // namespace utf8_to_latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf8_to_latin1/valid_utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf16_to_latin1/valid_utf16_to_latin1.h */
+#ifndef SIMDUTF_VALID_UTF16_TO_LATIN1_H
+#define SIMDUTF_VALID_UTF16_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf16_to_latin1 {
+
+template <endianness big_endian>
+inline size_t convert_valid(const char16_t *buf, size_t len,
+                            char *latin_output) {
+  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+  size_t pos = 0;
+  char *start{latin_output};
+  uint16_t word = 0;
+
+  while (pos < len) {
+    word = !match_system(big_endian) ? u16_swap_bytes(data[pos]) : data[pos];
+    *latin_output++ = char(word);
+    pos++;
+  }
+
+  return latin_output - start;
+}
+
+} // namespace utf16_to_latin1
+} // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf16_to_latin1/valid_utf16_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/scalar/utf32_to_latin1/valid_utf32_to_latin1.h */
+#ifndef SIMDUTF_VALID_UTF32_TO_LATIN1_H
+#define SIMDUTF_VALID_UTF32_TO_LATIN1_H
+
+namespace simdutf {
+namespace scalar {
+namespace {
+namespace utf32_to_latin1 {
+
+inline size_t convert_valid(const char32_t *buf, size_t len,
+                            char *latin1_output) {
+  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
+  char *start = latin1_output;
+  uint32_t utf32_char;
+  size_t pos = 0;
+
+  while (pos < len) {
+    utf32_char = (uint32_t)data[pos];
+
+    if (pos + 2 <=
+        len) { // if it is safe to read 8 more bytes, check that they are Latin1
+      uint64_t v;
+      ::memcpy(&v, data + pos, sizeof(uint64_t));
+      if ((v & 0xFFFFFF00FFFFFF00) == 0) {
+        *latin1_output++ = char(buf[pos]);
+        *latin1_output++ = char(buf[pos + 1]);
+        pos += 2;
+        continue;
+      } else {
+        // output can not be represented in latin1
+        return 0;
+      }
+    }
+    if ((utf32_char & 0xFFFFFF00) == 0) {
+      *latin1_output++ = char(utf32_char);
+    } else {
+      // output can not be represented in latin1
+      return 0;
+    }
+    pos++;
+  }
+  return latin1_output - start;
+}
+
+} // namespace utf32_to_latin1
 } // unnamed namespace
+} // namespace scalar
+} // namespace simdutf
+
+#endif
+/* end file src/scalar/utf32_to_latin1/valid_utf32_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+/* begin file src/implementation.cpp */
+#include <initializer_list>
+#include <climits>
+#include <type_traits>
+
+static_assert(sizeof(uint8_t) == sizeof(char),
+              "simdutf requires that uint8_t be a char");
+static_assert(sizeof(uint16_t) == sizeof(char16_t),
+              "simdutf requires that char16_t be 16 bits");
+static_assert(sizeof(uint32_t) == sizeof(char32_t),
+              "simdutf requires that char32_t be 32 bits");
+// next line is redundant, but it is kept to catch defective systems.
+static_assert(CHAR_BIT == 8, "simdutf requires 8-bit bytes");
+
+// Useful for debugging purposes
+namespace simdutf {
+namespace {
+
+template <typename T> std::string toBinaryString(T b) {
+  std::string binary = "";
+  T mask = T(1) << (sizeof(T) * CHAR_BIT - 1);
+  while (mask > 0) {
+    binary += ((b & mask) == 0) ? '0' : '1';
+    mask >>= 1;
+  }
+  return binary;
+}
+} // namespace
+} // namespace simdutf
+
+namespace simdutf {
+bool implementation::supported_by_runtime_system() const {
+  uint32_t required_instruction_sets = this->required_instruction_sets();
+  uint32_t supported_instruction_sets =
+      internal::detect_supported_architectures();
+  return ((supported_instruction_sets & required_instruction_sets) ==
+          required_instruction_sets);
+}
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused encoding_type implementation::autodetect_encoding(
+    const char *input, size_t length) const noexcept {
+  // If there is a BOM, then we trust it.
+  auto bom_encoding = simdutf::BOM::check_bom(input, length);
+  if (bom_encoding != encoding_type::unspecified) {
+    return bom_encoding;
+  }
+  // UTF8 is common, it includes ASCII, and is commonly represented
+  // without a BOM, so if it fits, go with that. Note that it is still
+  // possible to get it wrong, we are only 'guessing'. If some has UTF-16
+  // data without a BOM, it could pass as UTF-8.
+  //
+  // An interesting twist might be to check for UTF-16 ASCII first (every
+  // other byte is zero).
+  if (validate_utf8(input, length)) {
+    return encoding_type::UTF8;
+  }
+  // The next most common encoding that might appear without BOM is probably
+  // UTF-16LE, so try that next.
+  if ((length % 2) == 0) {
+    // important: we need to divide by two
+    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
+                         length / 2)) {
+      return encoding_type::UTF16_LE;
+    }
+  }
+  if ((length % 4) == 0) {
+    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
+      return encoding_type::UTF32_LE;
+    }
+  }
+  return encoding_type::unspecified;
+}
+
+  #ifdef SIMDUTF_INTERNAL_TESTS
+std::vector<implementation::TestProcedure>
+implementation::internal_tests() const {
+  return {};
+}
+  #endif
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
+    const char *input, size_t length) const noexcept {
+  return scalar::base64::maximal_binary_length_from_base64(input, length);
+}
+
+simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::base64::maximal_binary_length_from_base64(input, length);
+}
+
+simdutf_warn_unused size_t implementation::base64_length_from_binary(
+    size_t length, base64_options options) const noexcept {
+  return scalar::base64::base64_length_from_binary(length, options);
+}
+#endif // SIMDUTF_FEATURE_BASE64
+
+namespace internal {
+// When there is a single implementation, we should not pay a price
+// for dispatching to the best implementation. We should just use the
+// one we have. This is a compile-time check.
+#define SIMDUTF_SINGLE_IMPLEMENTATION                                          \
+  (SIMDUTF_IMPLEMENTATION_ICELAKE + SIMDUTF_IMPLEMENTATION_HASWELL +           \
+       SIMDUTF_IMPLEMENTATION_WESTMERE + SIMDUTF_IMPLEMENTATION_ARM64 +        \
+       SIMDUTF_IMPLEMENTATION_PPC64 + SIMDUTF_IMPLEMENTATION_LSX +             \
+       SIMDUTF_IMPLEMENTATION_LASX + SIMDUTF_IMPLEMENTATION_FALLBACK ==        \
+   1)
+
+// Static array of known implementations. We are hoping these get baked into the
+// executable without requiring a static initializer.
+
+#if SIMDUTF_IMPLEMENTATION_ICELAKE
+static const icelake::implementation *get_icelake_singleton() {
+  static const icelake::implementation icelake_singleton{};
+  return &icelake_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_HASWELL
+static const haswell::implementation *get_haswell_singleton() {
+  static const haswell::implementation haswell_singleton{};
+  return &haswell_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_WESTMERE
+static const westmere::implementation *get_westmere_singleton() {
+  static const westmere::implementation westmere_singleton{};
+  return &westmere_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_ARM64
+static const arm64::implementation *get_arm64_singleton() {
+  static const arm64::implementation arm64_singleton{};
+  return &arm64_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_PPC64
+static const ppc64::implementation *get_ppc64_singleton() {
+  static const ppc64::implementation ppc64_singleton{};
+  return &ppc64_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_RVV
+static const rvv::implementation *get_rvv_singleton() {
+  static const rvv::implementation rvv_singleton{};
+  return &rvv_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_LSX
+static const lsx::implementation *get_lsx_singleton() {
+  static const lsx::implementation lsx_singleton{};
+  return &lsx_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_LASX
+static const lasx::implementation *get_lasx_singleton() {
+  static const lasx::implementation lasx_singleton{};
+  return &lasx_singleton;
+}
+#endif
+#if SIMDUTF_IMPLEMENTATION_FALLBACK
+static const fallback::implementation *get_fallback_singleton() {
+  static const fallback::implementation fallback_singleton{};
+  return &fallback_singleton;
+}
+#endif
+
+#if SIMDUTF_SINGLE_IMPLEMENTATION
+static const implementation *get_single_implementation() {
+  return
+  #if SIMDUTF_IMPLEMENTATION_ICELAKE
+      get_icelake_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_HASWELL
+  get_haswell_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_WESTMERE
+  get_westmere_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_ARM64
+  get_arm64_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_PPC64
+  get_ppc64_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_LSX
+  get_lsx_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_LASX
+  get_lasx_singleton();
+  #endif
+  #if SIMDUTF_IMPLEMENTATION_FALLBACK
+  get_fallback_singleton();
+  #endif
+}
+#endif
+
+/**
+ * @private Detects best supported implementation on first use, and sets it
+ */
+class detect_best_supported_implementation_on_first_use final
+    : public implementation {
+public:
+  std::string name() const noexcept final { return set_best()->name(); }
+  std::string description() const noexcept final {
+    return set_best()->description();
+  }
+  uint32_t required_instruction_sets() const noexcept final {
+    return set_best()->required_instruction_sets();
+  }
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int
+  detect_encodings(const char *input, size_t length) const noexcept override {
+    return set_best()->detect_encodings(input, length);
+  }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool
+  validate_utf8(const char *buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf8(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result validate_utf8_with_errors(
+      const char *buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf8_with_errors(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool
+  validate_ascii(const char *buf, size_t len) const noexcept final override {
+    return set_best()->validate_ascii(buf, len);
+  }
+
+  simdutf_warn_unused result validate_ascii_with_errors(
+      const char *buf, size_t len) const noexcept final override {
+    return set_best()->validate_ascii_with_errors(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool
+  validate_utf16le(const char16_t *buf,
+                   size_t len) const noexcept final override {
+    return set_best()->validate_utf16le(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool
+  validate_utf16be(const char16_t *buf,
+                   size_t len) const noexcept final override {
+    return set_best()->validate_utf16be(buf, len);
+  }
+
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf16le_with_errors(buf, len);
+  }
+
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf16be_with_errors(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool
+  validate_utf32(const char32_t *buf,
+                 size_t len) const noexcept final override {
+    return set_best()->validate_utf32(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf32_with_errors(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_latin1_to_utf8(const char *buf, size_t len,
+                         char *utf8_output) const noexcept final override {
+    return set_best()->convert_latin1_to_utf8(buf, len, utf8_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_latin1_to_utf16le(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_latin1_to_utf16be(buf, len, utf16_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *buf, size_t len,
+      char32_t *latin1_output) const noexcept final override {
+    return set_best()->convert_latin1_to_utf32(buf, len, latin1_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf8_to_latin1(const char *buf, size_t len,
+                         char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf8_to_latin1(buf, len, latin1_output);
+  }
+
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf8_to_latin1_with_errors(buf, len,
+                                                          latin1_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_valid_utf8_to_latin1(buf, len, latin1_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf16le(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf16be(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf16le_with_errors(buf, len,
+                                                           utf16_output);
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf16be_with_errors(buf, len,
+                                                           utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_valid_utf8_to_utf16le(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_valid_utf8_to_utf16be(buf, len, utf16_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf8_to_utf32(const char *buf, size_t len,
+                        char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf32(buf, len, utf32_output);
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf8_to_utf32_with_errors(buf, len,
+                                                         utf32_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_valid_utf8_to_utf32(buf, len, utf32_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf16le_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_latin1(buf, len, latin1_output);
+  }
+
+  simdutf_warn_unused size_t
+  convert_utf16be_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_latin1(buf, len, latin1_output);
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_latin1_with_errors(buf, len,
+                                                             latin1_output);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_latin1_with_errors(buf, len,
+                                                             latin1_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
+      const char16_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_valid_utf16le_to_latin1(buf, len, latin1_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
+      const char16_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_valid_utf16be_to_latin1(buf, len, latin1_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  convert_utf16le_to_utf8(const char16_t *buf, size_t len,
+                          char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_utf8(buf, len, utf8_output);
+  }
+
+  simdutf_warn_unused size_t
+  convert_utf16be_to_utf8(const char16_t *buf, size_t len,
+                          char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_utf8(buf, len, utf8_output);
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *buf, size_t len,
+      char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_utf8_with_errors(buf, len,
+                                                           utf8_output);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *buf, size_t len,
+      char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_utf8_with_errors(buf, len,
+                                                           utf8_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *buf, size_t len,
+      char *utf8_output) const noexcept final override {
+    return set_best()->convert_valid_utf16le_to_utf8(buf, len, utf8_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *buf, size_t len,
+      char *utf8_output) const noexcept final override {
+    return set_best()->convert_valid_utf16be_to_utf8(buf, len, utf8_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                          char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf32_to_latin1(buf, len, latin1_output);
+  }
+
+  simdutf_warn_unused result convert_utf32_to_latin1_with_errors(
+      const char32_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf32_to_latin1_with_errors(buf, len,
+                                                           latin1_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
+      const char32_t *buf, size_t len,
+      char *latin1_output) const noexcept final override {
+    return set_best()->convert_utf32_to_latin1(buf, len, latin1_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  convert_utf32_to_utf8(const char32_t *buf, size_t len,
+                        char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf8(buf, len, utf8_output);
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *buf, size_t len,
+      char *utf8_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
+  }
+
+  simdutf_warn_unused size_t
+  convert_valid_utf32_to_utf8(const char32_t *buf, size_t len,
+                              char *utf8_output) const noexcept final override {
+    return set_best()->convert_valid_utf32_to_utf8(buf, len, utf8_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf16le(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf16le(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_utf32_to_utf16be(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf16be(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf16le_with_errors(buf, len,
+                                                            utf16_output);
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_utf32_to_utf16be_with_errors(buf, len,
+                                                            utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_valid_utf32_to_utf16le(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
+      const char32_t *buf, size_t len,
+      char16_t *utf16_output) const noexcept final override {
+    return set_best()->convert_valid_utf32_to_utf16be(buf, len, utf16_output);
+  }
+
+  simdutf_warn_unused size_t convert_utf16le_to_utf32(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_utf32(buf, len, utf32_output);
+  }
+
+  simdutf_warn_unused size_t convert_utf16be_to_utf32(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_utf32(buf, len, utf32_output);
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf16le_to_utf32_with_errors(buf, len,
+                                                            utf32_output);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_utf16be_to_utf32_with_errors(buf, len,
+                                                            utf32_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_valid_utf16le_to_utf32(buf, len, utf32_output);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
+      const char16_t *buf, size_t len,
+      char32_t *utf32_output) const noexcept final override {
+    return set_best()->convert_valid_utf16be_to_utf32(buf, len, utf32_output);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *buf, size_t len,
+                               char16_t *output) const noexcept final override {
+    set_best()->change_endianness_utf16(buf, len, output);
+  }
+
+  simdutf_warn_unused size_t
+  count_utf16le(const char16_t *buf, size_t len) const noexcept final override {
+    return set_best()->count_utf16le(buf, len);
+  }
+
+  simdutf_warn_unused size_t
+  count_utf16be(const char16_t *buf, size_t len) const noexcept final override {
+    return set_best()->count_utf16be(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t
+  count_utf8(const char *buf, size_t len) const noexcept final override {
+    return set_best()->count_utf8(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *buf, size_t len) const noexcept override {
+    return set_best()->latin1_length_from_utf8(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *buf, size_t len) const noexcept override {
+    return set_best()->utf8_length_from_latin1(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t utf8_length_from_utf16le(
+      const char16_t *buf, size_t len) const noexcept override {
+    return set_best()->utf8_length_from_utf16le(buf, len);
+  }
+
+  simdutf_warn_unused size_t utf8_length_from_utf16be(
+      const char16_t *buf, size_t len) const noexcept override {
+    return set_best()->utf8_length_from_utf16be(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf32_length_from_utf16le(
+      const char16_t *buf, size_t len) const noexcept override {
+    return set_best()->utf32_length_from_utf16le(buf, len);
+  }
+
+  simdutf_warn_unused size_t utf32_length_from_utf16be(
+      const char16_t *buf, size_t len) const noexcept override {
+    return set_best()->utf32_length_from_utf16be(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *buf, size_t len) const noexcept override {
+    return set_best()->utf16_length_from_utf8(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf8_length_from_utf32(
+      const char32_t *buf, size_t len) const noexcept override {
+    return set_best()->utf8_length_from_utf32(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t utf16_length_from_utf32(
+      const char32_t *buf, size_t len) const noexcept override {
+    return set_best()->utf16_length_from_utf32(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *buf, size_t len) const noexcept override {
+    return set_best()->utf32_length_from_utf8(buf, len);
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result base64_to_binary(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_handling_options =
+          last_chunk_handling_options::loose) const noexcept override {
+    return set_best()->base64_to_binary(input, length, output, options,
+                                        last_chunk_handling_options);
+  }
+
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *input, size_t length, char *output, base64_options options,
+      last_chunk_handling_options last_chunk_handling_options =
+          last_chunk_handling_options::loose) const noexcept override {
+    return set_best()->base64_to_binary_details(input, length, output, options,
+                                                last_chunk_handling_options);
+  }
+
+  simdutf_warn_unused result base64_to_binary(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_handling_options =
+          last_chunk_handling_options::loose) const noexcept override {
+    return set_best()->base64_to_binary(input, length, output, options,
+                                        last_chunk_handling_options);
+  }
+
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *input, size_t length, char *output,
+      base64_options options,
+      last_chunk_handling_options last_chunk_handling_options =
+          last_chunk_handling_options::loose) const noexcept override {
+    return set_best()->base64_to_binary_details(input, length, output, options,
+                                                last_chunk_handling_options);
+  }
+
+  size_t binary_to_base64(const char *input, size_t length, char *output,
+                          base64_options options) const noexcept override {
+    return set_best()->binary_to_base64(input, length, output, options);
+  }
+#endif // SIMDUTF_FEATURE_BASE64
+
+  simdutf_really_inline
+  detect_best_supported_implementation_on_first_use() noexcept
+      : implementation("best_supported_detector",
+                       "Detects the best supported implementation and sets it",
+                       0) {}
+
+private:
+  const implementation *set_best() const noexcept;
+};
+
+static_assert(std::is_trivially_destructible<
+                  detect_best_supported_implementation_on_first_use>::value,
+              "detect_best_supported_implementation_on_first_use should be "
+              "trivially destructible");
+
+static const std::initializer_list<const implementation *> &
+get_available_implementation_pointers() {
+  static const std::initializer_list<const implementation *>
+      available_implementation_pointers{
+#if SIMDUTF_IMPLEMENTATION_ICELAKE
+          get_icelake_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_HASWELL
+          get_haswell_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_WESTMERE
+          get_westmere_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_ARM64
+          get_arm64_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_PPC64
+          get_ppc64_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_RVV
+          get_rvv_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_LSX
+          get_lsx_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_LASX
+          get_lasx_singleton(),
+#endif
+#if SIMDUTF_IMPLEMENTATION_FALLBACK
+          get_fallback_singleton(),
+#endif
+      }; // available_implementation_pointers
+  return available_implementation_pointers;
+}
+
+// So we can return UNSUPPORTED_ARCHITECTURE from the parser when there is no
+// support
+class unsupported_implementation final : public implementation {
+public:
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused int detect_encodings(const char *,
+                                           size_t) const noexcept override {
+    return encoding_type::unspecified;
+  }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool validate_utf8(const char *,
+                                         size_t) const noexcept final override {
+    return false; // Just refuse to validate. Given that we have a fallback
+                  // implementation
+    // it seems unlikely that unsupported_implementation will ever be used. If
+    // it is used, then it will flag all strings as invalid. The alternative is
+    // to return an error_code from which the user has to figure out whether the
+    // string is valid UTF-8... which seems like a lot of work just to handle
+    // the very unlikely case that we have an unsupported implementation. And,
+    // when it does happen (that we have an unsupported implementation), what
+    // are the chances that the programmer has a fallback? Given that *we*
+    // provide the fallback, it implies that the programmer would need a
+    // fallback for our fallback.
+  }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused result validate_utf8_with_errors(
+      const char *, size_t) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+  simdutf_warn_unused bool
+  validate_ascii(const char *, size_t) const noexcept final override {
+    return false;
+  }
+
+  simdutf_warn_unused result validate_ascii_with_errors(
+      const char *, size_t) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool
+  validate_utf16le(const char16_t *, size_t) const noexcept final override {
+    return false;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused bool
+  validate_utf16be(const char16_t *, size_t) const noexcept final override {
+    return false;
+  }
+
+  simdutf_warn_unused result validate_utf16le_with_errors(
+      const char16_t *, size_t) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result validate_utf16be_with_errors(
+      const char16_t *, size_t) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+  simdutf_warn_unused bool
+  validate_utf32(const char32_t *, size_t) const noexcept final override {
+    return false;
+  }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused result validate_utf32_with_errors(
+      const char32_t *, size_t) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf8(
+      const char *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf16le(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_latin1_to_utf16be(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_latin1_to_utf32(
+      const char *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf8_to_latin1(
+      const char *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+      const char *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+      const char *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf8_to_utf16le(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf8_to_utf16be(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+      const char *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf8_to_utf32(
+      const char *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+      const char *, size_t, char32_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+      const char *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf16le_to_latin1(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf16be_to_latin1(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t convert_utf16le_to_utf8(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf16be_to_utf8(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+      const char16_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t convert_utf32_to_latin1(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf32_to_latin1_with_errors(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf8(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+      const char32_t *, size_t, char *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t convert_utf32_to_utf16le(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf32_to_utf16be(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
+      const char32_t *, size_t, char16_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf16le_to_utf32(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_utf16be_to_utf32(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
+      const char16_t *, size_t, char32_t *) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+  void change_endianness_utf16(const char16_t *, size_t,
+                               char16_t *) const noexcept final override {}
+
+  simdutf_warn_unused size_t
+  count_utf16le(const char16_t *, size_t) const noexcept final override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t
+  count_utf16be(const char16_t *, size_t) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+  simdutf_warn_unused size_t count_utf8(const char *,
+                                        size_t) const noexcept final override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  latin1_length_from_utf8(const char *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  simdutf_warn_unused size_t
+  utf8_length_from_latin1(const char *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16le(const char16_t *, size_t) const noexcept override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t
+  utf8_length_from_utf16be(const char16_t *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf16le(const char16_t *, size_t) const noexcept override {
+    return 0;
+  }
+
+  simdutf_warn_unused size_t
+  utf32_length_from_utf16be(const char16_t *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  simdutf_warn_unused size_t
+  utf16_length_from_utf8(const char *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf8_length_from_utf32(const char32_t *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf16_length_from_utf32(const char32_t *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  simdutf_warn_unused size_t
+  utf32_length_from_utf8(const char *, size_t) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+  simdutf_warn_unused result
+  base64_to_binary(const char *, size_t, char *, base64_options,
+                   last_chunk_handling_options) const noexcept override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char *, size_t, char *, base64_options,
+      last_chunk_handling_options) const noexcept override {
+    return full_result(error_code::OTHER, 0, 0);
+  }
+
+  simdutf_warn_unused result
+  base64_to_binary(const char16_t *, size_t, char *, base64_options,
+                   last_chunk_handling_options) const noexcept override {
+    return result(error_code::OTHER, 0);
+  }
+
+  simdutf_warn_unused full_result base64_to_binary_details(
+      const char16_t *, size_t, char *, base64_options,
+      last_chunk_handling_options) const noexcept override {
+    return full_result(error_code::OTHER, 0, 0);
+  }
+
+  size_t binary_to_base64(const char *, size_t, char *,
+                          base64_options) const noexcept override {
+    return 0;
+  }
+#endif // SIMDUTF_FEATURE_BASE64
+
+  unsupported_implementation()
+      : implementation("unsupported",
+                       "Unsupported CPU (no detected SIMD instructions)", 0) {}
+};
+
+const unsupported_implementation *get_unsupported_singleton() {
+  static const unsupported_implementation unsupported_singleton{};
+  return &unsupported_singleton;
+}
+static_assert(std::is_trivially_destructible<unsupported_implementation>::value,
+              "unsupported_singleton should be trivially destructible");
+
+size_t available_implementation_list::size() const noexcept {
+  return internal::get_available_implementation_pointers().size();
+}
+const implementation *const *
+available_implementation_list::begin() const noexcept {
+  return internal::get_available_implementation_pointers().begin();
+}
+const implementation *const *
+available_implementation_list::end() const noexcept {
+  return internal::get_available_implementation_pointers().end();
+}
+const implementation *
+available_implementation_list::detect_best_supported() const noexcept {
+  // They are prelisted in priority order, so we just go down the list
+  uint32_t supported_instruction_sets =
+      internal::detect_supported_architectures();
+  for (const implementation *impl :
+       internal::get_available_implementation_pointers()) {
+    uint32_t required_instruction_sets = impl->required_instruction_sets();
+    if ((supported_instruction_sets & required_instruction_sets) ==
+        required_instruction_sets) {
+      return impl;
+    }
+  }
+  return get_unsupported_singleton(); // this should never happen?
+}
+
+const implementation *
+detect_best_supported_implementation_on_first_use::set_best() const noexcept {
+  SIMDUTF_PUSH_DISABLE_WARNINGS
+  SIMDUTF_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC:
+                                     // manually verified this is safe
+      char *force_implementation_name = getenv("SIMDUTF_FORCE_IMPLEMENTATION");
+  SIMDUTF_POP_DISABLE_WARNINGS
+
+  if (force_implementation_name) {
+    auto force_implementation =
+        get_available_implementations()[force_implementation_name];
+    if (force_implementation) {
+      return get_active_implementation() = force_implementation;
+    } else {
+      // Note: abort() and stderr usage within the library is forbidden.
+      return get_active_implementation() = get_unsupported_singleton();
+    }
+  }
+  return get_active_implementation() =
+             get_available_implementations().detect_best_supported();
+}
+
+} // namespace internal
+
+/**
+ * The list of available implementations compiled into simdutf.
+ */
+SIMDUTF_DLLIMPORTEXPORT const internal::available_implementation_list &
+get_available_implementations() {
+  static const internal::available_implementation_list
+      available_implementations{};
+  return available_implementations;
+}
+
+/**
+ * The active implementation.
+ */
+SIMDUTF_DLLIMPORTEXPORT internal::atomic_ptr<const implementation> &
+get_active_implementation() {
+#if SIMDUTF_SINGLE_IMPLEMENTATION
+  // skip runtime detection
+  static internal::atomic_ptr<const implementation> active_implementation{
+      internal::get_single_implementation()};
+  return active_implementation;
+#else
+  static const internal::detect_best_supported_implementation_on_first_use
+      detect_best_supported_implementation_on_first_use_singleton;
+  static internal::atomic_ptr<const implementation> active_implementation{
+      &detect_best_supported_implementation_on_first_use_singleton};
+  return active_implementation;
+#endif
+}
+
+#if SIMDUTF_SINGLE_IMPLEMENTATION
+const implementation *get_default_implementation() {
+  return internal::get_single_implementation();
+}
+#else
+internal::atomic_ptr<const implementation> &get_default_implementation() {
+  return get_active_implementation();
+}
+#endif
+#define SIMDUTF_GET_CURRENT_IMPLEMENTION
+
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept {
+  return get_default_implementation()->validate_utf8(buf, len);
+}
+simdutf_warn_unused result validate_utf8_with_errors(const char *buf,
+                                                     size_t len) noexcept {
+  return get_default_implementation()->validate_utf8_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_ASCII
+simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) noexcept {
+  return get_default_implementation()->validate_ascii(buf, len);
+}
+simdutf_warn_unused result validate_ascii_with_errors(const char *buf,
+                                                      size_t len) noexcept {
+  return get_default_implementation()->validate_ascii_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_utf8_to_utf16(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf8_to_utf16be(input, length, utf16_output);
+  #else
+  return convert_utf8_to_utf16le(input, length, utf16_output);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_latin1_to_utf8(const char *buf, size_t len,
+                                                  char *utf8_output) noexcept {
+  return get_default_implementation()->convert_latin1_to_utf8(buf, len,
+                                                              utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_latin1_to_utf16le(buf, len,
+                                                                 utf16_output);
+}
+simdutf_warn_unused size_t convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_latin1_to_utf16be(buf, len,
+                                                                 utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *latin1_output) noexcept {
+  return get_default_implementation()->convert_latin1_to_utf32(buf, len,
+                                                               latin1_output);
+}
+simdutf_warn_unused size_t latin1_length_from_utf32(size_t length) noexcept {
+  return length;
+}
+simdutf_warn_unused size_t utf32_length_from_latin1(size_t length) noexcept {
+  return length;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_latin1(buf, len,
+                                                              latin1_output);
+}
+simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_latin1_with_errors(
+      buf, len, latin1_output);
+}
+simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) noexcept {
+  return get_default_implementation()->convert_valid_utf8_to_latin1(
+      buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_utf8_to_utf16le(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf16le(input, length,
+                                                               utf16_output);
+}
+simdutf_warn_unused size_t convert_utf8_to_utf16be(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf16be(input, length,
+                                                               utf16_output);
+}
+simdutf_warn_unused result convert_utf8_to_utf16_with_errors(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf8_to_utf16be_with_errors(input, length, utf16_output);
+  #else
+  return convert_utf8_to_utf16le_with_errors(input, length, utf16_output);
+  #endif
+}
+simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf16le_with_errors(
+      input, length, utf16_output);
+}
+simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
+    const char *input, size_t length, char16_t *utf16_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf16be_with_errors(
+      input, length, utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t convert_utf8_to_utf32(
+    const char *input, size_t length, char32_t *utf32_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf32(input, length,
+                                                             utf32_output);
+}
+simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
+    const char *input, size_t length, char32_t *utf32_output) noexcept {
+  return get_default_implementation()->convert_utf8_to_utf32_with_errors(
+      input, length, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool validate_utf16(const char16_t *buf,
+                                        size_t len) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return validate_utf16be(buf, len);
+  #else
+  return validate_utf16le(buf, len);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
+                                          size_t len) noexcept {
+  return get_default_implementation()->validate_utf16le(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
+                                          size_t len) noexcept {
+  return get_default_implementation()->validate_utf16be(buf, len);
+}
+simdutf_warn_unused result validate_utf16_with_errors(const char16_t *buf,
+                                                      size_t len) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return validate_utf16be_with_errors(buf, len);
+  #else
+  return validate_utf16le_with_errors(buf, len);
+  #endif
+}
+simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf,
+                                                        size_t len) noexcept {
+  return get_default_implementation()->validate_utf16le_with_errors(buf, len);
+}
+simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf,
+                                                        size_t len) noexcept {
+  return get_default_implementation()->validate_utf16be_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused bool validate_utf32(const char32_t *buf,
+                                        size_t len) noexcept {
+  return get_default_implementation()->validate_utf32(buf, len);
+}
+simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf,
+                                                      size_t len) noexcept {
+  return get_default_implementation()->validate_utf32_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
+    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_valid_utf8_to_utf16be(input, length, utf16_buffer);
+  #else
+  return convert_valid_utf8_to_utf16le(input, length, utf16_buffer);
+  #endif
+}
+simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
+    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf8_to_utf16le(
+      input, length, utf16_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
+    const char *input, size_t length, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf8_to_utf16be(
+      input, length, utf16_buffer);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
+    const char *input, size_t length, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf8_to_utf32(
+      input, length, utf32_buffer);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_utf16_to_utf8(const char16_t *buf,
+                                                 size_t len,
+                                                 char *utf8_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_utf8(buf, len, utf8_buffer);
+  #else
+  return convert_utf16le_to_utf8(buf, len, utf8_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_utf16_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_latin1(buf, len, latin1_buffer);
+  #else
+  return convert_utf16le_to_latin1(buf, len, latin1_buffer);
+  #endif
+}
+simdutf_warn_unused size_t convert_latin1_to_utf16(
+    const char *buf, size_t len, char16_t *utf16_output) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_latin1_to_utf16be(buf, len, utf16_output);
+  #else
+  return convert_latin1_to_utf16le(buf, len, utf16_output);
+  #endif
+}
+simdutf_warn_unused size_t convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_latin1(buf, len,
+                                                                 latin1_buffer);
+}
+simdutf_warn_unused size_t convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_latin1(buf, len,
+                                                                 latin1_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16be_to_latin1(
+      buf, len, latin1_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16le_to_latin1(
+      buf, len, latin1_buffer);
+}
+simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_latin1_with_errors(
+      buf, len, latin1_buffer);
+}
+simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_latin1_with_errors(
+      buf, len, latin1_buffer);
+}
+simdutf_warn_unused size_t latin1_length_from_utf16(size_t length) noexcept {
+  return length;
+}
+simdutf_warn_unused size_t utf16_length_from_latin1(size_t length) noexcept {
+  return length;
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t *buf,
+                                                   size_t len,
+                                                   char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_utf8(buf, len,
+                                                               utf8_buffer);
+}
+simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t *buf,
+                                                   size_t len,
+                                                   char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_utf8(buf, len,
+                                                               utf8_buffer);
+}
+simdutf_warn_unused result convert_utf16_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_utf8_with_errors(buf, len, utf8_buffer);
+  #else
+  return convert_utf16le_to_utf8_with_errors(buf, len, utf8_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused result convert_utf16_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_latin1_with_errors(buf, len, latin1_buffer);
+  #else
+  return convert_utf16le_to_latin1_with_errors(buf, len, latin1_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_utf8_with_errors(
+      buf, len, utf8_buffer);
+}
+simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_utf8_with_errors(
+      buf, len, utf8_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_valid_utf16be_to_utf8(buf, len, utf8_buffer);
+  #else
+  return convert_valid_utf16le_to_utf8(buf, len, utf8_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_valid_utf16be_to_latin1(buf, len, latin1_buffer);
+  #else
+  return convert_valid_utf16le_to_latin1(buf, len, latin1_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16le_to_utf8(
+      buf, len, utf8_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16be_to_utf8(
+      buf, len, utf8_buffer);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t *buf,
+                                                 size_t len,
+                                                 char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf8(buf, len,
+                                                             utf8_buffer);
+}
+simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
+    const char32_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf8_with_errors(
+      buf, len, utf8_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf32_to_utf8(buf, len,
+                                                                   utf8_buffer);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t convert_utf32_to_utf16(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf32_to_utf16be(buf, len, utf16_buffer);
+  #else
+  return convert_utf32_to_utf16le(buf, len, utf16_buffer);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_utf32_to_latin1(
+    const char32_t *input, size_t length, char *latin1_output) noexcept {
+  return get_default_implementation()->convert_utf32_to_latin1(input, length,
+                                                               latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t convert_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf16le(buf, len,
+                                                                utf16_buffer);
+}
+simdutf_warn_unused size_t convert_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf16be(buf, len,
+                                                                utf16_buffer);
+}
+simdutf_warn_unused result convert_utf32_to_utf16_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf32_to_utf16be_with_errors(buf, len, utf16_buffer);
+  #else
+  return convert_utf32_to_utf16le_with_errors(buf, len, utf16_buffer);
+  #endif
+}
+simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf16le_with_errors(
+      buf, len, utf16_buffer);
+}
+simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_utf32_to_utf16be_with_errors(
+      buf, len, utf16_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf32_to_utf16(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_valid_utf32_to_utf16be(buf, len, utf16_buffer);
+  #else
+  return convert_valid_utf32_to_utf16le(buf, len, utf16_buffer);
+  #endif
+}
+simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf32_to_utf16le(
+      buf, len, utf16_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf32_to_utf16be(
+      buf, len, utf16_buffer);
+}
+simdutf_warn_unused size_t convert_utf16_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_utf32(buf, len, utf32_buffer);
+  #else
+  return convert_utf16le_to_utf32(buf, len, utf32_buffer);
+  #endif
+}
+simdutf_warn_unused size_t convert_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_utf32(buf, len,
+                                                                utf32_buffer);
+}
+simdutf_warn_unused size_t convert_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_utf32(buf, len,
+                                                                utf32_buffer);
+}
+simdutf_warn_unused result convert_utf16_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_utf16be_to_utf32_with_errors(buf, len, utf32_buffer);
+  #else
+  return convert_utf16le_to_utf32_with_errors(buf, len, utf32_buffer);
+  #endif
+}
+simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_utf16le_to_utf32_with_errors(
+      buf, len, utf32_buffer);
+}
+simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_utf16be_to_utf32_with_errors(
+      buf, len, utf32_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return convert_valid_utf16be_to_utf32(buf, len, utf32_buffer);
+  #else
+  return convert_valid_utf16le_to_utf32(buf, len, utf32_buffer);
+  #endif
+}
+simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16le_to_utf32(
+      buf, len, utf32_buffer);
+}
+simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_buffer) noexcept {
+  return get_default_implementation()->convert_valid_utf16be_to_utf32(
+      buf, len, utf32_buffer);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16
+void change_endianness_utf16(const char16_t *input, size_t length,
+                             char16_t *output) noexcept {
+  get_default_implementation()->change_endianness_utf16(input, length, output);
+}
+simdutf_warn_unused size_t count_utf16(const char16_t *input,
+                                       size_t length) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return count_utf16be(input, length);
+  #else
+  return count_utf16le(input, length);
+  #endif
+}
+simdutf_warn_unused size_t count_utf16le(const char16_t *input,
+                                         size_t length) noexcept {
+  return get_default_implementation()->count_utf16le(input, length);
+}
+simdutf_warn_unused size_t count_utf16be(const char16_t *input,
+                                         size_t length) noexcept {
+  return get_default_implementation()->count_utf16be(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t count_utf8(const char *input,
+                                      size_t length) noexcept {
+  return get_default_implementation()->count_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t latin1_length_from_utf8(const char *buf,
+                                                   size_t len) noexcept {
+  return get_default_implementation()->latin1_length_from_utf8(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t utf8_length_from_latin1(const char *buf,
+                                                   size_t len) noexcept {
+  return get_default_implementation()->utf8_length_from_latin1(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t utf8_length_from_utf16(const char16_t *input,
+                                                  size_t length) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return utf8_length_from_utf16be(input, length);
+  #else
+  return utf8_length_from_utf16le(input, length);
+  #endif
+}
+simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t *input,
+                                                    size_t length) noexcept {
+  return get_default_implementation()->utf8_length_from_utf16le(input, length);
+}
+simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t *input,
+                                                    size_t length) noexcept {
+  return get_default_implementation()->utf8_length_from_utf16be(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t utf32_length_from_utf16(const char16_t *input,
+                                                   size_t length) noexcept {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return utf32_length_from_utf16be(input, length);
+  #else
+  return utf32_length_from_utf16le(input, length);
+  #endif
+}
+simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t *input,
+                                                     size_t length) noexcept {
+  return get_default_implementation()->utf32_length_from_utf16le(input, length);
+}
+simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t *input,
+                                                     size_t length) noexcept {
+  return get_default_implementation()->utf32_length_from_utf16be(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t utf16_length_from_utf8(const char *input,
+                                                  size_t length) noexcept {
+  return get_default_implementation()->utf16_length_from_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t *input,
+                                                  size_t length) noexcept {
+  return get_default_implementation()->utf8_length_from_utf32(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t *input,
+                                                   size_t length) noexcept {
+  return get_default_implementation()->utf16_length_from_utf32(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t utf32_length_from_utf8(const char *input,
+                                                  size_t length) noexcept {
+  return get_default_implementation()->utf32_length_from_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused size_t
+maximal_binary_length_from_base64(const char *input, size_t length) noexcept {
+  return get_default_implementation()->maximal_binary_length_from_base64(
+      input, length);
+}
+
+simdutf_warn_unused result base64_to_binary(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_handling_options) noexcept {
+  return get_default_implementation()->base64_to_binary(
+      input, length, output, options, last_chunk_handling_options);
+}
+
+simdutf_warn_unused size_t maximal_binary_length_from_base64(
+    const char16_t *input, size_t length) noexcept {
+  return get_default_implementation()->maximal_binary_length_from_base64(
+      input, length);
+}
+
+simdutf_warn_unused result base64_to_binary(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_handling_options) noexcept {
+  return get_default_implementation()->base64_to_binary(
+      input, length, output, options, last_chunk_handling_options);
+}
+
+template <typename chartype>
+simdutf_warn_unused result base64_to_binary_safe_impl(
+    const chartype *input, size_t length, char *output, size_t &outlen,
+    base64_options options,
+    last_chunk_handling_options last_chunk_handling_options) noexcept {
+  static_assert(std::is_same<chartype, char>::value ||
+                    std::is_same<chartype, char16_t>::value,
+                "Only char and char16_t are supported.");
+  // The implementation could be nicer, but we expect that most times, the user
+  // will provide us with a buffer that is large enough.
+  size_t max_length = maximal_binary_length_from_base64(input, length);
+  if (outlen >= max_length) {
+    // fast path
+    full_result r = get_default_implementation()->base64_to_binary_details(
+        input, length, output, options, last_chunk_handling_options);
+    if (r.error != error_code::INVALID_BASE64_CHARACTER &&
+        r.error != error_code::BASE64_EXTRA_BITS) {
+      outlen = r.output_count;
+      if (last_chunk_handling_options == stop_before_partial) {
+        if ((r.output_count % 3) != 0) {
+          bool empty_trail = true;
+          for (size_t i = r.input_count; i < length; i++) {
+            if (!scalar::base64::is_ascii_white_space_or_padding(input[i])) {
+              empty_trail = false;
+              break;
+            }
+          }
+          if (empty_trail) {
+            r.input_count = length;
+          }
+        }
+        return {r.error, r.input_count};
+      }
+      return {r.error, length};
+    }
+    return r;
+  }
+  // The output buffer is maybe too small. We will decode a truncated version of
+  // the input.
+  size_t outlen3 = outlen / 3 * 3; // round down to multiple of 3
+  size_t safe_input = base64_length_from_binary(outlen3, options);
+  full_result r = get_default_implementation()->base64_to_binary_details(
+      input, safe_input, output, options, loose);
+  if (r.error == error_code::INVALID_BASE64_CHARACTER) {
+    return r;
+  }
+  size_t offset =
+      (r.error == error_code::BASE64_INPUT_REMAINDER)
+          ? 1
+          : ((r.output_count % 3) == 0 ? 0 : (r.output_count % 3) + 1);
+  size_t output_index = r.output_count - (r.output_count % 3);
+  size_t input_index = safe_input;
+  // offset is a value that is no larger than 3. We backtrack
+  // by up to offset characters + an undetermined number of
+  // white space characters. It is expected that the next loop
+  // runs at most 3 times + the number of white space characters
+  // in between them, so we are not worried about performance.
+  while (offset > 0 && input_index > 0) {
+    chartype c = input[--input_index];
+    if (scalar::base64::is_ascii_white_space(c)) {
+      // skipping
+    } else {
+      offset--;
+    }
+  }
+  size_t remaining_out = outlen - output_index;
+  const chartype *tail_input = input + input_index;
+  size_t tail_length = length - input_index;
+  while (tail_length > 0 &&
+         scalar::base64::is_ascii_white_space(tail_input[tail_length - 1])) {
+    tail_length--;
+  }
+  size_t padding_characts = 0;
+  if (tail_length > 0 && tail_input[tail_length - 1] == '=') {
+    tail_length--;
+    padding_characts++;
+    while (tail_length > 0 &&
+           scalar::base64::is_ascii_white_space(tail_input[tail_length - 1])) {
+      tail_length--;
+    }
+    if (tail_length > 0 && tail_input[tail_length - 1] == '=') {
+      tail_length--;
+      padding_characts++;
+    }
+  }
+  // this will advance tail_input and tail_length
+  result rr = scalar::base64::base64_tail_decode_safe(
+      output + output_index, remaining_out, tail_input, tail_length,
+      padding_characts, options, last_chunk_handling_options);
+  outlen = output_index + remaining_out;
+  if (last_chunk_handling_options != stop_before_partial &&
+      rr.error == error_code::SUCCESS && padding_characts > 0) {
+    // additional checks
+    if ((outlen % 3 == 0) || ((outlen % 3) + 1 + padding_characts != 4)) {
+      rr.error = error_code::INVALID_BASE64_CHARACTER;
+    }
+  }
+  if (rr.error == error_code::SUCCESS &&
+      last_chunk_handling_options == stop_before_partial) {
+    if (tail_input > input + input_index) {
+      rr.count = tail_input - input;
+    } else if (r.input_count > 0) {
+      rr.count = r.input_count + rr.count;
+    }
+    return rr;
+  }
+  rr.count += input_index;
+  return rr;
+}
+
+  #if SIMDUTF_ATOMIC_REF
+size_t atomic_binary_to_base64(const char *input, size_t length, char *output,
+                               base64_options options) noexcept {
+  static_assert(std::atomic_ref<char>::required_alignment == 1);
+  size_t retval = 0;
+  // Arbitrary block sizes: 3KB for input, 4KB for output. Total is 7KB.
+  constexpr size_t input_block_size = 1024 * 3;
+  constexpr size_t output_block_size = input_block_size * 4 / 3;
+  std::array<char, input_block_size> inbuf;
+  std::array<char, output_block_size> outbuf;
+
+  // std::atomic_ref<T> must not have a const T, see
+  // https://cplusplus.github.io/LWG/issue3508
+  // we instead provide a mutable input, which is ok since we are only reading
+  // from it.
+  char *mutable_input = const_cast<char *>(input);
+
+  for (size_t i = 0; i < length; i += input_block_size) {
+    const size_t current_block_size = std::min(input_block_size, length - i);
+    // This copy is inefficient.
+    // Under x64, we could use 16-byte aligned loads.
+    // Note that we warn users that the performance might be poor.
+    for (size_t j = 0; j < current_block_size; ++j) {
+      inbuf[j] = std::atomic_ref<char>(mutable_input[i + j])
+                     .load(std::memory_order_relaxed);
+    }
+    const size_t written = binary_to_base64(inbuf.data(), current_block_size,
+                                            outbuf.data(), options);
+    // This copy is inefficient.
+    // Under x64, we could use 16-byte aligned stores.
+    for (size_t j = 0; j < written; ++j) {
+      std::atomic_ref<char>(output[retval + j])
+          .store(outbuf[j], std::memory_order_relaxed);
+    }
+    retval += written;
+  }
+  return retval;
+}
+  #endif // SIMDUTF_ATOMIC_REF
+
+#endif // SIMDUTF_FEATURE_BASE64
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t convert_latin1_to_utf8_safe(
+    const char *buf, size_t len, char *utf8_output, size_t utf8_len) noexcept {
+  const auto start{utf8_output};
+
+  while (true) {
+    // convert_latin1_to_utf8 will never write more than input length * 2
+    auto read_len = std::min(len, utf8_len >> 1);
+    if (read_len <= 16) {
+      break;
+    }
+
+    const auto write_len =
+        simdutf::convert_latin1_to_utf8(buf, read_len, utf8_output);
+
+    utf8_output += write_len;
+    utf8_len -= write_len;
+    buf += read_len;
+    len -= read_len;
+  }
+
+  utf8_output +=
+      scalar::latin1_to_utf8::convert_safe(buf, len, utf8_output, utf8_len);
+
+  return utf8_output - start;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused result base64_to_binary_safe(
+    const char *input, size_t length, char *output, size_t &outlen,
+    base64_options options,
+    last_chunk_handling_options last_chunk_handling_options) noexcept {
+  return base64_to_binary_safe_impl<char>(input, length, output, outlen,
+                                          options, last_chunk_handling_options);
+}
+simdutf_warn_unused result base64_to_binary_safe(
+    const char16_t *input, size_t length, char *output, size_t &outlen,
+    base64_options options,
+    last_chunk_handling_options last_chunk_handling_options) noexcept {
+  return base64_to_binary_safe_impl<char16_t>(
+      input, length, output, outlen, options, last_chunk_handling_options);
+}
+
+simdutf_warn_unused size_t
+base64_length_from_binary(size_t length, base64_options options) noexcept {
+  return get_default_implementation()->base64_length_from_binary(length,
+                                                                 options);
+}
+
+size_t binary_to_base64(const char *input, size_t length, char *output,
+                        base64_options options) noexcept {
+  return get_default_implementation()->binary_to_base64(input, length, output,
+                                                        options);
+}
+#endif // SIMDUTF_FEATURE_BASE64
+
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused simdutf::encoding_type
+autodetect_encoding(const char *buf, size_t length) noexcept {
+  return get_default_implementation()->autodetect_encoding(buf, length);
+}
+
+simdutf_warn_unused int detect_encodings(const char *buf,
+                                         size_t length) noexcept {
+  return get_default_implementation()->detect_encodings(buf, length);
+}
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
+
+const implementation *builtin_implementation() {
+  static const implementation *builtin_impl =
+      get_available_implementations()[SIMDUTF_STRINGIFY(
+          SIMDUTF_BUILTIN_IMPLEMENTATION)];
+  return builtin_impl;
+}
+
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused size_t trim_partial_utf8(const char *input, size_t length) {
+  return scalar::utf8::trim_partial_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t trim_partial_utf16be(const char16_t *input,
+                                                size_t length) {
+  return scalar::utf16::trim_partial_utf16<BIG>(input, length);
+}
+
+simdutf_warn_unused size_t trim_partial_utf16le(const char16_t *input,
+                                                size_t length) {
+  return scalar::utf16::trim_partial_utf16<LITTLE>(input, length);
+}
+
+simdutf_warn_unused size_t trim_partial_utf16(const char16_t *input,
+                                              size_t length) {
+  #if SIMDUTF_IS_BIG_ENDIAN
+  return trim_partial_utf16be(input, length);
+  #else
+  return trim_partial_utf16le(input, length);
+  #endif
+}
+#endif // SIMDUTF_FEATURE_UTF16
+
 } // namespace simdutf
+/* end file src/implementation.cpp */
+
+SIMDUTF_PUSH_DISABLE_WARNINGS
+SIMDUTF_DISABLE_UNDESIRED_WARNINGS
+
+#if SIMDUTF_IMPLEMENTATION_ARM64
+/* begin file src/arm64/implementation.cpp */
+/* begin file src/simdutf/arm64/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "arm64"
+// #define SIMDUTF_IMPLEMENTATION arm64
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
+/* end file src/simdutf/arm64/begin.h */
+namespace simdutf {
+namespace arm64 {
+namespace {
+#ifndef SIMDUTF_ARM64_H
+  #error "arm64.h must be included"
+#endif
+using namespace simd;
+
+#if SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    SIMDUTF_FEATURE_UTF8
+simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
+  simd8<uint8_t> bits = input.reduce_or();
+  return bits.max_val() < 0b10000000u;
+}
+#endif // SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_really_inline simd8<bool>
+must_be_2_3_continuation(const simd8<uint8_t> prev2,
+                         const simd8<uint8_t> prev3) {
+  simd8<bool> is_third_byte = prev2 >= uint8_t(0b11100000u);
+  simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
+  return is_third_byte ^ is_fourth_byte;
+}
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32)
+// common functions for utf8 conversions
+simdutf_really_inline uint16x4_t convert_utf8_3_byte_to_utf16(uint8x16_t in) {
+  // Low half contains  10cccccc|1110aaaa
+  // High half contains 10bbbbbb|10bbbbbb
+  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  const uint8x16_t sh = simdutf_make_uint8x16_t(0, 2, 3, 5, 6, 8, 9, 11, 1, 1,
+                                                4, 4, 7, 7, 10, 10);
+  #else
+  const uint8x16_t sh = {0, 2, 3, 5, 6, 8, 9, 11, 1, 1, 4, 4, 7, 7, 10, 10};
+  #endif
+  uint8x16_t perm = vqtbl1q_u8(in, sh);
+  // Split into half vectors.
+  // 10cccccc|1110aaaa
+  uint8x8_t perm_low = vget_low_u8(perm); // no-op
+  // 10bbbbbb|10bbbbbb
+  uint8x8_t perm_high = vget_high_u8(perm);
+  // xxxxxxxx 10bbbbbb
+  uint16x4_t mid = vreinterpret_u16_u8(perm_high); // no-op
+  // xxxxxxxx 1110aaaa
+  uint16x4_t high = vreinterpret_u16_u8(perm_low); // no-op
+  // Assemble with shift left insert.
+  // xxxxxxaa aabbbbbb
+  uint16x4_t mid_high = vsli_n_u16(mid, high, 6);
+  // (perm_low << 8) | (perm_low >> 8)
+  // xxxxxxxx 10cccccc
+  uint16x4_t low = vreinterpret_u16_u8(vrev16_u8(perm_low));
+  // Shift left insert into the low bits
+  // aaaabbbb bbcccccc
+  uint16x4_t composed = vsli_n_u16(low, mid_high, 6);
+  return composed;
+}
+
+simdutf_really_inline uint16x8_t convert_utf8_2_byte_to_utf16(uint8x16_t in) {
+  // Converts 6 2 byte UTF-8 characters to 6 UTF-16 characters.
+  // Technically this calculates 8, but 6 does better and happens more often
+  // (The languages which use these codepoints use ASCII spaces so 8 would need
+  // to be in the middle of a very long word).
+
+  // 10bbbbbb 110aaaaa
+  uint16x8_t upper = vreinterpretq_u16_u8(in);
+  // (in << 8) | (in >> 8)
+  // 110aaaaa 10bbbbbb
+  uint16x8_t lower = vreinterpretq_u16_u8(vrev16q_u8(in));
+  // 00000000 000aaaaa
+  uint16x8_t upper_masked = vandq_u16(upper, vmovq_n_u16(0x1F));
+  // Assemble with shift left insert.
+  // 00000aaa aabbbbbb
+  uint16x8_t composed = vsliq_n_u16(lower, upper_masked, 6);
+  return composed;
+}
+
+simdutf_really_inline uint16x8_t
+convert_utf8_1_to_2_byte_to_utf16(uint8x16_t in, size_t shufutf8_idx) {
+  // Converts 6 1-2 byte UTF-8 characters to 6 UTF-16 characters.
+  // This is a relatively easy scenario
+  // we process SIX (6) input code-code units. The max length in bytes of six
+  // code code units spanning between 1 and 2 bytes each is 12 bytes.
+  uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+      simdutf::tables::utf8_to_utf16::shufutf8[shufutf8_idx]));
+  // Shuffle
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 110aaaaa 10bbbbbb
+  uint16x8_t perm = vreinterpretq_u16_u8(vqtbl1q_u8(in, sh));
+  // Mask
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 00000000 00bbbbbb
+  uint16x8_t ascii = vandq_u16(perm, vmovq_n_u16(0x7f)); // 6 or 7 bits
+  // 1 byte: 00000000 00000000
+  // 2 byte: 000aaaaa 00000000
+  uint16x8_t highbyte = vandq_u16(perm, vmovq_n_u16(0x1f00)); // 5 bits
+  // Combine with a shift right accumulate
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 00000aaa aabbbbbb
+  uint16x8_t composed = vsraq_n_u16(ascii, highbyte, 2);
+  return composed;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32)
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/arm64/arm_validate_utf16.cpp */
+template <endianness big_endian>
+const char16_t *arm_validate_utf16(const char16_t *input, size_t size) {
+  const char16_t *end = input + size;
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+  while (end - input >= 16) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
+    if (!match_system(big_endian)) {
+      in0 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in0)));
+      in1 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in1)));
+    }
+    const auto t0 = in0.shr<8>();
+    const auto t1 = in1.shr<8>();
+    const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
+    if (surrogates_wordmask == 0) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
+
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint64_t V = ~surrogates_wordmask;
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = ((in & v_fc) == v_dc);
+      const uint64_t H = vH.to_bitmask64();
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint64_t L = ~H & surrogates_wordmask;
+
+      const uint64_t a =
+          L & (H >> 4); // A low surrogate must be followed by high one.
+                        // (A low surrogate placed in the 7th register's word
+                        // is an exception we handle.)
+      const uint64_t b =
+          a << 4; // Just mark that the opposite fact is hold,
+                  // thanks to that we have only two masks for valid case.
+      const uint64_t c = V | a | b; // Combine all the masks into the final one.
+      if (c == ~0ull) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0xfffffffffffffffull) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return nullptr;
+      }
+    }
+  }
+  return input;
+}
+
+template <endianness big_endian>
+const result arm_validate_utf16_with_errors(const char16_t *input,
+                                            size_t size) {
+  const char16_t *start = input;
+  const char16_t *end = input + size;
+
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+  while (input + 16 < end) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
+
+    if (!match_system(big_endian)) {
+      in0 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in0)));
+      in1 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in1)));
+    }
+    const auto t0 = in0.shr<8>();
+    const auto t1 = in1.shr<8>();
+    const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
+    if (surrogates_wordmask == 0) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
+
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint64_t V = ~surrogates_wordmask;
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = ((in & v_fc) == v_dc);
+      const uint64_t H = vH.to_bitmask64();
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint64_t L = ~H & surrogates_wordmask;
+
+      const uint64_t a =
+          L & (H >> 4); // A low surrogate must be followed by high one.
+                        // (A low surrogate placed in the 7th register's word
+                        // is an exception we handle.)
+      const uint64_t b =
+          a << 4; // Just mark that the opposite fact is hold,
+                  // thanks to that we have only two masks for valid case.
+      const uint64_t c = V | a | b; // Combine all the masks into the final one.
+      if (c == ~0ull) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0xfffffffffffffffull) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return result(error_code::SURROGATE, input - start);
+      }
+    }
+  }
+  return result(error_code::SUCCESS, input - start);
+}
+/* end file src/arm64/arm_validate_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/arm64/arm_validate_utf32le.cpp */
+
+const char32_t *arm_validate_utf32le(const char32_t *input, size_t size) {
+  const char32_t *end = input + size;
+
+  const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
+  const uint32x4_t offset = vmovq_n_u32(0xffff2000);
+  const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
+  uint32x4_t currentmax = vmovq_n_u32(0x0);
+  uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);
+
+  while (end - input >= 4) {
+    const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input));
+    currentmax = vmaxq_u32(in, currentmax);
+    currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);
+    input += 4;
+  }
+
+  uint32x4_t is_zero =
+      veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
+  if (vmaxvq_u32(is_zero) != 0) {
+    return nullptr;
+  }
+
+  is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax),
+                      standardoffsetmax);
+  if (vmaxvq_u32(is_zero) != 0) {
+    return nullptr;
+  }
+
+  return input;
+}
+
+const result arm_validate_utf32le_with_errors(const char32_t *input,
+                                              size_t size) {
+  const char32_t *start = input;
+  const char32_t *end = input + size;
+
+  const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
+  const uint32x4_t offset = vmovq_n_u32(0xffff2000);
+  const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
+  uint32x4_t currentmax = vmovq_n_u32(0x0);
+  uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);
+
+  while (end - input >= 4) {
+    const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input));
+    currentmax = vmaxq_u32(in, currentmax);
+    currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);
+
+    uint32x4_t is_zero =
+        veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
+    if (vmaxvq_u32(is_zero) != 0) {
+      return result(error_code::TOO_LARGE, input - start);
+    }
+
+    is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax),
+                        standardoffsetmax);
+    if (vmaxvq_u32(is_zero) != 0) {
+      return result(error_code::SURROGATE, input - start);
+    }
+
+    input += 4;
+  }
+
+  return result(error_code::SUCCESS, input - start);
+}
+/* end file src/arm64/arm_validate_utf32le.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_latin1_to_utf16.cpp */
+template <endianness big_endian>
+std::pair<const char *, char16_t *>
+arm_convert_latin1_to_utf16(const char *buf, size_t len,
+                            char16_t *utf16_output) {
+  const char *end = buf + len;
+
+  while (end - buf >= 16) {
+    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(buf));
+    uint16x8_t inlow = vmovl_u8(vget_low_u8(in8));
+    if (!match_system(big_endian)) {
+      inlow = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(inlow)));
+    }
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), inlow);
+    uint16x8_t inhigh = vmovl_u8(vget_high_u8(in8));
+    if (!match_system(big_endian)) {
+      inhigh = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(inhigh)));
+    }
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output + 8), inhigh);
+    utf16_output += 16;
+    buf += 16;
+  }
+
+  return std::make_pair(buf, utf16_output);
+}
+/* end file src/arm64/arm_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_latin1_to_utf32.cpp */
+std::pair<const char *, char32_t *>
+arm_convert_latin1_to_utf32(const char *buf, size_t len,
+                            char32_t *utf32_output) {
+  const char *end = buf + len;
+
+  while (end - buf >= 16) {
+    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(buf));
+    uint16x8_t in8low = vmovl_u8(vget_low_u8(in8));
+    uint32x4_t in16lowlow = vmovl_u16(vget_low_u16(in8low));
+    uint32x4_t in16lowhigh = vmovl_u16(vget_high_u16(in8low));
+    uint16x8_t in8high = vmovl_u8(vget_high_u8(in8));
+    uint32x4_t in8highlow = vmovl_u16(vget_low_u16(in8high));
+    uint32x4_t in8highhigh = vmovl_u16(vget_high_u16(in8high));
+    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output), in16lowlow);
+    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 4), in16lowhigh);
+    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 8), in8highlow);
+    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 12), in8highhigh);
+
+    utf32_output += 16;
+    buf += 16;
+  }
+
+  return std::make_pair(buf, utf32_output);
+}
+/* end file src/arm64/arm_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_latin1_to_utf8.cpp */
+/*
+  Returns a pair: the first unprocessed byte from buf and utf8_output
+  A scalar routing should carry on the conversion of the tail.
+*/
+std::pair<const char *, char *>
+arm_convert_latin1_to_utf8(const char *latin1_input, size_t len,
+                           char *utf8_out) {
+  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
+  const char *end = latin1_input + len;
+  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
+  // We always write 16 bytes, of which more than the first 8 bytes
+  // are valid. A safety margin of 8 is more than sufficient.
+  while (end - latin1_input >= 16 + 8) {
+    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(latin1_input));
+    if (vmaxvq_u8(in8) <= 0x7F) { // ASCII fast path!!!!
+      vst1q_u8(utf8_output, in8);
+      utf8_output += 16;
+      latin1_input += 16;
+      continue;
+    }
+
+    // We just fallback on UTF-16 code. This could be optimized/simplified
+    // further.
+    uint16x8_t in16 = vmovl_u8(vget_low_u8(in8));
+    // 1. prepare 2-byte values
+    // input 8-bit word : [aabb|bbbb] x 8
+    // expected output   : [1100|00aa|10bb|bbbb] x 8
+    const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
+    const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+
+    // t0 = [0000|00aa|bbbb|bb00]
+    const uint16x8_t t0 = vshlq_n_u16(in16, 2);
+    // t1 = [0000|00aa|0000|0000]
+    const uint16x8_t t1 = vandq_u16(t0, v_1f00);
+    // t2 = [0000|0000|00bb|bbbb]
+    const uint16x8_t t2 = vandq_u16(in16, v_003f);
+    // t3 = [0000|00aa|00bb|bbbb]
+    const uint16x8_t t3 = vorrq_u16(t1, t2);
+    // t4 = [1100|00aa|10bb|bbbb]
+    const uint16x8_t t4 = vorrq_u16(t3, v_c080);
+    // 2. merge ASCII and 2-byte codewords
+    const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+    const uint16x8_t one_byte_bytemask = vcleq_u16(in16, v_007f);
+    const uint8x16_t utf8_unpacked =
+        vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in16, t4));
+    // 3. prepare bitmask for 8-bit lookup
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+    const uint16x8_t mask = simdutf_make_uint16x8_t(
+        0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
+#else
+    const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
+                             0x0002, 0x0008, 0x0020, 0x0080};
+#endif
+    uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
+    // 4. pack the bytes
+    const uint8_t *row =
+        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+    const uint8x16_t shuffle = vld1q_u8(row + 1);
+    const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+
+    // 5. store bytes
+    vst1q_u8(utf8_output, utf8_packed);
+    // 6. adjust pointers
+    latin1_input += 8;
+    utf8_output += row[0];
+
+  } // while
+
+  return std::make_pair(latin1_input, reinterpret_cast<char *>(utf8_output));
+}
+/* end file src/arm64/arm_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_utf8_to_latin1.cpp */
+// Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 16, usually 12).
+size_t convert_masked_utf8_to_latin1(const char *input,
+                                     uint64_t utf8_end_of_code_point_mask,
+                                     char *&latin1_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+
+  // We first try a few fast paths.
+  // The obvious first test is ASCII, which actually consumes the full 16.
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process in chunks of 12 bytes
+    vst1q_u8(reinterpret_cast<uint8_t *>(latin1_output), in);
+    latin1_output += 12; // We wrote 12 18-bit characters.
+    return 12;           // We consumed 12 bytes.
+  }
+  /// We do not have a fast path available, or the fast path is unimportant, so
+  /// we fallback.
+  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][0];
+
+  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][1];
+  // this indicates an invalid input:
+  if (idx >= 64) {
+    return consumed;
+  }
+  // Here we should have (idx < 64), if not, there is a bug in the validation or
+  // elsewhere. SIX (6) input code-code units this is a relatively easy scenario
+  // we process SIX (6) input code-code units. The max length in bytes of six
+  // code code units spanning between 1 and 2 bytes each is 12 bytes. Converts 6
+  // 1-2 byte UTF-8 characters to 6 UTF-16 characters. This is a relatively easy
+  // scenario we process SIX (6) input code-code units. The max length in bytes
+  // of six code code units spanning between 1 and 2 bytes each is 12 bytes.
+  uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+      simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+  // Shuffle
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 110aaaaa 10bbbbbb
+  uint16x8_t perm = vreinterpretq_u16_u8(vqtbl1q_u8(in, sh));
+  // Mask
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 00000000 00bbbbbb
+  uint16x8_t ascii = vandq_u16(perm, vmovq_n_u16(0x7f)); // 6 or 7 bits
+  // 1 byte: 00000000 00000000
+  // 2 byte: 000aaaaa 00000000
+  uint16x8_t highbyte = vandq_u16(perm, vmovq_n_u16(0x1f00)); // 5 bits
+  // Combine with a shift right accumulate
+  // 1 byte: 00000000 0bbbbbbb
+  // 2 byte: 00000aaa aabbbbbb
+  uint16x8_t composed = vsraq_n_u16(ascii, highbyte, 2);
+  // writing 8 bytes even though we only care about the first 6 bytes.
+  uint8x8_t latin1_packed = vmovn_u16(composed);
+  vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
+  latin1_output += 6; // We wrote 6 bytes.
+  return consumed;
+}
+/* end file src/arm64/arm_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/arm64/arm_convert_utf8_to_utf16.cpp */
+// Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 16, usually 12).
+template <endianness big_endian>
+size_t convert_masked_utf8_to_utf16(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char16_t *&utf16_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+
+  // We first try a few fast paths.
+  // The obvious first test is ASCII, which actually consumes the full 16.
+  if ((utf8_end_of_code_point_mask & 0xFFFF) == 0xffff) {
+    // We process in chunks of 16 bytes
+    // The routine in simd.h is reused.
+    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
+    temp.store_ascii_as_utf16<big_endian>(utf16_output);
+    utf16_output += 16; // We wrote 16 16-bit characters.
+    return 16;          // We consumed 16 bytes.
+  }
+
+  // 3 byte sequences are the next most common, as seen in CJK, which has long
+  // sequences of these.
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte
+    // UTF-16 code units.
+    uint16x4_t composed = convert_utf8_3_byte_to_utf16(in);
+    // Byte swap if necessary
+    if (!match_system(big_endian)) {
+      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+    }
+    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+    utf16_output += 4; // We wrote 4 16-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
+  if ((utf8_end_of_code_point_mask & 0xFFF) == 0xaaa) {
+    // We want to take 6 2-byte UTF-8 code units and turn them into 6 2-byte
+    // UTF-16 code units.
+    uint16x8_t composed = convert_utf8_2_byte_to_utf16(in);
+    // Byte swap if necessary
+    if (!match_system(big_endian)) {
+      composed =
+          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+    }
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+
+    utf16_output += 6; // We wrote 6 16-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  /// We do not have a fast path available, or the fast path is unimportant, so
+  /// we fallback.
+  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][0];
+
+  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][1];
+
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // Convert to UTF-16
+    uint16x8_t composed = convert_utf8_1_to_2_byte_to_utf16(in, idx);
+    // Byte swap if necessary
+    if (!match_system(big_endian)) {
+      composed =
+          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+    }
+    // Store
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+    utf16_output += 6; // We wrote 6 16-bit characters.
+    return consumed;
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // XXX: depending on the system scalar instructions might be faster.
+    // 1 byte: 00000000 00000000 0ccccccc
+    // 2 byte: 00000000 110bbbbb 10cccccc
+    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: xx0bbbbb x0cccccc
+    // 3 byte: xxbbbbbb x0cccccc
+    uint16x4_t lowperm = vmovn_u32(perm);
+    // Partially mask with bic (doesn't require a temporary register unlike and)
+    // The shift left insert below will clear the top bits.
+    // 1 byte: 00000000 00000000
+    // 2 byte: xx0bbbbb 00000000
+    // 3 byte: xxbbbbbb 00000000
+    uint16x4_t middlebyte = vbic_u16(lowperm, vmov_n_u16(uint16_t(~0xFF00)));
+    // ASCII
+    // 1 byte: 00000000 0ccccccc
+    // 2+byte: 00000000 00cccccc
+    uint16x4_t ascii = vand_u16(lowperm, vmov_n_u16(0x7F));
+    // Split into narrow vectors.
+    // 2 byte: 00000000 00000000
+    // 3 byte: 00000000 xxxxaaaa
+    uint16x4_t highperm = vshrn_n_u32(perm, 16);
+    // Shift right accumulate the middle byte
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: 00xx0bbb bbcccccc
+    // 3 byte: 00xxbbbb bbcccccc
+    uint16x4_t middlelow = vsra_n_u16(ascii, middlebyte, 2);
+    // Shift left and insert the top 4 bits, overwriting the garbage
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: 00000bbb bbcccccc
+    // 3 byte: aaaabbbb bbcccccc
+    uint16x4_t composed = vsli_n_u16(middlelow, highperm, 12);
+    // Byte swap if necessary
+    if (!match_system(big_endian)) {
+      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+    }
+    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+
+    utf16_output += 4; // We wrote 4 16-bit codepoints
+    return consumed;
+  } else if (idx < 209) {
+    // THREE (3) input code-code units
+    if (input_utf8_end_of_code_point_mask == 0x888) {
+      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
+      // UTF-16 pairs. Generating surrogate pairs is a little tricky though, but
+      // it is easier when we can assume they are all pairs. This version does
+      // not use the LUT, but 4 byte sequences are less common and the overhead
+      // of the extra memory access is less important than the early branch
+      // overhead in shorter sequences.
+
+      // Swap byte pairs
+      // 10dddddd 10cccccc|10bbbbbb 11110aaa
+      // 10cccccc 10dddddd|11110aaa 10bbbbbb
+      uint8x16_t swap = vrev16q_u8(in);
+      // Shift left 2 bits
+      // cccccc00 dddddd00 xxxxxxxx bbbbbb00
+      uint32x4_t shift = vreinterpretq_u32_u8(vshlq_n_u8(swap, 2));
+      // Create a magic number containing the low 2 bits of the trail surrogate
+      // and all the corrections needed to create the pair. UTF-8 4b prefix   =
+      // -0x0000|0xF000 surrogate offset  = -0x0000|0x0040 (0x10000 << 6)
+      // surrogate high    = +0x0000|0xD800
+      // surrogate low     = +0xDC00|0x0000
+      // -------------------------------
+      //                   = +0xDC00|0xE7C0
+      uint32x4_t magic = vmovq_n_u32(0xDC00E7C0);
+      // Generate unadjusted trail surrogate minus lowest 2 bits
+      // xxxxxxxx xxxxxxxx|11110aaa bbbbbb00
+      uint32x4_t trail =
+          vbslq_u32(vmovq_n_u32(0x0000FF00), vreinterpretq_u32_u8(swap), shift);
+      // Insert low 2 bits of trail surrogate to magic number for later
+      // 11011100 00000000 11100111 110000cc
+      uint16x8_t magic_with_low_2 =
+          vreinterpretq_u16_u32(vsraq_n_u32(magic, shift, 30));
+      // Generate lead surrogate
+      // xxxxcccc ccdddddd|xxxxxxxx xxxxxxxx
+      uint32x4_t lead = vreinterpretq_u32_u16(
+          vsliq_n_u16(vreinterpretq_u16_u8(swap), vreinterpretq_u16_u8(in), 6));
+      // Mask out lead
+      // 000000cc ccdddddd|xxxxxxxx xxxxxxxx
+      lead = vbicq_u32(lead, vmovq_n_u32(uint32_t(~0x03FFFFFF)));
+      // Blend pairs
+      // 000000cc ccdddddd|11110aaa bbbbbb00
+      uint16x8_t blend = vreinterpretq_u16_u32(
+          vbslq_u32(vmovq_n_u32(0x0000FFFF), trail, lead));
+      // Add magic number to finish the result
+      // 110111CC CCDDDDDD|110110AA BBBBBBCC
+      uint16x8_t composed = vaddq_u16(blend, magic_with_low_2);
+      // Byte swap if necessary
+      if (!match_system(big_endian)) {
+        composed =
+            vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+      }
+      uint16_t buffer[8];
+      vst1q_u16(reinterpret_cast<uint16_t *>(buffer), composed);
+      for (int k = 0; k < 6; k++) {
+        utf16_output[k] = buffer[k];
+      } // the loop might compiler to a couple of instructions.
+      // We need some validation. See
+      // https://github.com/simdutf/simdutf/pull/631
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      uint8x16_t expected_mask = simdutf_make_uint8x16_t(
+          0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0,
+          0xc0, 0x0, 0x0, 0x0, 0x0);
+#else
+      uint8x16_t expected_mask = {0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0,
+                                  0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0,
+                                  0x0,  0x0,  0x0,  0x0};
+#endif
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      uint8x16_t expected = simdutf_make_uint8x16_t(
+          0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80,
+          0x80, 0x0, 0x0, 0x0, 0x0);
+#else
+      uint8x16_t expected = {0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80,
+                             0xf0, 0x80, 0x80, 0x80, 0x0,  0x0,  0x0,  0x0};
+#endif
+      uint8x16_t check = vceqq_u8(vandq_u8(in, expected_mask), expected);
+      bool correct = (vminvq_u32(vreinterpretq_u32_u8(check)) == 0xFFFFFFFF);
+      // The validation is just three instructions and it is not on a critical
+      // path.
+      if (correct) {
+        utf16_output += 6; // We wrote 3 32-bit surrogate pairs.
+      }
+      return 12; // We consumed 12 bytes.
+    }
+    // 3 1-4 byte sequences
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+
+    // 1 byte: 00000000 00000000 00000000 0ddddddd
+    // 3 byte: 00000000 00000000 110ccccc 10dddddd
+    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
+    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // added to fix issue https://github.com/simdutf/simdutf/issues/514
+    // We only want to write 2 * 16-bit code units when that is actually what we
+    // have. Unfortunately, we cannot trust the input. So it is possible to get
+    // 0xff as an input byte and it should not result in a surrogate pair. We
+    // need to check for that.
+    uint32_t permbuffer[4];
+    vst1q_u32(permbuffer, perm);
+    // Mask the low and middle bytes
+    // 00000000 00000000 00000000 0ddddddd
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7f));
+    // Because the surrogates need more work, the high surrogate is computed
+    // first.
+    uint32x4_t middlehigh = vshlq_n_u32(perm, 2);
+    // 00000000 00000000 00cccccc 00000000
+    uint32x4_t middlebyte = vandq_u32(perm, vmovq_n_u32(0x3F00));
+    // Start assembling the sequence. Since the 4th byte is in the same position
+    // as it would be in a surrogate and there is no dependency, shift left
+    // instead of right. 3 byte: 00000000 10bbbbxx xxxxxxxx xxxxxxxx 4 byte:
+    // 11110aaa bbbbbbxx xxxxxxxx xxxxxxxx
+    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0xFF000000), perm, middlehigh);
+    // Top 16 bits contains the high ten bits of the surrogate pair before
+    // correction 3 byte: 00000000 10bbbbcc|cccc0000 00000000 4 byte: 11110aaa
+    // bbbbbbcc|cccc0000 00000000 - high 10 bits correct w/o correction
+    uint32x4_t abc =
+        vbslq_u32(vmovq_n_u32(0xFFFC0000), ab, vshlq_n_u32(middlebyte, 4));
+    // Combine the low 6 or 7 bits by a shift right accumulate
+    // 3 byte: 00000000 00000010|bbbbcccc ccdddddd - low 16 bits correct
+    // 4 byte: 00000011 110aaabb|bbbbcccc ccdddddd - low 10 bits correct w/o
+    // correction
+    uint32x4_t composed = vsraq_n_u32(ascii, abc, 6);
+    // After this is for surrogates
+    // Blend the low and high surrogates
+    // 4 byte: 11110aaa bbbbbbcc|bbbbcccc ccdddddd
+    uint32x4_t mixed = vbslq_u32(vmovq_n_u32(0xFFFF0000), abc, composed);
+    // Clear the upper 6 bits of the low surrogate. Don't clear the upper bits
+    // yet as 0x10000 was not subtracted from the codepoint yet. 4 byte:
+    // 11110aaa bbbbbbcc|000000cc ccdddddd
+    uint16x8_t masked_pair = vreinterpretq_u16_u32(
+        vbicq_u32(mixed, vmovq_n_u32(uint32_t(~0xFFFF03FF))));
+    // Correct the remaining UTF-8 prefix, surrogate offset, and add the
+    // surrogate prefixes in one magic 16-bit addition. similar magic number but
+    // without the continue byte adjust and halfword swapped UTF-8 4b prefix   =
+    // -0xF000|0x0000 surrogate offset  = -0x0040|0x0000 (0x10000 << 6)
+    // surrogate high    = +0xD800|0x0000
+    // surrogate low     = +0x0000|0xDC00
+    // -----------------------------------
+    //                   = +0xE7C0|0xDC00
+    uint16x8_t magic = vreinterpretq_u16_u32(vmovq_n_u32(0xE7C0DC00));
+    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD - surrogate pair complete
+    uint32x4_t surrogates =
+        vreinterpretq_u32_u16(vaddq_u16(masked_pair, magic));
+    // If the high bit is 1 (s32 less than zero), this needs a surrogate pair
+    uint32x4_t is_pair = vcltzq_s32(vreinterpretq_s32_u32(perm));
+
+    // Select either the 4 byte surrogate pair or the 2 byte solo codepoint
+    // 3 byte: 0xxxxxxx xxxxxxxx|bbbbcccc ccdddddd
+    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD
+    uint32x4_t selected = vbslq_u32(is_pair, surrogates, composed);
+    // Byte swap if necessary
+    if (!match_system(big_endian)) {
+      selected =
+          vreinterpretq_u32_u8(vrev16q_u8(vreinterpretq_u8_u32(selected)));
+    }
+    // Attempting to shuffle and store would be complex, just scalarize.
+    uint32_t buffer[4];
+    vst1q_u32(buffer, selected);
+    // Test for the top bit of the surrogate mask. Remove due to issue 514
+    // const uint32_t SURROGATE_MASK = match_system(big_endian) ? 0x80000000 :
+    // 0x00800000;
+    for (size_t i = 0; i < 3; i++) {
+      // Surrogate
+      // Used to be if (buffer[i] & SURROGATE_MASK) {
+      // See discussion above.
+      // patch for issue https://github.com/simdutf/simdutf/issues/514
+      if ((permbuffer[i] & 0xf8000000) == 0xf0000000) {
+        utf16_output[0] = uint16_t(buffer[i] >> 16);
+        utf16_output[1] = uint16_t(buffer[i] & 0xFFFF);
+        utf16_output += 2;
+      } else {
+        utf16_output[0] = uint16_t(buffer[i] & 0xFFFF);
+        utf16_output++;
+      }
+    }
+    return consumed;
+  } else {
+    // here we know that there is an error but we do not handle errors
+    return 12;
+  }
+}
+/* end file src/arm64/arm_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/arm64/arm_convert_utf8_to_utf32.cpp */
+// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
+size_t convert_masked_utf8_to_utf32(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char32_t *&utf32_out) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  uint32_t *&utf32_output = reinterpret_cast<uint32_t *&>(utf32_out);
+  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xFFF;
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  // We first try a few fast paths.
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process in chunks of 12 bytes.
+    // use fast implementation in src/simdutf/arm64/simd.h
+    // Ideally the compiler can keep the tables in registers.
+    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
+    temp.store_ascii_as_utf32_tbl(utf32_out);
+    utf32_output += 12; // We wrote 12 32-bit characters.
+    return 12;          // We consumed 12 bytes.
+  }
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte
+    // UTF-32 code units. Convert to UTF-16
+    uint16x4_t composed_utf16 = convert_utf8_3_byte_to_utf16(in);
+    // Zero extend and store via ST2 with a zero.
+    uint16x4x2_t interleaver = {{composed_utf16, vmov_n_u16(0)}};
+    vst2_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 4; // We wrote 4 32-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
+  if (input_utf8_end_of_code_point_mask == 0xaaa) {
+    // We want to take 6 2-byte UTF-8 code units and turn them into 6 4-byte
+    // UTF-32 code units. Convert to UTF-16
+    uint16x8_t composed_utf16 = convert_utf8_2_byte_to_utf16(in);
+    // Zero extend and store via ST2 with a zero.
+    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 6; // We wrote 6 32-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+  /// Either no fast path or an unimportant fast path.
+
+  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][1];
+
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // Convert to UTF-16
+    uint16x8_t composed_utf16 = convert_utf8_1_to_2_byte_to_utf16(in, idx);
+    // Zero extend and store with ST2 and zero
+    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 6; // We wrote 6 32-bit characters.
+    return consumed;
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // Shuffle
+    // 1 byte: 00000000 00000000 0ccccccc
+    // 2 byte: 00000000 110bbbbb 10cccccc
+    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // Split
+    // 00000000 00000000 0ccccccc
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F)); // 6 or 7 bits
+    // Note: unmasked
+    // xxxxxxxx aaaaxxxx xxxxxxxx
+    uint32x4_t high = vshrq_n_u32(perm, 4); // 4 bits
+    // Use 16 bit bic instead of and.
+    // The top bits will be corrected later in the bsl
+    // 00000000 10bbbbbb 00000000
+    uint32x4_t middle = vreinterpretq_u32_u16(
+        vbicq_u16(vreinterpretq_u16_u32(perm),
+                  vmovq_n_u16(uint16_t(~0xff00)))); // 5 or 6 bits
+    // Combine low and middle with shift right accumulate
+    // 00000000 00xxbbbb bbcccccc
+    uint32x4_t lowmid = vsraq_n_u32(ascii, middle, 2);
+    // Insert top 4 bits from high byte with bitwise select
+    // 00000000 aaaabbbb bbcccccc
+    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0x0000F000), high, lowmid);
+    vst1q_u32(utf32_output, composed);
+    utf32_output += 4; // We wrote 4 32-bit characters.
+    return consumed;
+  } else if (idx < 209) {
+    // THREE (3) input code-code units
+    if (input_utf8_end_of_code_point_mask == 0x888) {
+      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
+      // UTF-32 code units. This uses the same method as the fixed 3 byte
+      // version, reversing and shift left insert. However, there is no need for
+      // a shuffle mask now, just rev16 and rev32.
+      //
+      // This version does not use the LUT, but 4 byte sequences are less common
+      // and the overhead of the extra memory access is less important than the
+      // early branch overhead in shorter sequences, so it comes last.
+
+      // Swap pairs of bytes
+      // 10dddddd|10cccccc|10bbbbbb|11110aaa
+      // 10cccccc 10dddddd|11110aaa 10bbbbbb
+      uint16x8_t swap1 = vreinterpretq_u16_u8(vrev16q_u8(in));
+      // Shift left and insert
+      // xxxxcccc ccdddddd|xxxxxxxa aabbbbbb
+      uint16x8_t merge1 = vsliq_n_u16(swap1, vreinterpretq_u16_u8(in), 6);
+      // Swap 16-bit lanes
+      // xxxxcccc ccdddddd xxxxxxxa aabbbbbb
+      // xxxxxxxa aabbbbbb xxxxcccc ccdddddd
+      uint32x4_t swap2 = vreinterpretq_u32_u16(vrev32q_u16(merge1));
+      // Shift insert again
+      // xxxxxxxx xxxaaabb bbbbcccc ccdddddd
+      uint32x4_t merge2 = vsliq_n_u32(swap2, vreinterpretq_u32_u16(merge1), 12);
+      // Clear the garbage
+      // 00000000 000aaabb bbbbcccc ccdddddd
+      uint32x4_t composed = vandq_u32(merge2, vmovq_n_u32(0x1FFFFF));
+      // Store
+      vst1q_u32(utf32_output, composed);
+
+      utf32_output += 3; // We wrote 3 32-bit characters.
+      return 12;         // We consumed 12 bytes.
+    }
+    // Unlike UTF-16, doing a fast codepath doesn't have nearly as much benefit
+    // due to surrogates no longer being involved.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // 1 byte: 00000000 00000000 00000000 0ddddddd
+    // 2 byte: 00000000 00000000 110ccccc 10dddddd
+    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
+    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // Ascii
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F));
+    uint32x4_t middle = vandq_u32(perm, vmovq_n_u32(0x3f00));
+    // When converting the way we do, the 3 byte prefix will be interpreted as
+    // the 18th bit being set, since the code would interpret the lead byte
+    // (0b1110bbbb) as a continuation byte (0b10bbbbbb). To fix this, we can
+    // either xor or do an 8 bit add of the 6th bit shifted right by 1. Since
+    // NEON has shift right accumulate, we use that.
+    //  4 byte   3 byte
+    // 10bbbbbb 1110bbbb
+    // 00000000 01000000 6th bit
+    // 00000000 00100000 shift right
+    // 10bbbbbb 0000bbbb add
+    // 00bbbbbb 0000bbbb mask
+    uint8x16_t correction =
+        vreinterpretq_u8_u32(vandq_u32(perm, vmovq_n_u32(0x00400000)));
+    uint32x4_t corrected = vreinterpretq_u32_u8(
+        vsraq_n_u8(vreinterpretq_u8_u32(perm), correction, 1));
+    // 00000000 00000000 0000cccc ccdddddd
+    uint32x4_t cd = vsraq_n_u32(ascii, middle, 2);
+    // Insert twice
+    // xxxxxxxx xxxaaabb bbbbxxxx xxxxxxxx
+    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0x01C0000), vshrq_n_u32(corrected, 6),
+                              vshrq_n_u32(corrected, 4));
+    // 00000000 000aaabb bbbbcccc ccdddddd
+    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0xFFE00FFF), cd, ab);
+    // Store
+    vst1q_u32(utf32_output, composed);
+    utf32_output += 3; // We wrote 3 32-bit characters.
+    return consumed;
+  } else {
+    // here we know that there is an error but we do not handle errors
+    return 12;
+  }
+}
+/* end file src/arm64/arm_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_utf16_to_latin1.cpp */
+
+template <endianness big_endian>
+std::pair<const char16_t *, char *>
+arm_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                            char *latin1_output) {
+  const char16_t *end = buf + len;
+  while (end - buf >= 8) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+    if (vmaxvq_u16(in) <= 0xff) {
+      // 1. pack the bytes
+      uint8x8_t latin1_packed = vmovn_u16(in);
+      // 2. store (8 bytes)
+      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
+      // 3. adjust pointers
+      buf += 8;
+      latin1_output += 8;
+    } else {
+      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
+    }
+  } // while
+  return std::make_pair(buf, latin1_output);
+}
+
+template <endianness big_endian>
+std::pair<result, char *>
+arm_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
+                                        char *latin1_output) {
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+  while (end - buf >= 8) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+    if (vmaxvq_u16(in) <= 0xff) {
+      // 1. pack the bytes
+      uint8x8_t latin1_packed = vmovn_u16(in);
+      // 2. store (8 bytes)
+      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
+      // 3. adjust pointers
+      buf += 8;
+      latin1_output += 8;
+    } else {
+      // Let us do a scalar fallback.
+      for (int k = 0; k < 8; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if (word <= 0xff) {
+          *latin1_output++ = char(word);
+        } else {
+          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
+                                latin1_output);
+        }
+      }
+    }
+  } // while
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        latin1_output);
+}
+/* end file src/arm64/arm_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/arm64/arm_convert_utf16_to_utf32.cpp */
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
+
+    We consider three cases:
+    1. an input register contains no surrogates and each value
+       is in range 0x0000 .. 0x07ff.
+    2. an input register contains no surrogates and values are
+       is in range 0x0000 .. 0xffff.
+    3. an input register contains surrogates --- i.e. codepoints
+       can have 16 or 32 bits.
+
+    Ad 1.
+
+    When values are less than 0x0800, it means that a 16-bit code unit
+    can be converted into: 1) single UTF8 byte (when it is an ASCII
+    char) or 2) two UTF8 bytes.
+
+    For this case we do only some shuffle to obtain these 2-byte
+    codes and finally compress the whole SSE register with a single
+    shuffle.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+    Ad 2.
+
+    When values fit in 16-bit code units, but are above 0x07ff, then
+    a single word may produce one, two or three UTF8 bytes.
+
+    We prepare data for all these three cases in two registers.
+    The first register contains lower two UTF8 bytes (used in all
+    cases), while the second one contains just the third byte for
+    the three-UTF8-bytes case.
+
+    Finally these two registers are interleaved forming eight-element
+    array of 32-bit values. The array spans two SSE registers.
+    The bytes from the registers are compressed using two shuffles.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+
+    To summarize:
+    - We need two 256-entry tables that have 8704 bytes in total.
+*/
+/*
+  Returns a pair: the first unprocessed byte from buf and utf8_output
+  A scalar routing should carry on the conversion of the tail.
+*/
+template <endianness big_endian>
+std::pair<const char16_t *, char32_t *>
+arm_convert_utf16_to_utf32(const char16_t *buf, size_t len,
+                           char32_t *utf32_out) {
+  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
+  const char16_t *end = buf + len;
+
+  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
+  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+
+  while (end - buf >= 8) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+
+    const uint16x8_t surrogates_bytemask =
+        vceqq_u16(vandq_u16(in, v_f800), v_d800);
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (vmaxvq_u16(surrogates_bytemask) == 0) {
+      // case: no surrogate pairs, extend all 16-bit code units to 32-bit code
+      // units
+      vst1q_u32(utf32_output, vmovl_u16(vget_low_u16(in)));
+      vst1q_u32(utf32_output + 4, vmovl_high_u16(in));
+      utf32_output += 8;
+      buf += 8;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xF800) != 0xD800) {
+          *utf32_output++ = char32_t(word);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = !match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char32_t *>(utf32_output));
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf32_output++ = char32_t(value);
+        }
+      }
+      buf += k;
+    }
+  } // while
+  return std::make_pair(buf, reinterpret_cast<char32_t *>(utf32_output));
+}
+
+/*
+  Returns a pair: a result struct and utf8_output.
+  If there is an error, the count field of the result is the position of the
+  error. Otherwise, it is the position of the first unprocessed byte in buf
+  (even if finished). A scalar routing should carry on the conversion of the
+  tail if needed.
+*/
+template <endianness big_endian>
+std::pair<result, char32_t *>
+arm_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
+                                       char32_t *utf32_out) {
+  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+
+  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
+  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+
+  while ((end - buf) >= 8) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+
+    const uint16x8_t surrogates_bytemask =
+        vceqq_u16(vandq_u16(in, v_f800), v_d800);
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (vmaxvq_u16(surrogates_bytemask) == 0) {
+      // case: no surrogate pairs, extend all 16-bit code units to 32-bit code
+      // units
+      vst1q_u32(utf32_output, vmovl_u16(vget_low_u16(in)));
+      vst1q_u32(utf32_output + 4, vmovl_high_u16(in));
+      utf32_output += 8;
+      buf += 8;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xF800) != 0xD800) {
+          *utf32_output++ = char32_t(word);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = !match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k - 1),
+                reinterpret_cast<char32_t *>(utf32_output));
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf32_output++ = char32_t(value);
+        }
+      }
+      buf += k;
+    }
+  } // while
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        reinterpret_cast<char32_t *>(utf32_output));
+}
+/* end file src/arm64/arm_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
+/* begin file src/arm64/arm_convert_utf16_to_utf8.cpp */
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
+
+    We consider three cases:
+    1. an input register contains no surrogates and each value
+       is in range 0x0000 .. 0x07ff.
+    2. an input register contains no surrogates and values are
+       is in range 0x0000 .. 0xffff.
+    3. an input register contains surrogates --- i.e. codepoints
+       can have 16 or 32 bits.
+
+    Ad 1.
+
+    When values are less than 0x0800, it means that a 16-bit code unit
+    can be converted into: 1) single UTF8 byte (when it is an ASCII
+    char) or 2) two UTF8 bytes.
+
+    For this case we do only some shuffle to obtain these 2-byte
+    codes and finally compress the whole SSE register with a single
+    shuffle.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+    Ad 2.
+
+    When values fit in 16-bit code units, but are above 0x07ff, then
+    a single word may produce one, two or three UTF8 bytes.
+
+    We prepare data for all these three cases in two registers.
+    The first register contains lower two UTF8 bytes (used in all
+    cases), while the second one contains just the third byte for
+    the three-UTF8-bytes case.
+
+    Finally these two registers are interleaved forming eight-element
+    array of 32-bit values. The array spans two SSE registers.
+    The bytes from the registers are compressed using two shuffles.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+
+    To summarize:
+    - We need two 256-entry tables that have 8704 bytes in total.
+*/
+/*
+  Returns a pair: the first unprocessed byte from buf and utf8_output
+  A scalar routing should carry on the conversion of the tail.
+*/
+template <endianness big_endian>
+std::pair<const char16_t *, char *>
+arm_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
+  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
+  const char16_t *end = buf + len;
+
+  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
+  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+    if (vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
+      // It is common enough that we have sequences of 16 consecutive ASCII
+      // characters.
+      uint16x8_t nextin =
+          vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
+      if (!match_system(big_endian)) {
+        nextin = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(nextin)));
+      }
+      if (vmaxvq_u16(nextin) > 0x7F) {
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x8_t utf8_packed = vmovn_u16(in);
+        // 2. store (8 bytes)
+        vst1_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 8;
+        utf8_output += 8;
+        in = nextin;
+      } else {
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
+        // 2. store (16 bytes)
+        vst1q_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 16;
+        utf8_output += 16;
+        continue; // we are done for this round!
+      }
+    }
+
+    if (vmaxvq_u16(in) <= 0x7FF) {
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
+      const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const uint16x8_t t0 = vshlq_n_u16(in, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const uint16x8_t t1 = vandq_u16(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const uint16x8_t t2 = vandq_u16(in, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const uint16x8_t t3 = vorrq_u16(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const uint16x8_t t4 = vorrq_u16(t3, v_c080);
+      // 2. merge ASCII and 2-byte codewords
+      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
+      const uint8x16_t utf8_unpacked =
+          vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
+      // 3. prepare bitmask for 8-bit lookup
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t mask = simdutf_make_uint16x8_t(
+          0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
+#else
+      const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
+                               0x0002, 0x0008, 0x0020, 0x0080};
+#endif
+      uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
+      // 4. pack the bytes
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+      const uint8x16_t shuffle = vld1q_u8(row + 1);
+      const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+
+      // 5. store bytes
+      vst1q_u8(utf8_output, utf8_packed);
+
+      // 6. adjust pointers
+      buf += 8;
+      utf8_output += row[0];
+      continue;
+    }
+    const uint16x8_t surrogates_bytemask =
+        vceqq_u16(vandq_u16(in, v_f800), v_d800);
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (vmaxvq_u16(surrogates_bytemask) == 0) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t dup_even = simdutf_make_uint16x8_t(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#else
+      const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
+                                   0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
+#endif
+      /* In this branch we handle three cases:
+         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
+
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
+
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
+
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
+
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const uint16x8_t t0 = vreinterpretq_u16_u8(
+          vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+
+      // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
+      const uint16x8_t s0 = vshrq_n_u16(in, 12);
+      // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
+      const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
+      // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
+      const uint16x8_t s1s = vshlq_n_u16(s1, 2);
+      // [00bb|bbbb|0000|aaaa]
+      const uint16x8_t s2 = vorrq_u16(s0, s1s);
+      // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
+      const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
+      const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
+      const uint16x8_t m0 =
+          vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
+      const uint16x8_t s4 = veorq_u16(s3, m0);
+#undef simdutf_vec
+
+      // 4. expand code units 16-bit => 32-bit
+      const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
+      const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t onemask = simdutf_make_uint16x8_t(
+          0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
+      const uint16x8_t twomask = simdutf_make_uint16x8_t(
+          0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
+#else
+      const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                  0x0100, 0x0400, 0x1000, 0x4000};
+      const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
+                                  0x0200, 0x0800, 0x2000, 0x8000};
+#endif
+      const uint16x8_t combined =
+          vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
+                    vandq_u16(one_or_two_bytes_bytemask, twomask));
+      const uint16_t mask = vaddvq_u16(combined);
+      // The following fast path may or may not be beneficial.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
+        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
+        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
+        vst1q_u8(utf8_output, utf8_0);
+        utf8_output += 12;
+        vst1q_u8(utf8_output, utf8_1);
+        utf8_output += 12;
+        buf += 8;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
+      const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
+      const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+
+      vst1q_u8(utf8_output, utf8_0);
+      utf8_output += row0[0];
+      vst1q_u8(utf8_output, utf8_1);
+      utf8_output += row1[0];
+
+      buf += 8;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xF800) != 0xD800) {
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = !match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char *>(utf8_output));
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf8_output++ = char((value >> 18) | 0b11110000);
+          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((value & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
+    }
+  } // while
+
+  return std::make_pair(buf, reinterpret_cast<char *>(utf8_output));
+}
+
+/*
+  Returns a pair: a result struct and utf8_output.
+  If there is an error, the count field of the result is the position of the
+  error. Otherwise, it is the position of the first unprocessed byte in buf
+  (even if finished). A scalar routing should carry on the conversion of the
+  tail if needed.
+*/
+template <endianness big_endian>
+std::pair<result, char *>
+arm_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
+                                      char *utf8_out) {
+  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+
+  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
+  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
+    if (!match_system(big_endian)) {
+      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+    }
+    if (vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
+      // It is common enough that we have sequences of 16 consecutive ASCII
+      // characters.
+      uint16x8_t nextin =
+          vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
+      if (!match_system(big_endian)) {
+        nextin = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(nextin)));
+      }
+      if (vmaxvq_u16(nextin) > 0x7F) {
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x8_t utf8_packed = vmovn_u16(in);
+        // 2. store (8 bytes)
+        vst1_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 8;
+        utf8_output += 8;
+        in = nextin;
+      } else {
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
+        // 2. store (16 bytes)
+        vst1q_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 16;
+        utf8_output += 16;
+        continue; // we are done for this round!
+      }
+    }
+
+    if (vmaxvq_u16(in) <= 0x7FF) {
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
+      const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const uint16x8_t t0 = vshlq_n_u16(in, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const uint16x8_t t1 = vandq_u16(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const uint16x8_t t2 = vandq_u16(in, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const uint16x8_t t3 = vorrq_u16(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const uint16x8_t t4 = vorrq_u16(t3, v_c080);
+      // 2. merge ASCII and 2-byte codewords
+      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
+      const uint8x16_t utf8_unpacked =
+          vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
+      // 3. prepare bitmask for 8-bit lookup
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t mask = simdutf_make_uint16x8_t(
+          0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
+#else
+      const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
+                               0x0002, 0x0008, 0x0020, 0x0080};
+#endif
+      uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
+      // 4. pack the bytes
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+      const uint8x16_t shuffle = vld1q_u8(row + 1);
+      const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+
+      // 5. store bytes
+      vst1q_u8(utf8_output, utf8_packed);
+
+      // 6. adjust pointers
+      buf += 8;
+      utf8_output += row[0];
+      continue;
+    }
+    const uint16x8_t surrogates_bytemask =
+        vceqq_u16(vandq_u16(in, v_f800), v_d800);
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (vmaxvq_u16(surrogates_bytemask) == 0) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t dup_even = simdutf_make_uint16x8_t(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#else
+      const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
+                                   0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
+#endif
+      /* In this branch we handle three cases:
+         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
+
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
+
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
+
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
+
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const uint16x8_t t0 = vreinterpretq_u16_u8(
+          vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+
+      // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
+      const uint16x8_t s0 = vshrq_n_u16(in, 12);
+      // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
+      const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
+      // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
+      const uint16x8_t s1s = vshlq_n_u16(s1, 2);
+      // [00bb|bbbb|0000|aaaa]
+      const uint16x8_t s2 = vorrq_u16(s0, s1s);
+      // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
+      const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
+      const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
+      const uint16x8_t m0 =
+          vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
+      const uint16x8_t s4 = veorq_u16(s3, m0);
+#undef simdutf_vec
+
+      // 4. expand code units 16-bit => 32-bit
+      const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
+      const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      const uint16x8_t onemask = simdutf_make_uint16x8_t(
+          0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
+      const uint16x8_t twomask = simdutf_make_uint16x8_t(
+          0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
+#else
+      const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                  0x0100, 0x0400, 0x1000, 0x4000};
+      const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
+                                  0x0200, 0x0800, 0x2000, 0x8000};
+#endif
+      const uint16x8_t combined =
+          vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
+                    vandq_u16(one_or_two_bytes_bytemask, twomask));
+      const uint16_t mask = vaddvq_u16(combined);
+      // The following fast path may or may not be beneficial.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
+        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
+        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
+        vst1q_u8(utf8_output, utf8_0);
+        utf8_output += 12;
+        vst1q_u8(utf8_output, utf8_1);
+        utf8_output += 12;
+        buf += 8;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
+      const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
+      const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+
+      vst1q_u8(utf8_output, utf8_0);
+      utf8_output += row0[0];
+      vst1q_u8(utf8_output, utf8_1);
+      utf8_output += row1[0];
+
+      buf += 8;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xF800) != 0xD800) {
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = !match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k - 1),
+                reinterpret_cast<char *>(utf8_output));
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf8_output++ = char((value >> 18) | 0b11110000);
+          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((value & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
+    }
+  } // while
+
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        reinterpret_cast<char *>(utf8_output));
+}
+
+template <endianness big_endian>
+simdutf_really_inline size_t
+arm64_utf8_length_from_utf16_bytemask(const char16_t *in, size_t size) {
+  size_t pos = 0;
+
+  constexpr size_t N = 8;
+  const auto one = vmovq_n_u16(1);
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
+
+  for (; pos < size / N * N; pos += N) {
+    auto input = vld1q_u16(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(input)));
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate =
+        vceqq_u16(vandq_u16(input, vmovq_n_u16(0xf800)), vmovq_n_u16(0xd800));
+
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = vminq_u16(vandq_u16(input, vmovq_n_u16(0xff80)), one);
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = vminq_u16(vandq_u16(input, vmovq_n_u16(0xf800)), one);
+
+    /*
+        Explanation how the counting works.
+
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    auto v_count = vaddq_u16(c1, c0);
+    v_count = vaddq_u16(v_count, is_surrogate);
+    count += vaddlvq_u16(v_count);
+  }
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
+}
+/* end file src/arm64/arm_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/arm64/arm_base64.cpp */
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
+
+size_t encode_base64(char *dst, const char *src, size_t srclen,
+                     base64_options options) {
+  // credit: Wojciech Muła
+  uint8_t *out = (uint8_t *)dst;
+  constexpr static uint8_t source_table[64] = {
+      'A', 'Q', 'g', 'w', 'B', 'R', 'h', 'x', 'C', 'S', 'i', 'y', 'D',
+      'T', 'j', 'z', 'E', 'U', 'k', '0', 'F', 'V', 'l', '1', 'G', 'W',
+      'm', '2', 'H', 'X', 'n', '3', 'I', 'Y', 'o', '4', 'J', 'Z', 'p',
+      '5', 'K', 'a', 'q', '6', 'L', 'b', 'r', '7', 'M', 'c', 's', '8',
+      'N', 'd', 't', '9', 'O', 'e', 'u', '+', 'P', 'f', 'v', '/',
+  };
+  constexpr static uint8_t source_table_url[64] = {
+      'A', 'Q', 'g', 'w', 'B', 'R', 'h', 'x', 'C', 'S', 'i', 'y', 'D',
+      'T', 'j', 'z', 'E', 'U', 'k', '0', 'F', 'V', 'l', '1', 'G', 'W',
+      'm', '2', 'H', 'X', 'n', '3', 'I', 'Y', 'o', '4', 'J', 'Z', 'p',
+      '5', 'K', 'a', 'q', '6', 'L', 'b', 'r', '7', 'M', 'c', 's', '8',
+      'N', 'd', 't', '9', 'O', 'e', 'u', '-', 'P', 'f', 'v', '_',
+  };
+  const uint8x16_t v3f = vdupq_n_u8(0x3f);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  // When trying to load a uint8_t array, Visual Studio might
+  // error with: error C2664: '__n128x4 neon_ld4m_q8(const char *)':
+  // cannot convert argument 1 from 'const uint8_t [64]' to 'const char *
+  const uint8x16x4_t table = vld4q_u8(
+      (reinterpret_cast<const char *>(options & base64_url) ? source_table_url
+                                                            : source_table));
+#else
+  const uint8x16x4_t table =
+      vld4q_u8((options & base64_url) ? source_table_url : source_table);
+#endif
+  size_t i = 0;
+  for (; i + 16 * 3 <= srclen; i += 16 * 3) {
+    const uint8x16x3_t in = vld3q_u8((const uint8_t *)src + i);
+    uint8x16x4_t result;
+    result.val[0] = vshrq_n_u8(in.val[0], 2);
+    result.val[1] =
+        vandq_u8(vsliq_n_u8(vshrq_n_u8(in.val[1], 4), in.val[0], 4), v3f);
+    result.val[2] =
+        vandq_u8(vsliq_n_u8(vshrq_n_u8(in.val[2], 6), in.val[1], 2), v3f);
+    result.val[3] = vandq_u8(in.val[2], v3f);
+    result.val[0] = vqtbl4q_u8(table, result.val[0]);
+    result.val[1] = vqtbl4q_u8(table, result.val[1]);
+    result.val[2] = vqtbl4q_u8(table, result.val[2]);
+    result.val[3] = vqtbl4q_u8(table, result.val[3]);
+    vst4q_u8(out, result);
+    out += 64;
+  }
+
+  if (i + 24 <= srclen) {
+    const uint8x8_t v3f_d = vdup_n_u8(0x3f);
+    const uint8x8x3_t in = vld3_u8((const uint8_t *)src + i);
+    uint8x8x4_t result;
+    result.val[0] = vshr_n_u8(in.val[0], 2);
+    result.val[1] =
+        vand_u8(vsli_n_u8(vshr_n_u8(in.val[1], 4), in.val[0], 4), v3f_d);
+    result.val[2] =
+        vand_u8(vsli_n_u8(vshr_n_u8(in.val[2], 6), in.val[1], 2), v3f_d);
+    result.val[3] = vand_u8(in.val[2], v3f_d);
+    result.val[0] = vqtbl4_u8(table, result.val[0]);
+    result.val[1] = vqtbl4_u8(table, result.val[1]);
+    result.val[2] = vqtbl4_u8(table, result.val[2]);
+    result.val[3] = vqtbl4_u8(table, result.val[3]);
+    vst4_u8(out, result);
+    out += 32;
+    i += 24;
+  }
+
+  out += scalar::base64::tail_encode_base64((char *)out, src + i, srclen - i,
+                                            options);
+
+  return size_t((char *)out - dst);
+}
+
+static inline void compress(uint8x16_t data, uint16_t mask, char *output) {
+  if (mask == 0) {
+    vst1q_u8((uint8_t *)output, data);
+    return;
+  }
+  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+  uint64x2_t compactmasku64 = {tables::base64::thintable_epi8[mask1],
+                               tables::base64::thintable_epi8[mask2]};
+  uint8x16_t compactmask = vreinterpretq_u8_u64(compactmasku64);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  const uint8x16_t off =
+      simdutf_make_uint8x16_t(0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8);
+#else
+  const uint8x16_t off = {0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8};
+#endif
+
+  compactmask = vaddq_u8(compactmask, off);
+  uint8x16_t pruned = vqtbl1q_u8(data, compactmask);
+
+  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
+  // then load the corresponding mask, what it does is to write
+  // only the first pop1 bytes from the first 8 bytes, and then
+  // it fills in with the bytes from the second 8 bytes + some filling
+  // at the end.
+  compactmask = vld1q_u8(tables::base64::pshufb_combine_table + pop1 * 8);
+  uint8x16_t answer = vqtbl1q_u8(pruned, compactmask);
+  vst1q_u8((uint8_t *)output, answer);
+}
+
+struct block64 {
+  uint8x16_t chunks[4];
+};
+
+static_assert(sizeof(block64) == 64, "block64 is not 64 bytes");
+template <bool base64_url> uint64_t to_base64_mask(block64 *b, bool *error) {
+  uint8x16_t v0f = vdupq_n_u8(0xf);
+
+  uint8x16_t underscore0, underscore1, underscore2, underscore3;
+  if (base64_url) {
+    underscore0 = vceqq_u8(b->chunks[0], vdupq_n_u8(0x5f));
+    underscore1 = vceqq_u8(b->chunks[1], vdupq_n_u8(0x5f));
+    underscore2 = vceqq_u8(b->chunks[2], vdupq_n_u8(0x5f));
+    underscore3 = vceqq_u8(b->chunks[3], vdupq_n_u8(0x5f));
+  } else {
+    (void)underscore0;
+    (void)underscore1;
+    (void)underscore2;
+    (void)underscore3;
+  }
+
+  uint8x16_t lo_nibbles0 = vandq_u8(b->chunks[0], v0f);
+  uint8x16_t lo_nibbles1 = vandq_u8(b->chunks[1], v0f);
+  uint8x16_t lo_nibbles2 = vandq_u8(b->chunks[2], v0f);
+  uint8x16_t lo_nibbles3 = vandq_u8(b->chunks[3], v0f);
+
+  // Needed by the decoding step.
+  uint8x16_t hi_nibbles0 = vshrq_n_u8(b->chunks[0], 4);
+  uint8x16_t hi_nibbles1 = vshrq_n_u8(b->chunks[1], 4);
+  uint8x16_t hi_nibbles2 = vshrq_n_u8(b->chunks[2], 4);
+  uint8x16_t hi_nibbles3 = vshrq_n_u8(b->chunks[3], 4);
+  uint8x16_t lut_lo;
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  if (base64_url) {
+    lut_lo =
+        simdutf_make_uint8x16_t(0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
+                                0x70, 0x61, 0xe1, 0xf4, 0xe5, 0xa5, 0xf4, 0xf4);
+  } else {
+    lut_lo =
+        simdutf_make_uint8x16_t(0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
+                                0x70, 0x61, 0xe1, 0xb4, 0xe5, 0xe5, 0xf4, 0xb4);
+  }
+#else
+  if (base64_url) {
+    lut_lo = uint8x16_t{0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
+                        0x70, 0x61, 0xe1, 0xf4, 0xe5, 0xa5, 0xf4, 0xf4};
+  } else {
+    lut_lo = uint8x16_t{0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
+                        0x70, 0x61, 0xe1, 0xb4, 0xe5, 0xe5, 0xf4, 0xb4};
+  }
+#endif
+  uint8x16_t lo0 = vqtbl1q_u8(lut_lo, lo_nibbles0);
+  uint8x16_t lo1 = vqtbl1q_u8(lut_lo, lo_nibbles1);
+  uint8x16_t lo2 = vqtbl1q_u8(lut_lo, lo_nibbles2);
+  uint8x16_t lo3 = vqtbl1q_u8(lut_lo, lo_nibbles3);
+  uint8x16_t lut_hi;
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  if (base64_url) {
+    lut_hi =
+        simdutf_make_uint8x16_t(0x11, 0x20, 0x42, 0x80, 0x8, 0x4, 0x8, 0x4,
+                                0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20);
+  } else {
+    lut_hi =
+        simdutf_make_uint8x16_t(0x11, 0x20, 0x42, 0x80, 0x8, 0x4, 0x8, 0x4,
+                                0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20);
+  }
+#else
+  if (base64_url) {
+    lut_hi = uint8x16_t{0x11, 0x20, 0x42, 0x80, 0x8,  0x4,  0x8,  0x4,
+                        0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20};
+  } else {
+    lut_hi = uint8x16_t{0x11, 0x20, 0x42, 0x80, 0x8,  0x4,  0x8,  0x4,
+                        0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20};
+  }
+#endif
+  uint8x16_t hi0 = vqtbl1q_u8(lut_hi, hi_nibbles0);
+  uint8x16_t hi1 = vqtbl1q_u8(lut_hi, hi_nibbles1);
+  uint8x16_t hi2 = vqtbl1q_u8(lut_hi, hi_nibbles2);
+  uint8x16_t hi3 = vqtbl1q_u8(lut_hi, hi_nibbles3);
+
+  if (base64_url) {
+    hi0 = vbicq_u8(hi0, underscore0);
+    hi1 = vbicq_u8(hi1, underscore1);
+    hi2 = vbicq_u8(hi2, underscore2);
+    hi3 = vbicq_u8(hi3, underscore3);
+  }
+
+  uint8_t checks =
+      vmaxvq_u8(vorrq_u8(vorrq_u8(vandq_u8(lo0, hi0), vandq_u8(lo1, hi1)),
+                         vorrq_u8(vandq_u8(lo2, hi2), vandq_u8(lo3, hi3))));
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  const uint8x16_t bit_mask =
+      simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                              0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+  const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                               0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+  uint64_t badcharmask = 0;
+  *error = checks > 0x3;
+  if (checks) {
+    // Add each of the elements next to each other, successively, to stuff each
+    // 8 byte mask into one.
+    uint8x16_t test0 = vtstq_u8(lo0, hi0);
+    uint8x16_t test1 = vtstq_u8(lo1, hi1);
+    uint8x16_t test2 = vtstq_u8(lo2, hi2);
+    uint8x16_t test3 = vtstq_u8(lo3, hi3);
+    uint8x16_t sum0 =
+        vpaddq_u8(vandq_u8(test0, bit_mask), vandq_u8(test1, bit_mask));
+    uint8x16_t sum1 =
+        vpaddq_u8(vandq_u8(test2, bit_mask), vandq_u8(test3, bit_mask));
+    sum0 = vpaddq_u8(sum0, sum1);
+    sum0 = vpaddq_u8(sum0, sum0);
+    badcharmask = vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+  }
+  // This is the transformation step that can be done while we are waiting for
+  // sum0
+  uint8x16_t roll_lut;
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  if (base64_url) {
+    roll_lut =
+        simdutf_make_uint8x16_t(0xe0, 0x11, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
+                                0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0);
+  } else {
+    roll_lut =
+        simdutf_make_uint8x16_t(0x0, 0x10, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
+                                0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0);
+  }
+#else
+  if (base64_url) {
+    roll_lut = uint8x16_t{0xe0, 0x11, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
+                          0x0,  0x0,  0x0,  0x0, 0x0,  0x0,  0x0,  0x0};
+  } else {
+    roll_lut = uint8x16_t{0x0, 0x10, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
+                          0x0, 0x0,  0x0,  0x0, 0x0,  0x0,  0x0,  0x0};
+  }
+#endif
+  uint8x16_t vsecond_last = base64_url ? vdupq_n_u8(0x2d) : vdupq_n_u8(0x2f);
+  if (base64_url) {
+    hi_nibbles0 = vbicq_u8(hi_nibbles0, underscore0);
+    hi_nibbles1 = vbicq_u8(hi_nibbles1, underscore1);
+    hi_nibbles2 = vbicq_u8(hi_nibbles2, underscore2);
+    hi_nibbles3 = vbicq_u8(hi_nibbles3, underscore3);
+  }
+  uint8x16_t roll0 = vqtbl1q_u8(
+      roll_lut, vaddq_u8(vceqq_u8(b->chunks[0], vsecond_last), hi_nibbles0));
+  uint8x16_t roll1 = vqtbl1q_u8(
+      roll_lut, vaddq_u8(vceqq_u8(b->chunks[1], vsecond_last), hi_nibbles1));
+  uint8x16_t roll2 = vqtbl1q_u8(
+      roll_lut, vaddq_u8(vceqq_u8(b->chunks[2], vsecond_last), hi_nibbles2));
+  uint8x16_t roll3 = vqtbl1q_u8(
+      roll_lut, vaddq_u8(vceqq_u8(b->chunks[3], vsecond_last), hi_nibbles3));
+  b->chunks[0] = vaddq_u8(b->chunks[0], roll0);
+  b->chunks[1] = vaddq_u8(b->chunks[1], roll1);
+  b->chunks[2] = vaddq_u8(b->chunks[2], roll2);
+  b->chunks[3] = vaddq_u8(b->chunks[3], roll3);
+  return badcharmask;
+}
+
+void copy_block(block64 *b, char *output) {
+  vst1q_u8((uint8_t *)output, b->chunks[0]);
+  vst1q_u8((uint8_t *)output + 16, b->chunks[1]);
+  vst1q_u8((uint8_t *)output + 32, b->chunks[2]);
+  vst1q_u8((uint8_t *)output + 48, b->chunks[3]);
+}
+
+uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
+  uint64_t popcounts =
+      vget_lane_u64(vreinterpret_u64_u8(vcnt_u8(vcreate_u8(~mask))), 0);
+  uint64_t offsets = popcounts * 0x0101010101010101;
+  compress(b->chunks[0], uint16_t(mask), output);
+  compress(b->chunks[1], uint16_t(mask >> 16), &output[(offsets >> 8) & 0xFF]);
+  compress(b->chunks[2], uint16_t(mask >> 32), &output[(offsets >> 24) & 0xFF]);
+  compress(b->chunks[3], uint16_t(mask >> 48), &output[(offsets >> 40) & 0xFF]);
+  return offsets >> 56;
+}
+
+// The caller of this function is responsible to ensure that there are 64 bytes
+// available from reading at src. The data is read into a block64 structure.
+void load_block(block64 *b, const char *src) {
+  b->chunks[0] = vld1q_u8(reinterpret_cast<const uint8_t *>(src));
+  b->chunks[1] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 16);
+  b->chunks[2] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 32);
+  b->chunks[3] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 48);
+}
+
+// The caller of this function is responsible to ensure that there are 32 bytes
+// available from reading at data. It returns a 16-byte value, narrowing with
+// saturation the 16-bit words.
+inline uint8x16_t load_satured(const uint16_t *data) {
+  uint16x8_t in1 = vld1q_u16(data);
+  uint16x8_t in2 = vld1q_u16(data + 8);
+  return vqmovn_high_u16(vqmovn_u16(in1), in2);
+}
+
+// The caller of this function is responsible to ensure that there are 128 bytes
+// available from reading at src. The data is read into a block64 structure.
+void load_block(block64 *b, const char16_t *src) {
+  b->chunks[0] = load_satured(reinterpret_cast<const uint16_t *>(src));
+  b->chunks[1] = load_satured(reinterpret_cast<const uint16_t *>(src) + 16);
+  b->chunks[2] = load_satured(reinterpret_cast<const uint16_t *>(src) + 32);
+  b->chunks[3] = load_satured(reinterpret_cast<const uint16_t *>(src) + 48);
+}
+
+// decode 64 bytes and output 48 bytes
+void base64_decode_block(char *out, const char *src) {
+  uint8x16x4_t str = vld4q_u8((uint8_t *)src);
+  uint8x16x3_t outvec;
+  outvec.val[0] = vsliq_n_u8(vshrq_n_u8(str.val[1], 4), str.val[0], 2);
+  outvec.val[1] = vsliq_n_u8(vshrq_n_u8(str.val[2], 2), str.val[1], 4);
+  outvec.val[2] = vsliq_n_u8(str.val[3], str.val[2], 6);
+  vst3q_u8((uint8_t *)out, outvec);
+}
+
+static size_t compress_block_single(block64 *b, uint64_t mask, char *output) {
+  const size_t pos64 = trailing_zeroes(mask);
+  const int8_t pos = pos64 & 0xf;
+
+  // Predefine the index vector
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+  const uint8x16_t v1 = simdutf_make_uint8x16_t(0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
+                                                10, 11, 12, 13, 14, 15);
+#else  // SIMDUTF_REGULAR_VISUAL_STUDIO
+  const uint8x16_t v1 = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
+#endif // SIMDUTF_REGULAR_VISUAL_STUDIO
+
+  switch (pos64 >> 4) {
+  case 0b00: {
+    const uint8x16_t v0 = vmovq_n_u8((uint8_t)(pos - 1));
+    const uint8x16_t v2 =
+        vcgtq_s8(vreinterpretq_s8_u8(v1),
+                 vreinterpretq_s8_u8(v0));  // Compare greater than
+    const uint8x16_t sh = vsubq_u8(v1, v2); // Subtract
+    const uint8x16_t compressed =
+        vqtbl1q_u8(b->chunks[0], sh); // Table lookup (shuffle)
+
+    vst1q_u8((uint8_t *)(output + 0 * 16), compressed);
+    vst1q_u8((uint8_t *)(output + 1 * 16 - 1), b->chunks[1]);
+    vst1q_u8((uint8_t *)(output + 2 * 16 - 1), b->chunks[2]);
+    vst1q_u8((uint8_t *)(output + 3 * 16 - 1), b->chunks[3]);
+  } break;
+
+  case 0b01: {
+    vst1q_u8((uint8_t *)(output + 0 * 16), b->chunks[0]);
+
+    const uint8x16_t v0 = vmovq_n_u8((uint8_t)(pos - 1));
+    const uint8x16_t v2 =
+        vcgtq_s8(vreinterpretq_s8_u8(v1), vreinterpretq_s8_u8(v0));
+    const uint8x16_t sh = vsubq_u8(v1, v2);
+    const uint8x16_t compressed = vqtbl1q_u8(b->chunks[1], sh);
+
+    vst1q_u8((uint8_t *)(output + 1 * 16), compressed);
+    vst1q_u8((uint8_t *)(output + 2 * 16 - 1), b->chunks[2]);
+    vst1q_u8((uint8_t *)(output + 3 * 16 - 1), b->chunks[3]);
+  } break;
+
+  case 0b10: {
+    vst1q_u8((uint8_t *)(output + 0 * 16), b->chunks[0]);
+    vst1q_u8((uint8_t *)(output + 1 * 16), b->chunks[1]);
+
+    const uint8x16_t v0 = vmovq_n_u8((uint8_t)(pos - 1));
+    const uint8x16_t v2 =
+        vcgtq_s8(vreinterpretq_s8_u8(v1), vreinterpretq_s8_u8(v0));
+    const uint8x16_t sh = vsubq_u8(v1, v2);
+    const uint8x16_t compressed = vqtbl1q_u8(b->chunks[2], sh);
+
+    vst1q_u8((uint8_t *)(output + 2 * 16), compressed);
+    vst1q_u8((uint8_t *)(output + 3 * 16 - 1), b->chunks[3]);
+  } break;
+
+  case 0b11: {
+    vst1q_u8((uint8_t *)(output + 0 * 16), b->chunks[0]);
+    vst1q_u8((uint8_t *)(output + 1 * 16), b->chunks[1]);
+    vst1q_u8((uint8_t *)(output + 2 * 16), b->chunks[2]);
+
+    const uint8x16_t v0 = vmovq_n_u8((uint8_t)(pos - 1));
+    const uint8x16_t v2 =
+        vcgtq_s8(vreinterpretq_s8_u8(v1), vreinterpretq_s8_u8(v0));
+    const uint8x16_t sh = vsubq_u8(v1, v2);
+    const uint8x16_t compressed = vqtbl1q_u8(b->chunks[3], sh);
+
+    vst1q_u8((uint8_t *)(output + 3 * 16), compressed);
+  } break;
+  }
+  return 63;
+}
+
+template <typename T> bool is_power_of_two(T x) { return (x & (x - 1)) == 0; }
+
+template <bool base64_url, bool ignore_garbage, typename char_type>
+full_result
+compress_decode_base64(char *dst, const char_type *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
+                                        : tables::base64::to_base64_value;
+  size_t equallocation =
+      srclen; // location of the first padding character if any
+  // skip trailing spaces
+  while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+         to_base64[uint8_t(src[srclen - 1])] == 64) {
+    srclen--;
+  }
+  size_t equalsigns = 0;
+  if (srclen > 0 && src[srclen - 1] == '=') {
+    equallocation = srclen - 1;
+    srclen--;
+    equalsigns = 1;
+    // skip trailing spaces
+    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+           to_base64[uint8_t(src[srclen - 1])] == 64) {
+      srclen--;
+    }
+    if (srclen > 0 && src[srclen - 1] == '=') {
+      equallocation = srclen - 1;
+      srclen--;
+      equalsigns = 2;
+    }
+  }
+  if (srclen == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, 0, 0};
+  }
+  const char_type *const srcinit = src;
+  const char *const dstinit = dst;
+  const char_type *const srcend = src + srclen;
+
+  constexpr size_t block_size = 10;
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const char_type *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b;
+      load_block(&b, src);
+      src += 64;
+      bool error = false;
+      uint64_t badcharmask = to_base64_mask<base64_url>(&b, &error);
+      if (badcharmask) {
+        if (error && !ignore_garbage) {
+          src -= 64;
+          while (src < srcend && scalar::base64::is_eight_byte(*src) &&
+                 to_base64[uint8_t(*src)] <= 64) {
+            src++;
+          }
+          if (src < srcend) {
+            // should never happen
+          }
+          return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                  size_t(dst - dstinit)};
+        }
+      }
+
+      if (badcharmask != 0) {
+        // optimization opportunity: check for simple masks like those made of
+        // continuous 1s followed by continuous 0s. And masks containing a
+        // single bad character.
+        if (is_power_of_two(badcharmask)) {
+          bufferptr += compress_block_single(&b, badcharmask, bufferptr);
+        } else {
+          bufferptr += compress_block(&b, badcharmask, bufferptr);
+        }
+      } else {
+        // optimization opportunity: if bufferptr == buffer and mask == 0, we
+        // can avoid the call to compress_block and decode directly.
+        copy_block(&b, bufferptr);
+        bufferptr += 64;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 1); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
+  }
+  char *buffer_start = buffer;
+  // Optimization note: if this is almost full, then it is worth our
+  // time, otherwise, we should just decode directly.
+  int last_block = (int)((bufferptr - buffer_start) % 64);
+  if (last_block != 0 && srcend - src + last_block >= 64) {
+    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
+      uint8_t val = to_base64[uint8_t(*src)];
+      *bufferptr = char(val);
+      if ((!scalar::base64::is_eight_byte(*src) || val > 64) &&
+          !ignore_garbage) {
+        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      }
+      bufferptr += (val <= 63);
+      src++;
+    }
+  }
+
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    base64_decode_block(dst, buffer_start);
+    dst += 48;
+  }
+  if ((bufferptr - buffer_start) % 64 != 0) {
+    while (buffer_start + 4 < bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+      triple = scalar::u32_swap_bytes(triple);
+      std::memcpy(dst, &triple, 4);
+
+      dst += 3;
+      buffer_start += 4;
+    }
+    if (buffer_start + 4 <= bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+      triple = scalar::u32_swap_bytes(triple);
+      std::memcpy(dst, &triple, 3);
+
+      dst += 3;
+      buffer_start += 4;
+    }
+    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
+    // backtrack
+    int leftover = int(bufferptr - buffer_start);
+    while (leftover > 0) {
+      if (!ignore_garbage) {
+        while (to_base64[uint8_t(*(src - 1))] == 64) {
+          src--;
+        }
+      } else {
+        while (to_base64[uint8_t(*(src - 1))] >= 64) {
+          src--;
+        }
+      }
+      src--;
+      leftover--;
+    }
+  }
+  if (src < srcend + equalsigns) {
+    full_result r = scalar::base64::base64_tail_decode(
+        dst, src, srcend - src, equalsigns, options, last_chunk_options);
+    r.input_count += size_t(src - srcinit);
+    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
+        r.error == error_code::BASE64_EXTRA_BITS) {
+      return r;
+    } else {
+      r.output_count += size_t(dst - dstinit);
+    }
+    if (last_chunk_options != stop_before_partial &&
+        r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
+      // additional checks
+      if ((r.output_count % 3 == 0) ||
+          ((r.output_count % 3) + 1 + equalsigns != 4)) {
+        r.error = error_code::INVALID_BASE64_CHARACTER;
+        r.input_count = equallocation;
+      }
+    }
+    return r;
+  }
+  if (equalsigns > 0 && !ignore_garbage) {
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
+}
+/* end file src/arm64/arm_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/arm64/arm_convert_utf32_to_latin1.cpp */
+std::pair<const char32_t *, char *>
+arm_convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                            char *latin1_output) {
+  const char32_t *end = buf + len;
+  while (end - buf >= 8) {
+    uint32x4_t in1 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+    uint32x4_t in2 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
+
+    uint16x8_t utf16_packed = vcombine_u16(vqmovn_u32(in1), vqmovn_u32(in2));
+    if (vmaxvq_u16(utf16_packed) <= 0xff) {
+      // 1. pack the bytes
+      uint8x8_t latin1_packed = vmovn_u16(utf16_packed);
+      // 2. store (8 bytes)
+      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
+      // 3. adjust pointers
+      buf += 8;
+      latin1_output += 8;
+    } else {
+      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
+    }
+  } // while
+  return std::make_pair(buf, latin1_output);
+}
+
+std::pair<result, char *>
+arm_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                        char *latin1_output) {
+  const char32_t *start = buf;
+  const char32_t *end = buf + len;
+
+  while (end - buf >= 8) {
+    uint32x4_t in1 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+    uint32x4_t in2 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
+
+    uint16x8_t utf16_packed = vcombine_u16(vqmovn_u32(in1), vqmovn_u32(in2));
+
+    if (vmaxvq_u16(utf16_packed) <= 0xff) {
+      // 1. pack the bytes
+      uint8x8_t latin1_packed = vmovn_u16(utf16_packed);
+      // 2. store (8 bytes)
+      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
+      // 3. adjust pointers
+      buf += 8;
+      latin1_output += 8;
+    } else {
+      // Let us do a scalar fallback.
+      for (int k = 0; k < 8; k++) {
+        uint32_t word = buf[k];
+        if (word <= 0xff) {
+          *latin1_output++ = char(word);
+        } else {
+          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
+                                latin1_output);
+        }
+      }
+    }
+  } // while
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        latin1_output);
+}
+/* end file src/arm64/arm_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF16
+/* begin file src/arm64/arm_convert_utf32_to_utf16.cpp */
+template <endianness big_endian>
+std::pair<const char32_t *, char16_t *>
+arm_convert_utf32_to_utf16(const char32_t *buf, size_t len,
+                           char16_t *utf16_out) {
+  uint16_t *utf16_output = reinterpret_cast<uint16_t *>(utf16_out);
+  const char32_t *end = buf + len;
+
+  uint16x4_t forbidden_bytemask = vmov_n_u16(0x0);
+
+  while (end - buf >= 4) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+
+    // Check if no bits set above 16th
+    if (vmaxvq_u32(in) <= 0xFFFF) {
+      uint16x4_t utf16_packed = vmovn_u32(in);
+
+      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
+      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
+      forbidden_bytemask = vorr_u16(vand_u16(vcle_u16(utf16_packed, v_dfff),
+                                             vcge_u16(utf16_packed, v_d800)),
+                                    forbidden_bytemask);
+
+      if (!match_system(big_endian)) {
+        utf16_packed =
+            vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(utf16_packed)));
+      }
+      vst1_u16(utf16_output, utf16_packed);
+      utf16_output += 4;
+      buf += 4;
+    } else {
+      size_t forward = 3;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFF0000) == 0) {
+          // will not generate a surrogate pair
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char16_t *>(utf16_output));
+          }
+          *utf16_output++ = !match_system(big_endian)
+                                ? char16_t(word >> 8 | word << 8)
+                                : char16_t(word);
+        } else {
+          // will generate a surrogate pair
+          if (word > 0x10FFFF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char16_t *>(utf16_output));
+          }
+          word -= 0x10000;
+          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+          if (!match_system(big_endian)) {
+            high_surrogate =
+                uint16_t(high_surrogate >> 8 | high_surrogate << 8);
+            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
+          }
+          *utf16_output++ = char16_t(high_surrogate);
+          *utf16_output++ = char16_t(low_surrogate);
+        }
+      }
+      buf += k;
+    }
+  }
+
+  // check for invalid input
+  if (vmaxv_u16(forbidden_bytemask) != 0) {
+    return std::make_pair(nullptr, reinterpret_cast<char16_t *>(utf16_output));
+  }
+
+  return std::make_pair(buf, reinterpret_cast<char16_t *>(utf16_output));
+}
+
+template <endianness big_endian>
+std::pair<result, char16_t *>
+arm_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
+                                       char16_t *utf16_out) {
+  uint16_t *utf16_output = reinterpret_cast<uint16_t *>(utf16_out);
+  const char32_t *start = buf;
+  const char32_t *end = buf + len;
+
+  while (end - buf >= 4) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+
+    // Check if no bits set above 16th
+    if (vmaxvq_u32(in) <= 0xFFFF) {
+      uint16x4_t utf16_packed = vmovn_u32(in);
+
+      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
+      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
+      const uint16x4_t forbidden_bytemask = vand_u16(
+          vcle_u16(utf16_packed, v_dfff), vcge_u16(utf16_packed, v_d800));
+      if (vmaxv_u16(forbidden_bytemask) != 0) {
+        return std::make_pair(result(error_code::SURROGATE, buf - start),
+                              reinterpret_cast<char16_t *>(utf16_output));
+      }
+
+      if (!match_system(big_endian)) {
+        utf16_packed =
+            vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(utf16_packed)));
+      }
+      vst1_u16(utf16_output, utf16_packed);
+      utf16_output += 4;
+      buf += 4;
+    } else {
+      size_t forward = 3;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFF0000) == 0) {
+          // will not generate a surrogate pair
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k),
+                reinterpret_cast<char16_t *>(utf16_output));
+          }
+          *utf16_output++ = !match_system(big_endian)
+                                ? char16_t(word >> 8 | word << 8)
+                                : char16_t(word);
+        } else {
+          // will generate a surrogate pair
+          if (word > 0x10FFFF) {
+            return std::make_pair(
+                result(error_code::TOO_LARGE, buf - start + k),
+                reinterpret_cast<char16_t *>(utf16_output));
+          }
+          word -= 0x10000;
+          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+          if (!match_system(big_endian)) {
+            high_surrogate =
+                uint16_t(high_surrogate >> 8 | high_surrogate << 8);
+            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
+          }
+          *utf16_output++ = char16_t(high_surrogate);
+          *utf16_output++ = char16_t(low_surrogate);
+        }
+      }
+      buf += k;
+    }
+  }
 
-#endif // SIMDUTF_UTF8_TO_UTF16_TABLES_H
-/* end file src/tables/utf8_to_utf16_tables.h */
-/* begin file src/tables/utf16_to_utf8_tables.h */
-// file generated by scripts/sse_convert_utf16_to_utf8.py
-#ifndef SIMDUTF_UTF16_TO_UTF8_TABLES_H
-#define SIMDUTF_UTF16_TO_UTF8_TABLES_H
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        reinterpret_cast<char16_t *>(utf16_output));
+}
+/* end file src/arm64/arm_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF8
+/* begin file src/arm64/arm_convert_utf32_to_utf8.cpp */
+std::pair<const char32_t *, char *>
+arm_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
+  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
+  const char32_t *end = buf + len;
 
-namespace simdutf {
-namespace {
-namespace tables {
-namespace utf16_to_utf8 {
+  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
 
-// 1 byte for length, 16 bytes for mask
-const uint8_t pack_1_2_utf8_bytes[256][17] = {
-    {16, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14},
-    {15, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
-    {15, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80},
-    {14, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {15, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
-    {14, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {14, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80},
-    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {15, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
-    {14, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {14, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80},
-    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 7, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {15, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80},
-    {14, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {14, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 6, 8, 11, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 6, 9, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 8, 10, 13, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 5, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 6, 9, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 8, 11, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 3, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 9, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 4, 6, 8, 10, 12, 15, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {15, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80},
-    {14, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {14, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
-    {13, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 4, 7, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 7, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 5, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 7, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 7, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 3, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 7, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 4, 7, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {14, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80},
-    {13, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 2, 5, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 5, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {13, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 4, 6, 9, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 8, 11, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 3, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 9, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 4, 6, 8, 10, 13, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {13, 1, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80},
-    {12, 0, 3, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 2, 5, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 5, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 3, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 5, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 5, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {12, 1, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80},
-    {11, 0, 3, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 3, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {11, 1, 0, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 2, 4, 6, 9, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 4, 6, 8, 11, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {11, 1, 0, 3, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 3, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 3, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 3, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 1, 0, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 2, 4, 6, 9, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 1, 0, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 0, 2, 4, 6, 8, 10, 12, 14, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80}};
+  uint16x8_t forbidden_bytemask = vmovq_n_u16(0x0);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+
+  while (buf + 16 + safety_margin < end) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
+
+    // Check if no bits set above 16th
+    if (vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
+      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
+      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
+      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
+      if (vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
+        // 2. store (8 bytes)
+        vst1_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 8;
+        utf8_output += 8;
+        continue; // we are done for this round!
+      }
+
+      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
+        // 1. prepare 2-byte values
+        // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+        // expected output   : [110a|aaaa|10bb|bbbb] x 8
+        const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
+        const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+
+        // t0 = [000a|aaaa|bbbb|bb00]
+        const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
+        // t1 = [000a|aaaa|0000|0000]
+        const uint16x8_t t1 = vandq_u16(t0, v_1f00);
+        // t2 = [0000|0000|00bb|bbbb]
+        const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
+        // t3 = [000a|aaaa|00bb|bbbb]
+        const uint16x8_t t3 = vorrq_u16(t1, t2);
+        // t4 = [110a|aaaa|10bb|bbbb]
+        const uint16x8_t t4 = vorrq_u16(t3, v_c080);
+        // 2. merge ASCII and 2-byte codewords
+        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
+        const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(
+            vbslq_u16(one_byte_bytemask, utf16_packed, t4));
+        // 3. prepare bitmask for 8-bit lookup
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t mask = simdutf_make_uint16x8_t(
+            0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
+#else
+        const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                 0x0002, 0x0008, 0x0020, 0x0080};
+#endif
+        uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
+        // 4. pack the bytes
+        const uint8_t *row =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+        const uint8x16_t shuffle = vld1q_u8(row + 1);
+        const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+
+        // 5. store bytes
+        vst1q_u8(utf8_output, utf8_packed);
+
+        // 6. adjust pointers
+        buf += 8;
+        utf8_output += row[0];
+        continue;
+      } else {
+        // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
+        forbidden_bytemask =
+            vorrq_u16(vandq_u16(vcleq_u16(utf16_packed, v_dfff),
+                                vcgeq_u16(utf16_packed, v_d800)),
+                      forbidden_bytemask);
+
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t dup_even = simdutf_make_uint16x8_t(
+            0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#else
+        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
+                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
+#endif
+        /* In this branch we handle three cases:
+          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+          single UFT-8 byte
+          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              -
+          two UTF-8 bytes
+          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+          three UTF-8 bytes
+
+          We expand the input word (16-bit) into two code units (32-bit), thus
+          we have room for four bytes. However, we need five distinct bit
+          layouts. Note that the last byte in cases #2 and #3 is the same.
+
+          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+          in register t2.
+
+          We precompute byte 1 for case #3 and -- **conditionally** --
+          precompute either byte 1 for case #2 or byte 2 for case #3. Note that
+          they differ by exactly one bit.
+
+          Finally from these two code units we build proper UTF-8 sequence,
+          taking into account the case (i.e, the number of bytes to write).
+        */
+        /**
+         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+         * t2 => [0ccc|cccc] [10cc|cccc]
+         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+         */
+#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
+        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+        const uint16x8_t t0 =
+            vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed),
+                                            vreinterpretq_u8_u16(dup_even)));
+        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
+        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+        const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+
+        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
+        const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
+        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
+        const uint16x8_t s1 =
+            vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
+        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
+        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
+        // [00bb|bbbb|0000|aaaa]
+        const uint16x8_t s2 = vorrq_u16(s0, s1s);
+        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
+        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
+        const uint16x8_t one_or_two_bytes_bytemask =
+            vcleq_u16(utf16_packed, v_07ff);
+        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000),
+                                        one_or_two_bytes_bytemask);
+        const uint16x8_t s4 = veorq_u16(s3, m0);
+#undef simdutf_vec
+
+        // 4. expand code units 16-bit => 32-bit
+        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
+        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+
+        // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t onemask = simdutf_make_uint16x8_t(
+            0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
+        const uint16x8_t twomask = simdutf_make_uint16x8_t(
+            0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
+#else
+        const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                    0x0100, 0x0400, 0x1000, 0x4000};
+        const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
+                                    0x0200, 0x0800, 0x2000, 0x8000};
+#endif
+        const uint16x8_t combined =
+            vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
+                      vandq_u16(one_or_two_bytes_bytemask, twomask));
+        const uint16_t mask = vaddvq_u16(combined);
+        // The following fast path may or may not be beneficial.
+        /*if(mask == 0) {
+          // We only have three-byte code units. Use fast path.
+          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
+          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
+          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
+          vst1q_u8(utf8_output, utf8_0);
+          utf8_output += 12;
+          vst1q_u8(utf8_output, utf8_1);
+          utf8_output += 12;
+          buf += 8;
+          continue;
+        }*/
+        const uint8_t mask0 = uint8_t(mask);
+        const uint8_t *row0 =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
+        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+
+        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+        const uint8_t *row1 =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
+        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+
+        vst1q_u8(utf8_output, utf8_0);
+        utf8_output += row0[0];
+        vst1q_u8(utf8_output, utf8_1);
+        utf8_output += row1[0];
+
+        buf += 8;
+      }
+      // At least one 32-bit word will produce a surrogate pair in UTF-16 <=>
+      // will produce four UTF-8 bytes.
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) {
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char *>(utf8_output));
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          if (word > 0x10FFFF) {
+            return std::make_pair(nullptr,
+                                  reinterpret_cast<char *>(utf8_output));
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
+    }
+  } // while
 
-// 1 byte for length, 16 bytes for mask
-const uint8_t pack_1_2_3_utf8_bytes[256][17] = {
-    {12, 2, 3, 1, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80},
-    {9, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {11, 3, 1, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {10, 0, 6, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 2, 3, 1, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {11, 2, 3, 1, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {10, 3, 1, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 7, 5, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {10, 2, 3, 1, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 3, 1, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 0, 4, 10, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 2, 3, 1, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 6, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 2, 3, 1, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {8, 2, 3, 1, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 7, 5, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 4, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {11, 2, 3, 1, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {10, 3, 1, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 6, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {10, 2, 3, 1, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 3, 1, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 0, 7, 5, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 2, 3, 1, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 4, 11, 9, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {10, 2, 3, 1, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 3, 1, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 0, 6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 2, 3, 1, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 2, 3, 1, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 7, 5, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 4, 8, 14, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 2, 3, 1, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 6, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 2, 3, 1, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {8, 2, 3, 1, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 7, 5, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 4, 10, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 2, 3, 1, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {3, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 6, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 2, 3, 1, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {2, 3, 1, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {1, 0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {5, 2, 3, 1, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 3, 1, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 0, 7, 5, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 2, 3, 1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {3, 3, 1, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {2, 0, 4, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {8, 2, 3, 1, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 6, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 2, 3, 1, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 3, 1, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 0, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {7, 2, 3, 1, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 3, 1, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 0, 7, 5, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 2, 3, 1, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {3, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 4, 11, 9, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 3, 1, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 0, 6, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 2, 3, 1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {3, 3, 1, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {2, 0, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 2, 3, 1, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {3, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 7, 5, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 2, 3, 1, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 3, 1, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 0, 4, 8, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {11, 2, 3, 1, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {10, 3, 1, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {9, 0, 6, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {10, 2, 3, 1, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 3, 1, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 0, 7, 5, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 2, 3, 1, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 4, 10, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 6, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 2, 3, 1, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {4, 3, 1, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {3, 0, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {7, 2, 3, 1, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 7, 5, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 2, 3, 1, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 4, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {10, 2, 3, 1, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 3, 1, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 0, 6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 2, 3, 1, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 2, 3, 1, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 7, 5, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 4, 11, 9, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 2, 3, 1, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 6, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 2, 3, 1, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {8, 2, 3, 1, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 7, 5, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 4, 8, 15, 13, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {10, 2, 3, 1, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {7, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {9, 3, 1, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {8, 0, 6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 2, 3, 1, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {9, 2, 3, 1, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 7, 5, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {8, 2, 3, 1, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 4, 10, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 3, 1, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 0, 6, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 2, 3, 1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80, 0x80},
-    {3, 3, 1, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {2, 0, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 2, 3, 1, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {3, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 7, 5, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 2, 3, 1, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 3, 1, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 0, 4, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {9, 2, 3, 1, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
-    {6, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 3, 1, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {7, 0, 6, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 2, 3, 1, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {8, 2, 3, 1, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 7, 5, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 2, 3, 1, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {6, 3, 1, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {5, 0, 4, 11, 9, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {8, 2, 3, 1, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {7, 3, 1, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {6, 0, 6, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 2, 3, 1, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {2, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {4, 3, 1, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {3, 0, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {7, 2, 3, 1, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {4, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 3, 1, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {5, 0, 7, 5, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {6, 2, 3, 1, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80},
-    {3, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80, 0x80},
-    {5, 3, 1, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80},
-    {4, 0, 4, 8, 12, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
-     0x80, 0x80}};
+  // check for invalid input
+  if (vmaxvq_u16(forbidden_bytemask) != 0) {
+    return std::make_pair(nullptr, reinterpret_cast<char *>(utf8_output));
+  }
+  return std::make_pair(buf, reinterpret_cast<char *>(utf8_output));
+}
+
+std::pair<result, char *>
+arm_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
+                                      char *utf8_out) {
+  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
+  const char32_t *start = buf;
+  const char32_t *end = buf + len;
+
+  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-} // namespace utf16_to_utf8
-} // namespace tables
-} // unnamed namespace
-} // namespace simdutf
+  while (buf + 16 + safety_margin < end) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
 
-#endif // SIMDUTF_UTF16_TO_UTF8_TABLES_H
-/* end file src/tables/utf16_to_utf8_tables.h */
-// End of tables.
+    // Check if no bits set above 16th
+    if (vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
+      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
+      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
+      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
+      if (vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
+        // 1. pack the bytes
+        // obviously suboptimal.
+        uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
+        // 2. store (8 bytes)
+        vst1_u8(utf8_output, utf8_packed);
+        // 3. adjust pointers
+        buf += 8;
+        utf8_output += 8;
+        continue; // we are done for this round!
+      }
 
-// The scalar routines should be included once.
-/* begin file src/scalar/ascii.h */
-#ifndef SIMDUTF_ASCII_H
-#define SIMDUTF_ASCII_H
+      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
+        // 1. prepare 2-byte values
+        // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+        // expected output   : [110a|aaaa|10bb|bbbb] x 8
+        const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
+        const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace ascii {
-#if SIMDUTF_IMPLEMENTATION_FALLBACK
-// Only used by the fallback kernel.
-inline simdutf_warn_unused bool validate(const char *buf, size_t len) noexcept {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  uint64_t pos = 0;
-  // process in blocks of 16 bytes when possible
-  for (; pos + 16 <= len; pos += 16) {
-    uint64_t v1;
-    std::memcpy(&v1, data + pos, sizeof(uint64_t));
-    uint64_t v2;
-    std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-    uint64_t v{v1 | v2};
-    if ((v & 0x8080808080808080) != 0) {
-      return false;
-    }
-  }
-  // process the tail byte-by-byte
-  for (; pos < len; pos++) {
-    if (data[pos] >= 0b10000000) {
-      return false;
-    }
-  }
-  return true;
-}
+        // t0 = [000a|aaaa|bbbb|bb00]
+        const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
+        // t1 = [000a|aaaa|0000|0000]
+        const uint16x8_t t1 = vandq_u16(t0, v_1f00);
+        // t2 = [0000|0000|00bb|bbbb]
+        const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
+        // t3 = [000a|aaaa|00bb|bbbb]
+        const uint16x8_t t3 = vorrq_u16(t1, t2);
+        // t4 = [110a|aaaa|10bb|bbbb]
+        const uint16x8_t t4 = vorrq_u16(t3, v_c080);
+        // 2. merge ASCII and 2-byte codewords
+        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
+        const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(
+            vbslq_u16(one_byte_bytemask, utf16_packed, t4));
+        // 3. prepare bitmask for 8-bit lookup
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t mask = simdutf_make_uint16x8_t(
+            0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
+#else
+        const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                 0x0002, 0x0008, 0x0020, 0x0080};
 #endif
+        uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
+        // 4. pack the bytes
+        const uint8_t *row =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+        const uint8x16_t shuffle = vld1q_u8(row + 1);
+        const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
 
-inline simdutf_warn_unused result validate_with_errors(const char *buf,
-                                                       size_t len) noexcept {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  // process in blocks of 16 bytes when possible
-  for (; pos + 16 <= len; pos += 16) {
-    uint64_t v1;
-    std::memcpy(&v1, data + pos, sizeof(uint64_t));
-    uint64_t v2;
-    std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-    uint64_t v{v1 | v2};
-    if ((v & 0x8080808080808080) != 0) {
-      for (; pos < len; pos++) {
-        if (data[pos] >= 0b10000000) {
-          return result(error_code::TOO_LARGE, pos);
-        }
-      }
-    }
-  }
-  // process the tail byte-by-byte
-  for (; pos < len; pos++) {
-    if (data[pos] >= 0b10000000) {
-      return result(error_code::TOO_LARGE, pos);
-    }
-  }
-  return result(error_code::SUCCESS, pos);
-}
+        // 5. store bytes
+        vst1q_u8(utf8_output, utf8_packed);
 
-} // namespace ascii
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+        // 6. adjust pointers
+        buf += 8;
+        utf8_output += row[0];
+        continue;
+      } else {
+        // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+
+        // check for invalid input
+        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
+        const uint16x8_t forbidden_bytemask = vandq_u16(
+            vcleq_u16(utf16_packed, v_dfff), vcgeq_u16(utf16_packed, v_d800));
+        if (vmaxvq_u16(forbidden_bytemask) != 0) {
+          return std::make_pair(result(error_code::SURROGATE, buf - start),
+                                reinterpret_cast<char *>(utf8_output));
+        }
 
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t dup_even = simdutf_make_uint16x8_t(
+            0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#else
+        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
+                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
 #endif
-/* end file src/scalar/ascii.h */
-/* begin file src/scalar/latin1.h */
-#ifndef SIMDUTF_LATIN1_H
-#define SIMDUTF_LATIN1_H
+        /* In this branch we handle three cases:
+          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+          single UFT-8 byte
+          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              -
+          two UTF-8 bytes
+          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+          three UTF-8 bytes
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace latin1 {
+          We expand the input word (16-bit) into two code units (32-bit), thus
+          we have room for four bytes. However, we need five distinct bit
+          layouts. Note that the last byte in cases #2 and #3 is the same.
 
-inline size_t utf32_length_from_latin1(size_t len) {
-  // We are not BOM aware.
-  return len; // a utf32 unit will always represent 1 latin1 character
-}
+          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+          in register t2.
 
-inline size_t utf8_length_from_latin1(const char *buf, size_t len) {
-  const uint8_t *c = reinterpret_cast<const uint8_t *>(buf);
-  size_t answer = 0;
-  for (size_t i = 0; i < len; i++) {
-    if ((c[i] >> 7)) {
-      answer++;
-    }
-  }
-  return answer + len;
-}
+          We precompute byte 1 for case #3 and -- **conditionally** --
+          precompute either byte 1 for case #2 or byte 2 for case #3. Note that
+          they differ by exactly one bit.
+
+          Finally from these two code units we build proper UTF-8 sequence,
+          taking into account the case (i.e, the number of bytes to write).
+        */
+        /**
+         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+         * t2 => [0ccc|cccc] [10cc|cccc]
+         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+         */
+#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
+        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+        const uint16x8_t t0 =
+            vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed),
+                                            vreinterpretq_u8_u16(dup_even)));
+        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
+        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+        const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
 
-inline size_t utf16_length_from_latin1(size_t len) { return len; }
+        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
+        const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
+        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
+        const uint16x8_t s1 =
+            vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
+        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
+        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
+        // [00bb|bbbb|0000|aaaa]
+        const uint16x8_t s2 = vorrq_u16(s0, s1s);
+        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
+        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
+        const uint16x8_t one_or_two_bytes_bytemask =
+            vcleq_u16(utf16_packed, v_07ff);
+        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000),
+                                        one_or_two_bytes_bytemask);
+        const uint16x8_t s4 = veorq_u16(s3, m0);
+#undef simdutf_vec
 
-} // namespace latin1
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+        // 4. expand code units 16-bit => 32-bit
+        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
+        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
 
+        // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
+        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+        const uint16x8_t onemask = simdutf_make_uint16x8_t(
+            0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
+        const uint16x8_t twomask = simdutf_make_uint16x8_t(
+            0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
+#else
+        const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
+                                    0x0100, 0x0400, 0x1000, 0x4000};
+        const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
+                                    0x0200, 0x0800, 0x2000, 0x8000};
 #endif
-/* end file src/scalar/latin1.h */
+        const uint16x8_t combined =
+            vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
+                      vandq_u16(one_or_two_bytes_bytemask, twomask));
+        const uint16_t mask = vaddvq_u16(combined);
+        // The following fast path may or may not be beneficial.
+        /*if(mask == 0) {
+          // We only have three-byte code units. Use fast path.
+          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
+          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
+          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
+          vst1q_u8(utf8_output, utf8_0);
+          utf8_output += 12;
+          vst1q_u8(utf8_output, utf8_1);
+          utf8_output += 12;
+          buf += 8;
+          continue;
+        }*/
+        const uint8_t mask0 = uint8_t(mask);
 
-/* begin file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
-#ifndef SIMDUTF_VALID_UTF32_TO_UTF8_H
-#define SIMDUTF_VALID_UTF32_TO_UTF8_H
+        const uint8_t *row0 =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
+        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf32_to_utf8 {
+        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+        const uint8_t *row1 =
+            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
+        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
 
-#if SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64
-// only used by the fallback and POWER kernel
-inline size_t convert_valid(const char32_t *buf, size_t len,
-                            char *utf8_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 2 ASCII characters
-    if (pos + 2 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
-        *utf8_output++ = char(buf[pos]);
-        *utf8_output++ = char(buf[pos + 1]);
-        pos += 2;
-        continue;
+        vst1q_u8(utf8_output, utf8_0);
+        utf8_output += row0[0];
+        vst1q_u8(utf8_output, utf8_1);
+        utf8_output += row1[0];
+
+        buf += 8;
       }
-    }
-    uint32_t word = data[pos];
-    if ((word & 0xFFFFFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xFFFFF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xFFFF0000) == 0) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
+      // At least one 32-bit word will produce a surrogate pair in UTF-16 <=>
+      // will produce four UTF-8 bytes.
     } else {
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 18) | 0b11110000);
-      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) {
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k),
+                reinterpret_cast<char *>(utf8_output));
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          if (word > 0x10FFFF) {
+            return std::make_pair(
+                result(error_code::TOO_LARGE, buf - start + k),
+                reinterpret_cast<char *>(utf8_output));
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
     }
-  }
-  return utf8_output - start;
+  } // while
+
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        reinterpret_cast<char *>(utf8_output));
 }
-#endif // SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64
+/* end file src/arm64/arm_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF8
 
-} // namespace utf32_to_utf8
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
 
-#endif
-/* end file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
-/* begin file src/scalar/utf32_to_utf8/utf32_to_utf8.h */
-#ifndef SIMDUTF_UTF32_TO_UTF8_H
-#define SIMDUTF_UTF32_TO_UTF8_H
-
+/* begin file src/generic/buf_block_reader.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf32_to_utf8 {
 
-inline size_t convert(const char32_t *buf, size_t len, char *utf8_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 2 ASCII characters
-    if (pos + 2 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
-        *utf8_output++ = char(buf[pos]);
-        *utf8_output++ = char(buf[pos + 1]);
-        pos += 2;
-        continue;
-      }
-    }
-    uint32_t word = data[pos];
-    if ((word & 0xFFFFFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xFFFFF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xFFFF0000) == 0) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      if (word >= 0xD800 && word <= 0xDFFF) {
-        return 0;
-      }
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else {
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      if (word > 0x10FFFF) {
-        return 0;
-      }
-      *utf8_output++ = char((word >> 18) | 0b11110000);
-      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    }
-  }
-  return utf8_output - start;
-}
+// Walks through a buffer in block-sized increments, loading the last part with
+// spaces
+template <size_t STEP_SIZE> struct buf_block_reader {
+public:
+  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdutf_really_inline size_t block_index();
+  simdutf_really_inline bool has_full_block() const;
+  simdutf_really_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0
+   * (in which case this function fills the buffer with spaces and returns 0. In
+   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
+   * block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
+  simdutf_really_inline void advance();
 
-inline result convert_with_errors(const char32_t *buf, size_t len,
-                                  char *utf8_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 2 ASCII characters
-    if (pos + 2 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
-        *utf8_output++ = char(buf[pos]);
-        *utf8_output++ = char(buf[pos + 1]);
-        pos += 2;
-        continue;
-      }
-    }
-    uint32_t word = data[pos];
-    if ((word & 0xFFFFFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xFFFFF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xFFFF0000) == 0) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      if (word >= 0xD800 && word <= 0xDFFF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else {
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      if (word > 0x10FFFF) {
-        return result(error_code::TOO_LARGE, pos);
-      }
-      *utf8_output++ = char((word >> 18) | 0b11110000);
-      *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text_64(const uint8_t *text) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  in.store(reinterpret_cast<uint8_t *>(buf));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') {
+      buf[i] = '_';
     }
   }
-  return result(error_code::SUCCESS, utf8_output - start);
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
 }
 
-} // namespace utf32_to_utf8
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+simdutf_unused static char *format_mask(uint64_t mask) {
+  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
+  for (size_t i = 0; i < 64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
 
-#endif
-/* end file src/scalar/utf32_to_utf8/utf32_to_utf8.h */
+template <size_t STEP_SIZE>
+simdutf_really_inline
+buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
+    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
+      idx{0} {}
 
-/* begin file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
-#ifndef SIMDUTF_VALID_UTF32_TO_UTF16_H
-#define SIMDUTF_VALID_UTF32_TO_UTF16_H
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
+  return idx;
+}
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf32_to_utf16 {
+template <size_t STEP_SIZE>
+simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
 
-template <endianness big_endian>
-inline size_t convert_valid(const char32_t *buf, size_t len,
-                            char16_t *utf16_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  while (pos < len) {
-    uint32_t word = data[pos];
-    if ((word & 0xFFFF0000) == 0) {
-      // will not generate a surrogate pair
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(uint16_t(word)))
-                            : char16_t(word);
-      pos++;
-    } else {
-      // will generate a surrogate pair
-      word -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
-      }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
-      pos++;
-    }
-  }
-  return utf16_output - start;
+template <size_t STEP_SIZE>
+simdutf_really_inline const uint8_t *
+buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
 }
 
-} // namespace utf32_to_utf16
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t
+buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if (len == idx) {
+    return 0;
+  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20,
+              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
+                          // to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
 
-#endif
-/* end file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
-/* begin file src/scalar/utf32_to_utf16/utf32_to_utf16.h */
-#ifndef SIMDUTF_UTF32_TO_UTF16_H
-#define SIMDUTF_UTF32_TO_UTF16_H
+template <size_t STEP_SIZE>
+simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
 
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdutf
+/* end file src/generic/buf_block_reader.h */
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf32_to_utf16 {
+namespace utf8_validation {
 
-template <endianness big_endian>
-inline size_t convert(const char32_t *buf, size_t len, char16_t *utf16_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  while (pos < len) {
-    uint32_t word = data[pos];
-    if ((word & 0xFFFF0000) == 0) {
-      if (word >= 0xD800 && word <= 0xDFFF) {
-        return 0;
-      }
-      // will not generate a surrogate pair
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(uint16_t(word)))
-                            : char16_t(word);
-    } else {
-      // will generate a surrogate pair
-      if (word > 0x10FFFF) {
-        return 0;
-      }
-      word -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
-      }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
-    }
-    pos++;
-  }
-  return utf16_output - start;
+using namespace simd;
+
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
+
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-template <endianness big_endian>
-inline result convert_with_errors(const char32_t *buf, size_t len,
-                                  char16_t *utf16_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  while (pos < len) {
-    uint32_t word = data[pos];
-    if ((word & 0xFFFF0000) == 0) {
-      if (word >= 0xD800 && word <= 0xDFFF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      // will not generate a surrogate pair
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(uint16_t(word)))
-                            : char16_t(word);
+//
+// Return nonzero if there are incomplete multibyte characters at the end of the
+// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
+//
+simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  // If the previous input's last 3 bytes match this, they're too short (they
+  // ended at EOF):
+  // ... 1111____ 111_____ 11______
+  static const uint8_t max_array[32] = {255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        0b11110000u - 1,
+                                        0b11100000u - 1,
+                                        0b11000000u - 1};
+  const simd8<uint8_t> max_value(
+      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
+  return input.gt_bits(max_value);
+}
+
+struct utf8_checker {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+  // The last input we received
+  simd8<uint8_t> prev_input_block;
+  // Whether the last input we received was incomplete (used for ASCII fast
+  // path)
+  simd8<uint8_t> prev_incomplete;
+
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
+  }
+
+  // The only problem that can happen at EOF is that a multibyte character is
+  // too short or a byte value too large in the last bytes: check_special_cases
+  // only checks for bytes too large in the first of two bytes.
+  simdutf_really_inline void check_eof() {
+    // If the previous block had incomplete UTF-8 characters at the end, an
+    // ASCII block can't possibly finish them.
+    this->error |= this->prev_incomplete;
+  }
+
+  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
+    if (simdutf_likely(is_ascii(input))) {
+      this->error |= this->prev_incomplete;
     } else {
-      // will generate a surrogate pair
-      if (word > 0x10FFFF) {
-        return result(error_code::TOO_LARGE, pos);
-      }
-      word -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
+      // you might think that a for-loop would work, but under Visual Studio, it
+      // is not good enough.
+      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                    "We support either two or four chunks per 64-byte block.");
+      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
       }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
+      this->prev_incomplete =
+          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
+      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
     }
-    pos++;
   }
-  return result(error_code::SUCCESS, utf16_output - start);
-}
 
-} // namespace utf32_to_utf16
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+  // do not forget to call check_eof!
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
+  }
 
-#endif
-/* end file src/scalar/utf32_to_utf16/utf32_to_utf16.h */
+}; // struct utf8_checker
+} // namespace utf8_validation
 
-/* begin file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
-#ifndef SIMDUTF_VALID_UTF16_TO_UTF8_H
-#define SIMDUTF_VALID_UTF16_TO_UTF8_H
+using utf8_validation::utf8_checker;
 
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
+/* begin file src/generic/utf8_validation/utf8_validator.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf16_to_utf8 {
+namespace utf8_validation {
 
-template <endianness big_endian>
-inline size_t convert_valid(const char16_t *buf, size_t len,
-                            char *utf8_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 4 ASCII characters
-    if (pos + 4 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if (!match_system(big_endian)) {
-        v = (v >> 8) | (v << (64 - 8));
-      }
-      if ((v & 0xFF80FF80FF80FF80) == 0) {
-        size_t final_pos = pos + 4;
-        while (pos < final_pos) {
-          *utf8_output++ = !match_system(big_endian)
-                               ? char(utf16::swap_bytes(buf[pos]))
-                               : char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template <class checker>
+bool generic_validate_utf8(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    c.check_next_input(in);
+    reader.advance();
+  }
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
+  reader.advance();
+  c.check_eof();
+  return !c.errors();
+}
 
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xF800) != 0xD800) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((value >> 18) | 0b11110000);
-      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((value & 0b111111) | 0b10000000);
-      pos += 2;
+bool generic_validate_utf8(const char *input, size_t length) {
+  return generic_validate_utf8<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
+}
+
+/**
+ * Validates that the string is actual UTF-8 and stops on errors.
+ */
+template <class checker>
+result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
+  size_t count{0};
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    c.check_next_input(in);
+    if (c.errors()) {
+      if (count != 0) {
+        count--;
+      } // Sometimes the error is only detected in the next chunk
+      result res = scalar::utf8::rewind_and_validate_with_errors(
+          reinterpret_cast<const char *>(input),
+          reinterpret_cast<const char *>(input + count), length - count);
+      res.count += count;
+      return res;
     }
+    reader.advance();
+    count += 64;
+  }
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
+  reader.advance();
+  c.check_eof();
+  if (c.errors()) {
+    if (count != 0) {
+      count--;
+    } // Sometimes the error is only detected in the next chunk
+    result res = scalar::utf8::rewind_and_validate_with_errors(
+        reinterpret_cast<const char *>(input),
+        reinterpret_cast<const char *>(input) + count, length - count);
+    res.count += count;
+    return res;
+  } else {
+    return result(error_code::SUCCESS, length);
   }
-  return utf8_output - start;
 }
 
-} // namespace utf16_to_utf8
+result generic_validate_utf8_with_errors(const char *input, size_t length) {
+  return generic_validate_utf8_with_errors<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
+}
+
+} // namespace utf8_validation
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-#endif
-/* end file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
-/* begin file src/scalar/utf16_to_utf8/utf16_to_utf8.h */
-#ifndef SIMDUTF_UTF16_TO_UTF8_H
-#define SIMDUTF_UTF16_TO_UTF8_H
-
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf16_to_utf8 {
+namespace ascii_validation {
 
-template <endianness big_endian>
-inline size_t convert(const char16_t *buf, size_t len, char *utf8_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 8 bytes
-    if (pos + 4 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if (!match_system(big_endian)) {
-        v = (v >> 8) | (v << (64 - 8));
-      }
-      if ((v & 0xFF80FF80FF80FF80) == 0) {
-        size_t final_pos = pos + 4;
-        while (pos < final_pos) {
-          *utf8_output++ = !match_system(big_endian)
-                               ? char(utf16::swap_bytes(buf[pos]))
-                               : char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xF800) != 0xD800) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      if (pos + 1 >= len) {
-        return 0;
-      }
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return 0;
-      }
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return 0;
-      }
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((value >> 18) | 0b11110000);
-      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((value & 0b111111) | 0b10000000);
-      pos += 2;
-    }
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+  uint8_t blocks[64]{};
+  simd::simd8x64<uint8_t> running_or(blocks);
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    running_or |= in;
+    reader.advance();
   }
-  return utf8_output - start;
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  running_or |= in;
+  return running_or.is_ascii();
 }
 
-template <endianness big_endian>
-inline result convert_with_errors(const char16_t *buf, size_t len,
-                                  char *utf8_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char *start{utf8_output};
-  while (pos < len) {
-    // try to convert the next block of 8 bytes
-    if (pos + 4 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if (!match_system(big_endian))
-        v = (v >> 8) | (v << (64 - 8));
-      if ((v & 0xFF80FF80FF80FF80) == 0) {
-        size_t final_pos = pos + 4;
-        while (pos < final_pos) {
-          *utf8_output++ = !match_system(big_endian)
-                               ? char(utf16::swap_bytes(buf[pos]))
-                               : char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xFF80) == 0) {
-      // will generate one UTF-8 bytes
-      *utf8_output++ = char(word);
-      pos++;
-    } else if ((word & 0xF800) == 0) {
-      // will generate two UTF-8 bytes
-      // we have 0b110XXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 6) | 0b11000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else if ((word & 0xF800) != 0xD800) {
-      // will generate three UTF-8 bytes
-      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((word >> 12) | 0b11100000);
-      *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((word & 0b111111) | 0b10000000);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      if (pos + 1 >= len) {
-        return result(error_code::SURROGATE, pos);
-      }
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
-      }
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      // will generate four UTF-8 bytes
-      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
-      *utf8_output++ = char((value >> 18) | 0b11110000);
-      *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-      *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-      *utf8_output++ = char((value & 0b111111) | 0b10000000);
-      pos += 2;
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+  size_t count{0};
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    if (!in.is_ascii()) {
+      result res = scalar::ascii::validate_with_errors(
+          reinterpret_cast<const char *>(input + count), length - count);
+      return result(res.error, count + res.count);
     }
+    reader.advance();
+
+    count += 64;
+  }
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  if (!in.is_ascii()) {
+    result res = scalar::ascii::validate_with_errors(
+        reinterpret_cast<const char *>(input + count), length - count);
+    return result(res.error, count + res.count);
+  } else {
+    return result(error_code::SUCCESS, length);
   }
-  return result(error_code::SUCCESS, utf8_output - start);
 }
 
-} // namespace utf16_to_utf8
+} // namespace ascii_validation
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  // transcoding from UTF-8 to UTF-16
+/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+namespace simdutf {
+namespace arm64 {
+namespace {
+namespace utf8_to_utf16 {
+using namespace simd;
 
-#endif
-/* end file src/scalar/utf16_to_utf8/utf16_to_utf8.h */
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
 
-/* begin file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
-#ifndef SIMDUTF_VALID_UTF16_TO_UTF32_H
-#define SIMDUTF_VALID_UTF16_TO_UTF32_H
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf16_to_utf32 {
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
 
-template <endianness big_endian>
-inline size_t convert_valid(const char16_t *buf, size_t len,
-                            char32_t *utf32_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xF800) != 0xD800) {
-      // No surrogate pair, extend 16-bit word to 32-bit word
-      *utf32_output++ = char32_t(word);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      *utf32_output++ = char32_t(value);
-      pos += 2;
-    }
-  }
-  return utf32_output - start;
-}
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
 
-} // namespace utf16_to_utf32
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
 
-#endif
-/* end file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
-/* begin file src/scalar/utf16_to_utf32/utf16_to_utf32.h */
-#ifndef SIMDUTF_UTF16_TO_UTF32_H
-#define SIMDUTF_UTF16_TO_UTF32_H
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
+}
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf16_to_utf32 {
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
 
-template <endianness big_endian>
-inline size_t convert(const char16_t *buf, size_t len, char32_t *utf32_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xF800) != 0xD800) {
-      // No surrogate pair, extend 16-bit word to 32-bit word
-      *utf32_output++ = char32_t(word);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return 0;
-      }
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return 0;
-      }
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      *utf32_output++ = char32_t(value);
-      pos += 2;
-    }
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
-  return utf32_output - start;
-}
 
-template <endianness big_endian>
-inline result convert_with_errors(const char16_t *buf, size_t len,
-                                  char32_t *utf32_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    uint16_t word =
-        !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xF800) != 0xD800) {
-      // No surrogate pair, extend 16-bit word to 32-bit word
-      *utf32_output++ = char32_t(word);
-      pos++;
-    } else {
-      // must be a surrogate pair
-      uint16_t diff = uint16_t(word - 0xD800);
-      if (diff > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
+  template <endianness endian>
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
       }
-      if (pos + 1 >= len) {
-        return result(error_code::SURROGATE, pos);
-      } // minimal bound checking
-      uint16_t next_word = !match_system(big_endian)
-                               ? utf16::swap_bytes(data[pos + 1])
-                               : data[pos + 1];
-      uint16_t diff2 = uint16_t(next_word - 0xDC00);
-      if (diff2 > 0x3FF) {
-        return result(error_code::SURROGATE, pos);
+    }
+    if (errors()) {
+      return 0;
+    }
+    if (pos < size) {
+      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
+          in + pos, size - pos, utf16_output);
+      if (howmany == 0) {
+        return 0;
       }
-      uint32_t value = (diff << 10) + diff2 + 0x10000;
-      *utf32_output++ = char32_t(value);
-      pos += 2;
+      utf16_output += howmany;
     }
+    return utf16_output - start;
   }
-  return result(error_code::SUCCESS, utf32_output - start);
-}
-
-} // namespace utf16_to_utf32
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf16_to_utf32/utf16_to_utf32.h */
-
-/* begin file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
-#ifndef SIMDUTF_VALID_UTF8_TO_UTF16_H
-#define SIMDUTF_VALID_UTF8_TO_UTF16_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf8_to_utf16 {
 
-template <endianness big_endian>
-inline size_t convert_valid(const char *buf, size_t len,
-                            char16_t *utf16_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  while (pos < len) {
-    // try to convert the next block of 8 ASCII bytes
-    if (pos + 8 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 8;
-        while (pos < final_pos) {
-          *utf16_output++ = !match_system(big_endian)
-                                ? char16_t(utf16::swap_bytes(buf[pos]))
-                                : char16_t(buf[pos]);
-          pos++;
+  template <endianness endian>
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
         }
-        continue;
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res =
+              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+                  pos, in + pos, size - pos, utf16_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
       }
     }
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(leading_byte))
-                            : char16_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 1 >= len) {
-        break;
-      } // minimal bound checking
-      uint16_t code_point = uint16_t(((leading_byte & 0b00011111) << 6) |
-                                     (data[pos + 1] & 0b00111111));
-      if (!match_system(big_endian)) {
-        code_point = utf16::swap_bytes(uint16_t(code_point));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 2 >= len) {
-        break;
-      } // minimal bound checking
-      uint16_t code_point = uint16_t(((leading_byte & 0b00001111) << 12) |
-                                     ((data[pos + 1] & 0b00111111) << 6) |
-                                     (data[pos + 2] & 0b00111111));
-      if (!match_system(big_endian)) {
-        code_point = utf16::swap_bytes(uint16_t(code_point));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        break;
-      } // minimal bound checking
-      uint32_t code_point = ((leading_byte & 0b00000111) << 18) |
-                            ((data[pos + 1] & 0b00111111) << 12) |
-                            ((data[pos + 2] & 0b00111111) << 6) |
-                            (data[pos + 3] & 0b00111111);
-      code_point -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
+    if (errors()) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf16_output += res.count;
       }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
-      pos += 4;
-    } else {
-      // we may have a continuation but we do not do error checking
-      return 0;
     }
+    return result(error_code::SUCCESS, utf16_output - start);
+  }
+
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
-  return utf16_output - start;
-}
 
+}; // struct utf8_checker
 } // namespace utf8_to_utf16
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
-
-#endif
-/* end file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
-/* begin file src/scalar/utf8_to_utf16/utf8_to_utf16.h */
-#ifndef SIMDUTF_UTF8_TO_UTF16_H
-#define SIMDUTF_UTF8_TO_UTF16_H
-
+/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
 namespace utf8_to_utf16 {
 
-template <endianness big_endian>
-inline size_t convert(const char *buf, size_t len, char16_t *utf16_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *utf16_output++ = !match_system(big_endian)
-                                ? char16_t(utf16::swap_bytes(buf[pos]))
-                                : char16_t(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(leading_byte))
-                            : char16_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      // range check
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if (code_point < 0x80 || 0x7ff < code_point) {
-        return 0;
-      }
-      if (!match_system(big_endian)) {
-        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 2 >= len) {
-        return 0;
-      } // minimal bound checking
-
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      // range check
-      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
-                            (data[pos + 1] & 0b00111111) << 6 |
-                            (data[pos + 2] & 0b00111111);
-      if (code_point < 0x800 || 0xffff < code_point ||
-          (0xd7ff < code_point && code_point < 0xe000)) {
-        return 0;
-      }
-      if (!match_system(big_endian)) {
-        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        return 0;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-
-      // range check
-      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
-                            (data[pos + 1] & 0b00111111) << 12 |
-                            (data[pos + 2] & 0b00111111) << 6 |
-                            (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff || 0x10ffff < code_point) {
-        return 0;
-      }
-      code_point -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
-      }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
-      pos += 4;
-    } else {
-      return 0;
-    }
-  }
-  return utf16_output - start;
-}
+using namespace simd;
 
-template <endianness big_endian>
-inline result convert_with_errors(const char *buf, size_t len,
-                                  char16_t *utf16_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+template <endianness endian>
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char16_t *utf16_output) noexcept {
+  // The implementation is not specific to haswell and should be moved to the
+  // generic directory.
   size_t pos = 0;
   char16_t *start{utf16_output};
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *utf16_output++ = !match_system(big_endian)
-                                ? char16_t(utf16::swap_bytes(buf[pos]))
-                                : char16_t(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf16_output++ = !match_system(big_endian)
-                            ? char16_t(utf16::swap_bytes(leading_byte))
-                            : char16_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 1 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if (code_point < 0x80 || 0x7ff < code_point) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (!match_system(big_endian)) {
-        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8, it should become
-      // a single UTF-16 word.
-      if (pos + 2 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
-                            (data[pos + 1] & 0b00111111) << 6 |
-                            (data[pos + 2] & 0b00111111);
-      if ((code_point < 0x800) || (0xffff < code_point)) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0xd7ff < code_point && code_point < 0xe000) {
-        return result(error_code::SURROGATE, pos);
-      }
-      if (!match_system(big_endian)) {
-        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
-      }
-      *utf16_output++ = char16_t(code_point);
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-
-      // range check
-      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
-                            (data[pos + 1] & 0b00111111) << 12 |
-                            (data[pos + 2] & 0b00111111) << 6 |
-                            (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0x10ffff < code_point) {
-        return result(error_code::TOO_LARGE, pos);
-      }
-      code_point -= 0x10000;
-      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
-      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
-      if (!match_system(big_endian)) {
-        high_surrogate = utf16::swap_bytes(high_surrogate);
-        low_surrogate = utf16::swap_bytes(low_surrogate);
-      }
-      *utf16_output++ = char16_t(high_surrogate);
-      *utf16_output++ = char16_t(low_surrogate);
-      pos += 4;
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    // this loop could be unrolled further. For example, we could process the
+    // mask far more than 64 bytes.
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf16<endian>(utf16_output);
+      utf16_output += 64;
+      pos += 64;
     } else {
-      // we either have too many continuation bytes or an invalid leading byte
-      if ((leading_byte & 0b11000000) == 0b10000000) {
-        return result(error_code::TOO_LONG, pos);
-      } else {
-        return result(error_code::HEADER_BITS, pos);
-      }
-    }
-  }
-  return result(error_code::SUCCESS, utf16_output - start);
-}
-
-/**
- * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and
- * we have up to len input bytes left, and we encountered some error. It is
- * possible that the error is at 'buf' exactly, but it could also be in the
- * previous bytes  (up to 3 bytes back).
- *
- * prior_bytes indicates how many bytes, prior to 'buf' may belong to the
- * current memory section and can be safely accessed. We prior_bytes to access
- * safely up to three bytes before 'buf'.
- *
- * The caller is responsible to ensure that len > 0.
- *
- * If the error is believed to have occurred prior to 'buf', the count value
- * contain in the result will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
- */
-template <endianness endian>
-inline result rewind_and_convert_with_errors(size_t prior_bytes,
-                                             const char *buf, size_t len,
-                                             char16_t *utf16_output) {
-  size_t extra_len{0};
-  // We potentially need to go back in time and find a leading byte.
-  // In theory '3' would be sufficient, but sometimes the error can go back
-  // quite far.
-  size_t how_far_back = prior_bytes;
-  // size_t how_far_back = 3; // 3 bytes in the past + current position
-  // if(how_far_back >= prior_bytes) { how_far_back = prior_bytes; }
-  bool found_leading_bytes{false};
-  // important: it is i <= how_far_back and not 'i < how_far_back'.
-  for (size_t i = 0; i <= how_far_back; i++) {
-    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
-    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
-    if (found_leading_bytes) {
-      if (i > 0 && byte < 128) {
-        // If we had to go back and the leading byte is ascii
-        // then we can stop right away.
-        return result(error_code::TOO_LONG, 0 - i + 1);
+      // Slow path. We hope that the compiler will recognize that this is a slow
+      // path. Anything that is not a continuation mask is a 'leading byte',
+      // that is, the start of a new code point.
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      // The *start* of code points is not so useful, rather, we want the *end*
+      // of code points.
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
+      size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times when using solely
+      // the slow/regular path, and at least four times if there are fast paths.
+      while (pos < max_starting_point) {
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        //
+        // Thus we may allow convert_masked_utf8_to_utf16 to process
+        // more bytes at a time under a fast-path mode where 16 bytes
+        // are consumed at once (e.g., when encountering ASCII).
+        size_t consumed = convert_masked_utf8_to_utf16<endian>(
+            input + pos, utf8_end_of_code_point_mask, utf16_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
       }
-      buf -= i;
-      extra_len = i;
-      break;
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
     }
   }
-  //
-  // It is possible for this function to return a negative count in its result.
-  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
-  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
-  // unsigned integral type of the result of the sizeof operator
-  //
-  // An unsigned type will simply wrap round arithmetically (well defined).
-  //
-  if (!found_leading_bytes) {
-    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
-    // [....] [continuation] [continuation] [continuation] | [buf is
-    // continuation] Or we possibly have a stream that does not start with a
-    // leading byte.
-    return result(error_code::TOO_LONG, 0 - how_far_back);
-  }
-  result res = convert_with_errors<endian>(buf, len + extra_len, utf16_output);
-  if (res.error) {
-    res.count -= extra_len;
-  }
-  return res;
+  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
+      input + pos, size - pos, utf16_output);
+  return utf16_output - start;
 }
 
 } // namespace utf8_to_utf16
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
-
-#endif
-/* end file src/scalar/utf8_to_utf16/utf8_to_utf16.h */
-
-/* begin file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
-#ifndef SIMDUTF_VALID_UTF8_TO_UTF32_H
-#define SIMDUTF_VALID_UTF8_TO_UTF32_H
-
+/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  // transcoding from UTF-8 to UTF-32
+/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
 namespace utf8_to_utf32 {
+using namespace simd;
 
-inline size_t convert_valid(const char *buf, size_t len,
-                            char32_t *utf32_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    // try to convert the next block of 8 ASCII bytes
-    if (pos + 8 <=
-        len) { // if it is safe to read 8 more bytes, check that they are ascii
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 8;
-        while (pos < final_pos) {
-          *utf32_output++ = char32_t(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf32_output++ = char32_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        break;
-      } // minimal bound checking
-      *utf32_output++ = char32_t(((leading_byte & 0b00011111) << 6) |
-                                 (data[pos + 1] & 0b00111111));
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8
-      if (pos + 2 >= len) {
-        break;
-      } // minimal bound checking
-      *utf32_output++ = char32_t(((leading_byte & 0b00001111) << 12) |
-                                 ((data[pos + 1] & 0b00111111) << 6) |
-                                 (data[pos + 2] & 0b00111111));
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        break;
-      } // minimal bound checking
-      uint32_t code_word = ((leading_byte & 0b00000111) << 18) |
-                           ((data[pos + 1] & 0b00111111) << 12) |
-                           ((data[pos + 2] & 0b00111111) << 6) |
-                           (data[pos + 3] & 0b00111111);
-      *utf32_output++ = char32_t(code_word);
-      pos += 4;
-    } else {
-      // we may have a continuation but we do not do error checking
-      return 0;
-    }
-  }
-  return utf32_output - start;
-}
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
 
-} // namespace utf8_to_utf32
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-#endif
-/* end file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
-/* begin file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
-#ifndef SIMDUTF_UTF8_TO_UTF32_H
-#define SIMDUTF_UTF8_TO_UTF32_H
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf8_to_utf32 {
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
 
-inline size_t convert(const char *buf, size_t len, char32_t *utf32_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *utf32_output++ = char32_t(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf32_output++ = char32_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      // range check
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if (code_point < 0x80 || 0x7ff < code_point) {
-        return 0;
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8
-      if (pos + 2 >= len) {
-        return 0;
-      } // minimal bound checking
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
 
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      // range check
-      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
-                            (data[pos + 1] & 0b00111111) << 6 |
-                            (data[pos + 2] & 0b00111111);
-      if (code_point < 0x800 || 0xffff < code_point ||
-          (0xd7ff < code_point && code_point < 0xe000)) {
-        return 0;
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        return 0;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return 0;
-      }
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
+}
 
-      // range check
-      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
-                            (data[pos + 1] & 0b00111111) << 12 |
-                            (data[pos + 2] & 0b00111111) << 6 |
-                            (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff || 0x10ffff < code_point) {
-        return 0;
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 4;
-    } else {
-      return 0;
-    }
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
-  return utf32_output - start;
-}
 
-inline result convert_with_errors(const char *buf, size_t len,
-                                  char32_t *utf32_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2};
-      if ((v & 0x8080808080808080) == 0) {
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *utf32_output++ = char32_t(buf[pos]);
-          pos++;
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 16 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
         }
-        continue;
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // we have an error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
       }
     }
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *utf32_output++ = char32_t(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) == 0b11000000) {
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
-      if (code_point < 0x80 || 0x7ff < code_point) {
-        return result(error_code::OVERLONG, pos);
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8
-      if (pos + 2 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      // range check
-      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
-                            (data[pos + 1] & 0b00111111) << 6 |
-                            (data[pos + 2] & 0b00111111);
-      if (code_point < 0x800 || 0xffff < code_point) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0xd7ff < code_point && code_point < 0xe000) {
-        return result(error_code::SURROGATE, pos);
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 3;
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      if (pos + 3 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      }
-
-      // range check
-      uint32_t code_point = (leading_byte & 0b00000111) << 18 |
-                            (data[pos + 1] & 0b00111111) << 12 |
-                            (data[pos + 2] & 0b00111111) << 6 |
-                            (data[pos + 3] & 0b00111111);
-      if (code_point <= 0xffff) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0x10ffff < code_point) {
-        return result(error_code::TOO_LARGE, pos);
-      }
-      *utf32_output++ = char32_t(code_point);
-      pos += 4;
-    } else {
-      // we either have too many continuation bytes or an invalid leading byte
-      if ((leading_byte & 0b11000000) == 0b10000000) {
-        return result(error_code::TOO_LONG, pos);
-      } else {
-        return result(error_code::HEADER_BITS, pos);
+    if (errors()) {
+      return 0;
+    }
+    if (pos < size) {
+      size_t howmany =
+          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
+      if (howmany == 0) {
+        return 0;
       }
+      utf32_output += howmany;
     }
+    return utf32_output - start;
   }
-  return result(error_code::SUCCESS, utf32_output - start);
-}
 
-/**
- * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and
- * we have up to len input bytes left, and we encountered some error. It is
- * possible that the error is at 'buf' exactly, but it could also be in the
- * previous bytes location (up to 3 bytes back).
- *
- * prior_bytes indicates how many bytes, prior to 'buf' may belong to the
- * current memory section and can be safely accessed. We prior_bytes to access
- * safely up to three bytes before 'buf'.
- *
- * The caller is responsible to ensure that len > 0.
- *
- * If the error is believed to have occurred prior to 'buf', the count value
- * contain in the result will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
- */
-inline result rewind_and_convert_with_errors(size_t prior_bytes,
-                                             const char *buf, size_t len,
-                                             char32_t *utf32_output) {
-  size_t extra_len{0};
-  // We potentially need to go back in time and find a leading byte.
-  size_t how_far_back = 3; // 3 bytes in the past + current position
-  if (how_far_back > prior_bytes) {
-    how_far_back = prior_bytes;
-  }
-  bool found_leading_bytes{false};
-  // important: it is i <= how_far_back and not 'i < how_far_back'.
-  for (size_t i = 0; i <= how_far_back; i++) {
-    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
-    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
-    if (found_leading_bytes) {
-      if (i > 0 && byte < 128) {
-        // If we had to go back and the leading byte is ascii
-        // then we can stop right away.
-        return result(error_code::TOO_LONG, 0 - i + 1);
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, utf32_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
       }
-      buf -= i;
-      extra_len = i;
-      break;
     }
-  }
-  //
-  // It is possible for this function to return a negative count in its result.
-  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
-  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
-  // unsigned integral type of the result of the sizeof operator
-  //
-  // An unsigned type will simply wrap round arithmetically (well defined).
-  //
-  if (!found_leading_bytes) {
-    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
-    // [....] [continuation] [continuation] [continuation] | [buf is
-    // continuation] Or we possibly have a stream that does not start with a
-    // leading byte.
-    return result(error_code::TOO_LONG, 0 - how_far_back);
+    if (errors()) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf32_output += res.count;
+      }
+    }
+    return result(error_code::SUCCESS, utf32_output - start);
   }
 
-  result res = convert_with_errors(buf, len + extra_len, utf32_output);
-  if (res.error) {
-    res.count -= extra_len;
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
-  return res;
-}
 
+}; // struct utf8_checker
 } // namespace utf8_to_utf32
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
+/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+namespace simdutf {
+namespace arm64 {
+namespace {
+namespace utf8_to_utf32 {
 
-#endif
-/* end file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
+using namespace simd;
 
-/* begin file src/scalar/latin1_to_utf16/latin1_to_utf16.h */
-#ifndef SIMDUTF_LATIN1_TO_UTF16_H
-#define SIMDUTF_LATIN1_TO_UTF16_H
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char32_t *utf32_output) noexcept {
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf32(utf32_output);
+      utf32_output += 64;
+      pos += 64;
+    } else {
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      size_t max_starting_point = (pos + 64) - 12;
+      while (pos < max_starting_point) {
+        size_t consumed = convert_masked_utf8_to_utf32(
+            input + pos, utf8_end_of_code_point_mask, utf32_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
+    }
+  }
+  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
+                                                       utf32_output);
+  return utf32_output - start;
+}
 
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+// other functions
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/generic/utf16.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace latin1_to_utf16 {
+namespace utf16 {
 
 template <endianness big_endian>
-inline size_t convert(const char *buf, size_t len, char16_t *utf16_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+simdutf_really_inline size_t count_code_points(const char16_t *in,
+                                               size_t size) {
   size_t pos = 0;
-  char16_t *start{utf16_output};
-
-  while (pos < len) {
-    uint16_t word =
-        uint16_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point
-    *utf16_output++ =
-        char16_t(match_system(big_endian) ? word : utf16::swap_bytes(word));
-    pos++;
+  size_t count = 0;
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
+    }
+    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
+    count += count_ones(not_pair) / 2;
   }
-
-  return utf16_output - start;
+  return count +
+         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
 }
 
 template <endianness big_endian>
-inline result convert_with_errors(const char *buf, size_t len,
-                                  char16_t *utf16_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
+                                                    size_t size) {
   size_t pos = 0;
-  char16_t *start{utf16_output};
+  size_t count = 0;
+  // This algorithm could no doubt be improved!
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
+    }
+    uint64_t ascii_mask = input.lteq(0x7F);
+    uint64_t twobyte_mask = input.lteq(0x7FF);
+    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
 
-  while (pos < len) {
-    uint16_t word =
-        uint16_t(data[pos]); // extend Latin-1 char to 16-bit Unicode code point
-    *utf16_output++ =
-        char16_t(match_system(big_endian) ? word : utf16::swap_bytes(word));
-    pos++;
+    size_t ascii_count = count_ones(ascii_mask) / 2;
+    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
+    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
+    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
+    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
+             ascii_count;
   }
-
-  return result(error_code::SUCCESS, utf16_output - start);
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
 
-} // namespace latin1_to_utf16
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
 
-#endif
-/* end file src/scalar/latin1_to_utf16/latin1_to_utf16.h */
-/* begin file src/scalar/latin1_to_utf32/latin1_to_utf32.h */
-#ifndef SIMDUTF_LATIN1_TO_UTF32_H
-#define SIMDUTF_LATIN1_TO_UTF32_H
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace latin1_to_utf32 {
+  const auto one = vector_u16::splat(1);
 
-inline size_t convert(const char *buf, size_t len, char32_t *utf32_output) {
-  const unsigned char *data = reinterpret_cast<const unsigned char *>(buf);
-  char32_t *start{utf32_output};
-  for (size_t i = 0; i < len; i++) {
-    *utf32_output++ = (char32_t)data[i];
-  }
-  return utf32_output - start;
-}
+  auto v_count = vector_u16::zero();
 
-} // namespace latin1_to_utf32
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
 
-#endif
-/* end file src/scalar/latin1_to_utf32/latin1_to_utf32.h */
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
 
-/* begin file src/scalar/utf8_to_latin1/utf8_to_latin1.h */
-#ifndef SIMDUTF_UTF8_TO_LATIN1_H
-#define SIMDUTF_UTF8_TO_LATIN1_H
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf8_to_latin1 {
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
 
-inline size_t convert(const char *buf, size_t len, char *latin_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char *start{latin_output};
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
 
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000
-                           // 1000 1000 .... etc
-      if ((v & 0x8080808080808080) ==
-          0) { // if NONE of these are set, e.g. all of them are zero, then
-               // everything is ASCII
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *latin_output++ = char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
+    /*
+        Explanation how the counting works.
 
-    // suppose it is not an all ASCII byte sequence
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *latin_output++ = char(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) ==
-               0b11000000) { // the first three bits indicate:
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        return 0;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      } // checks if the next byte is a valid continuation byte in UTF-8. A
-        // valid continuation byte starts with 10.
-      // range check -
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 |
-          (data[pos + 1] &
-           0b00111111); // assembles the Unicode code point from the two bytes.
-                        // It does this by discarding the leading 110 and 10
-                        // bits from the two bytes, shifting the remaining bits
-                        // of the first byte, and then combining the results
-                        // with a bitwise OR operation.
-      if (code_point < 0x80 || 0xFF < code_point) {
-        return 0; // We only care about the range 129-255 which is Non-ASCII
-                  // latin1 characters. A code_point beneath 0x80 is invalid as
-                  // it is already covered by bytes whose leading bit is zero.
-      }
-      *latin_output++ = char(code_point);
-      pos += 2;
-    } else {
-      return 0;
-    }
-  }
-  return latin_output - start;
-}
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
 
-inline result convert_with_errors(const char *buf, size_t len,
-                                  char *latin_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
-  size_t pos = 0;
-  char *start{latin_output};
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
 
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 | v2}; // We are only interested in these bits: 1000 1000
-                           // 1000 1000...etc
-      if ((v & 0x8080808080808080) ==
-          0) { // if NONE of these are set, e.g. all of them are zero, then
-               // everything is ASCII
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *latin_output++ = char(buf[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    // suppose it is not an all ASCII byte sequence
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *latin_output++ = char(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) ==
-               0b11000000) { // the first three bits indicate:
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        return result(error_code::TOO_SHORT, pos);
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return result(error_code::TOO_SHORT, pos);
-      } // checks if the next byte is a valid continuation byte in UTF-8. A
-        // valid continuation byte starts with 10.
-      // range check -
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 |
-          (data[pos + 1] &
-           0b00111111); // assembles the Unicode code point from the two bytes.
-                        // It does this by discarding the leading 110 and 10
-                        // bits from the two bytes, shifting the remaining bits
-                        // of the first byte, and then combining the results
-                        // with a bitwise OR operation.
-      if (code_point < 0x80) {
-        return result(error_code::OVERLONG, pos);
-      }
-      if (0xFF < code_point) {
-        return result(error_code::TOO_LARGE, pos);
-      } // We only care about the range 129-255 which is Non-ASCII latin1
-        // characters
-      *latin_output++ = char(code_point);
-      pos += 2;
-    } else if ((leading_byte & 0b11110000) == 0b11100000) {
-      // We have a three-byte UTF-8
-      return result(error_code::TOO_LARGE, pos);
-    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
-      // we have a 4-byte UTF-8 word.
-      return result(error_code::TOO_LARGE, pos);
-    } else {
-      // we either have too many continuation bytes or an invalid leading byte
-      if ((leading_byte & 0b11000000) == 0b10000000) {
-        return result(error_code::TOO_LONG, pos);
-      }
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
 
-      return result(error_code::HEADER_BITS, pos);
-    }
-  }
-  return result(error_code::SUCCESS, latin_output - start);
-}
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
 
-inline result rewind_and_convert_with_errors(size_t prior_bytes,
-                                             const char *buf, size_t len,
-                                             char *latin1_output) {
-  size_t extra_len{0};
-  // We potentially need to go back in time and find a leading byte.
-  // In theory '3' would be sufficient, but sometimes the error can go back
-  // quite far.
-  size_t how_far_back = prior_bytes;
-  // size_t how_far_back = 3; // 3 bytes in the past + current position
-  // if(how_far_back >= prior_bytes) { how_far_back = prior_bytes; }
-  bool found_leading_bytes{false};
-  // important: it is i <= how_far_back and not 'i < how_far_back'.
-  for (size_t i = 0; i <= how_far_back; i++) {
-    unsigned char byte = buf[-static_cast<std::ptrdiff_t>(i)];
-    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
-    if (found_leading_bytes) {
-      if (i > 0 && byte < 128) {
-        // If we had to go back and the leading byte is ascii
-        // then we can stop right away.
-        return result(error_code::TOO_LONG, 0 - i + 1);
-      }
-      buf -= i;
-      extra_len = i;
-      break;
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
     }
   }
-  //
-  // It is possible for this function to return a negative count in its result.
-  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described
-  // in C Standard as <stddef.h>. C Standard Section 4.1.5 defines size_t as an
-  // unsigned integral type of the result of the sizeof operator
-  //
-  // An unsigned type will simply wrap round arithmetically (well defined).
-  //
-  if (!found_leading_bytes) {
-    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
-    // [....] [continuation] [continuation] [continuation] | [buf is
-    // continuation] Or we possibly have a stream that does not start with a
-    // leading byte.
-    return result(error_code::TOO_LONG, 0 - how_far_back);
+
+  if (iteration > 0) {
+    count += v_count.sum();
   }
-  result res = convert_with_errors(buf, len + extra_len, latin1_output);
-  if (res.error) {
-    res.count -= extra_len;
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
+}
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
+
+template <endianness big_endian>
+simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
+                                                     size_t size) {
+  return count_code_points<big_endian>(in, size);
+}
+
+simdutf_really_inline void
+change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
+  size_t pos = 0;
+
+  while (pos < size / 32 * 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    input.swap_bytes();
+    input.store(reinterpret_cast<uint16_t *>(output));
+    pos += 32;
+    output += 32;
   }
-  return res;
+
+  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
 }
 
-} // namespace utf8_to_latin1
+} // namespace utf16
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
-
-#endif
-/* end file src/scalar/utf8_to_latin1/utf8_to_latin1.h */
-/* begin file src/scalar/utf16_to_latin1/utf16_to_latin1.h */
-#ifndef SIMDUTF_UTF16_TO_LATIN1_H
-#define SIMDUTF_UTF16_TO_LATIN1_H
-
+/* end file src/generic/utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/generic/utf8.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf16_to_latin1 {
+namespace utf8 {
 
-#include <cstring> // for std::memcpy
+using namespace simd;
 
-template <endianness big_endian>
-inline size_t convert(const char16_t *buf, size_t len, char *latin_output) {
-  if (len == 0) {
-    return 0;
-  }
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
   size_t pos = 0;
-  char *current_write = latin_output;
-  uint16_t word = 0;
-  uint16_t too_large = 0;
-
-  while (pos < len) {
-    word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    too_large |= word;
-    *current_write++ = char(word & 0xFF);
-    pos++;
-  }
-  if ((too_large & 0xFF00) != 0) {
-    return 0;
+  size_t count = 0;
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.gt(-65);
+    count += count_ones(utf8_continuation_mask);
   }
-
-  return current_write - latin_output;
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
-template <endianness big_endian>
-inline result convert_with_errors(const char16_t *buf, size_t len,
-                                  char *latin_output) {
-  if (len == 0) {
-    return result(error_code::SUCCESS, 0);
-  }
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
+
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
+
   size_t pos = 0;
-  char *start{latin_output};
-  uint16_t word;
+  size_t count = 0;
 
-  while (pos < len) {
-    if (pos + 16 <= len) { // if it is safe to read 32 more bytes, check that
-                           // they are Latin1
-      uint64_t v1, v2, v3, v4;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      ::memcpy(&v2, data + pos + 4, sizeof(uint64_t));
-      ::memcpy(&v3, data + pos + 8, sizeof(uint64_t));
-      ::memcpy(&v4, data + pos + 12, sizeof(uint64_t));
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
 
-      if (!match_system(big_endian)) {
-        v1 = (v1 >> 8) | (v1 << (64 - 8));
-      }
-      if (!match_system(big_endian)) {
-        v2 = (v2 >> 8) | (v2 << (64 - 8));
-      }
-      if (!match_system(big_endian)) {
-        v3 = (v3 >> 8) | (v3 << (64 - 8));
-      }
-      if (!match_system(big_endian)) {
-        v4 = (v4 >> 8) | (v4 << (64 - 8));
-      }
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
 
-      if (((v1 | v2 | v3 | v4) & 0xFF00FF00FF00FF00) == 0) {
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *latin_output++ = !match_system(big_endian)
-                                ? char(utf16::swap_bytes(data[pos]))
-                                : char(data[pos]);
-          pos++;
-        }
-        continue;
-      }
-    }
-    word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    if ((word & 0xFF00) == 0) {
-      *latin_output++ = char(word & 0xFF);
-      pos++;
-    } else {
-      return result(error_code::TOO_LARGE, pos);
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
     }
   }
-  return result(error_code::SUCCESS, latin_output - start);
-}
-
-} // namespace utf16_to_latin1
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf16_to_latin1/utf16_to_latin1.h */
-/* begin file src/scalar/utf32_to_latin1/utf32_to_latin1.h */
-#ifndef SIMDUTF_UTF32_TO_LATIN1_H
-#define SIMDUTF_UTF32_TO_LATIN1_H
 
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf32_to_latin1 {
+  if (iterations > 0) {
+    count += local.sum_bytes();
+  }
 
-inline size_t convert(const char32_t *buf, size_t len, char *latin1_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  char *start = latin1_output;
-  uint32_t utf32_char;
-  size_t pos = 0;
-  uint32_t too_large = 0;
+  count += counters.sum();
 
-  while (pos < len) {
-    utf32_char = (uint32_t)data[pos];
-    too_large |= utf32_char;
-    *latin1_output++ = (char)(utf32_char & 0xFF);
-    pos++;
-  }
-  if ((too_large & 0xFFFFFF00) != 0) {
-    return 0;
-  }
-  return latin1_output - start;
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
+#endif
 
-inline result convert_with_errors(const char32_t *buf, size_t len,
-                                  char *latin1_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  char *start{latin1_output};
+simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
+                                                    size_t size) {
   size_t pos = 0;
-  while (pos < len) {
-    if (pos + 2 <=
-        len) { // if it is safe to read 8 more bytes, check that they are Latin1
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0xFFFFFF00FFFFFF00) == 0) {
-        *latin1_output++ = char(buf[pos]);
-        *latin1_output++ = char(buf[pos + 1]);
-        pos += 2;
-        continue;
-      }
-    }
-    uint32_t utf32_char = data[pos];
-    if ((utf32_char & 0xFFFFFF00) ==
-        0) { // Check if the character can be represented in Latin-1
-      *latin1_output++ = (char)(utf32_char & 0xFF);
-      pos++;
-    } else {
-      return result(error_code::TOO_LARGE, pos);
-    };
+  size_t count = 0;
+  // This algorithm could no doubt be improved!
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+    // We count one word for anything that is not a continuation (so
+    // leading bytes).
+    count += 64 - count_ones(utf8_continuation_mask);
+    int64_t utf8_4byte = input.gteq_unsigned(240);
+    count += count_ones(utf8_4byte);
   }
-  return result(error_code::SUCCESS, latin1_output - start);
+  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
 }
-
-} // namespace utf32_to_latin1
+} // namespace utf8
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
-
-#endif
-/* end file src/scalar/utf32_to_latin1/utf32_to_latin1.h */
-
-/* begin file src/scalar/utf8_to_latin1/valid_utf8_to_latin1.h */
-#ifndef SIMDUTF_VALID_UTF8_TO_LATIN1_H
-#define SIMDUTF_VALID_UTF8_TO_LATIN1_H
-
+/* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  // transcoding from UTF-8 to Latin 1
+/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
 namespace utf8_to_latin1 {
+using namespace simd;
 
-inline size_t convert_valid(const char *buf, size_t len, char *latin_output) {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // For UTF-8 to Latin 1, we can allow any ASCII character, and any
+  // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or
+  // 0b11000010 and nothing else.
+  //
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  constexpr const uint8_t FORBIDDEN = 0xff;
 
-  size_t pos = 0;
-  char *start{latin_output};
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      FORBIDDEN,
+      // 1110____ ________ <three byte lead in byte 1>
+      FORBIDDEN,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      FORBIDDEN);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-  while (pos < len) {
-    // try to convert the next block of 16 ASCII bytes
-    if (pos + 16 <=
-        len) { // if it is safe to read 16 more bytes, check that they are ascii
-      uint64_t v1;
-      ::memcpy(&v1, data + pos, sizeof(uint64_t));
-      uint64_t v2;
-      ::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
-      uint64_t v{v1 |
-                 v2}; // We are only interested in these bits: 1000 1000 1000
-                      // 1000, so it makes sense to concatenate everything
-      if ((v & 0x8080808080808080) ==
-          0) { // if NONE of these are set, e.g. all of them are zero, then
-               // everything is ASCII
-        size_t final_pos = pos + 16;
-        while (pos < final_pos) {
-          *latin_output++ = char(buf[pos]);
-          pos++;
+              // ____0100 ________
+              FORBIDDEN,
+              // ____0101 ________
+              FORBIDDEN,
+              // ____011_ ________
+              FORBIDDEN, FORBIDDEN,
+
+              // ____1___ ________
+              FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN,
+              // ____1101 ________
+              FORBIDDEN, FORBIDDEN, FORBIDDEN);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    this->error |= check_special_cases(input, prev1);
+  }
+
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char *latin1_output) {
+    size_t pos = 0;
+    char *start{latin1_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 16 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 16; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) >
+                       -65); // twos complement of -65 is 1011 1111 ...
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
         }
-        continue;
+        uint64_t utf8_continuation_mask =
+            input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                               // this case, we also have ASCII to account for.
+        if (utf8_continuation_mask & 1) {
+          return 0; // error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
       }
     }
-
-    // suppose it is not an all ASCII byte sequence
-    uint8_t leading_byte = data[pos]; // leading byte
-    if (leading_byte < 0b10000000) {
-      // converting one ASCII byte !!!
-      *latin_output++ = char(leading_byte);
-      pos++;
-    } else if ((leading_byte & 0b11100000) ==
-               0b11000000) { // the first three bits indicate:
-      // We have a two-byte UTF-8
-      if (pos + 1 >= len) {
-        break;
-      } // minimal bound checking
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) {
-        return 0;
-      } // checks if the next byte is a valid continuation byte in UTF-8. A
-        // valid continuation byte starts with 10.
-      // range check -
-      uint32_t code_point =
-          (leading_byte & 0b00011111) << 6 |
-          (data[pos + 1] &
-           0b00111111); // assembles the Unicode code point from the two bytes.
-                        // It does this by discarding the leading 110 and 10
-                        // bits from the two bytes, shifting the remaining bits
-                        // of the first byte, and then combining the results
-                        // with a bitwise OR operation.
-      *latin_output++ = char(code_point);
-      pos += 2;
-    } else {
-      // we may have a continuation but we do not do error checking
+    if (errors()) {
       return 0;
     }
+    if (pos < size) {
+      size_t howmany =
+          scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      latin1_output += howmany;
+    }
+    return latin1_output - start;
+  }
+
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char *latin1_output) {
+    size_t pos = 0;
+    char *start{latin1_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        if (errors()) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, latin1_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        latin1_output += res.count;
+      }
+    }
+    return result(error_code::SUCCESS, latin1_output - start);
   }
-  return latin_output - start;
-}
-
-} // namespace utf8_to_latin1
-} // unnamed namespace
-} // namespace scalar
-} // namespace simdutf
-
-#endif
-/* end file src/scalar/utf8_to_latin1/valid_utf8_to_latin1.h */
-/* begin file src/scalar/utf16_to_latin1/valid_utf16_to_latin1.h */
-#ifndef SIMDUTF_VALID_UTF16_TO_LATIN1_H
-#define SIMDUTF_VALID_UTF16_TO_LATIN1_H
-
-namespace simdutf {
-namespace scalar {
-namespace {
-namespace utf16_to_latin1 {
-
-template <endianness big_endian>
-inline size_t convert_valid(const char16_t *buf, size_t len,
-                            char *latin_output) {
-  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
-  size_t pos = 0;
-  char *start{latin_output};
-  uint16_t word = 0;
 
-  while (pos < len) {
-    word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
-    *latin_output++ = char(word);
-    pos++;
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
 
-  return latin_output - start;
-}
-
-} // namespace utf16_to_latin1
+}; // struct utf8_checker
+} // namespace utf8_to_latin1
 } // unnamed namespace
-} // namespace scalar
+} // namespace arm64
 } // namespace simdutf
-
-#endif
-/* end file src/scalar/utf16_to_latin1/valid_utf16_to_latin1.h */
-/* begin file src/scalar/utf32_to_latin1/valid_utf32_to_latin1.h */
-#ifndef SIMDUTF_VALID_UTF32_TO_LATIN1_H
-#define SIMDUTF_VALID_UTF32_TO_LATIN1_H
-
+/* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+/* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
 namespace simdutf {
-namespace scalar {
+namespace arm64 {
 namespace {
-namespace utf32_to_latin1 {
+namespace utf8_to_latin1 {
+using namespace simd;
 
-inline size_t convert_valid(const char32_t *buf, size_t len,
-                            char *latin1_output) {
-  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
-  char *start = latin1_output;
-  uint32_t utf32_char;
+simdutf_really_inline size_t convert_valid(const char *in, size_t size,
+                                           char *latin1_output) {
   size_t pos = 0;
-
-  while (pos < len) {
-    utf32_char = (uint32_t)data[pos];
-
-    if (pos + 2 <=
-        len) { // if it is safe to read 8 more bytes, check that they are Latin1
-      uint64_t v;
-      ::memcpy(&v, data + pos, sizeof(uint64_t));
-      if ((v & 0xFFFFFF00FFFFFF00) == 0) {
-        *latin1_output++ = char(buf[pos]);
-        *latin1_output++ = char(buf[pos + 1]);
-        pos += 2;
-        continue;
-      } else {
-        // output can not be represented in latin1
-        return 0;
-      }
-    }
-    if ((utf32_char & 0xFFFFFF00) == 0) {
-      *latin1_output++ = char(utf32_char);
+  char *start{latin1_output};
+  // In the worst case, we have the haswell kernel which can cause an overflow
+  // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the last
+  // 16 bytes, and if the data is valid, then it is entirely safe because 16
+  // UTF-8 bytes generate much more than 8 bytes. However, you cannot generally
+  // assume that you have valid UTF-8 input, so we are going to go back from the
+  // end counting 8 leading bytes, to give us a good margin.
+  size_t leading_byte = 0;
+  size_t margin = size;
+  for (; margin > 0 && leading_byte < 8; margin--) {
+    leading_byte += (int8_t(in[margin - 1]) >
+                     -65); // twos complement of -65 is 1011 1111 ...
+  }
+  // If the input is long enough, then we have that margin-1 is the eight last
+  // leading byte.
+  const size_t safety_margin = size - margin + 1; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    if (input.is_ascii()) {
+      input.store((int8_t *)latin1_output);
+      latin1_output += 64;
+      pos += 64;
     } else {
-      // output can not be represented in latin1
-      return 0;
+      // you might think that a for-loop would work, but under Visual Studio, it
+      // is not good enough.
+      uint64_t utf8_continuation_mask =
+          input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                             // this case, we also have ASCII to account for.
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
+      size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times.
+      while (pos < max_starting_point) {
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        size_t consumed = convert_masked_utf8_to_latin1(
+            in + pos, utf8_end_of_code_point_mask, latin1_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
     }
-    pos++;
+  }
+  if (pos < size) {
+    size_t howmany = scalar::utf8_to_latin1::convert_valid(in + pos, size - pos,
+                                                           latin1_output);
+    latin1_output += howmany;
   }
   return latin1_output - start;
 }
 
-} // namespace utf32_to_latin1
-} // unnamed namespace
-} // namespace scalar
+} // namespace utf8_to_latin1
+} // namespace
+} // namespace arm64
 } // namespace simdutf
+  // namespace simdutf
+/* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-#endif
-/* end file src/scalar/utf32_to_latin1/valid_utf32_to_latin1.h */
-
-SIMDUTF_PUSH_DISABLE_WARNINGS
-SIMDUTF_DISABLE_UNDESIRED_WARNINGS
-
-#if SIMDUTF_IMPLEMENTATION_ARM64
-/* begin file src/arm64/implementation.cpp */
-/* begin file src/simdutf/arm64/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "arm64"
-// #define SIMDUTF_IMPLEMENTATION arm64
-/* end file src/simdutf/arm64/begin.h */
+//
+// Implementation-specific overrides
+//
 namespace simdutf {
 namespace arm64 {
-namespace {
-#ifndef SIMDUTF_ARM64_H
-  #error "arm64.h must be included"
-#endif
-using namespace simd;
 
-simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
-  simd8<uint8_t> bits = input.reduce_or();
-  return bits.max_val() < 0b10000000u;
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused int
+implementation::detect_encodings(const char *input,
+                                 size_t length) const noexcept {
+  // If there is a BOM, then we trust it.
+  auto bom_encoding = simdutf::BOM::check_bom(input, length);
+  if (bom_encoding != encoding_type::unspecified) {
+    return bom_encoding;
+  }
+  // todo: reimplement as a one-pass algorithm.
+  int out = 0;
+  if (validate_utf8(input, length)) {
+    out |= encoding_type::UTF8;
+  }
+  if ((length % 2) == 0) {
+    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
+                         length / 2)) {
+      out |= encoding_type::UTF16_LE;
+    }
+  }
+  if ((length % 4) == 0) {
+    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
+      out |= encoding_type::UTF32_LE;
+    }
+  }
+  return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<bool> is_second_byte = prev1 >= uint8_t(0b11000000u);
-  simd8<bool> is_third_byte = prev2 >= uint8_t(0b11100000u);
-  simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
-  // Use ^ instead of | for is_*_byte, because ^ is commutative, and the caller
-  // is using ^ as well. This will work fine because we only have to report
-  // errors for cases with 0-1 lead bytes. Multiple lead bytes implies 2
-  // overlapping multibyte characters, and if that happens, there is guaranteed
-  // to be at least *one* lead byte that is part of only 1 other multibyte
-  // character. The error will be detected there.
-  return is_second_byte ^ is_third_byte ^ is_fourth_byte;
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return arm64::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-simdutf_really_inline simd8<bool>
-must_be_2_3_continuation(const simd8<uint8_t> prev2,
-                         const simd8<uint8_t> prev3) {
-  simd8<bool> is_third_byte = prev2 >= uint8_t(0b11100000u);
-  simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
-  return is_third_byte ^ is_fourth_byte;
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused result implementation::validate_utf8_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return arm64::utf8_validation::generic_validate_utf8_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
-// common functions for utf8 conversions
-simdutf_really_inline uint16x4_t convert_utf8_3_byte_to_utf16(uint8x16_t in) {
-  // Low half contains  10cccccc|1110aaaa
-  // High half contains 10bbbbbb|10bbbbbb
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  const uint8x16_t sh = simdutf_make_uint8x16_t(0, 2, 3, 5, 6, 8, 9, 11, 1, 1,
-                                                4, 4, 7, 7, 10, 10);
-#else
-  const uint8x16_t sh = {0, 2, 3, 5, 6, 8, 9, 11, 1, 1, 4, 4, 7, 7, 10, 10};
-#endif
-  uint8x16_t perm = vqtbl1q_u8(in, sh);
-  // Split into half vectors.
-  // 10cccccc|1110aaaa
-  uint8x8_t perm_low = vget_low_u8(perm); // no-op
-  // 10bbbbbb|10bbbbbb
-  uint8x8_t perm_high = vget_high_u8(perm);
-  // xxxxxxxx 10bbbbbb
-  uint16x4_t mid = vreinterpret_u16_u8(perm_high); // no-op
-  // xxxxxxxx 1110aaaa
-  uint16x4_t high = vreinterpret_u16_u8(perm_low); // no-op
-  // Assemble with shift left insert.
-  // xxxxxxaa aabbbbbb
-  uint16x4_t mid_high = vsli_n_u16(mid, high, 6);
-  // (perm_low << 8) | (perm_low >> 8)
-  // xxxxxxxx 10cccccc
-  uint16x4_t low = vreinterpret_u16_u8(vrev16_u8(perm_low));
-  // Shift left insert into the low bits
-  // aaaabbbb bbcccccc
-  uint16x4_t composed = vsli_n_u16(low, mid_high, 6);
-  return composed;
+#if SIMDUTF_FEATURE_ASCII
+simdutf_warn_unused bool
+implementation::validate_ascii(const char *buf, size_t len) const noexcept {
+  return arm64::ascii_validation::generic_validate_ascii(buf, len);
 }
 
-simdutf_really_inline uint16x8_t convert_utf8_2_byte_to_utf16(uint8x16_t in) {
-  // Converts 6 2 byte UTF-8 characters to 6 UTF-16 characters.
-  // Technically this calculates 8, but 6 does better and happens more often
-  // (The languages which use these codepoints use ASCII spaces so 8 would need
-  // to be in the middle of a very long word).
-
-  // 10bbbbbb 110aaaaa
-  uint16x8_t upper = vreinterpretq_u16_u8(in);
-  // (in << 8) | (in >> 8)
-  // 110aaaaa 10bbbbbb
-  uint16x8_t lower = vreinterpretq_u16_u8(vrev16q_u8(in));
-  // 00000000 000aaaaa
-  uint16x8_t upper_masked = vandq_u16(upper, vmovq_n_u16(0x1F));
-  // Assemble with shift left insert.
-  // 00000aaa aabbbbbb
-  uint16x8_t composed = vsliq_n_u16(lower, upper_masked, 6);
-  return composed;
+simdutf_warn_unused result implementation::validate_ascii_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return arm64::ascii_validation::generic_validate_ascii_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
-simdutf_really_inline uint16x8_t
-convert_utf8_1_to_2_byte_to_utf16(uint8x16_t in, size_t shufutf8_idx) {
-  // Converts 6 1-2 byte UTF-8 characters to 6 UTF-16 characters.
-  // This is a relatively easy scenario
-  // we process SIX (6) input code-code units. The max length in bytes of six
-  // code code units spanning between 1 and 2 bytes each is 12 bytes.
-  uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-      simdutf::tables::utf8_to_utf16::shufutf8[shufutf8_idx]));
-  // Shuffle
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 110aaaaa 10bbbbbb
-  uint16x8_t perm = vreinterpretq_u16_u8(vqtbl1q_u8(in, sh));
-  // Mask
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 00000000 00bbbbbb
-  uint16x8_t ascii = vandq_u16(perm, vmovq_n_u16(0x7f)); // 6 or 7 bits
-  // 1 byte: 00000000 00000000
-  // 2 byte: 000aaaaa 00000000
-  uint16x8_t highbyte = vandq_u16(perm, vmovq_n_u16(0x1f00)); // 5 bits
-  // Combine with a shift right accumulate
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 00000aaa aabbbbbb
-  uint16x8_t composed = vsraq_n_u16(ascii, highbyte, 2);
-  return composed;
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf16le(const char16_t *buf,
+                                 size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    // empty input is valid. protected the implementation from nullptr.
+    return true;
+  }
+  const char16_t *tail = arm_validate_utf16<endianness::LITTLE>(buf, len);
+  if (tail) {
+    return scalar::utf16::validate<endianness::LITTLE>(tail,
+                                                       len - (tail - buf));
+  } else {
+    return false;
+  }
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-/* begin file src/arm64/arm_validate_utf16.cpp */
-template <endianness big_endian>
-const char16_t *arm_validate_utf16(const char16_t *input, size_t size) {
-  const char16_t *end = input + size;
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-  while (end - input >= 16) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-    if (!match_system(big_endian)) {
-      in0 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in0)));
-      in1 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in1)));
-    }
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
-    const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
-    if (surrogates_wordmask == 0) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint64_t V = ~surrogates_wordmask;
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = ((in & v_fc) == v_dc);
-      const uint64_t H = vH.to_bitmask64();
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint64_t L = ~H & surrogates_wordmask;
-
-      const uint64_t a =
-          L & (H >> 4); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint64_t b =
-          a << 4; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint64_t c = V | a | b; // Combine all the masks into the final one.
-      if (c == ~0ull) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0xfffffffffffffffull) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return nullptr;
-      }
-    }
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool
+implementation::validate_utf16be(const char16_t *buf,
+                                 size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    // empty input is valid. protected the implementation from nullptr.
+    return true;
+  }
+  const char16_t *tail = arm_validate_utf16<endianness::BIG>(buf, len);
+  if (tail) {
+    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
+  } else {
+    return false;
   }
-  return input;
 }
 
-template <endianness big_endian>
-const result arm_validate_utf16_with_errors(const char16_t *input,
-                                            size_t size) {
-  const char16_t *start = input;
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-  while (input + 16 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-
-    if (!match_system(big_endian)) {
-      in0 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in0)));
-      in1 = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in1)));
-    }
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
-    const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
-    if (surrogates_wordmask == 0) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint64_t V = ~surrogates_wordmask;
+simdutf_warn_unused result implementation::validate_utf16le_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+  result res = arm_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.count != len) {
+    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
+        buf + res.count, len - res.count);
+    return result(scalar_res.error, res.count + scalar_res.count);
+  } else {
+    return res;
+  }
+}
 
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = ((in & v_fc) == v_dc);
-      const uint64_t H = vH.to_bitmask64();
+simdutf_warn_unused result implementation::validate_utf16be_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+  result res = arm_validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.count != len) {
+    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
+        buf + res.count, len - res.count);
+    return result(scalar_res.error, res.count + scalar_res.count);
+  } else {
+    return res;
+  }
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint64_t L = ~H & surrogates_wordmask;
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    // empty input is valid. protected the implementation from nullptr.
+    return true;
+  }
+  const char32_t *tail = arm_validate_utf32le(buf, len);
+  if (tail) {
+    return scalar::utf32::validate(tail, len - (tail - buf));
+  } else {
+    return false;
+  }
+}
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-      const uint64_t a =
-          L & (H >> 4); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint64_t b =
-          a << 4; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint64_t c = V | a | b; // Combine all the masks into the final one.
-      if (c == ~0ull) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0xfffffffffffffffull) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return result(error_code::SURROGATE, input - start);
-      }
-    }
+#if SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused result implementation::validate_utf32_with_errors(
+    const char32_t *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+  result res = arm_validate_utf32le_with_errors(buf, len);
+  if (res.count != len) {
+    result scalar_res =
+        scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
+    return result(scalar_res.error, res.count + scalar_res.count);
+  } else {
+    return res;
   }
-  return result(error_code::SUCCESS, input - start);
 }
-/* end file src/arm64/arm_validate_utf16.cpp */
-/* begin file src/arm64/arm_validate_utf32le.cpp */
+#endif // SIMDUTF_FEATURE_UTF32
 
-const char32_t *arm_validate_utf32le(const char32_t *input, size_t size) {
-  const char32_t *end = input + size;
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
+    const char *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char *, char *> ret =
+      arm_convert_latin1_to_utf8(buf, len, utf8_output);
+  size_t converted_chars = ret.second - utf8_output;
 
-  const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
-  const uint32x4_t offset = vmovq_n_u32(0xffff2000);
-  const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
-  uint32x4_t currentmax = vmovq_n_u32(0x0);
-  uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
+  }
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-  while (end - input >= 4) {
-    const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input));
-    currentmax = vmaxq_u32(in, currentmax);
-    currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);
-    input += 4;
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char *, char16_t *> ret =
+      arm_convert_latin1_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  size_t converted_chars = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars =
+        scalar::latin1_to_utf16::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
   }
+  return converted_chars;
+}
 
-  uint32x4_t is_zero =
-      veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
-  if (vmaxvq_u32(is_zero) != 0) {
-    return nullptr;
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char *, char16_t *> ret =
+      arm_convert_latin1_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  size_t converted_chars = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars =
+        scalar::latin1_to_utf16::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
   }
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-  is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax),
-                      standardoffsetmax);
-  if (vmaxvq_u32(is_zero) != 0) {
-    return nullptr;
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char *, char32_t *> ret =
+      arm_convert_latin1_to_utf32(buf, len, utf32_output);
+  size_t converted_chars = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
   }
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-  return input;
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert(buf, len, latin1_output);
 }
 
-const result arm_validate_utf32le_with_errors(const char32_t *input,
-                                              size_t size) {
-  const char32_t *start = input;
-  const char32_t *end = input + size;
+simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, latin1_output);
+}
 
-  const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
-  const uint32x4_t offset = vmovq_n_u32(0xffff2000);
-  const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
-  uint32x4_t currentmax = vmovq_n_u32(0x0);
-  uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return arm64::utf8_to_latin1::convert_valid(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-  while (end - input >= 4) {
-    const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input));
-    currentmax = vmaxq_u32(in, currentmax);
-    currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    uint32x4_t is_zero =
-        veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
-    if (vmaxvq_u32(is_zero) != 0) {
-      return result(error_code::TOO_LARGE, input - start);
-    }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::BIG>(buf, len, utf16_output);
+}
 
-    is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax),
-                        standardoffsetmax);
-    if (vmaxvq_u32(is_zero) != 0) {
-      return result(error_code::SURROGATE, input - start);
-    }
+simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::LITTLE>(buf, len,
+                                                           utf16_output);
+}
 
-    input += 4;
-  }
+simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
+}
 
-  return result(error_code::SUCCESS, input - start);
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
+    const char *input, size_t size, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,
+                                                          utf16_output);
 }
-/* end file src/arm64/arm_validate_utf32le.cpp */
 
-/* begin file src/arm64/arm_convert_latin1_to_utf16.cpp */
-template <endianness big_endian>
-std::pair<const char *, char16_t *>
-arm_convert_latin1_to_utf16(const char *buf, size_t len,
-                            char16_t *utf16_output) {
-  const char *end = buf + len;
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
+    const char *input, size_t size, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
+                                                       utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  while (end - buf >= 16) {
-    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(buf));
-    uint16x8_t inlow = vmovl_u8(vget_low_u8(in8));
-    if (!match_system(big_endian)) {
-      inlow = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(inlow)));
-    }
-    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), inlow);
-    uint16x8_t inhigh = vmovl_u8(vget_high_u8(in8));
-    if (!match_system(big_endian)) {
-      inhigh = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(inhigh)));
-    }
-    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output + 8), inhigh);
-    utf16_output += 16;
-    buf += 16;
-  }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert(buf, len, utf32_output);
+}
 
-  return std::make_pair(buf, utf16_output);
+simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, utf32_output);
 }
-/* end file src/arm64/arm_convert_latin1_to_utf16.cpp */
-/* begin file src/arm64/arm_convert_latin1_to_utf32.cpp */
-std::pair<const char *, char32_t *>
-arm_convert_latin1_to_utf32(const char *buf, size_t len,
-                            char32_t *utf32_output) {
-  const char *end = buf + len;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-  while (end - buf >= 16) {
-    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(buf));
-    uint16x8_t in8low = vmovl_u8(vget_low_u8(in8));
-    uint32x4_t in16lowlow = vmovl_u16(vget_low_u16(in8low));
-    uint32x4_t in16lowhigh = vmovl_u16(vget_high_u16(in8low));
-    uint16x8_t in8high = vmovl_u8(vget_high_u8(in8));
-    uint32x4_t in8highlow = vmovl_u16(vget_low_u16(in8high));
-    uint32x4_t in8highhigh = vmovl_u16(vget_high_u16(in8high));
-    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output), in16lowlow);
-    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 4), in16lowhigh);
-    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 8), in8highlow);
-    vst1q_u32(reinterpret_cast<uint32_t *>(utf32_output + 12), in8highhigh);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
+    const char *input, size_t size, char32_t *utf32_output) const noexcept {
+  return utf8_to_utf32::convert_valid(input, size, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-    utf32_output += 16;
-    buf += 16;
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      arm_convert_utf16_to_latin1<endianness::LITTLE>(buf, len, latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
   }
+  size_t saved_bytes = ret.second - latin1_output;
 
-  return std::make_pair(buf, utf32_output);
-}
-/* end file src/arm64/arm_convert_latin1_to_utf32.cpp */
-/* begin file src/arm64/arm_convert_latin1_to_utf8.cpp */
-/*
-  Returns a pair: the first unprocessed byte from buf and utf8_output
-  A scalar routing should carry on the conversion of the tail.
-*/
-std::pair<const char *, char *>
-arm_convert_latin1_to_utf8(const char *latin1_input, size_t len,
-                           char *utf8_out) {
-  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
-  const char *end = latin1_input + len;
-  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
-  // We always write 16 bytes, of which more than the first 8 bytes
-  // are valid. A safety margin of 8 is more than sufficient.
-  while (end - latin1_input >= 16 + 8) {
-    uint8x16_t in8 = vld1q_u8(reinterpret_cast<const uint8_t *>(latin1_input));
-    if (vmaxvq_u8(in8) <= 0x7F) { // ASCII fast path!!!!
-      vst1q_u8(utf8_output, in8);
-      utf8_output += 16;
-      latin1_input += 16;
-      continue;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_latin1::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // We just fallback on UTF-16 code. This could be optimized/simplified
-    // further.
-    uint16x8_t in16 = vmovl_u8(vget_low_u8(in8));
-    // 1. prepare 2-byte values
-    // input 8-bit word : [aabb|bbbb] x 8
-    // expected output   : [1100|00aa|10bb|bbbb] x 8
-    const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
-    const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      arm_convert_utf16_to_latin1<endianness::BIG>(buf, len, latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - latin1_output;
 
-    // t0 = [0000|00aa|bbbb|bb00]
-    const uint16x8_t t0 = vshlq_n_u16(in16, 2);
-    // t1 = [0000|00aa|0000|0000]
-    const uint16x8_t t1 = vandq_u16(t0, v_1f00);
-    // t2 = [0000|0000|00bb|bbbb]
-    const uint16x8_t t2 = vandq_u16(in16, v_003f);
-    // t3 = [0000|00aa|00bb|bbbb]
-    const uint16x8_t t3 = vorrq_u16(t1, t2);
-    // t4 = [1100|00aa|10bb|bbbb]
-    const uint16x8_t t4 = vorrq_u16(t3, v_c080);
-    // 2. merge ASCII and 2-byte codewords
-    const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-    const uint16x8_t one_byte_bytemask = vcleq_u16(in16, v_007f);
-    const uint8x16_t utf8_unpacked =
-        vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in16, t4));
-    // 3. prepare bitmask for 8-bit lookup
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-    const uint16x8_t mask = simdutf_make_uint16x8_t(
-        0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
-#else
-    const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
-                             0x0002, 0x0008, 0x0020, 0x0080};
-#endif
-    uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
-    // 4. pack the bytes
-    const uint8_t *row =
-        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
-    const uint8x16_t shuffle = vld1q_u8(row + 1);
-    const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_latin1::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // 5. store bytes
-    vst1q_u8(utf8_output, utf8_packed);
-    // 6. adjust pointers
-    latin1_input += 8;
-    utf8_output += row[0];
+simdutf_warn_unused result
+implementation::convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<result, char *> ret =
+      arm_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
+          buf, len, latin1_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  } // while
+simdutf_warn_unused result
+implementation::convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<result, char *> ret =
+      arm_convert_utf16_to_latin1_with_errors<endianness::BIG>(buf, len,
+                                                               latin1_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  return std::make_pair(latin1_input, reinterpret_cast<char *>(utf8_output));
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement a custom function.
+  return convert_utf16be_to_latin1(buf, len, latin1_output);
 }
-/* end file src/arm64/arm_convert_latin1_to_utf8.cpp */
 
-/* begin file src/arm64/arm_convert_utf8_to_latin1.cpp */
-// Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
-// end of the code points. Only the least significant 12 bits of the mask
-// are accessed.
-// It returns how many bytes were consumed (up to 16, usually 12).
-size_t convert_masked_utf8_to_latin1(const char *input,
-                                     uint64_t utf8_end_of_code_point_mask,
-                                     char *&latin1_output) {
-  // we use an approach where we try to process up to 12 input bytes.
-  // Why 12 input bytes and not 16? Because we are concerned with the size of
-  // the lookup tables. Also 12 is nicely divisible by two and three.
-  //
-  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
-  const uint16_t input_utf8_end_of_code_point_mask =
-      utf8_end_of_code_point_mask & 0xfff;
-  //
-  // Optimization note: our main path below is load-latency dependent. Thus it
-  // is maybe beneficial to have fast paths that depend on branch prediction but
-  // have less latency. This results in more instructions but, potentially, also
-  // higher speeds.
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement a custom function.
+  return convert_utf16le_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-  // We first try a few fast paths.
-  // The obvious first test is ASCII, which actually consumes the full 16.
-  if (utf8_end_of_code_point_mask == 0xfff) {
-    // We process in chunks of 12 bytes
-    vst1q_u8(reinterpret_cast<uint8_t *>(latin1_output), in);
-    latin1_output += 12; // We wrote 12 18-bit characters.
-    return 12;           // We consumed 12 bytes.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      arm_convert_utf16_to_utf8<endianness::LITTLE>(buf, len, utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
   }
-  /// We do not have a fast path available, or the fast path is unimportant, so
-  /// we fallback.
-  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][0];
-
-  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][1];
-  // this indicates an invalid input:
-  if (idx >= 64) {
-    return consumed;
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf8::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
   }
-  // Here we should have (idx < 64), if not, there is a bug in the validation or
-  // elsewhere. SIX (6) input code-code units this is a relatively easy scenario
-  // we process SIX (6) input code-code units. The max length in bytes of six
-  // code code units spanning between 1 and 2 bytes each is 12 bytes. Converts 6
-  // 1-2 byte UTF-8 characters to 6 UTF-16 characters. This is a relatively easy
-  // scenario we process SIX (6) input code-code units. The max length in bytes
-  // of six code code units spanning between 1 and 2 bytes each is 12 bytes.
-  uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-      simdutf::tables::utf8_to_utf16::shufutf8[idx]));
-  // Shuffle
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 110aaaaa 10bbbbbb
-  uint16x8_t perm = vreinterpretq_u16_u8(vqtbl1q_u8(in, sh));
-  // Mask
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 00000000 00bbbbbb
-  uint16x8_t ascii = vandq_u16(perm, vmovq_n_u16(0x7f)); // 6 or 7 bits
-  // 1 byte: 00000000 00000000
-  // 2 byte: 000aaaaa 00000000
-  uint16x8_t highbyte = vandq_u16(perm, vmovq_n_u16(0x1f00)); // 5 bits
-  // Combine with a shift right accumulate
-  // 1 byte: 00000000 0bbbbbbb
-  // 2 byte: 00000aaa aabbbbbb
-  uint16x8_t composed = vsraq_n_u16(ascii, highbyte, 2);
-  // writing 8 bytes even though we only care about the first 6 bytes.
-  uint8x8_t latin1_packed = vmovn_u16(composed);
-  vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
-  latin1_output += 6; // We wrote 6 bytes.
-  return consumed;
+  return saved_bytes;
 }
-/* end file src/arm64/arm_convert_utf8_to_latin1.cpp */
-/* begin file src/arm64/arm_convert_utf8_to_utf16.cpp */
-// Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
-// end of the code points. Only the least significant 12 bits of the mask
-// are accessed.
-// It returns how many bytes were consumed (up to 16, usually 12).
-template <endianness big_endian>
-size_t convert_masked_utf8_to_utf16(const char *input,
-                                    uint64_t utf8_end_of_code_point_mask,
-                                    char16_t *&utf16_output) {
-  // we use an approach where we try to process up to 12 input bytes.
-  // Why 12 input bytes and not 16? Because we are concerned with the size of
-  // the lookup tables. Also 12 is nicely divisible by two and three.
-  //
-  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
-  const uint16_t input_utf8_end_of_code_point_mask =
-      utf8_end_of_code_point_mask & 0xfff;
-  //
-  // Optimization note: our main path below is load-latency dependent. Thus it
-  // is maybe beneficial to have fast paths that depend on branch prediction but
-  // have less latency. This results in more instructions but, potentially, also
-  // higher speeds.
 
-  // We first try a few fast paths.
-  // The obvious first test is ASCII, which actually consumes the full 16.
-  if ((utf8_end_of_code_point_mask & 0xFFFF) == 0xffff) {
-    // We process in chunks of 16 bytes
-    // The routine in simd.h is reused.
-    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
-    temp.store_ascii_as_utf16<big_endian>(utf16_output);
-    utf16_output += 16; // We wrote 16 16-bit characters.
-    return 16;          // We consumed 16 bytes.
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      arm_convert_utf16_to_utf8<endianness::BIG>(buf, len, utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf8::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
   }
+  return saved_bytes;
+}
 
-  // 3 byte sequences are the next most common, as seen in CJK, which has long
-  // sequences of these.
-  if (input_utf8_end_of_code_point_mask == 0x924) {
-    // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte
-    // UTF-16 code units.
-    uint16x4_t composed = convert_utf8_3_byte_to_utf16(in);
-    // Byte swap if necessary
-    if (!match_system(big_endian)) {
-      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      arm_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(buf, len,
+                                                                utf8_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
     }
-    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
-    utf16_output += 4; // We wrote 4 16-bit characters.
-    return 12;         // We consumed 12 bytes.
   }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
-  if ((utf8_end_of_code_point_mask & 0xFFF) == 0xaaa) {
-    // We want to take 6 2-byte UTF-8 code units and turn them into 6 2-byte
-    // UTF-16 code units.
-    uint16x8_t composed = convert_utf8_2_byte_to_utf16(in);
-    // Byte swap if necessary
-    if (!match_system(big_endian)) {
-      composed =
-          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      arm_convert_utf16_to_utf8_with_errors<endianness::BIG>(buf, len,
+                                                             utf8_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
     }
-    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
-
-    utf16_output += 6; // We wrote 6 16-bit characters.
-    return 12;         // We consumed 12 bytes.
   }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  /// We do not have a fast path available, or the fast path is unimportant, so
-  /// we fallback.
-  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][0];
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16le_to_utf8(buf, len, utf8_output);
+}
 
-  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][1];
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16be_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  if (idx < 64) {
-    // SIX (6) input code-code units
-    // Convert to UTF-16
-    uint16x8_t composed = convert_utf8_1_to_2_byte_to_utf16(in, idx);
-    // Byte swap if necessary
-    if (!match_system(big_endian)) {
-      composed =
-          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
-    }
-    // Store
-    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
-    utf16_output += 6; // We wrote 6 16-bit characters.
-    return consumed;
-  } else if (idx < 145) {
-    // FOUR (4) input code-code units
-    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
-    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
-    // XXX: depending on the system scalar instructions might be faster.
-    // 1 byte: 00000000 00000000 0ccccccc
-    // 2 byte: 00000000 110bbbbb 10cccccc
-    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
-    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
-    // 1 byte: 00000000 0ccccccc
-    // 2 byte: xx0bbbbb x0cccccc
-    // 3 byte: xxbbbbbb x0cccccc
-    uint16x4_t lowperm = vmovn_u32(perm);
-    // Partially mask with bic (doesn't require a temporary register unlike and)
-    // The shift left insert below will clear the top bits.
-    // 1 byte: 00000000 00000000
-    // 2 byte: xx0bbbbb 00000000
-    // 3 byte: xxbbbbbb 00000000
-    uint16x4_t middlebyte = vbic_u16(lowperm, vmov_n_u16(uint16_t(~0xFF00)));
-    // ASCII
-    // 1 byte: 00000000 0ccccccc
-    // 2+byte: 00000000 00cccccc
-    uint16x4_t ascii = vand_u16(lowperm, vmov_n_u16(0x7F));
-    // Split into narrow vectors.
-    // 2 byte: 00000000 00000000
-    // 3 byte: 00000000 xxxxaaaa
-    uint16x4_t highperm = vshrn_n_u32(perm, 16);
-    // Shift right accumulate the middle byte
-    // 1 byte: 00000000 0ccccccc
-    // 2 byte: 00xx0bbb bbcccccc
-    // 3 byte: 00xxbbbb bbcccccc
-    uint16x4_t middlelow = vsra_n_u16(ascii, middlebyte, 2);
-    // Shift left and insert the top 4 bits, overwriting the garbage
-    // 1 byte: 00000000 0ccccccc
-    // 2 byte: 00000bbb bbcccccc
-    // 3 byte: aaaabbbb bbcccccc
-    uint16x4_t composed = vsli_n_u16(middlelow, highperm, 12);
-    // Byte swap if necessary
-    if (!match_system(big_endian)) {
-      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return 0;
+  }
+  std::pair<const char32_t *, char *> ret =
+      arm_convert_utf32_to_utf8(buf, len, utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    utf16_output += 4; // We wrote 4 16-bit codepoints
-    return consumed;
-  } else if (idx < 209) {
-    // THREE (3) input code-code units
-    if (input_utf8_end_of_code_point_mask == 0x888) {
-      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
-      // UTF-16 pairs. Generating surrogate pairs is a little tricky though, but
-      // it is easier when we can assume they are all pairs. This version does
-      // not use the LUT, but 4 byte sequences are less common and the overhead
-      // of the extra memory access is less important than the early branch
-      // overhead in shorter sequences.
+simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      arm_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
+  if (ret.first.count != len) {
+    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
+        buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-      // Swap byte pairs
-      // 10dddddd 10cccccc|10bbbbbb 11110aaa
-      // 10cccccc 10dddddd|11110aaa 10bbbbbb
-      uint8x16_t swap = vrev16q_u8(in);
-      // Shift left 2 bits
-      // cccccc00 dddddd00 xxxxxxxx bbbbbb00
-      uint32x4_t shift = vreinterpretq_u32_u8(vshlq_n_u8(swap, 2));
-      // Create a magic number containing the low 2 bits of the trail surrogate
-      // and all the corrections needed to create the pair. UTF-8 4b prefix   =
-      // -0x0000|0xF000 surrogate offset  = -0x0000|0x0040 (0x10000 << 6)
-      // surrogate high    = +0x0000|0xD800
-      // surrogate low     = +0xDC00|0x0000
-      // -------------------------------
-      //                   = +0xDC00|0xE7C0
-      uint32x4_t magic = vmovq_n_u32(0xDC00E7C0);
-      // Generate unadjusted trail surrogate minus lowest 2 bits
-      // xxxxxxxx xxxxxxxx|11110aaa bbbbbb00
-      uint32x4_t trail =
-          vbslq_u32(vmovq_n_u32(0x0000FF00), vreinterpretq_u32_u8(swap), shift);
-      // Insert low 2 bits of trail surrogate to magic number for later
-      // 11011100 00000000 11100111 110000cc
-      uint16x8_t magic_with_low_2 =
-          vreinterpretq_u16_u32(vsraq_n_u32(magic, shift, 30));
-      // Generate lead surrogate
-      // xxxxcccc ccdddddd|xxxxxxxx xxxxxxxx
-      uint32x4_t lead = vreinterpretq_u32_u16(
-          vsliq_n_u16(vreinterpretq_u16_u8(swap), vreinterpretq_u16_u8(in), 6));
-      // Mask out lead
-      // 000000cc ccdddddd|xxxxxxxx xxxxxxxx
-      lead = vbicq_u32(lead, vmovq_n_u32(uint32_t(~0x03FFFFFF)));
-      // Blend pairs
-      // 000000cc ccdddddd|11110aaa bbbbbb00
-      uint16x8_t blend = vreinterpretq_u16_u32(
-          vbslq_u32(vmovq_n_u32(0x0000FFFF), trail, lead));
-      // Add magic number to finish the result
-      // 110111CC CCDDDDDD|110110AA BBBBBBCC
-      uint16x8_t composed = vaddq_u16(blend, magic_with_low_2);
-      // Byte swap if necessary
-      if (!match_system(big_endian)) {
-        composed =
-            vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
-      }
-      uint16_t buffer[8];
-      vst1q_u16(reinterpret_cast<uint16_t *>(buffer), composed);
-      for (int k = 0; k < 6; k++) {
-        utf16_output[k] = buffer[k];
-      } // the loop might compiler to a couple of instructions.
-      // We need some validation. See
-      // https://github.com/simdutf/simdutf/pull/631
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      uint8x16_t expected_mask = simdutf_make_uint8x16_t(
-          0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0,
-          0xc0, 0x0, 0x0, 0x0, 0x0);
-#else
-      uint8x16_t expected_mask = {0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0,
-                                  0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0,
-                                  0x0,  0x0,  0x0,  0x0};
-#endif
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      uint8x16_t expected = simdutf_make_uint8x16_t(
-          0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80,
-          0x80, 0x0, 0x0, 0x0, 0x0);
-#else
-      uint8x16_t expected = {0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80,
-                             0xf0, 0x80, 0x80, 0x80, 0x0,  0x0,  0x0,  0x0};
-#endif
-      uint8x16_t check = vceqq_u8(vandq_u8(in, expected_mask), expected);
-      bool correct = (vminvq_u32(vreinterpretq_u32_u8(check)) == 0xFFFFFFFF);
-      // The validation is just three instructions and it is not on a critical
-      // path.
-      if (correct) {
-        utf16_output += 6; // We wrote 3 32-bit surrogate pairs.
-      }
-      return 12; // We consumed 12 bytes.
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char16_t *, char32_t *> ret =
+      arm_convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    // 3 1-4 byte sequences
-    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // 1 byte: 00000000 00000000 00000000 0ddddddd
-    // 3 byte: 00000000 00000000 110ccccc 10dddddd
-    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
-    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
-    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
-    // added to fix issue https://github.com/simdutf/simdutf/issues/514
-    // We only want to write 2 * 16-bit code units when that is actually what we
-    // have. Unfortunately, we cannot trust the input. So it is possible to get
-    // 0xff as an input byte and it should not result in a surrogate pair. We
-    // need to check for that.
-    uint32_t permbuffer[4];
-    vst1q_u32(permbuffer, perm);
-    // Mask the low and middle bytes
-    // 00000000 00000000 00000000 0ddddddd
-    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7f));
-    // Because the surrogates need more work, the high surrogate is computed
-    // first.
-    uint32x4_t middlehigh = vshlq_n_u32(perm, 2);
-    // 00000000 00000000 00cccccc 00000000
-    uint32x4_t middlebyte = vandq_u32(perm, vmovq_n_u32(0x3F00));
-    // Start assembling the sequence. Since the 4th byte is in the same position
-    // as it would be in a surrogate and there is no dependency, shift left
-    // instead of right. 3 byte: 00000000 10bbbbxx xxxxxxxx xxxxxxxx 4 byte:
-    // 11110aaa bbbbbbxx xxxxxxxx xxxxxxxx
-    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0xFF000000), perm, middlehigh);
-    // Top 16 bits contains the high ten bits of the surrogate pair before
-    // correction 3 byte: 00000000 10bbbbcc|cccc0000 00000000 4 byte: 11110aaa
-    // bbbbbbcc|cccc0000 00000000 - high 10 bits correct w/o correction
-    uint32x4_t abc =
-        vbslq_u32(vmovq_n_u32(0xFFFC0000), ab, vshlq_n_u32(middlebyte, 4));
-    // Combine the low 6 or 7 bits by a shift right accumulate
-    // 3 byte: 00000000 00000010|bbbbcccc ccdddddd - low 16 bits correct
-    // 4 byte: 00000011 110aaabb|bbbbcccc ccdddddd - low 10 bits correct w/o
-    // correction
-    uint32x4_t composed = vsraq_n_u32(ascii, abc, 6);
-    // After this is for surrogates
-    // Blend the low and high surrogates
-    // 4 byte: 11110aaa bbbbbbcc|bbbbcccc ccdddddd
-    uint32x4_t mixed = vbslq_u32(vmovq_n_u32(0xFFFF0000), abc, composed);
-    // Clear the upper 6 bits of the low surrogate. Don't clear the upper bits
-    // yet as 0x10000 was not subtracted from the codepoint yet. 4 byte:
-    // 11110aaa bbbbbbcc|000000cc ccdddddd
-    uint16x8_t masked_pair = vreinterpretq_u16_u32(
-        vbicq_u32(mixed, vmovq_n_u32(uint32_t(~0xFFFF03FF))));
-    // Correct the remaining UTF-8 prefix, surrogate offset, and add the
-    // surrogate prefixes in one magic 16-bit addition. similar magic number but
-    // without the continue byte adjust and halfword swapped UTF-8 4b prefix   =
-    // -0xF000|0x0000 surrogate offset  = -0x0040|0x0000 (0x10000 << 6)
-    // surrogate high    = +0xD800|0x0000
-    // surrogate low     = +0x0000|0xDC00
-    // -----------------------------------
-    //                   = +0xE7C0|0xDC00
-    uint16x8_t magic = vreinterpretq_u16_u32(vmovq_n_u32(0xE7C0DC00));
-    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD - surrogate pair complete
-    uint32x4_t surrogates =
-        vreinterpretq_u32_u16(vaddq_u16(masked_pair, magic));
-    // If the high bit is 1 (s32 less than zero), this needs a surrogate pair
-    uint32x4_t is_pair = vcltzq_s32(vreinterpretq_s32_u32(perm));
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char16_t *, char32_t *> ret =
+      arm_convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // Select either the 4 byte surrogate pair or the 2 byte solo codepoint
-    // 3 byte: 0xxxxxxx xxxxxxxx|bbbbcccc ccdddddd
-    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD
-    uint32x4_t selected = vbslq_u32(is_pair, surrogates, composed);
-    // Byte swap if necessary
-    if (!match_system(big_endian)) {
-      selected =
-          vreinterpretq_u32_u8(vrev16q_u8(vreinterpretq_u8_u32(selected)));
+simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char32_t *> ret =
+      arm_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(buf, len,
+                                                                 utf32_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
     }
-    // Attempting to shuffle and store would be complex, just scalarize.
-    uint32_t buffer[4];
-    vst1q_u32(buffer, selected);
-    // Test for the top bit of the surrogate mask. Remove due to issue 514
-    // const uint32_t SURROGATE_MASK = match_system(big_endian) ? 0x80000000 :
-    // 0x00800000;
-    for (size_t i = 0; i < 3; i++) {
-      // Surrogate
-      // Used to be if (buffer[i] & SURROGATE_MASK) {
-      // See discussion above.
-      // patch for issue https://github.com/simdutf/simdutf/issues/514
-      if ((permbuffer[i] & 0xf8000000) == 0xf0000000) {
-        utf16_output[0] = uint16_t(buffer[i] >> 16);
-        utf16_output[1] = uint16_t(buffer[i] & 0xFFFF);
-        utf16_output += 2;
-      } else {
-        utf16_output[0] = uint16_t(buffer[i] & 0xFFFF);
-        utf16_output++;
-      }
+  }
+  ret.first.count =
+      ret.second -
+      utf32_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
+
+simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char32_t *> ret =
+      arm_convert_utf16_to_utf32_with_errors<endianness::BIG>(buf, len,
+                                                              utf32_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
     }
-    return consumed;
-  } else {
-    // here we know that there is an error but we do not handle errors
-    return 12;
   }
+  ret.first.count =
+      ret.second -
+      utf32_output; // Set count to the number of 8-bit code units written
+  return ret.first;
 }
-/* end file src/arm64/arm_convert_utf8_to_utf16.cpp */
-/* begin file src/arm64/arm_convert_utf8_to_utf32.cpp */
-// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
-// end of the code points. Only the least significant 12 bits of the mask
-// are accessed.
-// It returns how many bytes were consumed (up to 12).
-size_t convert_masked_utf8_to_utf32(const char *input,
-                                    uint64_t utf8_end_of_code_point_mask,
-                                    char32_t *&utf32_out) {
-  // we use an approach where we try to process up to 12 input bytes.
-  // Why 12 input bytes and not 16? Because we are concerned with the size of
-  // the lookup tables. Also 12 is nicely divisible by two and three.
-  //
-  uint32_t *&utf32_output = reinterpret_cast<uint32_t *&>(utf32_out);
-  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
-  const uint16_t input_utf8_end_of_code_point_mask =
-      utf8_end_of_code_point_mask & 0xFFF;
-  //
-  // Optimization note: our main path below is load-latency dependent. Thus it
-  // is maybe beneficial to have fast paths that depend on branch prediction but
-  // have less latency. This results in more instructions but, potentially, also
-  // higher speeds.
-  //
-  // We first try a few fast paths.
-  if (utf8_end_of_code_point_mask == 0xfff) {
-    // We process in chunks of 12 bytes.
-    // use fast implementation in src/simdutf/arm64/simd.h
-    // Ideally the compiler can keep the tables in registers.
-    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
-    temp.store_ascii_as_utf32_tbl(utf32_out);
-    utf32_output += 12; // We wrote 12 32-bit characters.
-    return 12;          // We consumed 12 bytes.
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char32_t *, char *> ret =
+      arm_convert_utf32_to_latin1(buf, len, latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
   }
-  if (input_utf8_end_of_code_point_mask == 0x924) {
-    // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte
-    // UTF-32 code units. Convert to UTF-16
-    uint16x4_t composed_utf16 = convert_utf8_3_byte_to_utf16(in);
-    // Zero extend and store via ST2 with a zero.
-    uint16x4x2_t interleaver = {{composed_utf16, vmov_n_u16(0)}};
-    vst2_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
-    utf32_output += 4; // We wrote 4 32-bit characters.
-    return 12;         // We consumed 12 bytes.
+  size_t saved_bytes = ret.second - latin1_output;
+
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
+
+simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<result, char *> ret =
+      arm_convert_utf32_to_latin1_with_errors(buf, len, latin1_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res = scalar::utf32_to_latin1::convert_with_errors(
+        buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
   }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
-  if (input_utf8_end_of_code_point_mask == 0xaaa) {
-    // We want to take 6 2-byte UTF-8 code units and turn them into 6 4-byte
-    // UTF-32 code units. Convert to UTF-16
-    uint16x8_t composed_utf16 = convert_utf8_2_byte_to_utf16(in);
-    // Zero extend and store via ST2 with a zero.
-    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
-    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
-    utf32_output += 6; // We wrote 6 32-bit characters.
-    return 12;         // We consumed 12 bytes.
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char32_t *, char *> ret =
+      arm_convert_utf32_to_latin1(buf, len, latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
   }
-  /// Either no fast path or an unimportant fast path.
-
-  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][0];
-  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][1];
+  size_t saved_bytes = ret.second - latin1_output;
 
-  if (idx < 64) {
-    // SIX (6) input code-code units
-    // Convert to UTF-16
-    uint16x8_t composed_utf16 = convert_utf8_1_to_2_byte_to_utf16(in, idx);
-    // Zero extend and store with ST2 and zero
-    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
-    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
-    utf32_output += 6; // We wrote 6 32-bit characters.
-    return consumed;
-  } else if (idx < 145) {
-    // FOUR (4) input code-code units
-    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
-    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
-    // Shuffle
-    // 1 byte: 00000000 00000000 0ccccccc
-    // 2 byte: 00000000 110bbbbb 10cccccc
-    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
-    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
-    // Split
-    // 00000000 00000000 0ccccccc
-    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F)); // 6 or 7 bits
-    // Note: unmasked
-    // xxxxxxxx aaaaxxxx xxxxxxxx
-    uint32x4_t high = vshrq_n_u32(perm, 4); // 4 bits
-    // Use 16 bit bic instead of and.
-    // The top bits will be corrected later in the bsl
-    // 00000000 10bbbbbb 00000000
-    uint32x4_t middle = vreinterpretq_u32_u16(
-        vbicq_u16(vreinterpretq_u16_u32(perm),
-                  vmovq_n_u16(uint16_t(~0xff00)))); // 5 or 6 bits
-    // Combine low and middle with shift right accumulate
-    // 00000000 00xxbbbb bbcccccc
-    uint32x4_t lowmid = vsraq_n_u32(ascii, middle, 2);
-    // Insert top 4 bits from high byte with bitwise select
-    // 00000000 aaaabbbb bbcccccc
-    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0x0000F000), high, lowmid);
-    vst1q_u32(utf32_output, composed);
-    utf32_output += 4; // We wrote 4 32-bit characters.
-    return consumed;
-  } else if (idx < 209) {
-    // THREE (3) input code-code units
-    if (input_utf8_end_of_code_point_mask == 0x888) {
-      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
-      // UTF-32 code units. This uses the same method as the fixed 3 byte
-      // version, reversing and shift left insert. However, there is no need for
-      // a shuffle mask now, just rev16 and rev32.
-      //
-      // This version does not use the LUT, but 4 byte sequences are less common
-      // and the overhead of the extra memory access is less important than the
-      // early branch overhead in shorter sequences, so it comes last.
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert_valid(
+        ret.first, len - (ret.first - buf), ret.second);
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-      // Swap pairs of bytes
-      // 10dddddd|10cccccc|10bbbbbb|11110aaa
-      // 10cccccc 10dddddd|11110aaa 10bbbbbb
-      uint16x8_t swap1 = vreinterpretq_u16_u8(vrev16q_u8(in));
-      // Shift left and insert
-      // xxxxcccc ccdddddd|xxxxxxxa aabbbbbb
-      uint16x8_t merge1 = vsliq_n_u16(swap1, vreinterpretq_u16_u8(in), 6);
-      // Swap 16-bit lanes
-      // xxxxcccc ccdddddd xxxxxxxa aabbbbbb
-      // xxxxxxxa aabbbbbb xxxxcccc ccdddddd
-      uint32x4_t swap2 = vreinterpretq_u32_u16(vrev32q_u16(merge1));
-      // Shift insert again
-      // xxxxxxxx xxxaaabb bbbbcccc ccdddddd
-      uint32x4_t merge2 = vsliq_n_u32(swap2, vreinterpretq_u32_u16(merge1), 12);
-      // Clear the garbage
-      // 00000000 000aaabb bbbbcccc ccdddddd
-      uint32x4_t composed = vandq_u32(merge2, vmovq_n_u32(0x1FFFFF));
-      // Store
-      vst1q_u32(utf32_output, composed);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  // optimization opportunity: implement a custom function.
+  return convert_utf32_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-      utf32_output += 3; // We wrote 3 32-bit characters.
-      return 12;         // We consumed 12 bytes.
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      arm_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf32_to_utf16::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    // Unlike UTF-16, doing a fast codepath doesn't have nearly as much benefit
-    // due to surrogates no longer being involved.
-    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
-        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
-    // 1 byte: 00000000 00000000 00000000 0ddddddd
-    // 2 byte: 00000000 00000000 110ccccc 10dddddd
-    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
-    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
-    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
-    // Ascii
-    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F));
-    uint32x4_t middle = vandq_u32(perm, vmovq_n_u32(0x3f00));
-    // When converting the way we do, the 3 byte prefix will be interpreted as
-    // the 18th bit being set, since the code would interpret the lead byte
-    // (0b1110bbbb) as a continuation byte (0b10bbbbbb). To fix this, we can
-    // either xor or do an 8 bit add of the 6th bit shifted right by 1. Since
-    // NEON has shift right accumulate, we use that.
-    //  4 byte   3 byte
-    // 10bbbbbb 1110bbbb
-    // 00000000 01000000 6th bit
-    // 00000000 00100000 shift right
-    // 10bbbbbb 0000bbbb add
-    // 00bbbbbb 0000bbbb mask
-    uint8x16_t correction =
-        vreinterpretq_u8_u32(vandq_u32(perm, vmovq_n_u32(0x00400000)));
-    uint32x4_t corrected = vreinterpretq_u32_u8(
-        vsraq_n_u8(vreinterpretq_u8_u32(perm), correction, 1));
-    // 00000000 00000000 0000cccc ccdddddd
-    uint32x4_t cd = vsraq_n_u32(ascii, middle, 2);
-    // Insert twice
-    // xxxxxxxx xxxaaabb bbbbxxxx xxxxxxxx
-    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0x01C0000), vshrq_n_u32(corrected, 6),
-                              vshrq_n_u32(corrected, 4));
-    // 00000000 000aaabb bbbbcccc ccdddddd
-    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0xFFE00FFF), cd, ab);
-    // Store
-    vst1q_u32(utf32_output, composed);
-    utf32_output += 3; // We wrote 3 32-bit characters.
-    return consumed;
-  } else {
-    // here we know that there is an error but we do not handle errors
-    return 12;
+    saved_bytes += scalar_saved_bytes;
   }
+  return saved_bytes;
 }
-/* end file src/arm64/arm_convert_utf8_to_utf32.cpp */
-
-/* begin file src/arm64/arm_convert_utf16_to_latin1.cpp */
 
-template <endianness big_endian>
-std::pair<const char16_t *, char *>
-arm_convert_utf16_to_latin1(const char16_t *buf, size_t len,
-                            char *latin1_output) {
-  const char16_t *end = buf + len;
-  while (end - buf >= 8) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      arm_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf32_to_utf16::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    if (vmaxvq_u16(in) <= 0xff) {
-      // 1. pack the bytes
-      uint8x8_t latin1_packed = vmovn_u16(in);
-      // 2. store (8 bytes)
-      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
-      // 3. adjust pointers
-      buf += 8;
-      latin1_output += 8;
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
+
+simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      arm_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len,
+                                                                 utf16_output);
+  if (ret.first.count != len) {
+    result scalar_res =
+        scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
     } else {
-      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
+      ret.second += scalar_res.count;
     }
-  } // while
-  return std::make_pair(buf, latin1_output);
+  }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
 }
 
-template <endianness big_endian>
-std::pair<result, char *>
-arm_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
-                                        char *latin1_output) {
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
-  while (end - buf >= 8) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
-    }
-    if (vmaxvq_u16(in) <= 0xff) {
-      // 1. pack the bytes
-      uint8x8_t latin1_packed = vmovn_u16(in);
-      // 2. store (8 bytes)
-      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
-      // 3. adjust pointers
-      buf += 8;
-      latin1_output += 8;
+simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      arm_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len,
+                                                              utf16_output);
+  if (ret.first.count != len) {
+    result scalar_res =
+        scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
     } else {
-      // Let us do a scalar fallback.
-      for (int k = 0; k < 8; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if (word <= 0xff) {
-          *latin1_output++ = char(word);
-        } else {
-          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
-                                latin1_output);
-        }
-      }
+      ret.second += scalar_res.count;
     }
-  } // while
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        latin1_output);
+  }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
 }
-/* end file src/arm64/arm_convert_utf16_to_latin1.cpp */
-/* begin file src/arm64/arm_convert_utf16_to_utf32.cpp */
-/*
-    The vectorized algorithm works on single SSE register i.e., it
-    loads eight 16-bit code units.
 
-    We consider three cases:
-    1. an input register contains no surrogates and each value
-       is in range 0x0000 .. 0x07ff.
-    2. an input register contains no surrogates and values are
-       is in range 0x0000 .. 0xffff.
-    3. an input register contains surrogates --- i.e. codepoints
-       can have 16 or 32 bits.
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16le(buf, len, utf16_output);
+}
 
-    Ad 1.
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16be(buf, len, utf16_output);
+}
 
-    When values are less than 0x0800, it means that a 16-bit code unit
-    can be converted into: 1) single UTF8 byte (when it is an ASCII
-    char) or 2) two UTF8 bytes.
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return convert_utf16le_to_utf32(buf, len, utf32_output);
+}
 
-    For this case we do only some shuffle to obtain these 2-byte
-    codes and finally compress the whole SSE register with a single
-    shuffle.
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return convert_utf16be_to_utf32(buf, len, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+#if SIMDUTF_FEATURE_UTF16
+void implementation::change_endianness_utf16(const char16_t *input,
+                                             size_t length,
+                                             char16_t *output) const noexcept {
+  utf16::change_endianness_utf16(input, length, output);
+}
 
-    Ad 2.
+simdutf_warn_unused size_t implementation::count_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::count_code_points<endianness::LITTLE>(input, length);
+}
 
-    When values fit in 16-bit code units, but are above 0x07ff, then
-    a single word may produce one, two or three UTF8 bytes.
+simdutf_warn_unused size_t implementation::count_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::count_code_points<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-    We prepare data for all these three cases in two registers.
-    The first register contains lower two UTF8 bytes (used in all
-    cases), while the second one contains just the third byte for
-    the three-UTF8-bytes case.
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused size_t
+implementation::count_utf8(const char *input, size_t length) const noexcept {
+  return utf8::count_code_points(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8
 
-    Finally these two registers are interleaved forming eight-element
-    array of 32-bit values. The array spans two SSE registers.
-    The bytes from the registers are compressed using two shuffles.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
+    const char *buf, size_t len) const noexcept {
+  return count_utf8(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
+    const char *input, size_t length) const noexcept {
+  // See
+  // https://lemire.me/blog/2023/05/15/computing-the-utf-8-size-of-a-latin-1-string-quickly-arm-neon-edition/
+  // credit to Pete Cawley
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
+  uint64_t result = 0;
+  const int lanes = sizeof(uint8x16_t);
+  uint8_t rem = length % lanes;
+  const uint8_t *simd_end = data + (length / lanes) * lanes;
+  const uint8x16_t threshold = vdupq_n_u8(0x80);
+  for (; data < simd_end; data += lanes) {
+    // load 16 bytes
+    uint8x16_t input_vec = vld1q_u8(data);
+    // compare to threshold (0x80)
+    uint8x16_t withhighbit = vcgeq_u8(input_vec, threshold);
+    // vertical addition
+    result -= vaddvq_s8(vreinterpretq_s8_u8(withhighbit));
+  }
+  return result + (length / lanes) * lanes +
+         scalar::latin1::utf8_length_from_latin1((const char *)simd_end, rem);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return arm64_utf8_length_from_utf16_bytemask<endianness::LITTLE>(input,
+                                                                   length);
+}
 
-    To summarize:
-    - We need two 256-entry tables that have 8704 bytes in total.
-*/
-/*
-  Returns a pair: the first unprocessed byte from buf and utf8_output
-  A scalar routing should carry on the conversion of the tail.
-*/
-template <endianness big_endian>
-std::pair<const char16_t *, char32_t *>
-arm_convert_utf16_to_utf32(const char16_t *buf, size_t len,
-                           char32_t *utf32_out) {
-  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
-  const char16_t *end = buf + len;
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return arm64_utf8_length_from_utf16_bytemask<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
-  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
+}
 
-  while (end - buf >= 8) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
-    }
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    const uint16x8_t surrogates_bytemask =
-        vceqq_u16(vandq_u16(in, v_f800), v_d800);
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (vmaxvq_u16(surrogates_bytemask) == 0) {
-      // case: no surrogate pairs, extend all 16-bit code units to 32-bit code
-      // units
-      vst1q_u32(utf32_output, vmovl_u16(vget_low_u16(in)));
-      vst1q_u32(utf32_output + 4, vmovl_high_u16(in));
-      utf32_output += 8;
-      buf += 8;
-      // surrogate pair(s) in a register
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return utf8::utf16_length_from_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  const uint32x4_t v_7f = vmovq_n_u32((uint32_t)0x7f);
+  const uint32x4_t v_7ff = vmovq_n_u32((uint32_t)0x7ff);
+  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
+  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos + 4 <= length; pos += 4) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
+    const uint32x4_t ascii_bytes_bytemask = vcleq_u32(in, v_7f);
+    const uint32x4_t one_two_bytes_bytemask = vcleq_u32(in, v_7ff);
+    const uint32x4_t two_bytes_bytemask =
+        veorq_u32(one_two_bytes_bytemask, ascii_bytes_bytemask);
+    const uint32x4_t three_bytes_bytemask =
+        veorq_u32(vcleq_u32(in, v_ffff), one_two_bytes_bytemask);
+
+    const uint16x8_t reduced_ascii_bytes_bytemask =
+        vreinterpretq_u16_u32(vandq_u32(ascii_bytes_bytemask, v_1));
+    const uint16x8_t reduced_two_bytes_bytemask =
+        vreinterpretq_u16_u32(vandq_u32(two_bytes_bytemask, v_1));
+    const uint16x8_t reduced_three_bytes_bytemask =
+        vreinterpretq_u16_u32(vandq_u32(three_bytes_bytemask, v_1));
+
+    const uint16x8_t compressed_bytemask0 =
+        vpaddq_u16(reduced_ascii_bytes_bytemask, reduced_two_bytes_bytemask);
+    const uint16x8_t compressed_bytemask1 =
+        vpaddq_u16(reduced_three_bytes_bytemask, reduced_three_bytes_bytemask);
+
+    size_t ascii_count = count_ones(
+        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 0));
+    size_t two_bytes_count = count_ones(
+        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 1));
+    size_t three_bytes_count = count_ones(
+        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask1), 0));
+
+    count += 16 - 3 * ascii_count - 2 * two_bytes_count - three_bytes_count;
+  }
+  return count +
+         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
+  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos + 4 <= length; pos += 4) {
+    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
+    const uint32x4_t surrogate_bytemask = vcgtq_u32(in, v_ffff);
+    const uint16x8_t reduced_bytemask =
+        vreinterpretq_u16_u32(vandq_u32(surrogate_bytemask, v_1));
+    const uint16x8_t compressed_bytemask =
+        vpaddq_u16(reduced_bytemask, reduced_bytemask);
+    size_t surrogate_count = count_ones(
+        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask), 0));
+    count += 4 + surrogate_count;
+  }
+  return count +
+         scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return utf8::count_code_points(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
     } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if ((word & 0xF800) != 0xD800) {
-          *utf32_output++ = char32_t(word);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
-                                   : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char32_t *>(utf32_output));
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf32_output++ = char32_t(value);
-        }
-      }
-      buf += k;
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
     }
-  } // while
-  return std::make_pair(buf, reinterpret_cast<char32_t *>(utf32_output));
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
 }
 
-/*
-  Returns a pair: a result struct and utf8_output.
-  If there is an error, the count field of the result is the position of the
-  error. Otherwise, it is the position of the first unprocessed byte in buf
-  (even if finished). A scalar routing should carry on the conversion of the
-  tail if needed.
-*/
-template <endianness big_endian>
-std::pair<result, char32_t *>
-arm_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
-                                       char32_t *utf32_out) {
-  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
-
-  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
-  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
+}
 
-  while ((end - buf) >= 8) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
     }
+  }
+}
 
-    const uint16x8_t surrogates_bytemask =
-        vceqq_u16(vandq_u16(in, v_f800), v_d800);
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (vmaxvq_u16(surrogates_bytemask) == 0) {
-      // case: no surrogate pairs, extend all 16-bit code units to 32-bit code
-      // units
-      vst1q_u32(utf32_output, vmovl_u16(vget_low_u16(in)));
-      vst1q_u32(utf32_output + 4, vmovl_high_u16(in));
-      utf32_output += 8;
-      buf += 8;
-      // surrogate pair(s) in a register
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
     } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if ((word & 0xF800) != 0xD800) {
-          *utf32_output++ = char32_t(word);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
-                                   : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k - 1),
-                reinterpret_cast<char32_t *>(utf32_output));
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf32_output++ = char32_t(value);
-        }
-      }
-      buf += k;
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
     }
-  } // while
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        reinterpret_cast<char32_t *>(utf32_output));
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
 }
-/* end file src/arm64/arm_convert_utf16_to_utf32.cpp */
-/* begin file src/arm64/arm_convert_utf16_to_utf8.cpp */
-/*
-    The vectorized algorithm works on single SSE register i.e., it
-    loads eight 16-bit code units.
 
-    We consider three cases:
-    1. an input register contains no surrogates and each value
-       is in range 0x0000 .. 0x07ff.
-    2. an input register contains no surrogates and values are
-       is in range 0x0000 .. 0xffff.
-    3. an input register contains surrogates --- i.e. codepoints
-       can have 16 or 32 bits.
+size_t implementation::binary_to_base64(const char *input, size_t length,
+                                        char *output,
+                                        base64_options options) const noexcept {
+  return encode_base64(output, input, length, options);
+}
+#endif // SIMDUTF_FEATURE_BASE64
 
-    Ad 1.
+} // namespace arm64
+} // namespace simdutf
 
-    When values are less than 0x0800, it means that a 16-bit code unit
-    can be converted into: 1) single UTF8 byte (when it is an ASCII
-    char) or 2) two UTF8 bytes.
+/* begin file src/simdutf/arm64/end.h */
+/* end file src/simdutf/arm64/end.h */
+/* end file src/arm64/implementation.cpp */
+#endif
+#if SIMDUTF_IMPLEMENTATION_FALLBACK
+/* begin file src/fallback/implementation.cpp */
+/* begin file src/simdutf/fallback/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "fallback"
+// #define SIMDUTF_IMPLEMENTATION fallback
+/* end file src/simdutf/fallback/begin.h */
 
-    For this case we do only some shuffle to obtain these 2-byte
-    codes and finally compress the whole SSE register with a single
-    shuffle.
+namespace simdutf {
+namespace fallback {
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused int
+implementation::detect_encodings(const char *input,
+                                 size_t length) const noexcept {
+  // If there is a BOM, then we trust it.
+  auto bom_encoding = simdutf::BOM::check_bom(input, length);
+  if (bom_encoding != encoding_type::unspecified) {
+    return bom_encoding;
+  }
+  int out = 0;
+  // todo: reimplement as a one-pass algorithm.
+  if (validate_utf8(input, length)) {
+    out |= encoding_type::UTF8;
+  }
+  if ((length % 2) == 0) {
+    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
+                         length / 2)) {
+      out |= encoding_type::UTF16_LE;
+    }
+  }
+  if ((length % 4) == 0) {
+    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
+      out |= encoding_type::UTF32_LE;
+    }
+  }
+  return out;
+}
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
-    Ad 2.
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return scalar::utf8::validate(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-    When values fit in 16-bit code units, but are above 0x07ff, then
-    a single word may produce one, two or three UTF8 bytes.
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused result implementation::validate_utf8_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return scalar::utf8::validate_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8
 
-    We prepare data for all these three cases in two registers.
-    The first register contains lower two UTF8 bytes (used in all
-    cases), while the second one contains just the third byte for
-    the three-UTF8-bytes case.
+#if SIMDUTF_FEATURE_ASCII
+simdutf_warn_unused bool
+implementation::validate_ascii(const char *buf, size_t len) const noexcept {
+  return scalar::ascii::validate(buf, len);
+}
 
-    Finally these two registers are interleaved forming eight-element
-    array of 32-bit values. The array spans two SSE registers.
-    The bytes from the registers are compressed using two shuffles.
+simdutf_warn_unused result implementation::validate_ascii_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return scalar::ascii::validate_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_ASCII
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf16le(const char16_t *buf,
+                                 size_t len) const noexcept {
+  return scalar::utf16::validate<endianness::LITTLE>(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool
+implementation::validate_utf16be(const char16_t *buf,
+                                 size_t len) const noexcept {
+  return scalar::utf16::validate<endianness::BIG>(buf, len);
+}
 
-    To summarize:
-    - We need two 256-entry tables that have 8704 bytes in total.
-*/
-/*
-  Returns a pair: the first unprocessed byte from buf and utf8_output
-  A scalar routing should carry on the conversion of the tail.
-*/
-template <endianness big_endian>
-std::pair<const char16_t *, char *>
-arm_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
-  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
-  const char16_t *end = buf + len;
+simdutf_warn_unused result implementation::validate_utf16le_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  return scalar::utf16::validate_with_errors<endianness::LITTLE>(buf, len);
+}
 
-  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
-  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
-  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
-    }
-    if (vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
-      // It is common enough that we have sequences of 16 consecutive ASCII
-      // characters.
-      uint16x8_t nextin =
-          vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
-      if (!match_system(big_endian)) {
-        nextin = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(nextin)));
-      }
-      if (vmaxvq_u16(nextin) > 0x7F) {
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x8_t utf8_packed = vmovn_u16(in);
-        // 2. store (8 bytes)
-        vst1_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 8;
-        utf8_output += 8;
-        in = nextin;
-      } else {
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
-        // 2. store (16 bytes)
-        vst1q_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 16;
-        utf8_output += 16;
-        continue; // we are done for this round!
-      }
-    }
+simdutf_warn_unused result implementation::validate_utf16be_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  return scalar::utf16::validate_with_errors<endianness::BIG>(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-    if (vmaxvq_u16(in) <= 0x7FF) {
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
-      const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
+  return scalar::utf32::validate(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const uint16x8_t t0 = vshlq_n_u16(in, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const uint16x8_t t1 = vandq_u16(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const uint16x8_t t2 = vandq_u16(in, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const uint16x8_t t3 = vorrq_u16(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const uint16x8_t t4 = vorrq_u16(t3, v_c080);
-      // 2. merge ASCII and 2-byte codewords
-      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
-      const uint8x16_t utf8_unpacked =
-          vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
-      // 3. prepare bitmask for 8-bit lookup
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t mask = simdutf_make_uint16x8_t(
-          0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
-#else
-      const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
-                               0x0002, 0x0008, 0x0020, 0x0080};
-#endif
-      uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
-      // 4. pack the bytes
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
-      const uint8x16_t shuffle = vld1q_u8(row + 1);
-      const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+#if SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused result implementation::validate_utf32_with_errors(
+    const char32_t *buf, size_t len) const noexcept {
+  return scalar::utf32::validate_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32
 
-      // 5. store bytes
-      vst1q_u8(utf8_output, utf8_packed);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
+    const char *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::latin1_to_utf8::convert(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-      // 6. adjust pointers
-      buf += 8;
-      utf8_output += row[0];
-      continue;
-    }
-    const uint16x8_t surrogates_bytemask =
-        vceqq_u16(vandq_u16(in, v_f800), v_d800);
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (vmaxvq_u16(surrogates_bytemask) == 0) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t dup_even = simdutf_make_uint16x8_t(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
-#else
-      const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
-                                   0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
-#endif
-      /* In this branch we handle three cases:
-         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::latin1_to_utf16::convert<endianness::LITTLE>(buf, len,
+                                                              utf16_output);
+}
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::latin1_to_utf16::convert<endianness::BIG>(buf, len,
+                                                           utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::latin1_to_utf32::convert(buf, len, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf8_to_latin1::convert(buf, len, latin1_output);
+}
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const uint16x8_t t0 = vreinterpretq_u16_u8(
-          vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf8_to_latin1::convert_with_errors(buf, len, latin1_output);
+}
 
-      // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
-      const uint16x8_t s0 = vshrq_n_u16(in, 12);
-      // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
-      const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
-      // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-      const uint16x8_t s1s = vshlq_n_u16(s1, 2);
-      // [00bb|bbbb|0000|aaaa]
-      const uint16x8_t s2 = vorrq_u16(s0, s1s);
-      // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
-      const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
-      const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
-      const uint16x8_t m0 =
-          vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
-      const uint16x8_t s4 = veorq_u16(s3, m0);
-#undef simdutf_vec
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf8_to_latin1::convert_valid(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-      // 4. expand code units 16-bit => 32-bit
-      const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
-      const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert<endianness::LITTLE>(buf, len,
+                                                            utf16_output);
+}
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t onemask = simdutf_make_uint16x8_t(
-          0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
-      const uint16x8_t twomask = simdutf_make_uint16x8_t(
-          0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
-#else
-      const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                  0x0100, 0x0400, 0x1000, 0x4000};
-      const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
-                                  0x0200, 0x0800, 0x2000, 0x8000};
-#endif
-      const uint16x8_t combined =
-          vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
-                    vandq_u16(one_or_two_bytes_bytemask, twomask));
-      const uint16_t mask = vaddvq_u16(combined);
-      // The following fast path may or may not be beneficial.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
-        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
-        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
-        vst1q_u8(utf8_output, utf8_0);
-        utf8_output += 12;
-        vst1q_u8(utf8_output, utf8_1);
-        utf8_output += 12;
-        buf += 8;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert<endianness::BIG>(buf, len,
+                                                         utf16_output);
+}
 
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
-      const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert_with_errors<endianness::LITTLE>(
+      buf, len, utf16_output);
+}
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
-      const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert_with_errors<endianness::BIG>(
+      buf, len, utf16_output);
+}
 
-      vst1q_u8(utf8_output, utf8_0);
-      utf8_output += row0[0];
-      vst1q_u8(utf8_output, utf8_1);
-      utf8_output += row1[0];
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(buf, len,
+                                                                  utf16_output);
+}
 
-      buf += 8;
-      // surrogate pair(s) in a register
-    } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if ((word & 0xFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xF800) != 0xD800) {
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
-                                   : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char *>(utf8_output));
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf8_output++ = char((value >> 18) | 0b11110000);
-          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((value & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf8_to_utf16::convert_valid<endianness::BIG>(buf, len,
+                                                               utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  return std::make_pair(buf, reinterpret_cast<char *>(utf8_output));
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf8_to_utf32::convert(buf, len, utf32_output);
 }
 
-/*
-  Returns a pair: a result struct and utf8_output.
-  If there is an error, the count field of the result is the position of the
-  error. Otherwise, it is the position of the first unprocessed byte in buf
-  (even if finished). A scalar routing should carry on the conversion of the
-  tail if needed.
-*/
-template <endianness big_endian>
-std::pair<result, char *>
-arm_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
-                                      char *utf8_out) {
-  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
+simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf8_to_utf32::convert_with_errors(buf, len, utf32_output);
+}
 
-  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
-  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
-  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
+    const char *input, size_t size, char32_t *utf32_output) const noexcept {
+  return scalar::utf8_to_utf32::convert_valid(input, size, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
-    if (!match_system(big_endian)) {
-      in = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(in)));
-    }
-    if (vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
-      // It is common enough that we have sequences of 16 consecutive ASCII
-      // characters.
-      uint16x8_t nextin =
-          vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
-      if (!match_system(big_endian)) {
-        nextin = vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(nextin)));
-      }
-      if (vmaxvq_u16(nextin) > 0x7F) {
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x8_t utf8_packed = vmovn_u16(in);
-        // 2. store (8 bytes)
-        vst1_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 8;
-        utf8_output += 8;
-        in = nextin;
-      } else {
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
-        // 2. store (16 bytes)
-        vst1q_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 16;
-        utf8_output += 16;
-        continue; // we are done for this round!
-      }
-    }
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert<endianness::LITTLE>(buf, len,
+                                                              latin1_output);
+}
 
-    if (vmaxvq_u16(in) <= 0x7FF) {
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
-      const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert<endianness::BIG>(buf, len,
+                                                           latin1_output);
+}
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const uint16x8_t t0 = vshlq_n_u16(in, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const uint16x8_t t1 = vandq_u16(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const uint16x8_t t2 = vandq_u16(in, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const uint16x8_t t3 = vorrq_u16(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const uint16x8_t t4 = vorrq_u16(t3, v_c080);
-      // 2. merge ASCII and 2-byte codewords
-      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
-      const uint8x16_t utf8_unpacked =
-          vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
-      // 3. prepare bitmask for 8-bit lookup
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t mask = simdutf_make_uint16x8_t(
-          0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
-#else
-      const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
-                               0x0002, 0x0008, 0x0020, 0x0080};
-#endif
-      uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
-      // 4. pack the bytes
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
-      const uint8x16_t shuffle = vld1q_u8(row + 1);
-      const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+simdutf_warn_unused result
+implementation::convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
+      buf, len, latin1_output);
+}
 
-      // 5. store bytes
-      vst1q_u8(utf8_output, utf8_packed);
+simdutf_warn_unused result
+implementation::convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
+      buf, len, latin1_output);
+}
 
-      // 6. adjust pointers
-      buf += 8;
-      utf8_output += row[0];
-      continue;
-    }
-    const uint16x8_t surrogates_bytemask =
-        vceqq_u16(vandq_u16(in, v_f800), v_d800);
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (vmaxvq_u16(surrogates_bytemask) == 0) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t dup_even = simdutf_make_uint16x8_t(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
-#else
-      const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
-                                   0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
-#endif
-      /* In this branch we handle three cases:
-         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert_valid<endianness::LITTLE>(
+      buf, len, latin1_output);
+}
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf16_to_latin1::convert_valid<endianness::BIG>(buf, len,
+                                                                 latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert<endianness::LITTLE>(buf, len,
+                                                            utf8_output);
+}
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert<endianness::BIG>(buf, len, utf8_output);
+}
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const uint16x8_t t0 = vreinterpretq_u16_u8(
-          vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
+      buf, len, utf8_output);
+}
 
-      // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
-      const uint16x8_t s0 = vshrq_n_u16(in, 12);
-      // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
-      const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
-      // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-      const uint16x8_t s1s = vshlq_n_u16(s1, 2);
-      // [00bb|bbbb|0000|aaaa]
-      const uint16x8_t s2 = vorrq_u16(s0, s1s);
-      // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
-      const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
-      const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
-      const uint16x8_t m0 =
-          vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
-      const uint16x8_t s4 = veorq_u16(s3, m0);
-#undef simdutf_vec
+simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
+      buf, len, utf8_output);
+}
 
-      // 4. expand code units 16-bit => 32-bit
-      const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
-      const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert_valid<endianness::LITTLE>(buf, len,
+                                                                  utf8_output);
+}
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-      const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-      const uint16x8_t onemask = simdutf_make_uint16x8_t(
-          0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
-      const uint16x8_t twomask = simdutf_make_uint16x8_t(
-          0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
-#else
-      const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                  0x0100, 0x0400, 0x1000, 0x4000};
-      const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
-                                  0x0200, 0x0800, 0x2000, 0x8000};
-#endif
-      const uint16x8_t combined =
-          vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
-                    vandq_u16(one_or_two_bytes_bytemask, twomask));
-      const uint16_t mask = vaddvq_u16(combined);
-      // The following fast path may or may not be beneficial.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
-        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
-        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
-        vst1q_u8(utf8_output, utf8_0);
-        utf8_output += 12;
-        vst1q_u8(utf8_output, utf8_1);
-        utf8_output += 12;
-        buf += 8;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf16_to_utf8::convert_valid<endianness::BIG>(buf, len,
+                                                               utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
-      const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf32_to_latin1::convert(buf, len, latin1_output);
+}
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
-      const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf32_to_latin1::convert_with_errors(buf, len, latin1_output);
+}
 
-      vst1q_u8(utf8_output, utf8_0);
-      utf8_output += row0[0];
-      vst1q_u8(utf8_output, utf8_1);
-      utf8_output += row1[0];
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return scalar::utf32_to_latin1::convert_valid(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-      buf += 8;
-      // surrogate pair(s) in a register
-    } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if ((word & 0xFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xF800) != 0xD800) {
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
-                                   : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k - 1),
-                reinterpret_cast<char *>(utf8_output));
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf8_output++ = char((value >> 18) | 0b11110000);
-          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((value & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf32_to_utf8::convert(buf, len, utf8_output);
+}
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        reinterpret_cast<char *>(utf8_output));
+simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf32_to_utf8::convert_with_errors(buf, len, utf8_output);
 }
-/* end file src/arm64/arm_convert_utf16_to_utf8.cpp */
 
-/* begin file src/arm64/arm_base64.cpp */
-/**
- * References and further reading:
- *
- * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
- * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
- * https://arxiv.org/abs/1910.05109
- *
- * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
- * Instructions, ACM Transactions on the Web 12 (3), 2018.
- * https://arxiv.org/abs/1704.00605
- *
- * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
- * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
- * Request for Comments: 4648.
- *
- * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
- * http://www.alfredklomp.com/programming/sse-base64/. (2014).
- *
- * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
- * acceleration. https://github.com/aklomp/base64. (2014).
- *
- * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
- * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
- *
- * Nick Kopp. 2013. Base64 Encoding on a GPU.
- * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
- */
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  return scalar::utf32_to_utf8::convert_valid(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-size_t encode_base64(char *dst, const char *src, size_t srclen,
-                     base64_options options) {
-  // credit: Wojciech Muła
-  uint8_t *out = (uint8_t *)dst;
-  constexpr static uint8_t source_table[64] = {
-      'A', 'Q', 'g', 'w', 'B', 'R', 'h', 'x', 'C', 'S', 'i', 'y', 'D',
-      'T', 'j', 'z', 'E', 'U', 'k', '0', 'F', 'V', 'l', '1', 'G', 'W',
-      'm', '2', 'H', 'X', 'n', '3', 'I', 'Y', 'o', '4', 'J', 'Z', 'p',
-      '5', 'K', 'a', 'q', '6', 'L', 'b', 'r', '7', 'M', 'c', 's', '8',
-      'N', 'd', 't', '9', 'O', 'e', 'u', '+', 'P', 'f', 'v', '/',
-  };
-  constexpr static uint8_t source_table_url[64] = {
-      'A', 'Q', 'g', 'w', 'B', 'R', 'h', 'x', 'C', 'S', 'i', 'y', 'D',
-      'T', 'j', 'z', 'E', 'U', 'k', '0', 'F', 'V', 'l', '1', 'G', 'W',
-      'm', '2', 'H', 'X', 'n', '3', 'I', 'Y', 'o', '4', 'J', 'Z', 'p',
-      '5', 'K', 'a', 'q', '6', 'L', 'b', 'r', '7', 'M', 'c', 's', '8',
-      'N', 'd', 't', '9', 'O', 'e', 'u', '-', 'P', 'f', 'v', '_',
-  };
-  const uint8x16_t v3f = vdupq_n_u8(0x3f);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  // When trying to load a uint8_t array, Visual Studio might
-  // error with: error C2664: '__n128x4 neon_ld4m_q8(const char *)':
-  // cannot convert argument 1 from 'const uint8_t [64]' to 'const char *
-  const uint8x16x4_t table = vld4q_u8(
-      (reinterpret_cast<const char *>(options & base64_url) ? source_table_url
-                                                            : source_table));
-#else
-  const uint8x16x4_t table =
-      vld4q_u8((options & base64_url) ? source_table_url : source_table);
-#endif
-  size_t i = 0;
-  for (; i + 16 * 3 <= srclen; i += 16 * 3) {
-    const uint8x16x3_t in = vld3q_u8((const uint8_t *)src + i);
-    uint8x16x4_t result;
-    result.val[0] = vshrq_n_u8(in.val[0], 2);
-    result.val[1] =
-        vandq_u8(vsliq_n_u8(vshrq_n_u8(in.val[1], 4), in.val[0], 4), v3f);
-    result.val[2] =
-        vandq_u8(vsliq_n_u8(vshrq_n_u8(in.val[2], 6), in.val[1], 2), v3f);
-    result.val[3] = vandq_u8(in.val[2], v3f);
-    result.val[0] = vqtbl4q_u8(table, result.val[0]);
-    result.val[1] = vqtbl4q_u8(table, result.val[1]);
-    result.val[2] = vqtbl4q_u8(table, result.val[2]);
-    result.val[3] = vqtbl4q_u8(table, result.val[3]);
-    vst4q_u8(out, result);
-    out += 64;
-  }
-  out += scalar::base64::tail_encode_base64((char *)out, src + i, srclen - i,
-                                            options);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert<endianness::LITTLE>(buf, len,
+                                                             utf16_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert<endianness::BIG>(buf, len,
+                                                          utf16_output);
+}
 
-  return size_t((char *)out - dst);
+simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
+      buf, len, utf16_output);
 }
 
-static inline void compress(uint8x16_t data, uint16_t mask, char *output) {
-  if (mask == 0) {
-    vst1q_u8((uint8_t *)output, data);
-    return;
-  }
-  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
-  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
-  uint64x2_t compactmasku64 = {tables::base64::thintable_epi8[mask1],
-                               tables::base64::thintable_epi8[mask2]};
-  uint8x16_t compactmask = vreinterpretq_u8_u64(compactmasku64);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  const uint8x16_t off =
-      simdutf_make_uint8x16_t(0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8);
-#else
-  const uint8x16_t off = {0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8};
-#endif
+simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
+      buf, len, utf16_output);
+}
 
-  compactmask = vaddq_u8(compactmask, off);
-  uint8x16_t pruned = vqtbl1q_u8(data, compactmask);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert_valid<endianness::LITTLE>(
+      buf, len, utf16_output);
+}
 
-  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
-  // then load the corresponding mask, what it does is to write
-  // only the first pop1 bytes from the first 8 bytes, and then
-  // it fills in with the bytes from the second 8 bytes + some filling
-  // at the end.
-  compactmask = vld1q_u8(tables::base64::pshufb_combine_table + pop1 * 8);
-  uint8x16_t answer = vqtbl1q_u8(pruned, compactmask);
-  vst1q_u8((uint8_t *)output, answer);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return scalar::utf32_to_utf16::convert_valid<endianness::BIG>(buf, len,
+                                                                utf16_output);
 }
 
-struct block64 {
-  uint8x16_t chunks[4];
-};
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert<endianness::LITTLE>(buf, len,
+                                                             utf32_output);
+}
 
-static_assert(sizeof(block64) == 64, "block64 is not 64 bytes");
-template <bool base64_url> uint64_t to_base64_mask(block64 *b, bool *error) {
-  uint8x16_t v0f = vdupq_n_u8(0xf);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert<endianness::BIG>(buf, len,
+                                                          utf32_output);
+}
 
-  uint8x16_t underscore0, underscore1, underscore2, underscore3;
-  if (base64_url) {
-    underscore0 = vceqq_u8(b->chunks[0], vdupq_n_u8(0x5f));
-    underscore1 = vceqq_u8(b->chunks[1], vdupq_n_u8(0x5f));
-    underscore2 = vceqq_u8(b->chunks[2], vdupq_n_u8(0x5f));
-    underscore3 = vceqq_u8(b->chunks[3], vdupq_n_u8(0x5f));
-  } else {
-    (void)underscore0;
-    (void)underscore1;
-    (void)underscore2;
-    (void)underscore3;
-  }
+simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
+      buf, len, utf32_output);
+}
 
-  uint8x16_t lo_nibbles0 = vandq_u8(b->chunks[0], v0f);
-  uint8x16_t lo_nibbles1 = vandq_u8(b->chunks[1], v0f);
-  uint8x16_t lo_nibbles2 = vandq_u8(b->chunks[2], v0f);
-  uint8x16_t lo_nibbles3 = vandq_u8(b->chunks[3], v0f);
+simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
+      buf, len, utf32_output);
+}
 
-  // Needed by the decoding step.
-  uint8x16_t hi_nibbles0 = vshrq_n_u8(b->chunks[0], 4);
-  uint8x16_t hi_nibbles1 = vshrq_n_u8(b->chunks[1], 4);
-  uint8x16_t hi_nibbles2 = vshrq_n_u8(b->chunks[2], 4);
-  uint8x16_t hi_nibbles3 = vshrq_n_u8(b->chunks[3], 4);
-  uint8x16_t lut_lo;
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  if (base64_url) {
-    lut_lo =
-        simdutf_make_uint8x16_t(0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
-                                0x70, 0x61, 0xe1, 0xf4, 0xe5, 0xa5, 0xf4, 0xf4);
-  } else {
-    lut_lo =
-        simdutf_make_uint8x16_t(0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
-                                0x70, 0x61, 0xe1, 0xb4, 0xe5, 0xe5, 0xf4, 0xb4);
-  }
-#else
-  if (base64_url) {
-    lut_lo = uint8x16_t{0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
-                        0x70, 0x61, 0xe1, 0xf4, 0xe5, 0xa5, 0xf4, 0xf4};
-  } else {
-    lut_lo = uint8x16_t{0x3a, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70, 0x70,
-                        0x70, 0x61, 0xe1, 0xb4, 0xe5, 0xe5, 0xf4, 0xb4};
-  }
-#endif
-  uint8x16_t lo0 = vqtbl1q_u8(lut_lo, lo_nibbles0);
-  uint8x16_t lo1 = vqtbl1q_u8(lut_lo, lo_nibbles1);
-  uint8x16_t lo2 = vqtbl1q_u8(lut_lo, lo_nibbles2);
-  uint8x16_t lo3 = vqtbl1q_u8(lut_lo, lo_nibbles3);
-  uint8x16_t lut_hi;
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  if (base64_url) {
-    lut_hi =
-        simdutf_make_uint8x16_t(0x11, 0x20, 0x42, 0x80, 0x8, 0x4, 0x8, 0x4,
-                                0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20);
-  } else {
-    lut_hi =
-        simdutf_make_uint8x16_t(0x11, 0x20, 0x42, 0x80, 0x8, 0x4, 0x8, 0x4,
-                                0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20);
-  }
-#else
-  if (base64_url) {
-    lut_hi = uint8x16_t{0x11, 0x20, 0x42, 0x80, 0x8,  0x4,  0x8,  0x4,
-                        0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20};
-  } else {
-    lut_hi = uint8x16_t{0x11, 0x20, 0x42, 0x80, 0x8,  0x4,  0x8,  0x4,
-                        0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20};
-  }
-#endif
-  uint8x16_t hi0 = vqtbl1q_u8(lut_hi, hi_nibbles0);
-  uint8x16_t hi1 = vqtbl1q_u8(lut_hi, hi_nibbles1);
-  uint8x16_t hi2 = vqtbl1q_u8(lut_hi, hi_nibbles2);
-  uint8x16_t hi3 = vqtbl1q_u8(lut_hi, hi_nibbles3);
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert_valid<endianness::LITTLE>(
+      buf, len, utf32_output);
+}
 
-  if (base64_url) {
-    hi0 = vbicq_u8(hi0, underscore0);
-    hi1 = vbicq_u8(hi1, underscore1);
-    hi2 = vbicq_u8(hi2, underscore2);
-    hi3 = vbicq_u8(hi3, underscore3);
-  }
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return scalar::utf16_to_utf32::convert_valid<endianness::BIG>(buf, len,
+                                                                utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-  uint8_t checks =
-      vmaxvq_u8(vorrq_u8(vorrq_u8(vandq_u8(lo0, hi0), vandq_u8(lo1, hi1)),
-                         vorrq_u8(vandq_u8(lo2, hi2), vandq_u8(lo3, hi3))));
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  const uint8x16_t bit_mask =
-      simdutf_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                              0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
-#else
-  const uint8x16_t bit_mask = {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
-                               0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
-#endif
-  uint64_t badcharmask = 0;
-  *error = checks > 0x3;
-  if (checks) {
-    // Add each of the elements next to each other, successively, to stuff each
-    // 8 byte mask into one.
-    uint8x16_t test0 = vtstq_u8(lo0, hi0);
-    uint8x16_t test1 = vtstq_u8(lo1, hi1);
-    uint8x16_t test2 = vtstq_u8(lo2, hi2);
-    uint8x16_t test3 = vtstq_u8(lo3, hi3);
-    uint8x16_t sum0 =
-        vpaddq_u8(vandq_u8(test0, bit_mask), vandq_u8(test1, bit_mask));
-    uint8x16_t sum1 =
-        vpaddq_u8(vandq_u8(test2, bit_mask), vandq_u8(test3, bit_mask));
-    sum0 = vpaddq_u8(sum0, sum1);
-    sum0 = vpaddq_u8(sum0, sum0);
-    badcharmask = vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
-  }
-  // This is the transformation step that can be done while we are waiting for
-  // sum0
-  uint8x16_t roll_lut;
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-  if (base64_url) {
-    roll_lut =
-        simdutf_make_uint8x16_t(0xe0, 0x11, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
-                                0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0);
-  } else {
-    roll_lut =
-        simdutf_make_uint8x16_t(0x0, 0x10, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
-                                0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0);
+#if SIMDUTF_FEATURE_UTF16
+void implementation::change_endianness_utf16(const char16_t *input,
+                                             size_t length,
+                                             char16_t *output) const noexcept {
+  scalar::utf16::change_endianness_utf16(input, length, output);
+}
+
+simdutf_warn_unused size_t implementation::count_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::count_code_points<endianness::LITTLE>(input, length);
+}
+
+simdutf_warn_unused size_t implementation::count_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::count_code_points<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused size_t
+implementation::count_utf8(const char *input, size_t length) const noexcept {
+  return scalar::utf8::count_code_points(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
+    const char *buf, size_t len) const noexcept {
+  return scalar::utf8::count_code_points(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
+    const char *input, size_t length) const noexcept {
+  size_t answer = length;
+  size_t i = 0;
+  auto pop = [](uint64_t v) {
+    return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) *
+                        UINT64_C(0x0101010101010101) >>
+                    56);
+  };
+  for (; i + 32 <= length; i += 32) {
+    uint64_t v;
+    memcpy(&v, input + i, 8);
+    answer += pop(v);
+    memcpy(&v, input + i + 8, sizeof(v));
+    answer += pop(v);
+    memcpy(&v, input + i + 16, sizeof(v));
+    answer += pop(v);
+    memcpy(&v, input + i + 24, sizeof(v));
+    answer += pop(v);
   }
-#else
-  if (base64_url) {
-    roll_lut = uint8x16_t{0xe0, 0x11, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
-                          0x0,  0x0,  0x0,  0x0, 0x0,  0x0,  0x0,  0x0};
-  } else {
-    roll_lut = uint8x16_t{0x0, 0x10, 0x13, 0x4, 0xbf, 0xbf, 0xb9, 0xb9,
-                          0x0, 0x0,  0x0,  0x0, 0x0,  0x0,  0x0,  0x0};
+  for (; i + 8 <= length; i += 8) {
+    uint64_t v;
+    memcpy(&v, input + i, sizeof(v));
+    answer += pop(v);
   }
-#endif
-  uint8x16_t vsecond_last = base64_url ? vdupq_n_u8(0x2d) : vdupq_n_u8(0x2f);
-  if (base64_url) {
-    hi_nibbles0 = vbicq_u8(hi_nibbles0, underscore0);
-    hi_nibbles1 = vbicq_u8(hi_nibbles1, underscore1);
-    hi_nibbles2 = vbicq_u8(hi_nibbles2, underscore2);
-    hi_nibbles3 = vbicq_u8(hi_nibbles3, underscore3);
+  for (; i + 1 <= length; i += 1) {
+    answer += static_cast<uint8_t>(input[i]) >> 7;
   }
-  uint8x16_t roll0 = vqtbl1q_u8(
-      roll_lut, vaddq_u8(vceqq_u8(b->chunks[0], vsecond_last), hi_nibbles0));
-  uint8x16_t roll1 = vqtbl1q_u8(
-      roll_lut, vaddq_u8(vceqq_u8(b->chunks[1], vsecond_last), hi_nibbles1));
-  uint8x16_t roll2 = vqtbl1q_u8(
-      roll_lut, vaddq_u8(vceqq_u8(b->chunks[2], vsecond_last), hi_nibbles2));
-  uint8x16_t roll3 = vqtbl1q_u8(
-      roll_lut, vaddq_u8(vceqq_u8(b->chunks[3], vsecond_last), hi_nibbles3));
-  b->chunks[0] = vaddq_u8(b->chunks[0], roll0);
-  b->chunks[1] = vaddq_u8(b->chunks[1], roll1);
-  b->chunks[2] = vaddq_u8(b->chunks[2], roll2);
-  b->chunks[3] = vaddq_u8(b->chunks[3], roll3);
-  return badcharmask;
+  return answer;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-void copy_block(block64 *b, char *output) {
-  vst1q_u8((uint8_t *)output, b->chunks[0]);
-  vst1q_u8((uint8_t *)output + 16, b->chunks[1]);
-  vst1q_u8((uint8_t *)output + 32, b->chunks[2]);
-  vst1q_u8((uint8_t *)output + 48, b->chunks[3]);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(input,
+                                                                   length);
 }
 
-uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
-  uint64_t popcounts =
-      vget_lane_u64(vreinterpret_u64_u8(vcnt_u8(vcreate_u8(~mask))), 0);
-  uint64_t offsets = popcounts * 0x0101010101010101;
-  compress(b->chunks[0], uint16_t(mask), output);
-  compress(b->chunks[1], uint16_t(mask >> 16), &output[(offsets >> 8) & 0xFF]);
-  compress(b->chunks[2], uint16_t(mask >> 32), &output[(offsets >> 24) & 0xFF]);
-  compress(b->chunks[3], uint16_t(mask >> 48), &output[(offsets >> 40) & 0xFF]);
-  return offsets >> 56;
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-// The caller of this function is responsible to ensure that there are 64 bytes
-// available from reading at src. The data is read into a block64 structure.
-void load_block(block64 *b, const char *src) {
-  b->chunks[0] = vld1q_u8(reinterpret_cast<const uint8_t *>(src));
-  b->chunks[1] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 16);
-  b->chunks[2] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 32);
-  b->chunks[3] = vld1q_u8(reinterpret_cast<const uint8_t *>(src) + 48);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::utf32_length_from_utf16<endianness::LITTLE>(input,
+                                                                    length);
 }
 
-// The caller of this function is responsible to ensure that there are 32 bytes
-// available from reading at data. It returns a 16-byte value, narrowing with
-// saturation the 16-bit words.
-inline uint8x16_t load_satured(const uint16_t *data) {
-  uint16x8_t in1 = vld1q_u16(data);
-  uint16x8_t in2 = vld1q_u16(data + 8);
-  return vqmovn_high_u16(vqmovn_u16(in1), in2);
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return scalar::utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-// The caller of this function is responsible to ensure that there are 128 bytes
-// available from reading at src. The data is read into a block64 structure.
-void load_block(block64 *b, const char16_t *src) {
-  b->chunks[0] = load_satured(reinterpret_cast<const uint16_t *>(src));
-  b->chunks[1] = load_satured(reinterpret_cast<const uint16_t *>(src) + 16);
-  b->chunks[2] = load_satured(reinterpret_cast<const uint16_t *>(src) + 32);
-  b->chunks[3] = load_satured(reinterpret_cast<const uint16_t *>(src) + 48);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return scalar::utf8::utf16_length_from_utf8(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-// decode 64 bytes and output 48 bytes
-void base64_decode_block(char *out, const char *src) {
-  uint8x16x4_t str = vld4q_u8((uint8_t *)src);
-  uint8x16x3_t outvec;
-  outvec.val[0] =
-      vorrq_u8(vshlq_n_u8(str.val[0], 2), vshrq_n_u8(str.val[1], 4));
-  outvec.val[1] =
-      vorrq_u8(vshlq_n_u8(str.val[1], 4), vshrq_n_u8(str.val[2], 2));
-  outvec.val[2] = vorrq_u8(vshlq_n_u8(str.val[2], 6), str.val[3]);
-  vst3q_u8((uint8_t *)out, outvec);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  return scalar::utf32::utf8_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-template <bool base64_url, bool ignore_garbage, typename char_type>
-full_result
-compress_decode_base64(char *dst, const char_type *src, size_t srclen,
-                       base64_options options,
-                       last_chunk_handling_options last_chunk_options) {
-  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
-                                        : tables::base64::to_base64_value;
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  return scalar::utf32::utf16_length_from_utf32(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return scalar::utf8::count_code_points(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+  while (length > 0 &&
+         scalar::base64::is_ascii_white_space(input[length - 1])) {
+    length--;
+  }
   size_t equallocation =
-      srclen; // location of the first padding character if any
-  // skip trailing spaces
-  while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
-         to_base64[uint8_t(src[srclen - 1])] == 64) {
-    srclen--;
+      length; // location of the first padding character if any
+  size_t equalsigns = 0;
+  if (length > 0 && input[length - 1] == '=') {
+    equallocation = length - 1;
+    length -= 1;
+    equalsigns++;
+    while (length > 0 &&
+           scalar::base64::is_ascii_white_space(input[length - 1])) {
+      length--;
+    }
+    if (length > 0 && input[length - 1] == '=') {
+      equallocation = length - 1;
+      equalsigns++;
+      length -= 1;
+    }
+  }
+  if (length == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation};
+    }
+    return {SUCCESS, 0};
+  }
+  result r = scalar::base64::base64_tail_decode(
+      output, input, length, equalsigns, options, last_chunk_options);
+  if (last_chunk_options != stop_before_partial &&
+      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
+    // additional checks
+    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation};
+    }
+  }
+  return r;
+}
+
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+  while (length > 0 &&
+         scalar::base64::is_ascii_white_space(input[length - 1])) {
+    length--;
+  }
+  size_t equallocation =
+      length; // location of the first padding character if any
+  size_t equalsigns = 0;
+  if (length > 0 && input[length - 1] == '=') {
+    equallocation = length - 1;
+    length -= 1;
+    equalsigns++;
+    while (length > 0 &&
+           scalar::base64::is_ascii_white_space(input[length - 1])) {
+      length--;
+    }
+    if (length > 0 && input[length - 1] == '=') {
+      equallocation = length - 1;
+      equalsigns++;
+      length -= 1;
+    }
+  }
+  if (length == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, 0, 0};
+  }
+  full_result r = scalar::base64::base64_tail_decode(
+      output, input, length, equalsigns, options, last_chunk_options);
+  if (last_chunk_options != stop_before_partial &&
+      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
+    // additional checks
+    if ((r.output_count % 3 == 0) ||
+        ((r.output_count % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, r.output_count};
+    }
+  }
+  return r;
+}
+
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+  while (length > 0 &&
+         scalar::base64::is_ascii_white_space(input[length - 1])) {
+    length--;
+  }
+  size_t equallocation =
+      length; // location of the first padding character if any
+  size_t equalsigns = 0;
+  if (length > 0 && input[length - 1] == '=') {
+    equallocation = length - 1;
+    length -= 1;
+    equalsigns++;
+    while (length > 0 &&
+           scalar::base64::is_ascii_white_space(input[length - 1])) {
+      length--;
+    }
+    if (length > 0 && input[length - 1] == '=') {
+      equallocation = length - 1;
+      equalsigns++;
+      length -= 1;
+    }
+  }
+  if (length == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation};
+    }
+    return {SUCCESS, 0};
+  }
+  result r = scalar::base64::base64_tail_decode(
+      output, input, length, equalsigns, options, last_chunk_options);
+  if (last_chunk_options != stop_before_partial &&
+      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
+    // additional checks
+    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation};
+    }
+  }
+  return r;
+}
+
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  const bool ignore_garbage =
+      (options == base64_options::base64_url_accept_garbage) ||
+      (options == base64_options::base64_default_accept_garbage);
+  while (length > 0 &&
+         scalar::base64::is_ascii_white_space(input[length - 1])) {
+    length--;
   }
+  size_t equallocation =
+      length; // location of the first padding character if any
   size_t equalsigns = 0;
-  if (srclen > 0 && src[srclen - 1] == '=') {
-    equallocation = srclen - 1;
-    srclen--;
-    equalsigns = 1;
-    // skip trailing spaces
-    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
-           to_base64[uint8_t(src[srclen - 1])] == 64) {
-      srclen--;
+  if (length > 0 && input[length - 1] == '=') {
+    equallocation = length - 1;
+    length -= 1;
+    equalsigns++;
+    while (length > 0 &&
+           scalar::base64::is_ascii_white_space(input[length - 1])) {
+      length--;
     }
-    if (srclen > 0 && src[srclen - 1] == '=') {
-      equallocation = srclen - 1;
-      srclen--;
-      equalsigns = 2;
+    if (length > 0 && input[length - 1] == '=') {
+      equallocation = length - 1;
+      equalsigns++;
+      length -= 1;
     }
   }
-  if (srclen == 0) {
+  if (length == 0) {
     if (!ignore_garbage && equalsigns > 0) {
       if (last_chunk_options == last_chunk_handling_options::strict) {
         return {BASE64_INPUT_REMAINDER, 0, 0};
@@ -18669,1165 +26491,1087 @@ compress_decode_base64(char *dst, const char_type *src, size_t srclen,
     }
     return {SUCCESS, 0, 0};
   }
-  const char_type *const srcinit = src;
-  const char *const dstinit = dst;
-  const char_type *const srcend = src + srclen;
-
-  constexpr size_t block_size = 10;
-  char buffer[block_size * 64];
-  char *bufferptr = buffer;
-  if (srclen >= 64) {
-    const char_type *const srcend64 = src + srclen - 64;
-    while (src <= srcend64) {
-      block64 b;
-      load_block(&b, src);
-      src += 64;
-      bool error = false;
-      uint64_t badcharmask = to_base64_mask<base64_url>(&b, &error);
-      if (badcharmask) {
-        if (error && !ignore_garbage) {
-          src -= 64;
-          while (src < srcend && scalar::base64::is_eight_byte(*src) &&
-                 to_base64[uint8_t(*src)] <= 64) {
-            src++;
-          }
-          if (src < srcend) {
-            // should never happen
-          }
-          return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
-                  size_t(dst - dstinit)};
-        }
-      }
-
-      if (badcharmask != 0) {
-        // optimization opportunity: check for simple masks like those made of
-        // continuous 1s followed by continuous 0s. And masks containing a
-        // single bad character.
-        bufferptr += compress_block(&b, badcharmask, bufferptr);
-      } else {
-        // optimization opportunity: if bufferptr == buffer and mask == 0, we
-        // can avoid the call to compress_block and decode directly.
-        copy_block(&b, bufferptr);
-        bufferptr += 64;
-      }
-      if (bufferptr >= (block_size - 1) * 64 + buffer) {
-        for (size_t i = 0; i < (block_size - 1); i++) {
-          base64_decode_block(dst, buffer + i * 64);
-          dst += 48;
-        }
-        std::memcpy(buffer, buffer + (block_size - 1) * 64,
-                    64); // 64 might be too much
-        bufferptr -= (block_size - 1) * 64;
-      }
-    }
-  }
-  char *buffer_start = buffer;
-  // Optimization note: if this is almost full, then it is worth our
-  // time, otherwise, we should just decode directly.
-  int last_block = (int)((bufferptr - buffer_start) % 64);
-  if (last_block != 0 && srcend - src + last_block >= 64) {
-    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
-      uint8_t val = to_base64[uint8_t(*src)];
-      *bufferptr = char(val);
-      if ((!scalar::base64::is_eight_byte(*src) || val > 64) &&
-          !ignore_garbage) {
-        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      }
-      bufferptr += (val <= 63);
-      src++;
+  full_result r = scalar::base64::base64_tail_decode(
+      output, input, length, equalsigns, options, last_chunk_options);
+  if (last_chunk_options != stop_before_partial &&
+      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
+    // additional checks
+    if ((r.output_count % 3 == 0) ||
+        ((r.output_count % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, r.output_count};
     }
   }
+  return r;
+}
 
-  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
-    base64_decode_block(dst, buffer_start);
-    dst += 48;
-  }
-  if ((bufferptr - buffer_start) % 64 != 0) {
-    while (buffer_start + 4 < bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 4);
+size_t implementation::binary_to_base64(const char *input, size_t length,
+                                        char *output,
+                                        base64_options options) const noexcept {
+  return scalar::base64::tail_encode_base64(output, input, length, options);
+}
+#endif // SIMDUTF_FEATURE_BASE64
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    if (buffer_start + 4 <= bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 3);
+} // namespace fallback
+} // namespace simdutf
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
-    // backtrack
-    int leftover = int(bufferptr - buffer_start);
-    while (leftover > 0) {
-      if (!ignore_garbage) {
-        while (to_base64[uint8_t(*(src - 1))] == 64) {
-          src--;
-        }
-      } else {
-        while (to_base64[uint8_t(*(src - 1))] >= 64) {
-          src--;
-        }
-      }
-      src--;
-      leftover--;
-    }
-  }
-  if (src < srcend + equalsigns) {
-    full_result r = scalar::base64::base64_tail_decode(
-        dst, src, srcend - src, equalsigns, options, last_chunk_options);
-    r.input_count += size_t(src - srcinit);
-    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
-        r.error == error_code::BASE64_EXTRA_BITS) {
-      return r;
-    } else {
-      r.output_count += size_t(dst - dstinit);
-    }
-    if (last_chunk_options != stop_before_partial &&
-        r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-      // additional checks
-      if ((r.output_count % 3 == 0) ||
-          ((r.output_count % 3) + 1 + equalsigns != 4)) {
-        r.error = error_code::INVALID_BASE64_CHARACTER;
-        r.input_count = equallocation;
-      }
-    }
-    return r;
-  }
-  if (equalsigns > 0 && !ignore_garbage) {
-    if ((size_t(dst - dstinit) % 3 == 0) ||
-        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
-    }
-  }
-  return {SUCCESS, srclen, size_t(dst - dstinit)};
-}
-/* end file src/arm64/arm_base64.cpp */
-/* begin file src/arm64/arm_convert_utf32_to_latin1.cpp */
-std::pair<const char32_t *, char *>
-arm_convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                            char *latin1_output) {
-  const char32_t *end = buf + len;
-  while (end - buf >= 8) {
-    uint32x4_t in1 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
-    uint32x4_t in2 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
+/* begin file src/simdutf/fallback/end.h */
+/* end file src/simdutf/fallback/end.h */
+/* end file src/fallback/implementation.cpp */
+#endif
+#if SIMDUTF_IMPLEMENTATION_ICELAKE
+/* begin file src/icelake/implementation.cpp */
+#include <tuple>
+#include <utility>
 
-    uint16x8_t utf16_packed = vcombine_u16(vqmovn_u32(in1), vqmovn_u32(in2));
-    if (vmaxvq_u16(utf16_packed) <= 0xff) {
-      // 1. pack the bytes
-      uint8x8_t latin1_packed = vmovn_u16(utf16_packed);
-      // 2. store (8 bytes)
-      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
-      // 3. adjust pointers
-      buf += 8;
-      latin1_output += 8;
-    } else {
-      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
-    }
-  } // while
-  return std::make_pair(buf, latin1_output);
-}
+/* begin file src/simdutf/icelake/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "icelake"
+// #define SIMDUTF_IMPLEMENTATION icelake
 
-std::pair<result, char *>
-arm_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                        char *latin1_output) {
-  const char32_t *start = buf;
-  const char32_t *end = buf + len;
+#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+// nothing needed.
+#else
+SIMDUTF_TARGET_ICELAKE
+#endif
 
-  while (end - buf >= 8) {
-    uint32x4_t in1 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
-    uint32x4_t in2 = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+// clang-format off
+SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
+// clang-format on
+#endif // end of workaround
+/* end file src/simdutf/icelake/begin.h */
+namespace simdutf {
+namespace icelake {
+namespace {
+#ifndef SIMDUTF_ICELAKE_H
+  #error "icelake.h must be included"
+#endif
+using namespace simd;
 
-    uint16x8_t utf16_packed = vcombine_u16(vqmovn_u32(in1), vqmovn_u32(in2));
+/* begin file src/icelake/icelake_macros.inl.cpp */
 
-    if (vmaxvq_u16(utf16_packed) <= 0xff) {
-      // 1. pack the bytes
-      uint8x8_t latin1_packed = vmovn_u16(utf16_packed);
-      // 2. store (8 bytes)
-      vst1_u8(reinterpret_cast<uint8_t *>(latin1_output), latin1_packed);
-      // 3. adjust pointers
-      buf += 8;
-      latin1_output += 8;
-    } else {
-      // Let us do a scalar fallback.
-      for (int k = 0; k < 8; k++) {
-        uint32_t word = buf[k];
-        if (word <= 0xff) {
-          *latin1_output++ = char(word);
-        } else {
-          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
-                                latin1_output);
-        }
-      }
-    }
-  } // while
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        latin1_output);
-}
-/* end file src/arm64/arm_convert_utf32_to_latin1.cpp */
-/* begin file src/arm64/arm_convert_utf32_to_utf16.cpp */
-template <endianness big_endian>
-std::pair<const char32_t *, char16_t *>
-arm_convert_utf32_to_utf16(const char32_t *buf, size_t len,
-                           char16_t *utf16_out) {
-  uint16_t *utf16_output = reinterpret_cast<uint16_t *>(utf16_out);
-  const char32_t *end = buf + len;
+/*
+    This upcoming macro (SIMDUTF_ICELAKE_TRANSCODE16) takes 16 + 4 bytes (of a
+   UTF-8 string) and loads all possible 4-byte substring into an AVX512
+   register.
 
-  uint16x4_t forbidden_bytemask = vmov_n_u16(0x0);
+    For example if we have bytes abcdefgh... we create following 32-bit lanes
 
-  while (end - buf >= 4) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+    [abcd|bcde|cdef|defg|efgh|...]
+     ^                          ^
+     byte 0 of reg              byte 63 of reg
+*/
+/** pshufb
+        # lane{0,1,2} have got bytes: [  0,  1,  2,  3,  4,  5,  6,  8,  9, 10,
+   11, 12, 13, 14, 15] # lane3 has got bytes:        [ 16, 17, 18, 19,  4,  5,
+   6,  8,  9, 10, 11, 12, 13, 14, 15]
 
-    // Check if no bits set above 16th
-    if (vmaxvq_u32(in) <= 0xFFFF) {
-      uint16x4_t utf16_packed = vmovn_u32(in);
+        expand_ver2 = [
+            # lane 0:
+            0, 1, 2, 3,
+            1, 2, 3, 4,
+            2, 3, 4, 5,
+            3, 4, 5, 6,
 
-      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
-      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
-      forbidden_bytemask = vorr_u16(vand_u16(vcle_u16(utf16_packed, v_dfff),
-                                             vcge_u16(utf16_packed, v_d800)),
-                                    forbidden_bytemask);
+            # lane 1:
+            4, 5, 6, 7,
+            5, 6, 7, 8,
+            6, 7, 8, 9,
+            7, 8, 9, 10,
 
-      if (!match_system(big_endian)) {
-        utf16_packed =
-            vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(utf16_packed)));
-      }
-      vst1_u16(utf16_output, utf16_packed);
-      utf16_output += 4;
-      buf += 4;
-    } else {
-      size_t forward = 3;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char16_t *>(utf16_output));
-          }
-          *utf16_output++ = !match_system(big_endian)
-                                ? char16_t(word >> 8 | word << 8)
-                                : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char16_t *>(utf16_output));
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (!match_system(big_endian)) {
-            high_surrogate =
-                uint16_t(high_surrogate >> 8 | high_surrogate << 8);
-            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
-        }
-      }
-      buf += k;
-    }
+            # lane 2:
+             8,  9, 10, 11,
+             9, 10, 11, 12,
+            10, 11, 12, 13,
+            11, 12, 13, 14,
+
+            # lane 3 order: 13, 14, 15, 16 14, 15, 16, 17, 15, 16, 17, 18, 16,
+   17, 18, 19 12, 13, 14, 15, 13, 14, 15,  0, 14, 15,  0,  1, 15,  0,  1,  2,
+        ]
+*/
+
+#define SIMDUTF_ICELAKE_TRANSCODE16(LANE0, LANE1, MASKED)                      \
+  {                                                                            \
+    const __m512i merged = _mm512_mask_mov_epi32(LANE0, 0x1000, LANE1);        \
+    const __m512i expand_ver2 = _mm512_setr_epi64(                             \
+        0x0403020103020100, 0x0605040305040302, 0x0807060507060504,            \
+        0x0a09080709080706, 0x0c0b0a090b0a0908, 0x0e0d0c0b0d0c0b0a,            \
+        0x000f0e0d0f0e0d0c, 0x0201000f01000f0e);                               \
+    const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);            \
+                                                                               \
+    __mmask16 leading_bytes;                                                   \
+    const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);                       \
+    const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);                   \
+    const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);                       \
+    leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);                 \
+                                                                               \
+    __m512i char_class;                                                        \
+    char_class = _mm512_srli_epi32(input, 4);                                  \
+    /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */                     \
+    const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);                       \
+    const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);                 \
+    char_class =                                                               \
+        _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea); \
+                                                                               \
+    const int valid_count = static_cast<int>(count_ones(leading_bytes));       \
+    const __m512i utf32 = expanded_utf8_to_utf32(char_class, input);           \
+                                                                               \
+    const __m512i out = _mm512_mask_compress_epi32(_mm512_setzero_si512(),     \
+                                                   leading_bytes, utf32);      \
+                                                                               \
+    if (UTF32) {                                                               \
+      if (MASKED) {                                                            \
+        const __mmask16 valid = uint16_t((1 << valid_count) - 1);              \
+        _mm512_mask_storeu_epi32((__m512i *)output, valid, out);               \
+      } else {                                                                 \
+        _mm512_storeu_si512((__m512i *)output, out);                           \
+      }                                                                        \
+      output += valid_count;                                                   \
+    } else {                                                                   \
+      if (MASKED) {                                                            \
+        output += utf32_to_utf16_masked<big_endian>(                           \
+            byteflip, out, valid_count, reinterpret_cast<char16_t *>(output)); \
+      } else {                                                                 \
+        output += utf32_to_utf16<big_endian>(                                  \
+            byteflip, out, valid_count, reinterpret_cast<char16_t *>(output)); \
+      }                                                                        \
+    }                                                                          \
   }
 
-  // check for invalid input
-  if (vmaxv_u16(forbidden_bytemask) != 0) {
-    return std::make_pair(nullptr, reinterpret_cast<char16_t *>(utf16_output));
+#define SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(INPUT, VALID_COUNT, MASKED)       \
+  {                                                                            \
+    if (UTF32) {                                                               \
+      if (MASKED) {                                                            \
+        const __mmask16 valid_mask = uint16_t((1 << VALID_COUNT) - 1);         \
+        _mm512_mask_storeu_epi32((__m512i *)output, valid_mask, INPUT);        \
+      } else {                                                                 \
+        _mm512_storeu_si512((__m512i *)output, INPUT);                         \
+      }                                                                        \
+      output += VALID_COUNT;                                                   \
+    } else {                                                                   \
+      if (MASKED) {                                                            \
+        output += utf32_to_utf16_masked<big_endian>(                           \
+            byteflip, INPUT, VALID_COUNT,                                      \
+            reinterpret_cast<char16_t *>(output));                             \
+      } else {                                                                 \
+        output +=                                                              \
+            utf32_to_utf16<big_endian>(byteflip, INPUT, VALID_COUNT,           \
+                                       reinterpret_cast<char16_t *>(output));  \
+      }                                                                        \
+    }                                                                          \
   }
 
-  return std::make_pair(buf, reinterpret_cast<char16_t *>(utf16_output));
+#define SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)                       \
+  if (UTF32) {                                                                 \
+    const __m128i t0 = _mm512_castsi512_si128(utf8);                           \
+    const __m128i t1 = _mm512_extracti32x4_epi32(utf8, 1);                     \
+    const __m128i t2 = _mm512_extracti32x4_epi32(utf8, 2);                     \
+    const __m128i t3 = _mm512_extracti32x4_epi32(utf8, 3);                     \
+    _mm512_storeu_si512((__m512i *)(output + 0 * 16),                          \
+                        _mm512_cvtepu8_epi32(t0));                             \
+    _mm512_storeu_si512((__m512i *)(output + 1 * 16),                          \
+                        _mm512_cvtepu8_epi32(t1));                             \
+    _mm512_storeu_si512((__m512i *)(output + 2 * 16),                          \
+                        _mm512_cvtepu8_epi32(t2));                             \
+    _mm512_storeu_si512((__m512i *)(output + 3 * 16),                          \
+                        _mm512_cvtepu8_epi32(t3));                             \
+  } else {                                                                     \
+    const __m256i h0 = _mm512_castsi512_si256(utf8);                           \
+    const __m256i h1 = _mm512_extracti64x4_epi64(utf8, 1);                     \
+    if (big_endian) {                                                          \
+      _mm512_storeu_si512(                                                     \
+          (__m512i *)(output + 0 * 16),                                        \
+          _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h0), byteflip));            \
+      _mm512_storeu_si512(                                                     \
+          (__m512i *)(output + 2 * 16),                                        \
+          _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h1), byteflip));            \
+    } else {                                                                   \
+      _mm512_storeu_si512((__m512i *)(output + 0 * 16),                        \
+                          _mm512_cvtepu8_epi16(h0));                           \
+      _mm512_storeu_si512((__m512i *)(output + 2 * 16),                        \
+                          _mm512_cvtepu8_epi16(h1));                           \
+    }                                                                          \
+  }
+/* end file src/icelake/icelake_macros.inl.cpp */
+/* begin file src/icelake/icelake_common.inl.cpp */
+// file included directly
+/**
+ * Store the last N bytes of previous followed by 512-N bytes from input.
+ */
+template <int N> __m512i prev(__m512i input, __m512i previous) {
+  static_assert(N <= 32, "N must be no larger than 32");
+  const __m512i movemask =
+      _mm512_setr_epi32(28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11);
+  const __m512i rotated = _mm512_permutex2var_epi32(input, movemask, previous);
+#if SIMDUTF_GCC8 || SIMDUTF_GCC9
+  constexpr int shift = 16 - N; // workaround for GCC8,9
+  return _mm512_alignr_epi8(input, rotated, shift);
+#else
+  return _mm512_alignr_epi8(input, rotated, 16 - N);
+#endif // SIMDUTF_GCC8 || SIMDUTF_GCC9
+}
+
+template <unsigned idx0, unsigned idx1, unsigned idx2, unsigned idx3>
+__m512i shuffle_epi128(__m512i v) {
+  static_assert((idx0 >= 0 && idx0 <= 3), "idx0 must be in range 0..3");
+  static_assert((idx1 >= 0 && idx1 <= 3), "idx1 must be in range 0..3");
+  static_assert((idx2 >= 0 && idx2 <= 3), "idx2 must be in range 0..3");
+  static_assert((idx3 >= 0 && idx3 <= 3), "idx3 must be in range 0..3");
+
+  constexpr unsigned shuffle = idx0 | (idx1 << 2) | (idx2 << 4) | (idx3 << 6);
+  return _mm512_shuffle_i32x4(v, v, shuffle);
+}
+
+template <unsigned idx> constexpr __m512i broadcast_epi128(__m512i v) {
+  return shuffle_epi128<idx, idx, idx, idx>(v);
 }
+/* end file src/icelake/icelake_common.inl.cpp */
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/icelake/icelake_utf8_common.inl.cpp */
+// Common procedures for both validating and non-validating conversions from
+// UTF-8.
+enum block_processing_mode { SIMDUTF_FULL, SIMDUTF_TAIL };
 
-template <endianness big_endian>
-std::pair<result, char16_t *>
-arm_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
-                                       char16_t *utf16_out) {
-  uint16_t *utf16_output = reinterpret_cast<uint16_t *>(utf16_out);
-  const char32_t *start = buf;
-  const char32_t *end = buf + len;
+using utf8_to_utf16_result = std::pair<const char *, char16_t *>;
+using utf8_to_utf32_result = std::pair<const char *, uint32_t *>;
 
-  while (end - buf >= 4) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
+/*
+    process_block_utf8_to_utf16 converts up to 64 bytes from 'in' from UTF-8
+    to UTF-16. When tail = SIMDUTF_FULL, then the full input buffer (64 bytes)
+    might be used. When tail = SIMDUTF_TAIL, we take into account 'gap' which
+    indicates how many input bytes are relevant.
 
-    // Check if no bits set above 16th
-    if (vmaxvq_u32(in) <= 0xFFFF) {
-      uint16x4_t utf16_packed = vmovn_u32(in);
+    Returns true when the result is correct, otherwise it returns false.
 
-      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
-      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
-      const uint16x4_t forbidden_bytemask = vand_u16(
-          vcle_u16(utf16_packed, v_dfff), vcge_u16(utf16_packed, v_d800));
-      if (vmaxv_u16(forbidden_bytemask) != 0) {
-        return std::make_pair(result(error_code::SURROGATE, buf - start),
-                              reinterpret_cast<char16_t *>(utf16_output));
+    The provided in and out pointers are advanced according to how many input
+    bytes have been processed, upon success.
+*/
+template <block_processing_mode tail, endianness big_endian>
+simdutf_really_inline bool
+process_block_utf8_to_utf16(const char *&in, char16_t *&out, size_t gap) {
+  // constants
+  __m512i mask_identity = _mm512_set_epi8(
+      63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46,
+      45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28,
+      27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9,
+      8, 7, 6, 5, 4, 3, 2, 1, 0);
+  __m512i mask_c0c0c0c0 = _mm512_set1_epi32(0xc0c0c0c0);
+  __m512i mask_80808080 = _mm512_set1_epi32(0x80808080);
+  __m512i mask_f0f0f0f0 = _mm512_set1_epi32(0xf0f0f0f0);
+  __m512i mask_dfdfdfdf_tail = _mm512_set_epi64(
+      0xffffdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf,
+      0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf,
+      0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf);
+  __m512i mask_c2c2c2c2 = _mm512_set1_epi32(0xc2c2c2c2);
+  __m512i mask_ffffffff = _mm512_set1_epi32(0xffffffff);
+  __m512i mask_d7c0d7c0 = _mm512_set1_epi32(0xd7c0d7c0);
+  __m512i mask_dc00dc00 = _mm512_set1_epi32(0xdc00dc00);
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  // Note that 'tail' is a compile-time constant !
+  __mmask64 b =
+      (tail == SIMDUTF_FULL) ? 0xFFFFFFFFFFFFFFFF : (uint64_t(1) << gap) - 1;
+  __m512i input = (tail == SIMDUTF_FULL) ? _mm512_loadu_si512(in)
+                                         : _mm512_maskz_loadu_epi8(b, in);
+  __mmask64 m1 = (tail == SIMDUTF_FULL)
+                     ? _mm512_cmplt_epu8_mask(input, mask_80808080)
+                     : _mm512_mask_cmplt_epu8_mask(b, input, mask_80808080);
+  if (_ktestc_mask64_u8(m1,
+                        b)) { // NOT(m1) AND b -- if all zeroes, then all ASCII
+                              // alternatively, we could do 'if (m1 == b) { '
+    if (tail == SIMDUTF_FULL) {
+      in += 64; // consumed 64 bytes
+      // we convert a full 64-byte block, writing 128 bytes.
+      __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
+      if (big_endian) {
+        input1 = _mm512_shuffle_epi8(input1, byteflip);
       }
-
-      if (!match_system(big_endian)) {
-        utf16_packed =
-            vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(utf16_packed)));
+      _mm512_storeu_si512(out, input1);
+      out += 32;
+      __m512i input2 =
+          _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
+      if (big_endian) {
+        input2 = _mm512_shuffle_epi8(input2, byteflip);
       }
-      vst1_u16(utf16_output, utf16_packed);
-      utf16_output += 4;
-      buf += 4;
+      _mm512_storeu_si512(out, input2);
+      out += 32;
+      return true; // we are done
     } else {
-      size_t forward = 3;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k),
-                reinterpret_cast<char16_t *>(utf16_output));
-          }
-          *utf16_output++ = !match_system(big_endian)
-                                ? char16_t(word >> 8 | word << 8)
-                                : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k),
-                reinterpret_cast<char16_t *>(utf16_output));
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (!match_system(big_endian)) {
-            high_surrogate =
-                uint16_t(high_surrogate >> 8 | high_surrogate << 8);
-            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
+      in += gap;
+      if (gap <= 32) {
+        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
+        if (big_endian) {
+          input1 = _mm512_shuffle_epi8(input1, byteflip);
+        }
+        _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << (gap)) - 1),
+                                 input1);
+        out += gap;
+      } else {
+        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
+        if (big_endian) {
+          input1 = _mm512_shuffle_epi8(input1, byteflip);
+        }
+        _mm512_storeu_si512(out, input1);
+        out += 32;
+        __m512i input2 =
+            _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
+        if (big_endian) {
+          input2 = _mm512_shuffle_epi8(input2, byteflip);
         }
+        _mm512_mask_storeu_epi16(
+            out, __mmask32((uint32_t(1) << (gap - 32)) - 1), input2);
+        out += gap - 32;
       }
-      buf += k;
+      return true; // we are done
     }
   }
+  // classify characters further
+  __mmask64 m234 = _mm512_cmp_epu8_mask(
+      mask_c0c0c0c0, input,
+      _MM_CMPINT_LE); // 0xc0 <= input, 2, 3, or 4 leading byte
+  __mmask64 m34 =
+      _mm512_cmp_epu8_mask(mask_dfdfdfdf_tail, input,
+                           _MM_CMPINT_LT); // 0xdf < input,  3 or 4 leading byte
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        reinterpret_cast<char16_t *>(utf16_output));
-}
-/* end file src/arm64/arm_convert_utf32_to_utf16.cpp */
-/* begin file src/arm64/arm_convert_utf32_to_utf8.cpp */
-std::pair<const char32_t *, char *>
-arm_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
-  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
-  const char32_t *end = buf + len;
-
-  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
-
-  uint16x8_t forbidden_bytemask = vmovq_n_u16(0x0);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-
-  while (buf + 16 + safety_margin < end) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
-    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
-
-    // Check if no bits set above 16th
-    if (vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
-      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
-      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
-      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
-      if (vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
-        // 2. store (8 bytes)
-        vst1_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 8;
-        utf8_output += 8;
-        continue; // we are done for this round!
-      }
-
-      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
-        // 1. prepare 2-byte values
-        // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-        // expected output   : [110a|aaaa|10bb|bbbb] x 8
-        const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
-        const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
-
-        // t0 = [000a|aaaa|bbbb|bb00]
-        const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
-        // t1 = [000a|aaaa|0000|0000]
-        const uint16x8_t t1 = vandq_u16(t0, v_1f00);
-        // t2 = [0000|0000|00bb|bbbb]
-        const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
-        // t3 = [000a|aaaa|00bb|bbbb]
-        const uint16x8_t t3 = vorrq_u16(t1, t2);
-        // t4 = [110a|aaaa|10bb|bbbb]
-        const uint16x8_t t4 = vorrq_u16(t3, v_c080);
-        // 2. merge ASCII and 2-byte codewords
-        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
-        const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(
-            vbslq_u16(one_byte_bytemask, utf16_packed, t4));
-        // 3. prepare bitmask for 8-bit lookup
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t mask = simdutf_make_uint16x8_t(
-            0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
-#else
-        const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                 0x0002, 0x0008, 0x0020, 0x0080};
-#endif
-        uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
-        // 4. pack the bytes
-        const uint8_t *row =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
-        const uint8x16_t shuffle = vld1q_u8(row + 1);
-        const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+  __mmask64 milltwobytes = _mm512_mask_cmp_epu8_mask(
+      m234, input, mask_c2c2c2c2,
+      _MM_CMPINT_LT); // 0xc0 <= input < 0xc2 (illegal two byte sequence)
+                      // Overlong 2-byte sequence
+  if (_ktestz_mask64_u8(milltwobytes, milltwobytes) == 0) {
+    // Overlong 2-byte sequence
+    return false;
+  }
+  if (_ktestz_mask64_u8(m34, m34) == 0) {
+    // We have a 3-byte sequence and/or a 2-byte sequence, or possibly even a
+    // 4-byte sequence!
+    __mmask64 m4 = _mm512_cmp_epu8_mask(
+        input, mask_f0f0f0f0,
+        _MM_CMPINT_NLT); // 0xf0 <= zmm0 (4 byte start bytes)
 
-        // 5. store bytes
-        vst1q_u8(utf8_output, utf8_packed);
+    __mmask64 mask_not_ascii = (tail == SIMDUTF_FULL)
+                                   ? _knot_mask64(m1)
+                                   : _kand_mask64(_knot_mask64(m1), b);
 
-        // 6. adjust pointers
-        buf += 8;
-        utf8_output += row[0];
-        continue;
+    __mmask64 mp1 = _kshiftli_mask64(m234, 1);
+    __mmask64 mp2 = _kshiftli_mask64(m34, 2);
+    // We could do it as follows...
+    // if (_kortestz_mask64_u8(m4,m4)) { // compute the bitwise OR of the 64-bit
+    // masks a and b and return 1 if all zeroes but GCC generates better code
+    // when we do:
+    if (m4 == 0) { // compute the bitwise OR of the 64-bit masks a and b and
+                   // return 1 if all zeroes
+      // Fast path with 1,2,3 bytes
+      __mmask64 mc = _kor_mask64(mp1, mp2); // expected continuation bytes
+      __mmask64 m1234 = _kor_mask64(m1, m234);
+      // mismatched continuation bytes:
+      if (tail == SIMDUTF_FULL) {
+        __mmask64 xnormcm1234 = _kxnor_mask64(
+            mc,
+            m1234); // XNOR of mc and m1234 should be all zero if they differ
+        // the presence of a 1 bit indicates that they overlap.
+        // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return
+        // 1 if all zeroes.
+        if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) {
+          return false;
+        }
       } else {
-        // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
-        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
-        forbidden_bytemask =
-            vorrq_u16(vandq_u16(vcleq_u16(utf16_packed, v_dfff),
-                                vcgeq_u16(utf16_packed, v_d800)),
-                      forbidden_bytemask);
+        __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
+        if (mc != bxorm1234) {
+          return false;
+        }
+      }
+      // mend: identifying the last bytes of each sequence to be decoded
+      __mmask64 mend = _kshiftri_mask64(m1234, 1);
+      if (tail != SIMDUTF_FULL) {
+        mend = _kor_mask64(mend, (uint64_t(1) << (gap - 1)));
+      }
 
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t dup_even = simdutf_make_uint16x8_t(
-            0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
-#else
-        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
-                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
-#endif
-        /* In this branch we handle three cases:
-          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-          single UFT-8 byte
-          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              -
-          two UTF-8 bytes
-          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-          three UTF-8 bytes
+      __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
+      __m512i last_and_thirdu16 =
+          _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));
 
-          We expand the input word (16-bit) into two code units (32-bit), thus
-          we have room for four bytes. However, we need five distinct bit
-          layouts. Note that the last byte in cases #2 and #3 is the same.
+      __m512i nonasciitags = _mm512_maskz_mov_epi8(
+          mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
+      __m512i clearedbytes = _mm512_andnot_si512(
+          nonasciitags, input); // high two bits cleared where not ASCII
+      __m512i lastbytes = _mm512_maskz_permutexvar_epi8(
+          0x5555555555555555, last_and_thirdu16,
+          clearedbytes); // the last byte of each character
 
-          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-          in register t2.
+      __mmask64 mask_before_non_ascii = _kshiftri_mask64(
+          mask_not_ascii, 1); // bytes that precede non-ASCII bytes
+      __m512i indexofsecondlastbytes = _mm512_add_epi16(
+          mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
+      __m512i beforeasciibytes =
+          _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
+      __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(
+          0x5555555555555555, indexofsecondlastbytes,
+          beforeasciibytes); // the second last bytes (of two, three byte seq,
+                             // surrogates)
+      secondlastbytes =
+          _mm512_slli_epi16(secondlastbytes, 6); // shifted into position
 
-          We precompute byte 1 for case #3 and -- **conditionally** --
-          precompute either byte 1 for case #2 or byte 2 for case #3. Note that
-          they differ by exactly one bit.
+      __m512i indexofthirdlastbytes = _mm512_add_epi16(
+          mask_ffffffff,
+          indexofsecondlastbytes); // indices of the second last bytes
+      __m512i thirdlastbyte =
+          _mm512_maskz_mov_epi8(m34,
+                                clearedbytes); // only those that are the third
+                                               // last byte of a sequence
+      __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(
+          0x5555555555555555, indexofthirdlastbytes,
+          thirdlastbyte); // the third last bytes (of three byte sequences, hi
+                          // surrogate)
+      thirdlastbytes =
+          _mm512_slli_epi16(thirdlastbytes, 12); // shifted into position
+      __m512i Wout = _mm512_ternarylogic_epi32(lastbytes, secondlastbytes,
+                                               thirdlastbytes, 254);
+      // the elements of Wout excluding the last element if it happens to be a
+      // high surrogate:
 
-          Finally from these two code units we build proper UTF-8 sequence,
-          taking into account the case (i.e, the number of bytes to write).
-        */
-        /**
-         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-         * t2 => [0ccc|cccc] [10cc|cccc]
-         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-         */
-#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
-        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-        const uint16x8_t t0 =
-            vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed),
-                                            vreinterpretq_u8_u16(dup_even)));
-        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
-        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+      __mmask64 mprocessed =
+          (tail == SIMDUTF_FULL)
+              ? _pdep_u64(0xFFFFFFFF, mend)
+              : _pdep_u64(
+                    0xFFFFFFFF,
+                    _kand_mask64(
+                        mend, b)); // we adjust mend at the end of the output.
 
-        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
-        const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
-        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
-        const uint16x8_t s1 =
-            vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
-        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
-        // [00bb|bbbb|0000|aaaa]
-        const uint16x8_t s2 = vorrq_u16(s0, s1s);
-        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
-        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
-        const uint16x8_t one_or_two_bytes_bytemask =
-            vcleq_u16(utf16_packed, v_07ff);
-        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000),
-                                        one_or_two_bytes_bytemask);
-        const uint16x8_t s4 = veorq_u16(s3, m0);
-#undef simdutf_vec
+      // Encodings out of range...
+      {
+        // the location of 3-byte sequence start bytes in the input
+        __mmask64 m3 = m34 & (b ^ m4);
+        // code units in Wout corresponding to 3-byte sequences.
+        __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
+        __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
+        __mmask32 Msmall800 =
+            _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
+        __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
+        __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
+        __mmask32 M3s =
+            _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
+        if (_kor_mask32(Msmall800, M3s)) {
+          return false;
+        }
+      }
+      int64_t nout = _mm_popcnt_u64(mprocessed);
+      in += 64 - _lzcnt_u64(mprocessed);
+      if (big_endian) {
+        Wout = _mm512_shuffle_epi8(Wout, byteflip);
+      }
+      _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
+      out += nout;
+      return true; // ok
+    }
+    //
+    // We have a 4-byte sequence, this is the general case.
+    // Slow!
+    __mmask64 mp3 = _kshiftli_mask64(m4, 3);
+    __mmask64 mc =
+        _kor_mask64(_kor_mask64(mp1, mp2), mp3); // expected continuation bytes
+    __mmask64 m1234 = _kor_mask64(m1, m234);
 
-        // 4. expand code units 16-bit => 32-bit
-        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
-        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+    // mend: identifying the last bytes of each sequence to be decoded
+    __mmask64 mend =
+        _kor_mask64(_kshiftri_mask64(_kor_mask64(mp3, m1234), 1), mp3);
+    if (tail != SIMDUTF_FULL) {
+      mend = _kor_mask64(mend, __mmask64(uint64_t(1) << (gap - 1)));
+    }
+    __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
+    __m512i last_and_thirdu16 =
+        _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));
 
-        // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t onemask = simdutf_make_uint16x8_t(
-            0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
-        const uint16x8_t twomask = simdutf_make_uint16x8_t(
-            0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
-#else
-        const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                    0x0100, 0x0400, 0x1000, 0x4000};
-        const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
-                                    0x0200, 0x0800, 0x2000, 0x8000};
-#endif
-        const uint16x8_t combined =
-            vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
-                      vandq_u16(one_or_two_bytes_bytemask, twomask));
-        const uint16_t mask = vaddvq_u16(combined);
-        // The following fast path may or may not be beneficial.
-        /*if(mask == 0) {
-          // We only have three-byte code units. Use fast path.
-          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
-          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
-          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
-          vst1q_u8(utf8_output, utf8_0);
-          utf8_output += 12;
-          vst1q_u8(utf8_output, utf8_1);
-          utf8_output += 12;
-          buf += 8;
-          continue;
-        }*/
-        const uint8_t mask0 = uint8_t(mask);
-        const uint8_t *row0 =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
-        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+    __m512i nonasciitags = _mm512_maskz_mov_epi8(
+        mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
+    __m512i clearedbytes = _mm512_andnot_si512(
+        nonasciitags, input); // high two bits cleared where not ASCII
+    __m512i lastbytes = _mm512_maskz_permutexvar_epi8(
+        0x5555555555555555, last_and_thirdu16,
+        clearedbytes); // the last byte of each character
 
-        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-        const uint8_t *row1 =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
-        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+    __mmask64 mask_before_non_ascii = _kshiftri_mask64(
+        mask_not_ascii, 1); // bytes that precede non-ASCII bytes
+    __m512i indexofsecondlastbytes = _mm512_add_epi16(
+        mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
+    __m512i beforeasciibytes =
+        _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
+    __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(
+        0x5555555555555555, indexofsecondlastbytes,
+        beforeasciibytes); // the second last bytes (of two, three byte seq,
+                           // surrogates)
+    secondlastbytes =
+        _mm512_slli_epi16(secondlastbytes, 6); // shifted into position
 
-        vst1q_u8(utf8_output, utf8_0);
-        utf8_output += row0[0];
-        vst1q_u8(utf8_output, utf8_1);
-        utf8_output += row1[0];
+    __m512i indexofthirdlastbytes = _mm512_add_epi16(
+        mask_ffffffff,
+        indexofsecondlastbytes); // indices of the second last bytes
+    __m512i thirdlastbyte = _mm512_maskz_mov_epi8(
+        m34,
+        clearedbytes); // only those that are the third last byte of a sequence
+    __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(
+        0x5555555555555555, indexofthirdlastbytes,
+        thirdlastbyte); // the third last bytes (of three byte sequences, hi
+                        // surrogate)
+    thirdlastbytes =
+        _mm512_slli_epi16(thirdlastbytes, 12); // shifted into position
+    __m512i thirdsecondandlastbytes = _mm512_ternarylogic_epi32(
+        lastbytes, secondlastbytes, thirdlastbytes, 254);
+    uint64_t Mlo_uint64 = _pext_u64(mp3, mend);
+    __mmask32 Mlo = __mmask32(Mlo_uint64);
+    __mmask32 Mhi = __mmask32(Mlo_uint64 >> 1);
+    __m512i lo_surr_mask = _mm512_maskz_mov_epi16(
+        Mlo,
+        mask_dc00dc00); // lo surr: 1101110000000000, other:  0000000000000000
+    __m512i shifted4_thirdsecondandlastbytes =
+        _mm512_srli_epi16(thirdsecondandlastbytes,
+                          4); // hi surr: 00000WVUTSRQPNML  vuts = WVUTS - 1
+    __m512i tagged_lo_surrogates = _mm512_or_si512(
+        thirdsecondandlastbytes,
+        lo_surr_mask); // lo surr: 110111KJHGFEDCBA, other:  unchanged
+    __m512i Wout = _mm512_mask_add_epi16(
+        tagged_lo_surrogates, Mhi, shifted4_thirdsecondandlastbytes,
+        mask_d7c0d7c0); // hi sur: 110110vutsRQPNML, other:  unchanged
+    // the elements of Wout excluding the last element if it happens to be a
+    // high surrogate:
+    __mmask32 Mout = ~(Mhi & 0x80000000);
+    __mmask64 mprocessed =
+        (tail == SIMDUTF_FULL)
+            ? _pdep_u64(Mout, mend)
+            : _pdep_u64(
+                  Mout,
+                  _kand_mask64(mend,
+                               b)); // we adjust mend at the end of the output.
 
-        buf += 8;
+    // mismatched continuation bytes:
+    if (tail == SIMDUTF_FULL) {
+      __mmask64 xnormcm1234 = _kxnor_mask64(
+          mc, m1234); // XNOR of mc and m1234 should be all zero if they differ
+      // the presence of a 1 bit indicates that they overlap.
+      // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return 1
+      // if all zeroes.
+      if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) {
+        return false;
       }
-      // At least one 32-bit word will produce a surrogate pair in UTF-16 <=>
-      // will produce four UTF-8 bytes.
     } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
+      __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
+      if (mc != bxorm1234) {
+        return false;
       }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) {
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char *>(utf8_output));
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr,
-                                  reinterpret_cast<char *>(utf8_output));
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
+    }
+    // Encodings out of range...
+    {
+      // the location of 3-byte sequence start bytes in the input
+      __mmask64 m3 = m34 & (b ^ m4);
+      // code units in Wout corresponding to 3-byte sequences.
+      __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
+      __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
+      __mmask32 Msmall800 =
+          _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
+      __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
+      __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
+      __mmask32 M3s =
+          _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
+      __m512i mask_04000400 = _mm512_set1_epi32(0x04000400);
+      __mmask32 M4s =
+          _mm512_mask_cmpge_epu16_mask(Mhi, Moutminusd800, mask_04000400);
+      if (!_kortestz_mask32_u8(M4s, _kor_mask32(Msmall800, M3s))) {
+        return false;
       }
-      buf += k;
     }
-  } // while
-
-  // check for invalid input
-  if (vmaxvq_u16(forbidden_bytemask) != 0) {
-    return std::make_pair(nullptr, reinterpret_cast<char *>(utf8_output));
+    in += 64 - _lzcnt_u64(mprocessed);
+    int64_t nout = _mm_popcnt_u64(mprocessed);
+    if (big_endian) {
+      Wout = _mm512_shuffle_epi8(Wout, byteflip);
+    }
+    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
+    out += nout;
+    return true; // ok
   }
-  return std::make_pair(buf, reinterpret_cast<char *>(utf8_output));
-}
-
-std::pair<result, char *>
-arm_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
-                                      char *utf8_out) {
-  uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
-  const char32_t *start = buf;
-  const char32_t *end = buf + len;
-
-  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-
-  while (buf + 16 + safety_margin < end) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
-    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf + 4));
-
-    // Check if no bits set above 16th
-    if (vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
-      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
-      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
-      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
-      if (vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
-        // 1. pack the bytes
-        // obviously suboptimal.
-        uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
-        // 2. store (8 bytes)
-        vst1_u8(utf8_output, utf8_packed);
-        // 3. adjust pointers
-        buf += 8;
-        utf8_output += 8;
-        continue; // we are done for this round!
-      }
+  // Fast path 2: all ASCII or 2 byte
+  __mmask64 continuation_or_ascii = (tail == SIMDUTF_FULL)
+                                        ? _knot_mask64(m234)
+                                        : _kand_mask64(_knot_mask64(m234), b);
+  // on top of -0xc0 we subtract -2 which we get back later of the
+  // continuation byte tags
+  __m512i leading2byte = _mm512_maskz_sub_epi8(m234, input, mask_c2c2c2c2);
+  __mmask64 leading = tail == (tail == SIMDUTF_FULL)
+                          ? _kor_mask64(m1, m234)
+                          : _kand_mask64(_kor_mask64(m1, m234),
+                                         b); // first bytes of each sequence
+  if (tail == SIMDUTF_FULL) {
+    __mmask64 xnor234leading =
+        _kxnor_mask64(_kshiftli_mask64(m234, 1), leading);
+    if (!_kortestz_mask64_u8(xnor234leading, xnor234leading)) {
+      return false;
+    }
+  } else {
+    __mmask64 bxorleading = _kxor_mask64(b, leading);
+    if (_kshiftli_mask64(m234, 1) != bxorleading) {
+      return false;
+    }
+  }
+  //
+  if (tail == SIMDUTF_FULL) {
+    // In the two-byte/ASCII scenario, we are easily latency bound, so we want
+    // to increment the input buffer as quickly as possible.
+    // We process 32 bytes unless the byte at index 32 is a continuation byte,
+    // in which case we include it as well for a total of 33 bytes.
+    // Note that if x is an ASCII byte, then the following is false:
+    // int8_t(x) <= int8_t(0xc0) under two's complement.
+    in += 32;
+    if (int8_t(*in) <= int8_t(0xc0))
+      in++;
+    // The alternative is to do
+    // in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
+    // but it requires loading the input, doing the mask computation, and
+    // converting back the mask to a general register. It just takes too long,
+    // leaving the processor likely to be idle.
+  } else {
+    in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
+  }
+  __m512i lead = _mm512_maskz_compress_epi8(
+      leading, leading2byte); // will contain zero for ascii, and the data
+  lead = _mm512_cvtepu8_epi16(
+      _mm512_castsi512_si256(lead)); // ... zero extended into code units
+  __m512i follow = _mm512_maskz_compress_epi8(
+      continuation_or_ascii, input); // the last bytes of each sequence
+  follow = _mm512_cvtepu8_epi16(
+      _mm512_castsi512_si256(follow)); // ... zero extended into code units
+  lead = _mm512_slli_epi16(lead, 6);   // shifted into position
+  __m512i final = _mm512_add_epi16(follow, lead); // combining lead and follow
 
-      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
-        // 1. prepare 2-byte values
-        // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-        // expected output   : [110a|aaaa|10bb|bbbb] x 8
-        const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
-        const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);
+  if (big_endian) {
+    final = _mm512_shuffle_epi8(final, byteflip);
+  }
+  if (tail == SIMDUTF_FULL) {
+    // Next part is UTF-16 specific and can be generalized to UTF-32.
+    int nout = _mm_popcnt_u32(uint32_t(leading));
+    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
+    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
+  } else {
+    int nout = int(_mm_popcnt_u64(_pdep_u64(0xFFFFFFFF, leading)));
+    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
+    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
+  }
 
-        // t0 = [000a|aaaa|bbbb|bb00]
-        const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
-        // t1 = [000a|aaaa|0000|0000]
-        const uint16x8_t t1 = vandq_u16(t0, v_1f00);
-        // t2 = [0000|0000|00bb|bbbb]
-        const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
-        // t3 = [000a|aaaa|00bb|bbbb]
-        const uint16x8_t t3 = vorrq_u16(t1, t2);
-        // t4 = [110a|aaaa|10bb|bbbb]
-        const uint16x8_t t4 = vorrq_u16(t3, v_c080);
-        // 2. merge ASCII and 2-byte codewords
-        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
-        const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(
-            vbslq_u16(one_byte_bytemask, utf16_packed, t4));
-        // 3. prepare bitmask for 8-bit lookup
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t mask = simdutf_make_uint16x8_t(
-            0x0001, 0x0004, 0x0010, 0x0040, 0x0002, 0x0008, 0x0020, 0x0080);
-#else
-        const uint16x8_t mask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                 0x0002, 0x0008, 0x0020, 0x0080};
-#endif
-        uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
-        // 4. pack the bytes
-        const uint8_t *row =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
-        const uint8x16_t shuffle = vld1q_u8(row + 1);
-        const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);
+  return true; // we are fine.
+}
 
-        // 5. store bytes
-        vst1q_u8(utf8_output, utf8_packed);
+/*
+    utf32_to_utf16_masked converts `count` lower UTF-32 code units
+    from input `utf32` into UTF-16. It differs from utf32_to_utf16
+    in that it 'masks' the writes.
 
-        // 6. adjust pointers
-        buf += 8;
-        utf8_output += row[0];
-        continue;
-      } else {
-        // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+    Returns how many 16-bit code units were stored.
 
-        // check for invalid input
-        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
-        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
-        const uint16x8_t forbidden_bytemask = vandq_u16(
-            vcleq_u16(utf16_packed, v_dfff), vcgeq_u16(utf16_packed, v_d800));
-        if (vmaxvq_u16(forbidden_bytemask) != 0) {
-          return std::make_pair(result(error_code::SURROGATE, buf - start),
-                                reinterpret_cast<char *>(utf8_output));
-        }
+    byteflip is used for flipping 16-bit code units, and it should be
+        __m512i byteflip = _mm512_setr_epi64(
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809
+        );
+    We pass it to the (always inlined) function to encourage the compiler to
+    keep the value in a (constant) register.
+*/
+template <endianness big_endian>
+simdutf_really_inline size_t utf32_to_utf16_masked(const __m512i byteflip,
+                                                   __m512i utf32,
+                                                   unsigned int count,
+                                                   char16_t *output) {
 
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t dup_even = simdutf_make_uint16x8_t(
-            0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
-#else
-        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
-                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
-#endif
-        /* In this branch we handle three cases:
-          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-          single UFT-8 byte
-          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              -
-          two UTF-8 bytes
-          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-          three UTF-8 bytes
+  const __mmask16 valid = uint16_t((1 << count) - 1);
+  // 1. check if we have any surrogate pairs
+  const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
+  const __mmask16 sp_mask =
+      _mm512_mask_cmpgt_epu32_mask(valid, utf32, v_0000_ffff);
 
-          We expand the input word (16-bit) into two code units (32-bit), thus
-          we have room for four bytes. However, we need five distinct bit
-          layouts. Note that the last byte in cases #2 and #3 is the same.
+  if (sp_mask == 0) {
+    if (big_endian) {
+      _mm256_mask_storeu_epi16(
+          (__m256i *)output, valid,
+          _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32),
+                              _mm512_castsi512_si256(byteflip)));
 
-          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-          in register t2.
+    } else {
+      _mm256_mask_storeu_epi16((__m256i *)output, valid,
+                               _mm512_cvtepi32_epi16(utf32));
+    }
+    return count;
+  }
 
-          We precompute byte 1 for case #3 and -- **conditionally** --
-          precompute either byte 1 for case #2 or byte 2 for case #3. Note that
-          they differ by exactly one bit.
+  {
+    // build surrogate pair code units in 32-bit lanes
 
-          Finally from these two code units we build proper UTF-8 sequence,
-          taking into account the case (i.e, the number of bytes to write).
-        */
-        /**
-         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-         * t2 => [0ccc|cccc] [10cc|cccc]
-         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-         */
-#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
-        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-        const uint16x8_t t0 =
-            vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed),
-                                            vreinterpretq_u8_u16(dup_even)));
-        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
-        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        const uint16x8_t t2 = vorrq_u16(t1, simdutf_vec(0b1000000000000000));
+    //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
+    const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
+    const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);
 
-        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
-        const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
-        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
-        const uint16x8_t s1 =
-            vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
-        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
-        // [00bb|bbbb|0000|aaaa]
-        const uint16x8_t s2 = vorrq_u16(s0, s1s);
-        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
-        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
-        const uint16x8_t one_or_two_bytes_bytemask =
-            vcleq_u16(utf16_packed, v_07ff);
-        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000),
-                                        one_or_two_bytes_bytemask);
-        const uint16x8_t s4 = veorq_u16(s3, m0);
-#undef simdutf_vec
+    //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
+    const __m512i t1 = _mm512_slli_epi32(t0, 6);
 
-        // 4. expand code units 16-bit => 32-bit
-        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
-        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));
+    //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1
+    //    to t0
+    //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
+    const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
+    const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);
 
-        // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
-        const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
-#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
-        const uint16x8_t onemask = simdutf_make_uint16x8_t(
-            0x0001, 0x0004, 0x0010, 0x0040, 0x0100, 0x0400, 0x1000, 0x4000);
-        const uint16x8_t twomask = simdutf_make_uint16x8_t(
-            0x0002, 0x0008, 0x0020, 0x0080, 0x0200, 0x0800, 0x2000, 0x8000);
-#else
-        const uint16x8_t onemask = {0x0001, 0x0004, 0x0010, 0x0040,
-                                    0x0100, 0x0400, 0x1000, 0x4000};
-        const uint16x8_t twomask = {0x0002, 0x0008, 0x0020, 0x0080,
-                                    0x0200, 0x0800, 0x2000, 0x8000};
-#endif
-        const uint16x8_t combined =
-            vorrq_u16(vandq_u16(one_byte_bytemask, onemask),
-                      vandq_u16(one_or_two_bytes_bytemask, twomask));
-        const uint16_t mask = vaddvq_u16(combined);
-        // The following fast path may or may not be beneficial.
-        /*if(mask == 0) {
-          // We only have three-byte code units. Use fast path.
-          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
-          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
-          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
-          vst1q_u8(utf8_output, utf8_0);
-          utf8_output += 12;
-          vst1q_u8(utf8_output, utf8_1);
-          utf8_output += 12;
-          buf += 8;
-          continue;
-        }*/
-        const uint8_t mask0 = uint8_t(mask);
+    //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1
+    //    to t0
+    //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
+    const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
+    const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
+    const __m512i t3 =
+        _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
+    const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
+    __m512i t5 = _mm512_ror_epi32(t4, 16);
+    // Here we want to trim all of the upper 16-bit code units from the 2-byte
+    // characters represented as 4-byte values. We can compute it from
+    // sp_mask or the following... It can be more optimized!
+    const __mmask32 nonzero = _kor_mask32(
+        0xaaaaaaaa, _mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
+    const __mmask32 nonzero_masked =
+        _kand_mask32(nonzero, __mmask32((uint64_t(1) << (2 * count)) - 1));
+    if (big_endian) {
+      t5 = _mm512_shuffle_epi8(t5, byteflip);
+    }
+    // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+    // (AMD Zen4 has terrible performance with it, it is effectively broken)
+    __m512i compressed = _mm512_maskz_compress_epi16(nonzero_masked, t5);
+    _mm512_mask_storeu_epi16(
+        output, _bzhi_u32(0xFFFFFFFF, count + _mm_popcnt_u32(sp_mask)),
+        compressed);
+    //_mm512_mask_compressstoreu_epi16(output, nonzero_masked, t5);
+  }
 
-        const uint8_t *row0 =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
-        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);
+  return count + static_cast<unsigned int>(count_ones(sp_mask));
+}
 
-        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-        const uint8_t *row1 =
-            &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
-        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);
+/*
+    utf32_to_utf16 converts `count` lower UTF-32 code units
+    from input `utf32` into UTF-16. It may overflow.
 
-        vst1q_u8(utf8_output, utf8_0);
-        utf8_output += row0[0];
-        vst1q_u8(utf8_output, utf8_1);
-        utf8_output += row1[0];
+    Returns how many 16-bit code units were stored.
 
-        buf += 8;
-      }
-      // At least one 32-bit word will produce a surrogate pair in UTF-16 <=>
-      // will produce four UTF-8 bytes.
-    } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) {
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k),
-                reinterpret_cast<char *>(utf8_output));
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k),
-                reinterpret_cast<char *>(utf8_output));
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+    byteflip is used for flipping 16-bit code units, and it should be
+        __m512i byteflip = _mm512_setr_epi64(
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809,
+            0x0607040502030001,
+            0x0e0f0c0d0a0b0809
+        );
+    We pass it to the (always inlined) function to encourage the compiler to
+    keep the value in a (constant) register.
+*/
+template <endianness big_endian>
+simdutf_really_inline size_t utf32_to_utf16(const __m512i byteflip,
+                                            __m512i utf32, unsigned int count,
+                                            char16_t *output) {
+  // check if we have any surrogate pairs
+  const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
+  const __mmask16 sp_mask = _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        reinterpret_cast<char *>(utf8_output));
-}
-/* end file src/arm64/arm_convert_utf32_to_utf8.cpp */
+  if (sp_mask == 0) {
+    // technically, it should be _mm256_storeu_epi16
+    if (big_endian) {
+      _mm256_storeu_si256(
+          (__m256i *)output,
+          _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32),
+                              _mm512_castsi512_si256(byteflip)));
+    } else {
+      _mm256_storeu_si256((__m256i *)output, _mm512_cvtepi32_epi16(utf32));
+    }
+    return count;
+  }
 
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* begin file src/generic/buf_block_reader.h */
-namespace simdutf {
-namespace arm64 {
-namespace {
+  {
+    // build surrogate pair code units in 32-bit lanes
 
-// Walks through a buffer in block-sized increments, loading the last part with
-// spaces
-template <size_t STEP_SIZE> struct buf_block_reader {
-public:
-  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdutf_really_inline size_t block_index();
-  simdutf_really_inline bool has_full_block() const;
-  simdutf_really_inline const uint8_t *full_block() const;
-  /**
-   * Get the last block, padded with spaces.
-   *
-   * There will always be a last block, with at least 1 byte, unless len == 0
-   * (in which case this function fills the buffer with spaces and returns 0. In
-   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
-   * block with STEP_SIZE bytes and no spaces for padding.
-   *
-   * @return the number of effective characters in the last block.
-   */
-  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdutf_really_inline void advance();
+    //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
+    const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
+    const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);
 
-private:
-  const uint8_t *buf;
-  const size_t len;
-  const size_t lenminusstep;
-  size_t idx;
-};
+    //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
+    const __m512i t1 = _mm512_slli_epi32(t0, 6);
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text_64(const uint8_t *text) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
-  }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
+    //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1
+    //    to t0
+    //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
+    const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
+    const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  in.store(reinterpret_cast<uint8_t *>(buf));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    if (buf[i] < ' ') {
-      buf[i] = '_';
+    //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1
+    //    to t0
+    //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
+    const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
+    const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
+    const __m512i t3 =
+        _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
+    const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
+    __m512i t5 = _mm512_ror_epi32(t4, 16);
+    const __mmask32 nonzero = _kor_mask32(
+        0xaaaaaaaa, _mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
+    if (big_endian) {
+      t5 = _mm512_shuffle_epi8(t5, byteflip);
     }
+    // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+    // (zen4)
+    __m512i compressed = _mm512_maskz_compress_epi16(nonzero, t5);
+    _mm512_mask_storeu_epi16(
+        output,
+        (1 << (count + static_cast<unsigned int>(count_ones(sp_mask)))) - 1,
+        compressed);
+    //_mm512_mask_compressstoreu_epi16(output, nonzero, t5);
   }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
 
-simdutf_unused static char *format_mask(uint64_t mask) {
-  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
-  for (size_t i = 0; i < 64; i++) {
-    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
-  }
-  buf[64] = '\0';
-  return buf;
+  return count + static_cast<unsigned int>(count_ones(sp_mask));
 }
 
-template <size_t STEP_SIZE>
-simdutf_really_inline
-buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
-    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
-      idx{0} {}
+/*
+    expanded_utf8_to_utf32 converts expanded UTF-8 characters (`utf8`)
+    stored at separate 32-bit lanes.
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
-  return idx;
-}
+    For each lane we have also a character class (`char_class), given in form
+    0x8080800N, where N is 4 highest bits from the leading byte; 0x80 resets
+    corresponding bytes during pshufb.
+*/
+simdutf_really_inline __m512i expanded_utf8_to_utf32(__m512i char_class,
+                                                     __m512i utf8) {
+  /*
+      Input:
+      - utf8: bytes stored at separate 32-bit code units
+      - valid: which code units have valid UTF-8 characters
 
-template <size_t STEP_SIZE>
-simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
-  return idx < lenminusstep;
-}
+      Bit layout of single word. We show 4 cases for each possible
+      UTF-8 character encoding. The `?` denotes bits we must not
+      assume their value.
 
-template <size_t STEP_SIZE>
-simdutf_really_inline const uint8_t *
-buf_block_reader<STEP_SIZE>::full_block() const {
-  return &buf[idx];
-}
+      |10dd.dddd|10cc.cccc|10bb.bbbb|1111.0aaa| 4-byte char
+      |????.????|10cc.cccc|10bb.bbbb|1110.aaaa| 3-byte char
+      |????.????|????.????|10bb.bbbb|110a.aaaa| 2-byte char
+      |????.????|????.????|????.????|0aaa.aaaa| ASCII char
+        byte 3    byte 2    byte 1     byte 0
+  */
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t
-buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
-  if (len == idx) {
-    return 0;
-  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
-  std::memset(dst, 0x20,
-              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
-                          // to write out 8 or 16 bytes at once.
-  std::memcpy(dst, buf + idx, len - idx);
-  return len - idx;
-}
+  /* 1. Reset control bits of continuation bytes and the MSB
+        of the leading byte; this makes all bytes unsigned (and
+        does not alter ASCII char).
 
-template <size_t STEP_SIZE>
-simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
-  idx += STEP_SIZE;
-}
+      |00dd.dddd|00cc.cccc|00bb.bbbb|0111.0aaa| 4-byte char
+      |00??.????|00cc.cccc|00bb.bbbb|0110.aaaa| 3-byte char
+      |00??.????|00??.????|00bb.bbbb|010a.aaaa| 2-byte char
+      |00??.????|00??.????|00??.????|0aaa.aaaa| ASCII char
+       ^^        ^^        ^^        ^
+  */
+  __m512i values;
+  const __m512i v_3f3f_3f7f = _mm512_set1_epi32(0x3f3f3f7f);
+  values = _mm512_and_si512(utf8, v_3f3f_3f7f);
 
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/buf_block_reader.h */
-/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_validation {
+  /* 2. Swap and join fields A-B and C-D
 
-using namespace simd;
+      |0000.cccc|ccdd.dddd|0001.110a|aabb.bbbb| 4-byte char
+      |0000.cccc|cc??.????|0001.10aa|aabb.bbbb| 3-byte char
+      |0000.????|????.????|0001.0aaa|aabb.bbbb| 2-byte char
+      |0000.????|????.????|000a.aaaa|aa??.????| ASCII char */
+  const __m512i v_0140_0140 = _mm512_set1_epi32(0x01400140);
+  values = _mm512_maddubs_epi16(values, v_0140_0140);
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  /* 3. Swap and join fields AB & CD
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+      |0000.0001|110a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char
+      |0000.0001|10aa.aabb|bbbb.cccc|cc??.????| 3-byte char
+      |0000.0001|0aaa.aabb|bbbb.????|????.????| 2-byte char
+      |0000.000a|aaaa.aa??|????.????|????.????| ASCII char */
+  const __m512i v_0001_1000 = _mm512_set1_epi32(0x00011000);
+  values = _mm512_madd_epi16(values, v_0001_1000);
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+  /* 4. Shift left the values by variable amounts to reset highest UTF-8 bits
+      |aaab.bbbb|bccc.cccd|dddd.d000|0000.0000| 4-byte char -- by 11
+      |aaaa.bbbb|bbcc.cccc|????.??00|0000.0000| 3-byte char -- by 10
+      |aaaa.abbb|bbb?.????|????.???0|0000.0000| 2-byte char -- by 9
+      |aaaa.aaa?|????.????|????.????|?000.0000| ASCII char -- by 7 */
+  {
+    /** pshufb
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+    continuation = 0
+    ascii    = 7
+    _2_bytes = 9
+    _3_bytes = 10
+    _4_bytes = 11
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+    shift_left_v3 = 4 * [
+        ascii, # 0000
+        ascii, # 0001
+        ascii, # 0010
+        ascii, # 0011
+        ascii, # 0100
+        ascii, # 0101
+        ascii, # 0110
+        ascii, # 0111
+        continuation, # 1000
+        continuation, # 1001
+        continuation, # 1010
+        continuation, # 1011
+        _2_bytes, # 1100
+        _2_bytes, # 1101
+        _3_bytes, # 1110
+        _4_bytes, # 1111
+    ] */
+    const __m512i shift_left_v3 = _mm512_setr_epi64(
+        0x0707070707070707, 0x0b0a090900000000, 0x0707070707070707,
+        0x0b0a090900000000, 0x0707070707070707, 0x0b0a090900000000,
+        0x0707070707070707, 0x0b0a090900000000);
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
+    const __m512i shift = _mm512_shuffle_epi8(shift_left_v3, char_class);
+    values = _mm512_sllv_epi32(values, shift);
+  }
+
+  /* 5. Shift right the values by variable amounts to reset lowest bits
+      |0000.0000|000a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char -- by 11
+      |0000.0000|0000.0000|aaaa.bbbb|bbcc.cccc| 3-byte char -- by 16
+      |0000.0000|0000.0000|0000.0aaa|aabb.bbbb| 2-byte char -- by 21
+      |0000.0000|0000.0000|0000.0000|0aaa.aaaa| ASCII char -- by 25 */
+  {
+    // 4 * [25, 25, 25, 25, 25, 25, 25, 25, 0, 0, 0, 0, 21, 21, 16, 11]
+    const __m512i shift_right = _mm512_setr_epi64(
+        0x1919191919191919, 0x0b10151500000000, 0x1919191919191919,
+        0x0b10151500000000, 0x1919191919191919, 0x0b10151500000000,
+        0x1919191919191919, 0x0b10151500000000);
+
+    const __m512i shift = _mm512_shuffle_epi8(shift_right, char_class);
+    values = _mm512_srlv_epi32(values, shift);
+  }
+
+  return values;
 }
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
+
+simdutf_really_inline __m512i expand_and_identify(__m512i lane0, __m512i lane1,
+                                                  int &count) {
+  const __m512i merged = _mm512_mask_mov_epi32(lane0, 0x1000, lane1);
+  const __m512i expand_ver2 = _mm512_setr_epi64(
+      0x0403020103020100, 0x0605040305040302, 0x0807060507060504,
+      0x0a09080709080706, 0x0c0b0a090b0a0908, 0x0e0d0c0b0d0c0b0a,
+      0x000f0e0d0f0e0d0c, 0x0201000f01000f0e);
+  const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);
+  const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);
+  const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);
+  const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);
+  const __mmask16 leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);
+  count = static_cast<int>(count_ones(leading_bytes));
+  return _mm512_mask_compress_epi32(_mm512_setzero_si512(), leading_bytes,
+                                    input);
+}
+
+simdutf_really_inline __m512i expand_utf8_to_utf32(__m512i input) {
+  __m512i char_class = _mm512_srli_epi32(input, 4);
+  /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */
+  const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);
+  const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);
+  char_class =
+      _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea);
+  return expanded_utf8_to_utf32(char_class, input);
+}
+/* end file src/icelake/icelake_utf8_common.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/icelake/icelake_utf8_validation.inl.cpp */
+// file included directly
+
+simdutf_really_inline __m512i check_special_cases(__m512i input,
+                                                  const __m512i prev1) {
+  __m512i mask1 = _mm512_setr_epi64(0x0202020202020202, 0x4915012180808080,
+                                    0x0202020202020202, 0x4915012180808080,
+                                    0x0202020202020202, 0x4915012180808080,
+                                    0x0202020202020202, 0x4915012180808080);
+  const __m512i v_0f = _mm512_set1_epi8(0x0f);
+  __m512i index1 = _mm512_and_si512(_mm512_srli_epi16(prev1, 4), v_0f);
+
+  __m512i byte_1_high = _mm512_shuffle_epi8(mask1, index1);
+  __m512i mask2 = _mm512_setr_epi64(0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
+                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
+                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
+                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb);
+  __m512i index2 = _mm512_and_si512(prev1, v_0f);
+
+  __m512i byte_1_low = _mm512_shuffle_epi8(mask2, index2);
+  __m512i mask3 =
+      _mm512_setr_epi64(0x101010101010101, 0x1010101babaaee6, 0x101010101010101,
+                        0x1010101babaaee6, 0x101010101010101, 0x1010101babaaee6,
+                        0x101010101010101, 0x1010101babaaee6);
+  __m512i index3 = _mm512_and_si512(_mm512_srli_epi16(input, 4), v_0f);
+  __m512i byte_2_high = _mm512_shuffle_epi8(mask3, index3);
+  return _mm512_ternarylogic_epi64(byte_1_high, byte_1_low, byte_2_high, 128);
 }
 
+simdutf_really_inline __m512i check_multibyte_lengths(const __m512i input,
+                                                      const __m512i prev_input,
+                                                      const __m512i sc) {
+  __m512i prev2 = prev<2>(input, prev_input);
+  __m512i prev3 = prev<3>(input, prev_input);
+  __m512i is_third_byte = _mm512_subs_epu8(
+      prev2, _mm512_set1_epi8(0b11100000u - 1)); // Only 111_____ will be > 0
+  __m512i is_fourth_byte = _mm512_subs_epu8(
+      prev3, _mm512_set1_epi8(0b11110000u - 1)); // Only 1111____ will be > 0
+  __m512i is_third_or_fourth_byte =
+      _mm512_or_si512(is_third_byte, is_fourth_byte);
+  const __m512i v_7f = _mm512_set1_epi8(char(0x7f));
+  is_third_or_fourth_byte = _mm512_adds_epu8(v_7f, is_third_or_fourth_byte);
+  // We want to compute (is_third_or_fourth_byte AND v80) XOR sc.
+  const __m512i v_80 = _mm512_set1_epi8(char(0x80));
+  return _mm512_ternarylogic_epi32(is_third_or_fourth_byte, v_80, sc,
+                                   0b1101010);
+  //__m512i is_third_or_fourth_byte_mask =
+  //_mm512_and_si512(is_third_or_fourth_byte, v_80); return
+  // _mm512_xor_si512(is_third_or_fourth_byte_mask, sc);
+}
 //
 // Return nonzero if there are incomplete multibyte characters at the end of the
 // block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
 //
-simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+simdutf_really_inline __m512i is_incomplete(const __m512i input) {
   // If the previous input's last 3 bytes match this, they're too short (they
   // ended at EOF):
   // ... 1111____ 111_____ 11______
-  static const uint8_t max_array[32] = {255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        0b11110000u - 1,
-                                        0b11100000u - 1,
-                                        0b11000000u - 1};
-  const simd8<uint8_t> max_value(
-      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
-  return input.gt_bits(max_value);
+  __m512i max_value = _mm512_setr_epi64(0xffffffffffffffff, 0xffffffffffffffff,
+                                        0xffffffffffffffff, 0xffffffffffffffff,
+                                        0xffffffffffffffff, 0xffffffffffffffff,
+                                        0xffffffffffffffff, 0xbfdfefffffffffff);
+  return _mm512_subs_epu8(input, max_value);
 }
 
-struct utf8_checker {
+struct avx512_utf8_checker {
   // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+  __m512i error{};
+
   // The last input we received
-  simd8<uint8_t> prev_input_block;
+  __m512i prev_input_block{};
   // Whether the last input we received was incomplete (used for ASCII fast
   // path)
-  simd8<uint8_t> prev_incomplete;
+  __m512i prev_incomplete{};
 
   //
   // Check whether the current bytes are valid UTF-8.
   //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
+  simdutf_really_inline void check_utf8_bytes(const __m512i input,
+                                              const __m512i prev_input) {
     // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
     // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
     // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
+    __m512i prev1 = prev<1>(input, prev_input);
+    __m512i sc = check_special_cases(input, prev1);
+    this->error = _mm512_or_si512(
+        check_multibyte_lengths(input, prev_input, sc), this->error);
   }
 
   // The only problem that can happen at EOF is that a multibyte character is
@@ -19836,5854 +27580,6251 @@ struct utf8_checker {
   simdutf_really_inline void check_eof() {
     // If the previous block had incomplete UTF-8 characters at the end, an
     // ASCII block can't possibly finish them.
-    this->error |= this->prev_incomplete;
+    this->error = _mm512_or_si512(this->error, this->prev_incomplete);
   }
 
-  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
-    if (simdutf_likely(is_ascii(input))) {
-      this->error |= this->prev_incomplete;
+  // returns true if ASCII.
+  simdutf_really_inline bool check_next_input(const __m512i input) {
+    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
+    const __mmask64 ascii = _mm512_test_epi8_mask(input, v_80);
+    if (ascii == 0) {
+      this->error = _mm512_or_si512(this->error, this->prev_incomplete);
+      return true;
     } else {
-      // you might think that a for-loop would work, but under Visual Studio, it
-      // is not good enough.
-      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-                    "We support either two or four chunks per 64-byte block.");
-      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-      }
-      this->prev_incomplete =
-          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
-      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
+      this->check_utf8_bytes(input, this->prev_input_block);
+      this->prev_incomplete = is_incomplete(input);
+      this->prev_input_block = input;
+      return false;
     }
   }
-
   // do not forget to call check_eof!
   simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+    return _mm512_test_epi8_mask(this->error, this->error) != 0;
   }
+}; // struct avx512_utf8_checker
+/* end file src/icelake/icelake_utf8_validation.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-}; // struct utf8_checker
-} // namespace utf8_validation
-
-using utf8_validation::utf8_checker;
-
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
-/* begin file src/generic/utf8_validation/utf8_validator.h */
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_validation {
+#if SIMDUTF_FEATURE_UTF8 &&                                                    \
+    (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_LATIN1)
+/* begin file src/icelake/icelake_from_valid_utf8.inl.cpp */
+// file included directly
 
-/**
- * Validates that the string is actual UTF-8.
- */
-template <class checker>
-bool generic_validate_utf8(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    reader.advance();
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  return !c.errors();
-}
+// File contains conversion procedure from VALID UTF-8 strings.
 
-bool generic_validate_utf8(const char *input, size_t length) {
-  return generic_validate_utf8<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+/*
+    valid_utf8_to_fixed_length converts a valid UTF-8 string into UTF-32.
 
-/**
- * Validates that the string is actual UTF-8 and stops on errors.
- */
-template <class checker>
-result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  size_t count{0};
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    if (c.errors()) {
-      if (count != 0) {
-        count--;
-      } // Sometimes the error is only detected in the next chunk
-      result res = scalar::utf8::rewind_and_validate_with_errors(
-          reinterpret_cast<const char *>(input),
-          reinterpret_cast<const char *>(input + count), length - count);
-      res.count += count;
-      return res;
-    }
-    reader.advance();
-    count += 64;
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  if (c.errors()) {
-    if (count != 0) {
-      count--;
-    } // Sometimes the error is only detected in the next chunk
-    result res = scalar::utf8::rewind_and_validate_with_errors(
-        reinterpret_cast<const char *>(input),
-        reinterpret_cast<const char *>(input) + count, length - count);
-    res.count += count;
-    return res;
-  } else {
-    return result(error_code::SUCCESS, length);
-  }
-}
+    The `OUTPUT` template type decides what to do with UTF-32: store
+    it directly or convert into UTF-16 (with AVX512).
 
-result generic_validate_utf8_with_errors(const char *input, size_t length) {
-  return generic_validate_utf8_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+    Input:
+    - str           - valid UTF-8 string
+    - len           - string length
+    - out_buffer    - output buffer
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
-  uint8_t blocks[64]{};
-  simd::simd8x64<uint8_t> running_or(blocks);
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    running_or |= in;
-    reader.advance();
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  running_or |= in;
-  return running_or.is_ascii();
-}
+    Result:
+    - pair.first    - the first unprocessed input byte
+    - pair.second   - the first unprocessed output word
+*/
+template <endianness big_endian, typename OUTPUT>
+std::pair<const char *, OUTPUT *>
+valid_utf8_to_fixed_length(const char *str, size_t len, OUTPUT *dwords) {
+  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
+  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
+  static_assert(
+      UTF32 or UTF16,
+      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
+  static_assert(!(UTF32 and big_endian),
+                "we do not currently support big-endian UTF-32");
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  const char *ptr = str;
+  const char *end = ptr + len;
 
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
-  size_t count{0};
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    if (!in.is_ascii()) {
-      result res = scalar::ascii::validate_with_errors(
-          reinterpret_cast<const char *>(input + count), length - count);
-      return result(res.error, count + res.count);
+  OUTPUT *output = dwords;
+  /**
+   * In the main loop, we consume 64 bytes per iteration,
+   * but we access 64 + 4 bytes.
+   * We check for ptr + 64 + 64 <= end because
+   * we want to be do maskless writes without overruns.
+   */
+  while (end - ptr >= 64 + 4) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
+    const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
+    if (ascii == 0) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+      continue;
     }
-    reader.advance();
 
-    count += 64;
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  if (!in.is_ascii()) {
-    result res = scalar::ascii::validate_with_errors(
-        reinterpret_cast<const char *>(input + count), length - count);
-    return result(res.error, count + res.count);
-  } else {
-    return result(error_code::SUCCESS, length);
+    const __m512i lane0 = broadcast_epi128<0>(utf8);
+    const __m512i lane1 = broadcast_epi128<1>(utf8);
+    int valid_count0;
+    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+    const __m512i lane2 = broadcast_epi128<2>(utf8);
+    int valid_count1;
+    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+    if (valid_count0 + valid_count1 <= 16) {
+      vec0 = _mm512_mask_expand_epi32(
+          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+      valid_count0 += valid_count1;
+      vec0 = expand_utf8_to_utf32(vec0);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+    } else {
+      vec0 = expand_utf8_to_utf32(vec0);
+      vec1 = expand_utf8_to_utf32(vec1);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+    }
+    const __m512i lane3 = broadcast_epi128<3>(utf8);
+    int valid_count2;
+    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
+    uint32_t tmp1;
+    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
+    const __m512i lane4 = _mm512_set1_epi32(tmp1);
+    int valid_count3;
+    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
+    if (valid_count2 + valid_count3 <= 16) {
+      vec2 = _mm512_mask_expand_epi32(
+          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
+      valid_count2 += valid_count3;
+      vec2 = expand_utf8_to_utf32(vec2);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+    } else {
+      vec2 = expand_utf8_to_utf32(vec2);
+      vec3 = expand_utf8_to_utf32(vec3);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
+    }
+    ptr += 4 * 16;
   }
-}
-
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
 
-} // namespace utf8_validation
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
-// transcoding from UTF-8 to UTF-16
-/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+  if (end - ptr >= 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
+    const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
+    if (ascii == 0) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+    } else {
+      const __m512i lane0 = broadcast_epi128<0>(utf8);
+      const __m512i lane1 = broadcast_epi128<1>(utf8);
+      int valid_count0;
+      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+      const __m512i lane2 = broadcast_epi128<2>(utf8);
+      int valid_count1;
+      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+      if (valid_count0 + valid_count1 <= 16) {
+        vec0 = _mm512_mask_expand_epi32(
+            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+        valid_count0 += valid_count1;
+        vec0 = expand_utf8_to_utf32(vec0);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+      } else {
+        vec0 = expand_utf8_to_utf32(vec0);
+        vec1 = expand_utf8_to_utf32(vec1);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+      }
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_utf16 {
-using namespace simd;
+      const __m512i lane3 = broadcast_epi128<3>(utf8);
+      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+      ptr += 3 * 16;
+    }
+  }
+  return {ptr, output};
+}
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+using utf8_to_utf16_result = std::pair<const char *, char16_t *>;
+/* end file src/icelake/icelake_from_valid_utf8.inl.cpp */
+/* begin file src/icelake/icelake_from_utf8.inl.cpp */
+// file included directly
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+// File contains conversion procedure from possibly invalid UTF-8 strings.
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+template <endianness big_endian, typename OUTPUT>
+// todo: replace with the utf-8 to utf-16 routine adapted to utf-32. This code
+// is legacy.
+std::pair<const char *, OUTPUT *>
+validating_utf8_to_fixed_length(const char *str, size_t len, OUTPUT *dwords) {
+  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
+  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
+  static_assert(
+      UTF32 or UTF16,
+      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
+  static_assert(!(UTF32 and big_endian),
+                "we do not currently support big-endian UTF-32");
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+  const char *ptr = str;
+  const char *end = ptr + len;
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  OUTPUT *output = dwords;
+  avx512_utf8_checker checker{};
+  /**
+   * In the main loop, we consume 64 bytes per iteration,
+   * but we access 64 + 4 bytes.
+   * We use masked writes to avoid overruns, see
+   * https://github.com/simdutf/simdutf/issues/471
+   */
+  while (end - ptr >= 64 + 4) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    if (checker.check_next_input(utf8)) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+      continue;
+    }
+    const __m512i lane0 = broadcast_epi128<0>(utf8);
+    const __m512i lane1 = broadcast_epi128<1>(utf8);
+    int valid_count0;
+    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+    const __m512i lane2 = broadcast_epi128<2>(utf8);
+    int valid_count1;
+    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+    if (valid_count0 + valid_count1 <= 16) {
+      vec0 = _mm512_mask_expand_epi32(
+          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+      valid_count0 += valid_count1;
+      vec0 = expand_utf8_to_utf32(vec0);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+    } else {
+      vec0 = expand_utf8_to_utf32(vec0);
+      vec1 = expand_utf8_to_utf32(vec1);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+    }
+    const __m512i lane3 = broadcast_epi128<3>(utf8);
+    int valid_count2;
+    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
+    uint32_t tmp1;
+    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
+    const __m512i lane4 = _mm512_set1_epi32(tmp1);
+    int valid_count3;
+    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
+    if (valid_count2 + valid_count3 <= 16) {
+      vec2 = _mm512_mask_expand_epi32(
+          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
+      valid_count2 += valid_count3;
+      vec2 = expand_utf8_to_utf32(vec2);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+    } else {
+      vec2 = expand_utf8_to_utf32(vec2);
+      vec3 = expand_utf8_to_utf32(vec3);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
+    }
+    ptr += 4 * 16;
+  }
+  const char *validatedptr = ptr; // validated up to ptr
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
+  // For the final pass, we validate 64 bytes, but we only transcode
+  // 3*16 bytes, so we may end up double-validating 16 bytes.
+  if (end - ptr >= 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    if (checker.check_next_input(utf8)) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+    } else {
+      const __m512i lane0 = broadcast_epi128<0>(utf8);
+      const __m512i lane1 = broadcast_epi128<1>(utf8);
+      int valid_count0;
+      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+      const __m512i lane2 = broadcast_epi128<2>(utf8);
+      int valid_count1;
+      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+      if (valid_count0 + valid_count1 <= 16) {
+        vec0 = _mm512_mask_expand_epi32(
+            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+        valid_count0 += valid_count1;
+        vec0 = expand_utf8_to_utf32(vec0);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+      } else {
+        vec0 = expand_utf8_to_utf32(vec0);
+        vec1 = expand_utf8_to_utf32(vec1);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+      }
 
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+      const __m512i lane3 = broadcast_epi128<3>(utf8);
+      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
 
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
+      ptr += 3 * 16;
+    }
+    validatedptr += 4 * 16;
+  }
+  if (end != validatedptr) {
+    const __m512i utf8 =
+        _mm512_maskz_loadu_epi8(~UINT64_C(0) >> (64 - (end - validatedptr)),
+                                (const __m512i *)validatedptr);
+    checker.check_next_input(utf8);
+  }
+  checker.check_eof();
+  if (checker.errors()) {
+    return {ptr, nullptr}; // We found an error.
   }
+  return {ptr, output};
+}
 
-  template <endianness endian>
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
+// Like validating_utf8_to_fixed_length but returns as soon as an error is
+// identified todo: replace with the utf-8 to utf-16 routine adapted to utf-32.
+// This code is legacy.
+template <endianness big_endian, typename OUTPUT>
+std::tuple<const char *, OUTPUT *, bool>
+validating_utf8_to_fixed_length_with_constant_checks(const char *str,
+                                                     size_t len,
+                                                     OUTPUT *dwords) {
+  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
+  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
+  static_assert(
+      UTF32 or UTF16,
+      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
+  static_assert(!(UTF32 and big_endian),
+                "we do not currently support big-endian UTF-32");
+
+  const char *ptr = str;
+  const char *end = ptr + len;
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  OUTPUT *output = dwords;
+  avx512_utf8_checker checker{};
+  /**
+   * In the main loop, we consume 64 bytes per iteration,
+   * but we access 64 + 4 bytes.
+   */
+  while (end - ptr >= 4 + 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    bool ascii = checker.check_next_input(utf8);
+    if (checker.errors()) {
+      return {ptr, output, false}; // We found an error.
     }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (utf8_continuation_mask & 1) {
-          return 0; // error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
+    if (ascii) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+      continue;
     }
-    if (errors()) {
-      return 0;
+    const __m512i lane0 = broadcast_epi128<0>(utf8);
+    const __m512i lane1 = broadcast_epi128<1>(utf8);
+    int valid_count0;
+    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+    const __m512i lane2 = broadcast_epi128<2>(utf8);
+    int valid_count1;
+    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+    if (valid_count0 + valid_count1 <= 16) {
+      vec0 = _mm512_mask_expand_epi32(
+          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+      valid_count0 += valid_count1;
+      vec0 = expand_utf8_to_utf32(vec0);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+    } else {
+      vec0 = expand_utf8_to_utf32(vec0);
+      vec1 = expand_utf8_to_utf32(vec1);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
     }
-    if (pos < size) {
-      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
-          in + pos, size - pos, utf16_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      utf16_output += howmany;
+    const __m512i lane3 = broadcast_epi128<3>(utf8);
+    int valid_count2;
+    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
+    uint32_t tmp1;
+    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
+    const __m512i lane4 = _mm512_set1_epi32(tmp1);
+    int valid_count3;
+    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
+    if (valid_count2 + valid_count3 <= 16) {
+      vec2 = _mm512_mask_expand_epi32(
+          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
+      valid_count2 += valid_count3;
+      vec2 = expand_utf8_to_utf32(vec2);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+    } else {
+      vec2 = expand_utf8_to_utf32(vec2);
+      vec3 = expand_utf8_to_utf32(vec3);
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
     }
-    return utf16_output - start;
+    ptr += 4 * 16;
   }
+  const char *validatedptr = ptr; // validated up to ptr
 
-  template <endianness endian>
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
+  // For the final pass, we validate 64 bytes, but we only transcode
+  // 3*16 bytes, so we may end up double-validating 16 bytes.
+  if (end - ptr >= 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    bool ascii = checker.check_next_input(utf8);
+    if (checker.errors()) {
+      return {ptr, output, false}; // We found an error.
     }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
+    if (ascii) {
+      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
+      output += 64;
+      ptr += 64;
+    } else {
+      const __m512i lane0 = broadcast_epi128<0>(utf8);
+      const __m512i lane1 = broadcast_epi128<1>(utf8);
+      int valid_count0;
+      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
+      const __m512i lane2 = broadcast_epi128<2>(utf8);
+      int valid_count1;
+      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
+      if (valid_count0 + valid_count1 <= 16) {
+        vec0 = _mm512_mask_expand_epi32(
+            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
+        valid_count0 += valid_count1;
+        vec0 = expand_utf8_to_utf32(vec0);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
       } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          // rewind_and_convert_with_errors will seek a potential error from
-          // in+pos onward, with the ability to go back up to pos bytes, and
-          // read size-pos bytes forward.
-          result res =
-              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-                  pos, in + pos, size - pos, utf16_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
+        vec0 = expand_utf8_to_utf32(vec0);
+        vec1 = expand_utf8_to_utf32(vec1);
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
       }
+
+      const __m512i lane3 = broadcast_epi128<3>(utf8);
+      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
+
+      ptr += 3 * 16;
     }
-    if (errors()) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      res.count += pos;
-      return res;
-    }
-    if (pos < size) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        utf16_output += res.count;
-      }
+    validatedptr += 4 * 16;
+  }
+  if (end != validatedptr) {
+    const __m512i utf8 =
+        _mm512_maskz_loadu_epi8(~UINT64_C(0) >> (64 - (end - validatedptr)),
+                                (const __m512i *)validatedptr);
+    checker.check_next_input(utf8);
+  }
+  checker.check_eof();
+  if (checker.errors()) {
+    return {ptr, output, false}; // We found an error.
+  }
+  return {ptr, output, true};
+}
+/* end file src/icelake/icelake_from_utf8.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_LATIN1)
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/icelake/icelake_convert_utf8_to_latin1.inl.cpp */
+// file included directly
+
+// File contains conversion procedure from possibly invalid UTF-8 strings.
+
+template <bool is_remaining>
+simdutf_really_inline size_t process_block_from_utf8_to_latin1(
+    const char *buf, size_t len, char *latin_output, __m512i minus64,
+    __m512i one, __mmask64 *next_leading_ptr, __mmask64 *next_bit6_ptr) {
+  __mmask64 load_mask =
+      is_remaining ? _bzhi_u64(~0ULL, (unsigned int)len) : ~0ULL;
+  __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)buf);
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
+  if (nonascii == 0) {
+    if (*next_leading_ptr) { // If we ended with a leading byte, it is an error.
+      return 0;              // Indicates error
     }
-    return result(error_code::SUCCESS, utf16_output - start);
+    is_remaining
+        ? _mm512_mask_storeu_epi8((__m512i *)latin_output, load_mask, input)
+        : _mm512_storeu_si512((__m512i *)latin_output, input);
+    return len;
   }
 
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+  const __mmask64 leading = _mm512_cmpge_epu8_mask(input, minus64);
+
+  __m512i highbits = _mm512_xor_si512(input, _mm512_set1_epi8(-62));
+  __mmask64 invalid_leading_bytes =
+      _mm512_mask_cmpgt_epu8_mask(leading, highbits, one);
+
+  if (invalid_leading_bytes) {
+    return 0; // Indicates error
   }
 
-}; // struct utf8_checker
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+  __mmask64 leading_shift = (leading << 1) | *next_leading_ptr;
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_utf16 {
+  if ((nonascii ^ leading) != leading_shift) {
+    return 0; // Indicates error
+  }
 
-using namespace simd;
+  const __mmask64 bit6 = _mm512_cmpeq_epi8_mask(highbits, one);
+  input =
+      _mm512_mask_sub_epi8(input, (bit6 << 1) | *next_bit6_ptr, input, minus64);
 
-template <endianness endian>
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char16_t *utf16_output) noexcept {
-  // The implementation is not specific to haswell and should be moved to the
-  // generic directory.
+  __mmask64 retain = ~leading & load_mask;
+  __m512i output = _mm512_maskz_compress_epi8(retain, input);
+  int64_t written_out = count_ones(retain);
+  if (written_out == 0) {
+    return 0; // Indicates error
+  }
+  *next_bit6_ptr = bit6 >> 63;
+  *next_leading_ptr = leading >> 63;
+
+  __mmask64 store_mask = ~UINT64_C(0) >> (64 - written_out);
+
+  _mm512_mask_storeu_epi8((__m512i *)latin_output, store_mask, output);
+
+  return written_out;
+}
+
+size_t utf8_to_latin1_avx512(const char *&inbuf, size_t len,
+                             char *&inlatin_output) {
+  const char *buf = inbuf;
+  char *latin_output = inlatin_output;
+  char *start = latin_output;
   size_t pos = 0;
-  char16_t *start{utf16_output};
-  const size_t safety_margin = 16; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    // this loop could be unrolled further. For example, we could process the
-    // mask far more than 64 bytes.
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf16<endian>(utf16_output);
-      utf16_output += 64;
-      pos += 64;
-    } else {
-      // Slow path. We hope that the compiler will recognize that this is a slow
-      // path. Anything that is not a continuation mask is a 'leading byte',
-      // that is, the start of a new code point.
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      // The *start* of code points is not so useful, rather, we want the *end*
-      // of code points.
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      // We process in blocks of up to 12 bytes except possibly
-      // for fast paths which may process up to 16 bytes. For the
-      // slow path to work, we should have at least 12 input bytes left.
-      size_t max_starting_point = (pos + 64) - 12;
-      // Next loop is going to run at least five times when using solely
-      // the slow/regular path, and at least four times if there are fast paths.
-      while (pos < max_starting_point) {
-        // Performance note: our ability to compute 'consumed' and
-        // then shift and recompute is critical. If there is a
-        // latency of, say, 4 cycles on getting 'consumed', then
-        // the inner loop might have a total latency of about 6 cycles.
-        // Yet we process between 6 to 12 inputs bytes, thus we get
-        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-        // for this section of the code. Hence, there is a limit
-        // to how much we can further increase this latency before
-        // it seriously harms performance.
-        //
-        // Thus we may allow convert_masked_utf8_to_utf16 to process
-        // more bytes at a time under a fast-path mode where 16 bytes
-        // are consumed at once (e.g., when encountering ASCII).
-        size_t consumed = convert_masked_utf8_to_utf16<endian>(
-            input + pos, utf8_end_of_code_point_mask, utf16_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
-      }
-      // At this point there may remain between 0 and 12 bytes in the
-      // 64-byte block. These bytes will be processed again. So we have an
-      // 80% efficiency (in the worst case). In practice we expect an
-      // 85% to 90% efficiency.
+  __m512i minus64 = _mm512_set1_epi8(-64); // 11111111111 ... 1100 0000
+  __m512i one = _mm512_set1_epi8(1);
+  __mmask64 next_leading = 0;
+  __mmask64 next_bit6 = 0;
+
+  while (pos + 64 <= len) {
+    size_t written = process_block_from_utf8_to_latin1<false>(
+        buf + pos, 64, latin_output, minus64, one, &next_leading, &next_bit6);
+    if (written == 0) {
+      inlatin_output = latin_output;
+      inbuf = buf + pos - next_leading;
+      return 0; // Indicates error at pos or after, or just before pos (too
+                // short error)
     }
+    latin_output += written;
+    pos += 64;
   }
-  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
-      input + pos, size - pos, utf16_output);
-  return utf16_output - start;
+
+  if (pos < len) {
+    size_t remaining = len - pos;
+    size_t written = process_block_from_utf8_to_latin1<true>(
+        buf + pos, remaining, latin_output, minus64, one, &next_leading,
+        &next_bit6);
+    if (written == 0) {
+      inbuf = buf + pos - next_leading;
+      inlatin_output = latin_output;
+      return 0; // Indicates error at pos or after, or just before pos (too
+                // short error)
+    }
+    latin_output += written;
+  }
+  if (next_leading) {
+    inbuf = buf + len - next_leading;
+    inlatin_output = latin_output;
+    return 0; // Indicates error at end of buffer
+  }
+  inlatin_output = latin_output;
+  inbuf += len;
+  return size_t(latin_output - start);
 }
+/* end file src/icelake/icelake_convert_utf8_to_latin1.inl.cpp */
+/* begin file src/icelake/icelake_convert_valid_utf8_to_latin1.inl.cpp */
+// file included directly
 
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
-/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+// File contains conversion procedure from valid UTF-8 strings.
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_utf32 {
-using namespace simd;
+template <bool is_remaining>
+simdutf_really_inline size_t process_valid_block_from_utf8_to_latin1(
+    const char *buf, size_t len, char *latin_output, __m512i minus64,
+    __m512i one, __mmask64 *next_leading_ptr, __mmask64 *next_bit6_ptr) {
+  __mmask64 load_mask =
+      is_remaining ? _bzhi_u64(~0ULL, (unsigned int)len) : ~0ULL;
+  __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)buf);
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  if (nonascii == 0) {
+    is_remaining
+        ? _mm512_mask_storeu_epi8((__m512i *)latin_output, load_mask, input)
+        : _mm512_storeu_si512((__m512i *)latin_output, input);
+    return len;
+  }
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+  __mmask64 leading = _mm512_cmpge_epu8_mask(input, minus64);
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+  __m512i highbits = _mm512_xor_si512(input, _mm512_set1_epi8(-62));
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+  *next_leading_ptr = leading >> 63;
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+  __mmask64 bit6 = _mm512_cmpeq_epi8_mask(highbits, one);
+  input =
+      _mm512_mask_sub_epi8(input, (bit6 << 1) | *next_bit6_ptr, input, minus64);
+  *next_bit6_ptr = bit6 >> 63;
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
+  __mmask64 retain = ~leading & load_mask;
+  __m512i output = _mm512_maskz_compress_epi8(retain, input);
+  int64_t written_out = count_ones(retain);
+  if (written_out == 0) {
+    return 0; // Indicates error
+  }
+  __mmask64 store_mask = ~UINT64_C(0) >> (64 - written_out);
+  // Optimization opportunity: sometimes, masked writes are not needed.
+  _mm512_mask_storeu_epi8((__m512i *)latin_output, store_mask, output);
+  return written_out;
 }
 
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+size_t valid_utf8_to_latin1_avx512(const char *buf, size_t len,
+                                   char *latin_output) {
+  char *start = latin_output;
+  size_t pos = 0;
+  __m512i minus64 = _mm512_set1_epi8(-64); // 11111111111 ... 1100 0000
+  __m512i one = _mm512_set1_epi8(1);
+  __mmask64 next_leading = 0;
+  __mmask64 next_bit6 = 0;
 
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
+  while (pos + 64 <= len) {
+    size_t written = process_valid_block_from_utf8_to_latin1<false>(
+        buf + pos, 64, latin_output, minus64, one, &next_leading, &next_bit6);
+    latin_output += written;
+    pos += 64;
   }
 
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char32_t *utf32_output) {
-    size_t pos = 0;
-    char32_t *start{utf32_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 16 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (utf8_continuation_mask & 1) {
-          return 0; // we have an error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
+  if (pos < len) {
+    size_t remaining = len - pos;
+    size_t written = process_valid_block_from_utf8_to_latin1<true>(
+        buf + pos, remaining, latin_output, minus64, one, &next_leading,
+        &next_bit6);
+    latin_output += written;
+  }
+
+  return (size_t)(latin_output - start);
+}
+/* end file src/icelake/icelake_convert_valid_utf8_to_latin1.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/icelake/icelake_convert_utf16_to_latin1.inl.cpp */
+// file included directly
+template <endianness big_endian>
+size_t icelake_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                                       char *latin1_output) {
+  const char16_t *end = buf + len;
+  __m512i v_0xFF = _mm512_set1_epi16(0xff);
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
+      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
+  while (end - buf >= 32) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
     }
-    if (errors()) {
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
       return 0;
     }
-    if (pos < size) {
-      size_t howmany =
-          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      utf32_output += howmany;
+    _mm256_storeu_si256(
+        (__m256i *)latin1_output,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 32;
+    buf += 32;
+  }
+  if (buf < end) {
+    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
     }
-    return utf32_output - start;
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm256_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
   }
+  return len;
+}
 
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char32_t *utf32_output) {
-    size_t pos = 0;
-    char32_t *start{utf32_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
+template <endianness big_endian>
+std::pair<result, char *>
+icelake_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
+                                            char *latin1_output) {
+  const char16_t *end = buf + len;
+  const char16_t *start = buf;
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  __m512i v_0xFF = _mm512_set1_epi16(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
+      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
+  while (end - buf >= 32) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
     }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-              pos, in + pos, size - pos, utf32_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+      uint16_t word;
+      while ((word = (big_endian ? scalar::u16_swap_bytes(uint16_t(*buf))
+                                 : uint16_t(*buf))) <= 0xff) {
+        *latin1_output++ = uint8_t(word);
+        buf++;
       }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
     }
-    if (errors()) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
-      res.count += pos;
-      return res;
+    _mm256_storeu_si256(
+        (__m256i *)latin1_output,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 32;
+    buf += 32;
+  }
+  if (buf < end) {
+    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
     }
-    if (pos < size) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        utf32_output += res.count;
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+
+      uint16_t word;
+      while ((word = (big_endian ? scalar::u16_swap_bytes(uint16_t(*buf))
+                                 : uint16_t(*buf))) <= 0xff) {
+        *latin1_output++ = uint8_t(word);
+        buf++;
       }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
     }
-    return result(error_code::SUCCESS, utf32_output - start);
-  }
-
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+    _mm256_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
   }
+  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
+}
+/* end file src/icelake/icelake_convert_utf16_to_latin1.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF16
 
-}; // struct utf8_checker
-} // namespace utf8_to_utf32
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_utf32 {
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */
+// file included directly
 
-using namespace simd;
+/**
+ * This function converts the input (inbuf, inlen), assumed to be valid
+ * UTF16 (little endian) into UTF-8 (to outbuf). The number of code units
+ * written is written to 'outlen' and the function reports the number of input
+ * word consumed.
+ */
+template <endianness big_endian>
+size_t utf16_to_utf8_avx512i(const char16_t *inbuf, size_t inlen,
+                             unsigned char *outbuf, size_t *outlen) {
+  __m512i in;
+  __mmask32 inmask = _cvtu32_mask32(0x7fffffff);
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  const char16_t *const inbuf_orig = inbuf;
+  const unsigned char *const outbuf_orig = outbuf;
+  int adjust = 0;
+  int carry = 0;
 
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char32_t *utf32_output) noexcept {
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  const size_t safety_margin = 16; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf32(utf32_output);
-      utf32_output += 64;
-      pos += 64;
-    } else {
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      size_t max_starting_point = (pos + 64) - 12;
-      while (pos < max_starting_point) {
-        size_t consumed = convert_masked_utf8_to_utf32(
-            input + pos, utf8_end_of_code_point_mask, utf32_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
-      }
+  while (inlen >= 32) {
+    in = _mm512_loadu_si512(inbuf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
     }
-  }
-  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
-                                                       utf32_output);
-  return utf32_output - start;
-}
+    inlen -= 31;
+  lastiteration:
+    inbuf += 31;
 
-} // namespace utf8_to_utf32
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-// other functions
-/* begin file src/generic/utf16.h */
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf16 {
+  failiteration:
+    const __mmask32 is234byte = _mm512_mask_cmp_epu16_mask(
+        inmask, in, _mm512_set1_epi16(0x0080), _MM_CMPINT_NLT);
+
+    if (_ktestz_mask32_u8(inmask, is234byte)) {
+      // fast path for ASCII only
+      _mm512_mask_cvtepi16_storeu_epi8(outbuf, inmask, in);
+      outbuf += 31;
+      carry = 0;
 
-template <endianness big_endian>
-simdutf_really_inline size_t count_code_points(const char16_t *in,
-                                               size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
+      if (inlen < 32) {
+        goto tail;
+      } else {
+        continue;
+      }
     }
-    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
-    count += count_ones(not_pair) / 2;
-  }
-  return count +
-         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
-}
 
-template <endianness big_endian>
-simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
-    }
-    uint64_t ascii_mask = input.lteq(0x7F);
-    uint64_t twobyte_mask = input.lteq(0x7FF);
-    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
+    const __mmask32 is12byte =
+        _mm512_cmp_epu16_mask(in, _mm512_set1_epi16(0x0800), _MM_CMPINT_LT);
 
-    size_t ascii_count = count_ones(ascii_mask) / 2;
-    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
-    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
-    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
-    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
-             ascii_count;
-  }
-  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
-                                                                   size - pos);
-}
+    if (_ktestc_mask32_u8(is12byte, inmask)) {
+      // fast path for 1 and 2 byte only
 
-template <endianness big_endian>
-simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
-                                                     size_t size) {
-  return count_code_points<big_endian>(in, size);
-}
+      const __m512i twobytes = _mm512_ternarylogic_epi32(
+          _mm512_slli_epi16(in, 8), _mm512_srli_epi16(in, 6),
+          _mm512_set1_epi16(0x3f3f), 0xa8); // (A|B)&C
+      in = _mm512_mask_add_epi16(in, is234byte, twobytes,
+                                 _mm512_set1_epi16(int16_t(0x80c0)));
+      const __m512i cmpmask =
+          _mm512_mask_blend_epi16(inmask, _mm512_set1_epi16(int16_t(0xffff)),
+                                  _mm512_set1_epi16(0x0800));
+      const __mmask64 smoosh =
+          _mm512_cmp_epu8_mask(in, cmpmask, _MM_CMPINT_NLT);
+      const __m512i out = _mm512_maskz_compress_epi8(smoosh, in);
+      _mm512_mask_storeu_epi8(outbuf,
+                              _cvtu64_mask64(_pext_u64(_cvtmask64_u64(smoosh),
+                                                       _cvtmask64_u64(smoosh))),
+                              out);
+      outbuf += 31 + _mm_popcnt_u32(_cvtmask32_u32(is234byte));
+      carry = 0;
 
-simdutf_really_inline void
-change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
-  size_t pos = 0;
+      if (inlen < 32) {
+        goto tail;
+      } else {
+        continue;
+      }
+    }
+    __m512i lo = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
+    __m512i hi = _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1));
 
-  while (pos < size / 32 * 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    input.swap_bytes();
-    input.store(reinterpret_cast<uint16_t *>(output));
-    pos += 32;
-    output += 32;
-  }
+    __m512i taglo = _mm512_set1_epi32(0x8080e000);
+    __m512i taghi = taglo;
 
-  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
-}
+    const __m512i fc00masked =
+        _mm512_and_epi32(in, _mm512_set1_epi16(int16_t(0xfc00)));
+    const __mmask32 hisurr = _mm512_mask_cmp_epu16_mask(
+        inmask, fc00masked, _mm512_set1_epi16(int16_t(0xd800)), _MM_CMPINT_EQ);
+    const __mmask32 losurr = _mm512_cmp_epu16_mask(
+        fc00masked, _mm512_set1_epi16(int16_t(0xdc00)), _MM_CMPINT_EQ);
 
-} // namespace utf16
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf16.h */
-/* begin file src/generic/utf8.h */
+    int carryout = 0;
+    if (!_kortestz_mask32_u8(hisurr, losurr)) {
+      // handle surrogates
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8 {
+      __m512i los = _mm512_alignr_epi32(hi, lo, 1);
+      __m512i his = _mm512_alignr_epi32(lo, hi, 1);
 
-using namespace simd;
+      const __mmask32 hisurrhi = _kshiftri_mask32(hisurr, 16);
+      taglo = _mm512_mask_mov_epi32(taglo, __mmask16(hisurr),
+                                    _mm512_set1_epi32(0x808080f0));
+      taghi = _mm512_mask_mov_epi32(taghi, __mmask16(hisurrhi),
+                                    _mm512_set1_epi32(0x808080f0));
 
-simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.gt(-65);
-    count += count_ones(utf8_continuation_mask);
-  }
-  return count + scalar::utf8::count_code_points(in + pos, size - pos);
-}
+      lo = _mm512_mask_slli_epi32(lo, __mmask16(hisurr), lo, 10);
+      hi = _mm512_mask_slli_epi32(hi, __mmask16(hisurrhi), hi, 10);
+      los = _mm512_add_epi32(los, _mm512_set1_epi32(0xfca02400));
+      his = _mm512_add_epi32(his, _mm512_set1_epi32(0xfca02400));
+      lo = _mm512_mask_add_epi32(lo, __mmask16(hisurr), lo, los);
+      hi = _mm512_mask_add_epi32(hi, __mmask16(hisurrhi), hi, his);
 
-simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-    // We count one word for anything that is not a continuation (so
-    // leading bytes).
-    count += 64 - count_ones(utf8_continuation_mask);
-    int64_t utf8_4byte = input.gteq_unsigned(240);
-    count += count_ones(utf8_4byte);
-  }
-  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
-}
-} // namespace utf8
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8.h */
-// transcoding from UTF-8 to Latin 1
-/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+      carryout = _cvtu32_mask32(_kshiftri_mask32(hisurr, 30));
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_latin1 {
-using namespace simd;
+      const uint32_t h = _cvtmask32_u32(hisurr);
+      const uint32_t l = _cvtmask32_u32(losurr);
+      // check for mismatched surrogates
+      if ((h + h + carry) ^ l) {
+        const uint32_t lonohi = l & ~(h + h + carry);
+        const uint32_t hinolo = h & ~(l >> 1);
+        inlen = _tzcnt_u32(hinolo | lonohi);
+        inmask = __mmask32(0x7fffffff & ((1U << inlen) - 1));
+        in = _mm512_maskz_mov_epi16(inmask, in);
+        adjust = (int)inlen - 31;
+        inlen = 0;
+        goto failiteration;
+      }
+    }
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // For UTF-8 to Latin 1, we can allow any ASCII character, and any
-  // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or
-  // 0b11000010 and nothing else.
-  //
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
-  constexpr const uint8_t FORBIDDEN = 0xff;
+    hi = _mm512_maskz_mov_epi32(_cvtu32_mask16(0x7fff), hi);
+    carry = carryout;
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      FORBIDDEN,
-      // 1110____ ________ <three byte lead in byte 1>
-      FORBIDDEN,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      FORBIDDEN);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+    __m512i mslo =
+        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), lo);
 
-              // ____0100 ________
-              FORBIDDEN,
-              // ____0101 ________
-              FORBIDDEN,
-              // ____011_ ________
-              FORBIDDEN, FORBIDDEN,
+    __m512i mshi =
+        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), hi);
 
-              // ____1___ ________
-              FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN,
-              // ____1101 ________
-              FORBIDDEN, FORBIDDEN, FORBIDDEN);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+    const __mmask32 outmask = __mmask32(_kandn_mask64(losurr, inmask));
+    const __mmask64 outmhi = _kshiftri_mask64(outmask, 16);
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+    const __mmask32 is1byte = __mmask32(_knot_mask64(is234byte));
+    const __mmask64 is1bhi = _kshiftri_mask64(is1byte, 16);
+    const __mmask64 is12bhi = _kshiftri_mask64(is12byte, 16);
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
+    taglo = _mm512_mask_mov_epi32(taglo, __mmask16(is12byte),
+                                  _mm512_set1_epi32(0x80c00000));
+    taghi = _mm512_mask_mov_epi32(taghi, __mmask16(is12bhi),
+                                  _mm512_set1_epi32(0x80c00000));
+    __m512i magiclo = _mm512_mask_blend_epi32(__mmask16(outmask),
+                                              _mm512_set1_epi32(0xffffffff),
+                                              _mm512_set1_epi32(0x00010101));
+    __m512i magichi = _mm512_mask_blend_epi32(__mmask16(outmhi),
+                                              _mm512_set1_epi32(0xffffffff),
+                                              _mm512_set1_epi32(0x00010101));
 
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+    magiclo = _mm512_mask_blend_epi32(__mmask16(outmask),
+                                      _mm512_set1_epi32(0xffffffff),
+                                      _mm512_set1_epi32(0x00010101));
+    magichi = _mm512_mask_blend_epi32(__mmask16(outmhi),
+                                      _mm512_set1_epi32(0xffffffff),
+                                      _mm512_set1_epi32(0x00010101));
 
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    this->error |= check_special_cases(input, prev1);
-  }
+    mslo = _mm512_ternarylogic_epi32(mslo, _mm512_set1_epi32(0x3f3f3f3f), taglo,
+                                     0xea); // A&B|C
+    mshi = _mm512_ternarylogic_epi32(mshi, _mm512_set1_epi32(0x3f3f3f3f), taghi,
+                                     0xea);
+    mslo = _mm512_mask_slli_epi32(mslo, __mmask16(is1byte), lo, 24);
 
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char *latin1_output) {
-    size_t pos = 0;
-    char *start{latin1_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 16 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 16; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) >
-                       -65); // twos complement of -65 is 1011 1111 ...
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store((int8_t *)latin1_output);
-        latin1_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask =
-            input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
-                               // this case, we also have ASCII to account for.
-        if (utf8_continuation_mask & 1) {
-          return 0; // error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_latin1(
-              in + pos, utf8_end_of_code_point_mask, latin1_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      return 0;
-    }
-    if (pos < size) {
-      size_t howmany =
-          scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      latin1_output += howmany;
-    }
-    return latin1_output - start;
-  }
+    mshi = _mm512_mask_slli_epi32(mshi, __mmask16(is1bhi), hi, 24);
 
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char *latin1_output) {
-    size_t pos = 0;
-    char *start{latin1_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store((int8_t *)latin1_output);
-        latin1_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        if (errors()) {
-          // rewind_and_convert_with_errors will seek a potential error from
-          // in+pos onward, with the ability to go back up to pos bytes, and
-          // read size-pos bytes forward.
-          result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-              pos, in + pos, size - pos, latin1_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_latin1(
-              in + pos, utf8_end_of_code_point_mask, latin1_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, latin1_output);
-      res.count += pos;
-      return res;
-    }
-    if (pos < size) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, latin1_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        latin1_output += res.count;
-      }
-    }
-    return result(error_code::SUCCESS, latin1_output - start);
+    const __mmask64 wantlo =
+        _mm512_cmp_epu8_mask(mslo, magiclo, _MM_CMPINT_NLT);
+    const __mmask64 wanthi =
+        _mm512_cmp_epu8_mask(mshi, magichi, _MM_CMPINT_NLT);
+    const __m512i outlo = _mm512_maskz_compress_epi8(wantlo, mslo);
+    const __m512i outhi = _mm512_maskz_compress_epi8(wanthi, mshi);
+    const uint64_t wantlo_uint64 = _cvtmask64_u64(wantlo);
+    const uint64_t wanthi_uint64 = _cvtmask64_u64(wanthi);
+
+    uint64_t advlo = _mm_popcnt_u64(wantlo_uint64);
+    uint64_t advhi = _mm_popcnt_u64(wanthi_uint64);
+
+    _mm512_mask_storeu_epi8(
+        outbuf, _cvtu64_mask64(_pext_u64(wantlo_uint64, wantlo_uint64)), outlo);
+    _mm512_mask_storeu_epi8(
+        outbuf + advlo, _cvtu64_mask64(_pext_u64(wanthi_uint64, wanthi_uint64)),
+        outhi);
+    outbuf += advlo + advhi;
   }
+  outbuf += -adjust;
 
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+tail:
+  if (inlen != 0) {
+    // We must have inlen < 31.
+    inmask = _cvtu32_mask32((1U << inlen) - 1);
+    in = _mm512_maskz_loadu_epi16(inmask, inbuf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    adjust = (int)inlen - 31;
+    inlen = 0;
+    goto lastiteration;
   }
+  *outlen = (outbuf - outbuf_orig) + adjust;
+  return ((inbuf - inbuf_orig) + adjust);
+}
+/* end file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */
+/* begin file src/icelake/icelake_convert_utf8_to_utf16.inl.cpp */
+// file included directly
 
-}; // struct utf8_checker
-} // namespace utf8_to_latin1
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdutf
-/* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
-/* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+// File contains conversion procedure from possibly invalid UTF-8 strings.
 
-namespace simdutf {
-namespace arm64 {
-namespace {
-namespace utf8_to_latin1 {
-using namespace simd;
+/**
+ * Attempts to convert up to len 1-byte code units from in (in UTF-8 format) to
+ * out.
+ * Returns the position of the input and output after the processing is
+ * completed. Upon error, the output is set to null.
+ */
 
-simdutf_really_inline size_t convert_valid(const char *in, size_t size,
-                                           char *latin1_output) {
-  size_t pos = 0;
-  char *start{latin1_output};
-  // In the worst case, we have the haswell kernel which can cause an overflow
-  // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the last
-  // 16 bytes, and if the data is valid, then it is entirely safe because 16
-  // UTF-8 bytes generate much more than 8 bytes. However, you cannot generally
-  // assume that you have valid UTF-8 input, so we are going to go back from the
-  // end counting 8 leading bytes, to give us a good margin.
-  size_t leading_byte = 0;
-  size_t margin = size;
-  for (; margin > 0 && leading_byte < 8; margin--) {
-    leading_byte += (int8_t(in[margin - 1]) >
-                     -65); // twos complement of -65 is 1011 1111 ...
-  }
-  // If the input is long enough, then we have that margin-1 is the eight last
-  // leading byte.
-  const size_t safety_margin = size - margin + 1; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    if (input.is_ascii()) {
-      input.store((int8_t *)latin1_output);
-      latin1_output += 64;
-      pos += 64;
+template <endianness big_endian>
+utf8_to_utf16_result
+fast_avx512_convert_utf8_to_utf16(const char *in, size_t len, char16_t *out) {
+  const char *const final_in = in + len;
+  bool result = true;
+  while (result) {
+    if (final_in - in >= 64) {
+      result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(
+          in, out, final_in - in);
+    } else if (in < final_in) {
+      result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(
+          in, out, final_in - in);
     } else {
-      // you might think that a for-loop would work, but under Visual Studio, it
-      // is not good enough.
-      uint64_t utf8_continuation_mask =
-          input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
-                             // this case, we also have ASCII to account for.
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      // We process in blocks of up to 12 bytes except possibly
-      // for fast paths which may process up to 16 bytes. For the
-      // slow path to work, we should have at least 12 input bytes left.
-      size_t max_starting_point = (pos + 64) - 12;
-      // Next loop is going to run at least five times.
-      while (pos < max_starting_point) {
-        // Performance note: our ability to compute 'consumed' and
-        // then shift and recompute is critical. If there is a
-        // latency of, say, 4 cycles on getting 'consumed', then
-        // the inner loop might have a total latency of about 6 cycles.
-        // Yet we process between 6 to 12 inputs bytes, thus we get
-        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-        // for this section of the code. Hence, there is a limit
-        // to how much we can further increase this latency before
-        // it seriously harms performance.
-        size_t consumed = convert_masked_utf8_to_latin1(
-            in + pos, utf8_end_of_code_point_mask, latin1_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
-      }
-      // At this point there may remain between 0 and 12 bytes in the
-      // 64-byte block. These bytes will be processed again. So we have an
-      // 80% efficiency (in the worst case). In practice we expect an
-      // 85% to 90% efficiency.
+      break;
     }
   }
-  if (pos < size) {
-    size_t howmany = scalar::utf8_to_latin1::convert_valid(in + pos, size - pos,
-                                                           latin1_output);
-    latin1_output += howmany;
+  if (!result) {
+    out = nullptr;
   }
-  return latin1_output - start;
+  return std::make_pair(in, out);
 }
 
-} // namespace utf8_to_latin1
-} // namespace
-} // namespace arm64
-} // namespace simdutf
-  // namespace simdutf
-/* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-
-// placeholder scalars
-
-//
-// Implementation-specific overrides
-//
-namespace simdutf {
-namespace arm64 {
-
-simdutf_warn_unused int
-implementation::detect_encodings(const char *input,
-                                 size_t length) const noexcept {
-  // If there is a BOM, then we trust it.
-  auto bom_encoding = simdutf::BOM::check_bom(input, length);
-  if (bom_encoding != encoding_type::unspecified) {
-    return bom_encoding;
-  }
-  // todo: reimplement as a one-pass algorithm.
-  int out = 0;
-  if (validate_utf8(input, length)) {
-    out |= encoding_type::UTF8;
-  }
-  if ((length % 2) == 0) {
-    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
-                         length / 2)) {
-      out |= encoding_type::UTF16_LE;
+template <endianness big_endian>
+simdutf::result fast_avx512_convert_utf8_to_utf16_with_errors(const char *in,
+                                                              size_t len,
+                                                              char16_t *out) {
+  const char *const init_in = in;
+  const char16_t *const init_out = out;
+  const char *const final_in = in + len;
+  bool result = true;
+  while (result) {
+    if (final_in - in >= 64) {
+      result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(
+          in, out, final_in - in);
+    } else if (in < final_in) {
+      result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(
+          in, out, final_in - in);
+    } else {
+      break;
     }
   }
-  if ((length % 4) == 0) {
-    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
-      out |= encoding_type::UTF32_LE;
+  if (!result) {
+    size_t pos = size_t(in - init_in);
+    if (pos < len && (init_in[pos] & 0xc0) == 0x80 && pos >= 64) {
+      // We must check whether we are the fourth continuation byte
+      bool c1 = (init_in[pos - 1] & 0xc0) == 0x80;
+      bool c2 = (init_in[pos - 2] & 0xc0) == 0x80;
+      bool c3 = (init_in[pos - 3] & 0xc0) == 0x80;
+      if (c1 && c2 && c3) {
+        return {simdutf::TOO_LONG, pos};
+      }
     }
+    // rewind_and_convert_with_errors will seek a potential error from in
+    // onward, with the ability to go back up to in - init_in bytes, and read
+    // final_in - in bytes forward.
+    simdutf::result res =
+        scalar::utf8_to_utf16::rewind_and_convert_with_errors<big_endian>(
+            in - init_in, in, final_in - in, out);
+    res.count += (in - init_in);
+    return res;
+  } else {
+    return simdutf::result(error_code::SUCCESS, out - init_out);
   }
-  return out;
 }
+/* end file src/icelake/icelake_convert_utf8_to_utf16.inl.cpp */
+/* begin file src/icelake/icelake_utf8_length_from_utf16.inl.cpp */
+// This is translation of `utf8_length_from_utf16_bytemask` from
+// `generic/utf16.h`
+template <endianness big_endian>
+simdutf_really_inline size_t icelake_utf8_length_from_utf16(const char16_t *in,
+                                                            size_t size) {
+  size_t pos = 0;
 
-simdutf_warn_unused bool
-implementation::validate_utf8(const char *buf, size_t len) const noexcept {
-  return arm64::utf8_validation::generic_validate_utf8(buf, len);
-}
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
 
-simdutf_warn_unused result implementation::validate_utf8_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return arm64::utf8_validation::generic_validate_utf8_with_errors(buf, len);
-}
+  const auto one = vector_u16::splat(1);
 
-simdutf_warn_unused bool
-implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return arm64::utf8_validation::generic_validate_ascii(buf, len);
-}
+  auto v_count = vector_u16::zero();
 
-simdutf_warn_unused result implementation::validate_ascii_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return arm64::utf8_validation::generic_validate_ascii_with_errors(buf, len);
-}
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
 
-simdutf_warn_unused bool
-implementation::validate_utf16le(const char16_t *buf,
-                                 size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid. protected the implementation from nullptr.
-    return true;
-  }
-  const char16_t *tail = arm_validate_utf16<endianness::LITTLE>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::LITTLE>(tail,
-                                                       len - (tail - buf));
-  } else {
-    return false;
-  }
-}
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
 
-simdutf_warn_unused bool
-implementation::validate_utf16be(const char16_t *buf,
-                                 size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid. protected the implementation from nullptr.
-    return true;
-  }
-  const char16_t *tail = arm_validate_utf16<endianness::BIG>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
-  } else {
-    return false;
-  }
-}
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
 
-simdutf_warn_unused result implementation::validate_utf16le_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  result res = arm_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
-  if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
-        buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
-    return res;
-  }
-}
+    // not_surrogate[i] = non-zero if i-th element is not a surrogate word
+    const auto not_surrogate = (input & uint16_t(0xf800)) ^ uint16_t(0xd800);
 
-simdutf_warn_unused result implementation::validate_utf16be_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  result res = arm_validate_utf16_with_errors<endianness::BIG>(buf, len);
-  if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
-        buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
-    return res;
-  }
-}
+    // not_surrogate[i] = 1 if surrogate word, 0 otherwise
+    const auto is_surrogate = min(not_surrogate, one) ^ one;
 
-simdutf_warn_unused bool
-implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid. protected the implementation from nullptr.
-    return true;
-  }
-  const char32_t *tail = arm_validate_utf32le(buf, len);
-  if (tail) {
-    return scalar::utf32::validate(tail, len - (tail - buf));
-  } else {
-    return false;
-  }
-}
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
 
-simdutf_warn_unused result implementation::validate_utf32_with_errors(
-    const char32_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  result res = arm_validate_utf32le_with_errors(buf, len);
-  if (res.count != len) {
-    result scalar_res =
-        scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
-    return res;
-  }
-}
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
-    const char *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char *, char *> ret =
-      arm_convert_latin1_to_utf8(buf, len, utf8_output);
-  size_t converted_chars = ret.second - utf8_output;
+    /*
+        Explanation how the counting works.
 
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    converted_chars += scalar_converted_chars;
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count -= is_surrogate;
+
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+
+      iteration = max_iterations;
+    }
   }
-  return converted_chars;
-}
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char *, char16_t *> ret =
-      arm_convert_latin1_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
-  size_t converted_chars = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars =
-        scalar::latin1_to_utf16::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    converted_chars += scalar_converted_chars;
+  if (iteration > 0) {
+    count += v_count.sum();
   }
-  return converted_chars;
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
+/* end file src/icelake/icelake_utf8_length_from_utf16.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char *, char16_t *> ret =
-      arm_convert_latin1_to_utf16<endianness::BIG>(buf, len, utf16_output);
-  size_t converted_chars = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars =
-        scalar::latin1_to_utf16::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    converted_chars += scalar_converted_chars;
-  }
-  return converted_chars;
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
+// file included directly
+
+/*
+  Returns a pair: the first unprocessed byte from buf and utf32_output
+  A scalar routing should carry on the conversion of the tail.
+*/
+template <endianness big_endian>
+std::tuple<const char16_t *, char32_t *, bool>
+convert_utf16_to_utf32(const char16_t *buf, size_t len,
+                       char32_t *utf32_output) {
+  const char16_t *end = buf + len;
+  const __m512i v_fc00 = _mm512_set1_epi16((uint16_t)0xfc00);
+  const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);
+  const __m512i v_dc00 = _mm512_set1_epi16((uint16_t)0xdc00);
+  __mmask32 carry{0};
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  while (std::distance(buf, end) >= 32) {
+    // Always safe because buf + 32 <= end so that end - buf >= 32 bytes:
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+
+    // H - bitmask for high surrogates
+    const __mmask32 H =
+        _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_d800);
+    // H - bitmask for low surrogates
+    const __mmask32 L =
+        _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_dc00);
+
+    if ((H | L)) {
+      // surrogate pair(s) in a register
+      const __mmask32 V =
+          (L ^
+           (carry | (H << 1))); // A high surrogate must be followed by low one
+                                // and a low one must be preceded by a high one.
+                                // If valid, V should be equal to 0
+
+      if (V == 0) {
+        // valid case
+        /*
+            Input surrogate pair:
+            |1101.11aa.aaaa.aaaa|1101.10bb.bbbb.bbbb|
+                low surrogate      high surrogate
+        */
+        /*  1. Expand all code units to 32-bit code units
+            in
+           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
+        */
+        const __m512i first = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
+        const __m512i second =
+            _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1));
+
+        /*  2. Shift by one 16-bit word to align low surrogates with high
+           surrogates in
+           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
+            shifted
+           |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
+        */
+        const __m512i shifted_first = _mm512_alignr_epi32(second, first, 1);
+        const __m512i shifted_second =
+            _mm512_alignr_epi32(_mm512_setzero_si512(), second, 1);
+
+        /*  3. Align all high surrogates in first and second by shifting to the
+           left by 10 bits
+            |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
+        */
+        const __m512i aligned_first =
+            _mm512_mask_slli_epi32(first, (__mmask16)H, first, 10);
+        const __m512i aligned_second =
+            _mm512_mask_slli_epi32(second, (__mmask16)(H >> 16), second, 10);
+
+        /*  4. Remove surrogate prefixes and add offset 0x10000 by adding in,
+           shifted and constant in
+           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
+            shifted
+           |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
+            constant|1111.1100.1010.0000.0010.0100.0000.0000|1111.1100.1010.0000.0010.0100.0000.0000|
+        */
+        const __m512i constant = _mm512_set1_epi32((uint32_t)0xfca02400);
+        const __m512i added_first = _mm512_mask_add_epi32(
+            aligned_first, (__mmask16)H, aligned_first, shifted_first);
+        const __m512i utf32_first = _mm512_mask_add_epi32(
+            added_first, (__mmask16)H, added_first, constant);
+
+        const __m512i added_second =
+            _mm512_mask_add_epi32(aligned_second, (__mmask16)(H >> 16),
+                                  aligned_second, shifted_second);
+        const __m512i utf32_second = _mm512_mask_add_epi32(
+            added_second, (__mmask16)(H >> 16), added_second, constant);
+
+        //  5. Store all valid UTF-32 code units (low surrogate positions and
+        //  32nd word are invalid)
+        const __mmask32 valid = ~L & 0x7fffffff;
+        // We deliberately do a _mm512_maskz_compress_epi32 followed by
+        // storeu_epi32 to ease performance portability to Zen 4.
+        const __m512i compressed_first =
+            _mm512_maskz_compress_epi32((__mmask16)(valid), utf32_first);
+        const size_t howmany1 = count_ones((uint16_t)(valid));
+        _mm512_storeu_si512((__m512i *)utf32_output, compressed_first);
+        utf32_output += howmany1;
+        const __m512i compressed_second =
+            _mm512_maskz_compress_epi32((__mmask16)(valid >> 16), utf32_second);
+        const size_t howmany2 = count_ones((uint16_t)(valid >> 16));
+        // The following could be unsafe in some cases?
+        //_mm512_storeu_epi32((__m512i *) utf32_output, compressed_second);
+        _mm512_mask_storeu_epi32((__m512i *)utf32_output,
+                                 __mmask16((1 << howmany2) - 1),
+                                 compressed_second);
+        utf32_output += howmany2;
+        // Only process 31 code units, but keep track if the 31st word is a high
+        // surrogate as a carry
+        buf += 31;
+        carry = (H >> 30) & 0x1;
+      } else {
+        // invalid case
+        return std::make_tuple(buf + carry, utf32_output, false);
+      }
+    } else {
+      // no surrogates
+      // extend all thirty-two 16-bit code units to thirty-two 32-bit code units
+      _mm512_storeu_si512((__m512i *)(utf32_output),
+                          _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in)));
+      _mm512_storeu_si512(
+          (__m512i *)(utf32_output) + 1,
+          _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1)));
+      utf32_output += 32;
+      buf += 32;
+      carry = 0;
+    }
+  } // while
+  return std::make_tuple(buf + carry, utf32_output, true);
 }
+/* end file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char *, char32_t *> ret =
-      arm_convert_latin1_to_utf32(buf, len, utf32_output);
-  size_t converted_chars = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    converted_chars += scalar_converted_chars;
+#if SIMDUTF_FEATURE_UTF32
+/* begin file src/icelake/icelake_convert_utf32_to_latin1.inl.cpp */
+// file included directly
+size_t icelake_convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                                       char *latin1_output) {
+  const char32_t *end = buf + len;
+  __m512i v_0xFF = _mm512_set1_epi32(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
+      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
+  while (end - buf >= 16) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm_storeu_si128(
+        (__m128i *)latin1_output,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 16;
+    buf += 16;
+  }
+  if (buf < end) {
+    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
   }
-  return converted_chars;
-}
-
-simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  utf8_to_latin1::validating_transcoder converter;
-  return converter.convert(buf, len, latin1_output);
-}
-
-simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  utf8_to_latin1::validating_transcoder converter;
-  return converter.convert_with_errors(buf, len, latin1_output);
+  return len;
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return arm64::utf8_to_latin1::convert_valid(buf, len, latin1_output);
+std::pair<result, char *>
+icelake_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                            char *latin1_output) {
+  const char32_t *end = buf + len;
+  const char32_t *start = buf;
+  __m512i v_0xFF = _mm512_set1_epi32(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
+      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
+  while (end - buf >= 16) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      while (uint32_t(*buf) <= 0xff) {
+        *latin1_output++ = uint8_t(*buf++);
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm_storeu_si128(
+        (__m128i *)latin1_output,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 16;
+    buf += 16;
+  }
+  if (buf < end) {
+    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      while (uint32_t(*buf) <= 0xff) {
+        *latin1_output++ = uint8_t(*buf++);
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+  }
+  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
 }
+/* end file src/icelake/icelake_convert_utf32_to_latin1.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
+// file included directly
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert<endianness::BIG>(buf, len, utf16_output);
-}
+// Todo: currently, this is just the haswell code, optimize for icelake kernel.
+std::pair<const char32_t *, char *>
+avx512_convert_utf32_to_utf8(const char32_t *buf, size_t len,
+                             char *utf8_output) {
+  const char32_t *end = buf + len;
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
+  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
+  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
+  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
+  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
+  __m256i running_max = _mm256_setzero_si256();
+  __m256i forbidden_bytemask = _mm256_setzero_si256();
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert_with_errors<endianness::LITTLE>(buf, len,
-                                                           utf16_output);
-}
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
-}
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
+    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
-    const char *input, size_t size, char16_t *utf16_output) const noexcept {
-  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,
-                                                          utf16_output);
-}
+    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
+    // saturation
+    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
+                                        _mm256_and_si256(nextin, v_7fffffff));
+    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
-    const char *input, size_t size, char16_t *utf16_output) const noexcept {
-  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
-                                                       utf16_output);
-}
+    // Try to apply UTF-16 => UTF-8 routine on 256 bits
+    // (haswell/avx2_convert_utf16_to_utf8.cpp)
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  utf8_to_utf32::validating_transcoder converter;
-  return converter.convert(buf, len, utf32_output);
-}
+    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
+      // 1. pack the bytes
+      const __m128i utf8_packed = _mm_packus_epi16(
+          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
+      // 2. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
+      // 3. adjust pointers
+      buf += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
+    }
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  utf8_to_utf32::validating_transcoder converter;
-  return converter.convert_with_errors(buf, len, utf32_output);
-}
+    // no bits set above 11th bit
+    const __m256i one_or_two_bytes_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
+    const uint32_t one_or_two_bytes_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
+    if (one_or_two_bytes_bitmask == 0xffffffff) {
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
-    const char *input, size_t size, char32_t *utf32_output) const noexcept {
-  return utf8_to_utf32::convert_valid(input, size, utf32_output);
-}
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const __m256i t3 = _mm256_or_si256(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const __m256i t4 = _mm256_or_si256(t3, v_c080);
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      arm_convert_utf16_to_latin1<endianness::LITTLE>(buf, len, latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
+      // 2. merge ASCII and 2-byte codewords
+      const __m256i utf8_unpacked =
+          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
 
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_latin1::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      // 3. prepare bitmask for 8-bit lookup
+      const uint32_t M0 = one_byte_bitmask & 0x55555555;
+      const uint32_t M1 = M0 >> 7;
+      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+      // 4. pack the bytes
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      arm_convert_utf16_to_latin1<endianness::BIG>(buf, len, latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+      const uint8_t *row_2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
+                                                                       16)][0];
 
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_latin1::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
 
-simdutf_warn_unused result
-implementation::convert_utf16le_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<result, char *> ret =
-      arm_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
-          buf, len, latin1_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+      const __m256i utf8_packed = _mm256_shuffle_epi8(
+          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+      // 5. store bytes
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_castsi256_si128(utf8_packed));
+      utf8_output += row[0];
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_extractf128_si256(utf8_packed, 1));
+      utf8_output += row_2[0];
 
-simdutf_warn_unused result
-implementation::convert_utf16be_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<result, char *> ret =
-      arm_convert_utf16_to_latin1_with_errors<endianness::BIG>(buf, len,
-                                                               latin1_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+      // 6. adjust pointers
+      buf += 16;
+      continue;
     }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement a custom function.
-  return convert_utf16be_to_latin1(buf, len, latin1_output);
-}
+    // Must check for overflow in packing
+    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
+        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
+    const uint32_t saturation_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+    if (saturation_bitmask == 0xffffffff) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
+      forbidden_bytemask = _mm256_or_si256(
+          forbidden_bytemask,
+          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement a custom function.
-  return convert_utf16le_to_latin1(buf, len, latin1_output);
-}
+      const __m256i dup_even = _mm256_setr_epi16(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      arm_convert_utf16_to_utf8<endianness::LITTLE>(buf, len, utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf8::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      /* In this branch we handle three cases:
+        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      arm_convert_utf16_to_utf8<endianness::BIG>(buf, len, utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf8::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      arm_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(buf, len,
-                                                                utf8_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      arm_convert_utf16_to_utf8_with_errors<endianness::BIG>(buf, len,
-                                                             utf8_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16le_to_utf8(buf, len, utf8_output);
-}
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16be_to_utf8(buf, len, utf8_output);
-}
+      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
+      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
+      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
+      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
+      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
+      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
+      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
+      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
+                                             simdutf_vec(0b0100000000000000));
+      const __m256i s4 = _mm256_xor_si256(s3, m0);
+#undef simdutf_vec
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return 0;
-  }
-  std::pair<const char32_t *, char *> ret =
-      arm_convert_utf32_to_utf8(buf, len, utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      arm_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
-  if (ret.first.count != len) {
-    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
-        buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char16_t *, char32_t *> ret =
-      arm_convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char16_t *, char32_t *> ret =
-      arm_convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char32_t *> ret =
-      arm_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(buf, len,
-                                                                 utf32_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf32_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char32_t *> ret =
-      arm_convert_utf16_to_utf32_with_errors<endianness::BIG>(buf, len,
-                                                              utf32_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
     } else {
-      ret.second += scalar_res.count;
+      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
+      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
+      // wasteful to use scalar code, but being efficient with SIMD may require
+      // large, non-trivial tables?
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(nullptr, utf8_output);
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else { // 4-byte
+          if (word > 0x10FFFF) {
+            return std::make_pair(nullptr, utf8_output);
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
     }
-  }
-  ret.first.count =
-      ret.second -
-      utf32_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+  } // while
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char32_t *, char *> ret =
-      arm_convert_utf32_to_latin1(buf, len, latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
+  // check for invalid input
+  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
+  if (static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(
+          _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
+    return std::make_pair(nullptr, utf8_output);
   }
-  size_t saved_bytes = ret.second - latin1_output;
 
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
+    return std::make_pair(nullptr, utf8_output);
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<result, char *> ret =
-      arm_convert_utf32_to_latin1_with_errors(buf, len, latin1_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res = scalar::utf32_to_latin1::convert_with_errors(
-        buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
+  return std::make_pair(buf, utf8_output);
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char32_t *, char *> ret =
-      arm_convert_utf32_to_latin1(buf, len, latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
+// Todo: currently, this is just the haswell code, optimize for icelake kernel.
+std::pair<result, char *>
+avx512_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
+                                         char *utf8_output) {
+  const char32_t *end = buf + len;
+  const char32_t *start = buf;
 
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert_valid(
-        ret.first, len - (ret.first - buf), ret.second);
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
+  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
+  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
+  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
+  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
+  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  // optimization opportunity: implement a custom function.
-  return convert_utf32_to_utf8(buf, len, utf8_output);
-}
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      arm_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
+    // Check for too large input
+    const __m256i max_input =
+        _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
+    if (static_cast<uint32_t>(_mm256_movemask_epi8(
+            _mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            utf8_output);
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      arm_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
+    // saturation
+    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
+                                        _mm256_and_si256(nextin, v_7fffffff));
+    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      arm_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len,
-                                                                 utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+    // Try to apply UTF-16 => UTF-8 routine on 256 bits
+    // (haswell/avx2_convert_utf16_to_utf8.cpp)
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      arm_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len,
-                                                              utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
+      // 1. pack the bytes
+      const __m128i utf8_packed = _mm_packus_epi16(
+          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
+      // 2. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
+      // 3. adjust pointers
+      buf += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
     }
-  }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16le(buf, len, utf16_output);
-}
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16be(buf, len, utf16_output);
-}
+    // no bits set above 11th bit
+    const __m256i one_or_two_bytes_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
+    const uint32_t one_or_two_bytes_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
+    if (one_or_two_bytes_bitmask == 0xffffffff) {
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return convert_utf16le_to_utf32(buf, len, utf32_output);
-}
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const __m256i t3 = _mm256_or_si256(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const __m256i t4 = _mm256_or_si256(t3, v_c080);
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return convert_utf16be_to_utf32(buf, len, utf32_output);
-}
+      // 2. merge ASCII and 2-byte codewords
+      const __m256i utf8_unpacked =
+          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
 
-void implementation::change_endianness_utf16(const char16_t *input,
-                                             size_t length,
-                                             char16_t *output) const noexcept {
-  utf16::change_endianness_utf16(input, length, output);
-}
+      // 3. prepare bitmask for 8-bit lookup
+      const uint32_t M0 = one_byte_bitmask & 0x55555555;
+      const uint32_t M1 = M0 >> 7;
+      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+      // 4. pack the bytes
 
-simdutf_warn_unused size_t implementation::count_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::count_code_points<endianness::LITTLE>(input, length);
-}
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+      const uint8_t *row_2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
+                                                                       16)][0];
 
-simdutf_warn_unused size_t implementation::count_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::count_code_points<endianness::BIG>(input, length);
-}
+      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
 
-simdutf_warn_unused size_t
-implementation::count_utf8(const char *input, size_t length) const noexcept {
-  return utf8::count_code_points(input, length);
-}
+      const __m256i utf8_packed = _mm256_shuffle_epi8(
+          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+      // 5. store bytes
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_castsi256_si128(utf8_packed));
+      utf8_output += row[0];
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_extractf128_si256(utf8_packed, 1));
+      utf8_output += row_2[0];
 
-simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
-    const char *buf, size_t len) const noexcept {
-  return count_utf8(buf, len);
-}
+      // 6. adjust pointers
+      buf += 16;
+      continue;
+    }
+    // Must check for overflow in packing
+    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
+        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
+    const uint32_t saturation_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+    if (saturation_bitmask == 0xffffffff) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return scalar::utf16::latin1_length_from_utf16(length);
-}
+      // Check for illegal surrogate code units
+      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
+      const __m256i forbidden_bytemask =
+          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
+      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
+          0x0) {
+        return std::make_pair(result(error_code::SURROGATE, buf - start),
+                              utf8_output);
+      }
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return scalar::utf32::latin1_length_from_utf32(length);
-}
+      const __m256i dup_even = _mm256_setr_epi16(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
 
-simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
-    const char *input, size_t length) const noexcept {
-  // See
-  // https://lemire.me/blog/2023/05/15/computing-the-utf-8-size-of-a-latin-1-string-quickly-arm-neon-edition/
-  // credit to Pete Cawley
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
-  uint64_t result = 0;
-  const int lanes = sizeof(uint8x16_t);
-  uint8_t rem = length % lanes;
-  const uint8_t *simd_end = data + (length / lanes) * lanes;
-  const uint8x16_t threshold = vdupq_n_u8(0x80);
-  for (; data < simd_end; data += lanes) {
-    // load 16 bytes
-    uint8x16_t input_vec = vld1q_u8(data);
-    // compare to threshold (0x80)
-    uint8x16_t withhighbit = vcgeq_u8(input_vec, threshold);
-    // vertical addition
-    result -= vaddvq_s8(vreinterpretq_s8_u8(withhighbit));
-  }
-  return result + (length / lanes) * lanes +
-         scalar::latin1::utf8_length_from_latin1((const char *)simd_end, rem);
-}
+      /* In this branch we handle three cases:
+        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
-}
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
-}
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf16_length_from_latin1(length);
-}
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
 
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf32_length_from_latin1(length);
-}
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
-}
+      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
+      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
+      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
+      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
+      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
+      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
+      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
+      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
+                                             simdutf_vec(0b0100000000000000));
+      const __m256i s4 = _mm256_xor_si256(s3, m0);
+#undef simdutf_vec
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
-}
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return utf8::utf16_length_from_utf8(input, length);
-}
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const uint32x4_t v_7f = vmovq_n_u32((uint32_t)0x7f);
-  const uint32x4_t v_7ff = vmovq_n_u32((uint32_t)0x7ff);
-  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
-  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 4 <= length; pos += 4) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
-    const uint32x4_t ascii_bytes_bytemask = vcleq_u32(in, v_7f);
-    const uint32x4_t one_two_bytes_bytemask = vcleq_u32(in, v_7ff);
-    const uint32x4_t two_bytes_bytemask =
-        veorq_u32(one_two_bytes_bytemask, ascii_bytes_bytemask);
-    const uint32x4_t three_bytes_bytemask =
-        veorq_u32(vcleq_u32(in, v_ffff), one_two_bytes_bytemask);
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-    const uint16x8_t reduced_ascii_bytes_bytemask =
-        vreinterpretq_u16_u32(vandq_u32(ascii_bytes_bytemask, v_1));
-    const uint16x8_t reduced_two_bytes_bytemask =
-        vreinterpretq_u16_u32(vandq_u32(two_bytes_bytemask, v_1));
-    const uint16x8_t reduced_three_bytes_bytemask =
-        vreinterpretq_u16_u32(vandq_u32(three_bytes_bytemask, v_1));
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
 
-    const uint16x8_t compressed_bytemask0 =
-        vpaddq_u16(reduced_ascii_bytes_bytemask, reduced_two_bytes_bytemask);
-    const uint16x8_t compressed_bytemask1 =
-        vpaddq_u16(reduced_three_bytes_bytemask, reduced_three_bytes_bytemask);
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
 
-    size_t ascii_count = count_ones(
-        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 0));
-    size_t two_bytes_count = count_ones(
-        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 1));
-    size_t three_bytes_count = count_ones(
-        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask1), 0));
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
+    } else {
+      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
+      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
+      // wasteful to use scalar code, but being efficient with SIMD may require
+      // large, non-trivial tables?
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k), utf8_output);
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else { // 4-byte
+          if (word > 0x10FFFF) {
+            return std::make_pair(
+                result(error_code::TOO_LARGE, buf - start + k), utf8_output);
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
+    }
+  } // while
 
-    count += 16 - 3 * ascii_count - 2 * two_bytes_count - three_bytes_count;
-  }
-  return count +
-         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
 }
+/* end file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
-  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 4 <= length; pos += 4) {
-    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
-    const uint32x4_t surrogate_bytemask = vcgtq_u32(in, v_ffff);
-    const uint16x8_t reduced_bytemask =
-        vreinterpretq_u16_u32(vandq_u32(surrogate_bytemask, v_1));
-    const uint16x8_t compressed_bytemask =
-        vpaddq_u16(reduced_bytemask, reduced_bytemask);
-    size_t surrogate_count = count_ones(
-        vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask), 0));
-    count += 4 + surrogate_count;
-  }
-  return count +
-         scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
-}
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
+// file included directly
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return utf8::count_code_points(input, length);
-}
+template <endianness big_endian>
+std::pair<const char32_t *, char16_t *>
+avx512_convert_utf32_to_utf16(const char32_t *buf, size_t len,
+                              char16_t *utf16_output) {
+  const char32_t *end = buf + len;
+  __mmask32 forbidden_bytemask = 0;
+  const __m512i v_00000000 = _mm512_setzero_si512();
+  const __m512i v_ffff0000 = _mm512_set1_epi32((int32_t)0xffff0000);
+  const __m512i v_f800 = _mm512_set1_epi32((uint32_t)0xf800);
+  const __m512i v_d800 = _mm512_set1_epi32((uint32_t)0xd800);
+  const __m512i v_10ffff = _mm512_set1_epi32(0x10FFFF);
+  const __m512i v_10000 = _mm512_set1_epi32(0x10000);
+  const __m512i v_3ff0000 = _mm512_set1_epi32(0x3FF0000);
+  const __m512i v_3ff = _mm512_set1_epi32(0x3FF);
+  const __m512i v_dc00d800 = _mm512_set1_epi32((int32_t)0xDC00D800);
+
+  while (end - buf >= std::ptrdiff_t(16)) {
+    __m512i in = _mm512_loadu_si512(buf);
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
+    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
+    const __mmask16 saturation_bitmask =
+        _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_ffff0000), v_00000000);
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
+    if (saturation_bitmask == 0xffff) {
+      forbidden_bytemask |=
+          _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_f800), v_d800);
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      __m256i utf16_packed = _mm512_cvtepi32_epi16(in);
+      if (big_endian) {
+        const __m256i swap = _mm256_setr_epi8(
+            1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 1, 0, 3, 2, 5,
+            4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm256_shuffle_epi8(utf16_packed, swap);
+      }
+      _mm256_storeu_si256((__m256i *)utf16_output, utf16_packed);
+      utf16_output += 16;
+      buf += 16;
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      // saturation_bitmask == 1 words will generate 1 utf16 char,
+      // and saturation_bitmask == 0 words will generate 2 utf16 chars assuming
+      // no errors. Thus we need a output_mask which has the structure b_2i = 1,
+      // b_2i+1 = !saturation_bitmask_i
+      const __mmask32 output_mask = ~_pdep_u32(saturation_bitmask, 0xAAAAAAAA);
+      const __mmask16 surrogate_bitmask = __mmask16(~saturation_bitmask);
+      __mmask32 error = _mm512_mask_cmpeq_epi32_mask(
+          saturation_bitmask, _mm512_and_si512(in, v_f800), v_d800);
+      error |= _mm512_mask_cmpgt_epu32_mask(surrogate_bitmask, in, v_10ffff);
+      if (simdutf_unlikely(error)) {
+        return std::make_pair(nullptr, utf16_output);
+      }
+      __m512i v1, v2, v;
+      // for the bits saturation_bitmask == 0, we need to unpack the 32-bit word
+      // into two 16 bit words corresponding to high_surrogate and
+      // low_surrogate. Once the bits are unpacked and merged, the output will
+      // be compressed as per output_mask.
+      in = _mm512_mask_sub_epi32(in, surrogate_bitmask, in, v_10000);
+      v1 = _mm512_mask_slli_epi32(in, surrogate_bitmask, in, 16);
+      v1 = _mm512_mask_and_epi32(in, surrogate_bitmask, v1, v_3ff0000);
+      v2 = _mm512_mask_srli_epi32(in, surrogate_bitmask, in, 10);
+      v2 = _mm512_mask_and_epi32(in, surrogate_bitmask, v2, v_3ff);
+      v = _mm512_or_si512(v1, v2);
+      in = _mm512_mask_add_epi32(in, surrogate_bitmask, v, v_dc00d800);
+      if (big_endian) {
+        const __m512i swap_512 = _mm512_set_epi8(
+            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12,
+            13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8,
+            9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5,
+            2, 3, 0, 1);
+        in = _mm512_shuffle_epi8(in, swap_512);
+      }
+      // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      __m512i compressed = _mm512_maskz_compress_epi16(output_mask, in);
+      auto written_out = _mm_popcnt_u32(output_mask);
+      _mm512_mask_storeu_epi16(utf16_output, _bzhi_u32(0xFFFFFFFF, written_out),
+                               compressed);
+      //_mm512_mask_compressstoreu_epi16(utf16_output, output_mask, in);
+      utf16_output += written_out;
+      buf += 16;
     }
   }
-}
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
+  size_t remaining_len = size_t(end - buf);
+  if (remaining_len) {
+    __mmask16 input_mask = __mmask16((1 << remaining_len) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(input_mask, buf);
+    const __mmask16 saturation_bitmask =
+        _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_ffff0000), v_00000000) &
+        input_mask;
+    if (saturation_bitmask == input_mask) {
+      forbidden_bytemask |=
+          _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_f800), v_d800);
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      __m256i utf16_packed = _mm512_cvtepi32_epi16(in);
+      if (big_endian) {
+        const __m256i swap = _mm256_setr_epi8(
+            1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 1, 0, 3, 2, 5,
+            4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm256_shuffle_epi8(utf16_packed, swap);
+      }
+      _mm256_mask_storeu_epi16(utf16_output, input_mask, utf16_packed);
+      utf16_output += remaining_len;
+      buf += remaining_len;
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      const __mmask32 output_max_mask = (1 << (remaining_len * 2)) - 1;
+      const __mmask32 output_mask =
+          (~_pdep_u32(saturation_bitmask, 0xAAAAAAAA)) & output_max_mask;
+      const __mmask16 surrogate_bitmask =
+          __mmask16(~saturation_bitmask) & input_mask;
+      __mmask32 error = _mm512_mask_cmpeq_epi32_mask(
+          saturation_bitmask, _mm512_and_si512(in, v_f800), v_d800);
+      error |= _mm512_mask_cmpgt_epu32_mask(surrogate_bitmask, in, v_10ffff);
+      if (simdutf_unlikely(error)) {
+        return std::make_pair(nullptr, utf16_output);
+      }
+      __m512i v1, v2, v;
+      in = _mm512_mask_sub_epi32(in, surrogate_bitmask, in, v_10000);
+      v1 = _mm512_mask_slli_epi32(in, surrogate_bitmask, in, 16);
+      v1 = _mm512_mask_and_epi32(in, surrogate_bitmask, v1, v_3ff0000);
+      v2 = _mm512_mask_srli_epi32(in, surrogate_bitmask, in, 10);
+      v2 = _mm512_mask_and_epi32(in, surrogate_bitmask, v2, v_3ff);
+      v = _mm512_or_si512(v1, v2);
+      in = _mm512_mask_add_epi32(in, surrogate_bitmask, v, v_dc00d800);
+      if (big_endian) {
+        const __m512i swap_512 = _mm512_set_epi8(
+            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12,
+            13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8,
+            9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5,
+            2, 3, 0, 1);
+        in = _mm512_shuffle_epi8(in, swap_512);
+      }
+      // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      __m512i compressed = _mm512_maskz_compress_epi16(output_mask, in);
+      auto written_out = _mm_popcnt_u32(output_mask);
+      _mm512_mask_storeu_epi16(utf16_output, _bzhi_u32(0xFFFFFFFF, written_out),
+                               compressed);
+      //_mm512_mask_compressstoreu_epi16(utf16_output, output_mask, in);
+      utf16_output += written_out;
+      buf += remaining_len;
     }
   }
-}
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
+  // check for invalid input
+  if (forbidden_bytemask != 0) {
+    return std::make_pair(nullptr, utf16_output);
   }
-}
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
+  return std::make_pair(buf, utf16_output);
 }
 
-size_t implementation::binary_to_base64(const char *input, size_t length,
-                                        char *output,
-                                        base64_options options) const noexcept {
-  return encode_base64(output, input, length, options);
-}
+template <endianness big_endian>
+std::pair<result, char16_t *>
+avx512_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
+                                          char16_t *utf16_output) {
+  const char32_t *start = buf;
+  const char32_t *end = buf + len;
+  const __m512i v_00000000 = _mm512_setzero_si512();
+  const __m512i v_ffff0000 = _mm512_set1_epi32((int32_t)0xffff0000);
+  const __m512i v_f800 = _mm512_set1_epi32((uint32_t)0xf800);
+  const __m512i v_d800 = _mm512_set1_epi32((uint32_t)0xd800);
+  const __m512i v_10ffff = _mm512_set1_epi32(0x10FFFF);
+  const __m512i v_10000 = _mm512_set1_epi32(0x10000);
+  const __m512i v_3ff0000 = _mm512_set1_epi32(0x3FF0000);
+  const __m512i v_3ff = _mm512_set1_epi32(0x3FF);
+  const __m512i v_dc00d800 = _mm512_set1_epi32((int32_t)0xDC00D800);
+  int error_idx = 0;
+  error_code code = error_code::SUCCESS;
+  bool err = false;
+
+  while (end - buf >= std::ptrdiff_t(16)) {
+    __m512i in = _mm512_loadu_si512(buf);
 
-} // namespace arm64
-} // namespace simdutf
+    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
+    const __mmask16 saturation_bitmask =
+        _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_ffff0000), v_00000000);
 
-/* begin file src/simdutf/arm64/end.h */
-/* end file src/simdutf/arm64/end.h */
-/* end file src/arm64/implementation.cpp */
-#endif
-#if SIMDUTF_IMPLEMENTATION_FALLBACK
-/* begin file src/fallback/implementation.cpp */
-/* begin file src/simdutf/fallback/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "fallback"
-// #define SIMDUTF_IMPLEMENTATION fallback
-/* end file src/simdutf/fallback/begin.h */
+    if (saturation_bitmask == 0xffff) {
+      __mmask32 forbidden_bytemask =
+          _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_f800), v_d800);
 
+      __m256i utf16_packed = _mm512_cvtepi32_epi16(in);
+      if (big_endian) {
+        const __m256i swap = _mm256_setr_epi8(
+            1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 1, 0, 3, 2, 5,
+            4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm256_shuffle_epi8(utf16_packed, swap);
+      }
+      if (simdutf_unlikely(forbidden_bytemask)) {
+        int idx = _tzcnt_u32(forbidden_bytemask);
+        _mm256_mask_storeu_epi16(
+            utf16_output, __mmask16(_blsmsk_u32(forbidden_bytemask) >> 1),
+            utf16_packed);
+        return std::make_pair(result(error_code::SURROGATE, buf - start + idx),
+                              utf16_output + idx);
+      }
+      _mm256_storeu_si256((__m256i *)utf16_output, utf16_packed);
+      utf16_output += 16;
+    } else {
+      __mmask32 output_mask = ~_pdep_u32(saturation_bitmask, 0xAAAAAAAA);
+      const __mmask16 surrogate_bitmask = __mmask16(~saturation_bitmask);
+      __mmask32 error_surrogate = _mm512_mask_cmpeq_epi32_mask(
+          saturation_bitmask, _mm512_and_si512(in, v_f800), v_d800);
+      __mmask32 error_too_large =
+          _mm512_mask_cmpgt_epu32_mask(surrogate_bitmask, in, v_10ffff);
+      if (simdutf_unlikely(error_surrogate || error_too_large)) {
+        // Need to find the lowest set bit between the two error masks
+        // Need to also write the partial chunk until the error index to output.
+        int large_idx = _tzcnt_u32(error_too_large);
+        int surrogate_idx = _tzcnt_u32(error_surrogate);
+        err = true;
+        if (large_idx < surrogate_idx) {
+          code = error_code::TOO_LARGE;
+          error_idx = large_idx;
+        } else {
+          code = error_code::SURROGATE;
+          error_idx = surrogate_idx;
+        }
+        output_mask &= ((1 << (2 * error_idx)) - 1);
+      }
+      __m512i v1, v2, v;
+      in = _mm512_mask_sub_epi32(in, surrogate_bitmask, in, v_10000);
+      v1 = _mm512_mask_slli_epi32(in, surrogate_bitmask, in, 16);
+      v1 = _mm512_mask_and_epi32(in, surrogate_bitmask, v1, v_3ff0000);
+      v2 = _mm512_mask_srli_epi32(in, surrogate_bitmask, in, 10);
+      v2 = _mm512_mask_and_epi32(in, surrogate_bitmask, v2, v_3ff);
+      v = _mm512_or_si512(v1, v2);
+      in = _mm512_mask_add_epi32(in, surrogate_bitmask, v, v_dc00d800);
+      if (big_endian) {
+        const __m512i swap_512 = _mm512_set_epi8(
+            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12,
+            13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8,
+            9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5,
+            2, 3, 0, 1);
+        in = _mm512_shuffle_epi8(in, swap_512);
+      }
+      // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      __m512i compressed = _mm512_maskz_compress_epi16(output_mask, in);
+      auto written_out = _mm_popcnt_u32(output_mask);
+      _mm512_mask_storeu_epi16(utf16_output, _bzhi_u32(0xFFFFFFFF, written_out),
+                               compressed);
+      //_mm512_mask_compressstoreu_epi16(utf16_output, output_mask, in);
+      utf16_output += written_out;
+      if (simdutf_unlikely(err)) {
+        return std::make_pair(result(code, buf - start + error_idx),
+                              utf16_output);
+      }
+    }
+    buf += 16;
+  }
 
+  size_t remaining_len = size_t(end - buf);
+  if (remaining_len) {
+    __mmask16 input_mask = __mmask16((1 << remaining_len) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(input_mask, buf);
+    const __mmask16 saturation_bitmask =
+        _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_ffff0000), v_00000000) &
+        input_mask;
+    if (saturation_bitmask == input_mask) {
+      __mmask32 forbidden_bytemask =
+          _mm512_cmpeq_epi32_mask(_mm512_and_si512(in, v_f800), v_d800);
+      __m256i utf16_packed = _mm512_cvtepi32_epi16(in);
+      if (big_endian) {
+        const __m256i swap = _mm256_setr_epi8(
+            1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 1, 0, 3, 2, 5,
+            4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm256_shuffle_epi8(utf16_packed, swap);
+      }
+      if (simdutf_unlikely(forbidden_bytemask)) {
+        int idx = _tzcnt_u32(forbidden_bytemask);
+        _mm256_mask_storeu_epi16(
+            utf16_output, __mmask16(_blsmsk_u32(forbidden_bytemask) >> 1),
+            utf16_packed);
+        return std::make_pair(result(error_code::SURROGATE, buf - start + idx),
+                              utf16_output + idx);
+      }
+      _mm256_mask_storeu_epi16(utf16_output, input_mask, utf16_packed);
+      utf16_output += remaining_len;
+    } else {
+      const __mmask32 output_max_mask = (1 << (remaining_len * 2)) - 1;
+      __mmask32 output_mask =
+          (~_pdep_u32(saturation_bitmask, 0xAAAAAAAA)) & output_max_mask;
+      const __mmask16 surrogate_bitmask =
+          __mmask16(~saturation_bitmask) & input_mask;
+      __mmask32 error_surrogate = _mm512_mask_cmpeq_epi32_mask(
+          saturation_bitmask, _mm512_and_si512(in, v_f800), v_d800);
+      __mmask32 error_too_large =
+          _mm512_mask_cmpgt_epu32_mask(surrogate_bitmask, in, v_10ffff);
+      if (simdutf_unlikely(error_surrogate || error_too_large)) {
+        int large_idx = _tzcnt_u32(error_too_large);
+        int surrogate_idx = _tzcnt_u32(error_surrogate);
+        err = true;
+        if (large_idx < surrogate_idx) {
+          code = error_code::TOO_LARGE;
+          error_idx = large_idx;
+        } else {
+          code = error_code::SURROGATE;
+          error_idx = surrogate_idx;
+        }
+        output_mask &= ((1 << (2 * error_idx)) - 1);
+      }
+      __m512i v1, v2, v;
+      in = _mm512_mask_sub_epi32(in, surrogate_bitmask, in, v_10000);
+      v1 = _mm512_mask_slli_epi32(in, surrogate_bitmask, in, 16);
+      v1 = _mm512_mask_and_epi32(in, surrogate_bitmask, v1, v_3ff0000);
+      v2 = _mm512_mask_srli_epi32(in, surrogate_bitmask, in, 10);
+      v2 = _mm512_mask_and_epi32(in, surrogate_bitmask, v2, v_3ff);
+      v = _mm512_or_si512(v1, v2);
+      in = _mm512_mask_add_epi32(in, surrogate_bitmask, v, v_dc00d800);
+      if (big_endian) {
+        const __m512i swap_512 = _mm512_set_epi8(
+            14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12,
+            13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8,
+            9, 6, 7, 4, 5, 2, 3, 0, 1, 14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5,
+            2, 3, 0, 1);
+        in = _mm512_shuffle_epi8(in, swap_512);
+      }
+      // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      __m512i compressed = _mm512_maskz_compress_epi16(output_mask, in);
+      auto written_out = _mm_popcnt_u32(output_mask);
+      _mm512_mask_storeu_epi16(utf16_output, _bzhi_u32(0xFFFFFFFF, written_out),
+                               compressed);
+      //_mm512_mask_compressstoreu_epi16(utf16_output, output_mask, in);
+      utf16_output += written_out;
+      if (simdutf_unlikely(err)) {
+        return std::make_pair(result(code, buf - start + error_idx),
+                              utf16_output);
+      }
+    }
+    buf += remaining_len;
+  }
 
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
+}
+/* end file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/icelake/icelake_ascii_validation.inl.cpp */
+// file included directly
 
+bool validate_ascii(const char *buf, size_t len) {
+  const char *end = buf + len;
+  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
+  __m512i running_or = _mm512_setzero_si512();
+  for (; end - buf >= 64; buf += 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)buf);
+    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
+                                           0xf8); // running_or | (utf8 & ascii)
+  }
+  if (buf < end) {
+    const __m512i utf8 = _mm512_maskz_loadu_epi8(
+        (uint64_t(1) << (end - buf)) - 1, (const __m512i *)buf);
+    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
+                                           0xf8); // running_or | (utf8 & ascii)
+  }
+  return (_mm512_test_epi8_mask(running_or, running_or) == 0);
+}
+/* end file src/icelake/icelake_ascii_validation.inl.cpp */
+#endif // SIMDUTF_FEATURE_ASCII
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/icelake/icelake_utf32_validation.inl.cpp */
+// file included directly
 
+bool validate_utf32(const char32_t *buf, size_t len) {
+  if (simdutf_unlikely(len == 0)) {
+    return true;
+  }
+  const char32_t *end = buf + len;
 
+  const __m512i offset = _mm512_set1_epi32((uint32_t)0xffff2000);
+  __m512i currentmax = _mm512_setzero_si512();
+  __m512i currentoffsetmax = _mm512_setzero_si512();
 
-#include <cstdint>
-#include <cstring>
+  while (buf < end - 16) {
+    __m512i utf32 = _mm512_loadu_si512((const __m512i *)buf);
+    buf += 16;
+    currentoffsetmax =
+        _mm512_max_epu32(_mm512_add_epi32(utf32, offset), currentoffsetmax);
+    currentmax = _mm512_max_epu32(utf32, currentmax);
+  }
 
-namespace simdutf {
-namespace fallback {
+  __m512i utf32 =
+      _mm512_maskz_loadu_epi32(__mmask16((1 << (end - buf)) - 1), buf);
+  currentoffsetmax =
+      _mm512_max_epu32(_mm512_add_epi32(utf32, offset), currentoffsetmax);
+  currentmax = _mm512_max_epu32(utf32, currentmax);
 
-simdutf_warn_unused int
-implementation::detect_encodings(const char *input,
-                                 size_t length) const noexcept {
-  // If there is a BOM, then we trust it.
-  auto bom_encoding = simdutf::BOM::check_bom(input, length);
-  if (bom_encoding != encoding_type::unspecified) {
-    return bom_encoding;
-  }
-  // todo: reimplement as a one-pass algorithm.
-  int out = 0;
-  if (validate_utf8(input, length)) {
-    out |= encoding_type::UTF8;
-  }
-  if ((length % 2) == 0) {
-    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
-                         length / 2)) {
-      out |= encoding_type::UTF16_LE;
-    }
-  }
-  if ((length % 4) == 0) {
-    if (validate_utf32(reinterpret_cast<const char32_t *>(input), length / 4)) {
-      out |= encoding_type::UTF32_LE;
-    }
+  const __m512i standardmax = _mm512_set1_epi32((uint32_t)0x10ffff);
+  const __m512i standardoffsetmax = _mm512_set1_epi32((uint32_t)0xfffff7ff);
+  const auto outside_range = _mm512_cmpgt_epu32_mask(currentmax, standardmax);
+  if (outside_range != 0) {
+    return false;
   }
-  return out;
-}
 
-simdutf_warn_unused bool
-implementation::validate_utf8(const char *buf, size_t len) const noexcept {
-  return scalar::utf8::validate(buf, len);
-}
+  const auto surrogate =
+      _mm512_cmpgt_epu32_mask(currentoffsetmax, standardoffsetmax);
+  if (surrogate != 0) {
+    return false;
+  }
 
-simdutf_warn_unused result implementation::validate_utf8_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return scalar::utf8::validate_with_errors(buf, len);
+  return true;
 }
+/* end file src/icelake/icelake_utf32_validation.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/icelake/icelake_convert_latin1_to_utf8.inl.cpp */
+// file included directly
 
-simdutf_warn_unused bool
-implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return scalar::ascii::validate(buf, len);
-}
+static inline size_t latin1_to_utf8_avx512_vec(__m512i input, size_t input_len,
+                                               char *utf8_output,
+                                               int mask_output) {
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
+  size_t output_size = input_len + (size_t)count_ones(nonascii);
 
-simdutf_warn_unused result implementation::validate_ascii_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return scalar::ascii::validate_with_errors(buf, len);
-}
+  // Mask to denote whether the byte is a leading byte that is not ascii
+  __mmask64 sixth = _mm512_cmpge_epu8_mask(
+      input, _mm512_set1_epi8(-64)); // binary representation of -64: 1100 0000
 
-simdutf_warn_unused bool
-implementation::validate_utf16le(const char16_t *buf,
-                                 size_t len) const noexcept {
-  return scalar::utf16::validate<endianness::LITTLE>(buf, len);
-}
+  const uint64_t alternate_bits = UINT64_C(0x5555555555555555);
+  uint64_t ascii = ~nonascii;
+  // the bits in ascii are inverted and zeros are interspersed in between them
+  uint64_t maskA = ~_pdep_u64(ascii, alternate_bits);
+  uint64_t maskB = ~_pdep_u64(ascii >> 32, alternate_bits);
 
-simdutf_warn_unused bool
-implementation::validate_utf16be(const char16_t *buf,
-                                 size_t len) const noexcept {
-  return scalar::utf16::validate<endianness::BIG>(buf, len);
-}
+  // interleave bytes from top and bottom halves (abcd...ABCD -> aAbBcCdD)
+  __m512i input_interleaved = _mm512_permutexvar_epi8(
+      _mm512_set_epi32(0x3f1f3e1e, 0x3d1d3c1c, 0x3b1b3a1a, 0x39193818,
+                       0x37173616, 0x35153414, 0x33133212, 0x31113010,
+                       0x2f0f2e0e, 0x2d0d2c0c, 0x2b0b2a0a, 0x29092808,
+                       0x27072606, 0x25052404, 0x23032202, 0x21012000),
+      input);
 
-simdutf_warn_unused result implementation::validate_utf16le_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  return scalar::utf16::validate_with_errors<endianness::LITTLE>(buf, len);
-}
+  // double size of each byte, and insert the leading byte 1100 0010
 
-simdutf_warn_unused result implementation::validate_utf16be_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  return scalar::utf16::validate_with_errors<endianness::BIG>(buf, len);
-}
+  /*
+  upscale the bytes to 16-bit value, adding the 0b11000000 leading byte in the
+  process. We adjust for the bytes that have their two most significant bits.
+  This takes care of the first 32 bytes, assuming we interleaved the bytes. */
+  __m512i outputA =
+      _mm512_shldi_epi16(input_interleaved, _mm512_set1_epi8(-62), 8);
+  outputA = _mm512_mask_add_epi16(
+      outputA, (__mmask32)sixth, outputA,
+      _mm512_set1_epi16(1 - 0x4000)); // 1- 0x4000 = 1100 0000 0000 0001????
 
-simdutf_warn_unused bool
-implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
-  return scalar::utf32::validate(buf, len);
-}
+  // in the second 32-bit half, set first or second option based on whether
+  // original input is leading byte (second case) or not (first case)
+  __m512i leadingB =
+      _mm512_mask_blend_epi16((__mmask32)(sixth >> 32),
+                              _mm512_set1_epi16(0x00c2),  // 0000 0000 1101 0010
+                              _mm512_set1_epi16(0x40c3)); // 0100 0000 1100 0011
+  __m512i outputB = _mm512_ternarylogic_epi32(
+      input_interleaved, leadingB, _mm512_set1_epi16((short)0xff00),
+      (240 & 170) ^ 204); // (input_interleaved & 0xff00) ^ leadingB
 
-simdutf_warn_unused result implementation::validate_utf32_with_errors(
-    const char32_t *buf, size_t len) const noexcept {
-  return scalar::utf32::validate_with_errors(buf, len);
-}
+  // prune redundant bytes
+  outputA = _mm512_maskz_compress_epi8(maskA, outputA);
+  outputB = _mm512_maskz_compress_epi8(maskB, outputB);
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
-    const char *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::latin1_to_utf8::convert(buf, len, utf8_output);
-}
+  size_t output_sizeA = (size_t)count_ones((uint32_t)nonascii) + 32;
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::latin1_to_utf16::convert<endianness::LITTLE>(buf, len,
-                                                              utf16_output);
+  if (mask_output) {
+    if (input_len > 32) { // is the second half of the input vector used?
+      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_sizeA);
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
+      utf8_output += output_sizeA;
+      write_mask = _bzhi_u64(~0ULL, (unsigned int)(output_size - output_sizeA));
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputB);
+    } else {
+      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_size);
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
+    }
+  } else {
+    _mm512_storeu_si512(utf8_output, outputA);
+    utf8_output += output_sizeA;
+    _mm512_storeu_si512(utf8_output, outputB);
+  }
+  return output_size;
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::latin1_to_utf16::convert<endianness::BIG>(buf, len,
-                                                           utf16_output);
+static inline size_t latin1_to_utf8_avx512_branch(__m512i input,
+                                                  char *utf8_output) {
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
+  if (nonascii) {
+    return latin1_to_utf8_avx512_vec(input, 64, utf8_output, 0);
+  } else {
+    _mm512_storeu_si512(utf8_output, input);
+    return 64;
+  }
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::latin1_to_utf32::convert(buf, len, utf32_output);
+size_t latin1_to_utf8_avx512_start(const char *buf, size_t len,
+                                   char *utf8_output) {
+  char *start = utf8_output;
+  size_t pos = 0;
+  // if there's at least 128 bytes remaining, we don't need to mask the output
+  for (; pos + 128 <= len; pos += 64) {
+    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_branch(input, utf8_output);
+  }
+  // in the last 128 bytes, the first 64 may require masking the output
+  if (pos + 64 <= len) {
+    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_vec(input, 64, utf8_output, 1);
+    pos += 64;
+  }
+  // with the last 64 bytes, the input also needs to be masked
+  if (pos < len) {
+    __mmask64 load_mask = _bzhi_u64(~0ULL, (unsigned int)(len - pos));
+    __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_vec(input, len - pos, utf8_output, 1);
+  }
+  return (size_t)(utf8_output - start);
 }
+/* end file src/icelake/icelake_convert_latin1_to_utf8.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/icelake/icelake_convert_latin1_to_utf16.inl.cpp */
+// file included directly
+template <endianness big_endian>
+size_t icelake_convert_latin1_to_utf16(const char *latin1_input, size_t len,
+                                       char16_t *utf16_output) {
+  size_t rounded_len = len & ~0x1F; // Round down to nearest multiple of 32
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf8_to_latin1::convert(buf, len, latin1_output);
-}
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  for (size_t i = 0; i < rounded_len; i += 32) {
+    // Load 32 Latin1 characters into a 256-bit register
+    __m256i in = _mm256_loadu_si256((__m256i *)&latin1_input[i]);
+    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
+    __m512i out = _mm512_cvtepu8_epi16(in);
+    if (big_endian) {
+      out = _mm512_shuffle_epi8(out, byteflip);
+    }
+    // Store the results back to memory
+    _mm512_storeu_si512((__m512i *)&utf16_output[i], out);
+  }
+  if (rounded_len != len) {
+    uint32_t mask = uint32_t(1 << (len - rounded_len)) - 1;
+    __m256i in = _mm256_maskz_loadu_epi8(mask, latin1_input + rounded_len);
 
-simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf8_to_latin1::convert_with_errors(buf, len, latin1_output);
-}
+    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
+    __m512i out = _mm512_cvtepu8_epi16(in);
+    if (big_endian) {
+      out = _mm512_shuffle_epi8(out, byteflip);
+    }
+    // Store the results back to memory
+    _mm512_mask_storeu_epi16(utf16_output + rounded_len, mask, out);
+  }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf8_to_latin1::convert_valid(buf, len, latin1_output);
+  return len;
 }
+/* end file src/icelake/icelake_convert_latin1_to_utf16.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF32
+/* begin file src/icelake/icelake_convert_latin1_to_utf32.inl.cpp */
+void avx512_convert_latin1_to_utf32(const char *buf, size_t len,
+                                    char32_t *utf32_output) {
+  while (len >= 16) {
+    // Load 16 Latin1 characters into a 128-bit register
+    __m128i in = _mm_loadu_si128((__m128i *)buf);
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert<endianness::LITTLE>(buf, len,
-                                                            utf16_output);
-}
+    // Zero extend each set of 8 Latin1 characters to 16 32-bit integers using
+    // vpmovzxbd
+    __m512i out = _mm512_cvtepu8_epi32(in);
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert<endianness::BIG>(buf, len,
-                                                         utf16_output);
-}
+    // Store the results back to memory
+    _mm512_storeu_si512((__m512i *)utf32_output, out);
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf16_output);
-}
+    len -= 16;
+    buf += 16;
+    utf32_output += 16;
+  }
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert_with_errors<endianness::BIG>(
-      buf, len, utf16_output);
+  __mmask16 mask = __mmask16((1 << len) - 1);
+  __m128i in = _mm_maskz_loadu_epi8(mask, buf);
+  __m512i out = _mm512_cvtepu8_epi32(in);
+  _mm512_mask_storeu_epi32((__m512i *)utf32_output, mask, out);
 }
+/* end file src/icelake/icelake_convert_latin1_to_utf32.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/icelake/icelake_base64.inl.cpp */
+// file included directly
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(buf, len,
-                                                                  utf16_output);
-}
+struct block64 {
+  __m512i chunks[1];
+};
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf8_to_utf16::convert_valid<endianness::BIG>(buf, len,
-                                                               utf16_output);
-}
+template <bool base64_url>
+size_t encode_base64(char *dst, const char *src, size_t srclen,
+                     base64_options options) {
+  // credit: Wojciech Muła
+  const uint8_t *input = (const uint8_t *)src;
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf8_to_utf32::convert(buf, len, utf32_output);
-}
+  uint8_t *out = (uint8_t *)dst;
+  static const char *lookup_tbl =
+      base64_url
+          ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
+          : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf8_to_utf32::convert_with_errors(buf, len, utf32_output);
+  const __m512i shuffle_input = _mm512_setr_epi32(
+      0x01020001, 0x04050304, 0x07080607, 0x0a0b090a, 0x0d0e0c0d, 0x10110f10,
+      0x13141213, 0x16171516, 0x191a1819, 0x1c1d1b1c, 0x1f201e1f, 0x22232122,
+      0x25262425, 0x28292728, 0x2b2c2a2b, 0x2e2f2d2e);
+  const __m512i lookup =
+      _mm512_loadu_si512(reinterpret_cast<const __m512i *>(lookup_tbl));
+  const __m512i multi_shifts = _mm512_set1_epi64(UINT64_C(0x3036242a1016040a));
+  size_t size = srclen;
+  __mmask64 input_mask = 0xffffffffffff; // (1 << 48) - 1
+  while (size >= 48) {
+    const __m512i v = _mm512_maskz_loadu_epi8(
+        input_mask, reinterpret_cast<const __m512i *>(input));
+    const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
+    const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
+    const __m512i result = _mm512_permutexvar_epi8(indices, lookup);
+    _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), result);
+    out += 64;
+    input += 48;
+    size -= 48;
+  }
+  input_mask = ((__mmask64)1 << size) - 1;
+  const __m512i v = _mm512_maskz_loadu_epi8(
+      input_mask, reinterpret_cast<const __m512i *>(input));
+  const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
+  const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
+  bool padding_needed =
+      (((options & base64_url) == 0) ^
+       ((options & base64_reverse_padding) == base64_reverse_padding));
+  size_t padding_amount = ((size % 3) > 0) ? (3 - (size % 3)) : 0;
+  size_t output_len = ((size + 2) / 3) * 4;
+  size_t non_padded_output_len = output_len - padding_amount;
+  if (!padding_needed) {
+    output_len = non_padded_output_len;
+  }
+  __mmask64 output_mask = output_len == 64 ? (__mmask64)UINT64_MAX
+                                           : ((__mmask64)1 << output_len) - 1;
+  __m512i result = _mm512_mask_permutexvar_epi8(
+      _mm512_set1_epi8('='), ((__mmask64)1 << non_padded_output_len) - 1,
+      indices, lookup);
+  _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out), output_mask,
+                          result);
+  return (size_t)(out - (uint8_t *)dst) + output_len;
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
-    const char *input, size_t size, char32_t *utf32_output) const noexcept {
-  return scalar::utf8_to_utf32::convert_valid(input, size, utf32_output);
-}
+template <bool base64_url, bool ignore_garbage>
+static inline uint64_t to_base64_mask(block64 *b, uint64_t *error,
+                                      uint64_t input_mask = UINT64_MAX) {
+  __m512i input = b->chunks[0];
+  const __m512i ascii_space_tbl = _mm512_set_epi8(
+      0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10,
+      9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0,
+      0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32);
+  __m512i lookup0;
+  if (base64_url) {
+    lookup0 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1,
+        -128, -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -1);
+  } else {
+    lookup0 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, 63, -128, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1, -128,
+        -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -128);
+  }
+  __m512i lookup1;
+  if (base64_url) {
+    lookup1 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
+        63, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15,
+        14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
+  } else {
+    lookup1 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
+        -128, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16,
+        15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
+  }
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert<endianness::LITTLE>(buf, len,
-                                                              latin1_output);
-}
+  const __m512i translated = _mm512_permutex2var_epi8(lookup0, input, lookup1);
+  const __m512i combined = _mm512_or_si512(translated, input);
+  const __mmask64 mask = _mm512_movepi8_mask(combined) & input_mask;
+  if (!ignore_garbage && mask) {
+    const __mmask64 spaces =
+        _mm512_cmpeq_epi8_mask(_mm512_shuffle_epi8(ascii_space_tbl, input),
+                               input) &
+        input_mask;
+    *error = (mask ^ spaces);
+  }
+  b->chunks[0] = translated;
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert<endianness::BIG>(buf, len,
-                                                           latin1_output);
+  return mask | (~input_mask);
 }
 
-simdutf_warn_unused result
-implementation::convert_utf16le_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
-      buf, len, latin1_output);
+static inline void copy_block(block64 *b, char *output) {
+  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), b->chunks[0]);
 }
 
-simdutf_warn_unused result
-implementation::convert_utf16be_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
-      buf, len, latin1_output);
+static inline uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
+  uint64_t nmask = ~mask;
+  __m512i c = _mm512_maskz_compress_epi8(nmask, b->chunks[0]);
+  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), c);
+  return _mm_popcnt_u64(nmask);
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert_valid<endianness::LITTLE>(
-      buf, len, latin1_output);
+// The caller of this function is responsible to ensure that there are 64 bytes
+// available from reading at src. The data is read into a block64 structure.
+static inline void load_block(block64 *b, const char *src) {
+  b->chunks[0] = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf16_to_latin1::convert_valid<endianness::BIG>(buf, len,
-                                                                 latin1_output);
+static inline void load_block_partial(block64 *b, const char *src,
+                                      __mmask64 input_mask) {
+  b->chunks[0] = _mm512_maskz_loadu_epi8(
+      input_mask, reinterpret_cast<const __m512i *>(src));
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert<endianness::LITTLE>(buf, len,
-                                                            utf8_output);
+// The caller of this function is responsible to ensure that there are 128 bytes
+// available from reading at src. The data is read into a block64 structure.
+static inline void load_block(block64 *b, const char16_t *src) {
+  __m512i m1 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
+  __m512i m2 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src + 32));
+  __m512i p = _mm512_packus_epi16(m1, m2);
+  b->chunks[0] =
+      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert<endianness::BIG>(buf, len, utf8_output);
+static inline void load_block_partial(block64 *b, const char16_t *src,
+                                      __mmask64 input_mask) {
+  __m512i m1 = _mm512_maskz_loadu_epi16((__mmask32)input_mask,
+                                        reinterpret_cast<const __m512i *>(src));
+  __m512i m2 =
+      _mm512_maskz_loadu_epi16((__mmask32)(input_mask >> 32),
+                               reinterpret_cast<const __m512i *>(src + 32));
+  __m512i p = _mm512_packus_epi16(m1, m2);
+  b->chunks[0] =
+      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
 }
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf8_output);
+static inline void base64_decode(char *out, __m512i str) {
+  const __m512i merge_ab_and_bc =
+      _mm512_maddubs_epi16(str, _mm512_set1_epi32(0x01400140));
+  const __m512i merged =
+      _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
+  const __m512i pack = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
+      52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
+      28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
+      5, 6, 0, 1, 2);
+  const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
+  _mm512_mask_storeu_epi8(
+      (__m512i *)out, 0xffffffffffff,
+      shuffled); // mask would be 0xffffffffffff since we write 48 bytes.
 }
-
-simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
-      buf, len, utf8_output);
+// decode 64 bytes and output 48 bytes
+static inline void base64_decode_block(char *out, const char *src) {
+  base64_decode(out,
+                _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src)));
 }
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_valid<endianness::LITTLE>(buf, len,
-                                                                  utf8_output);
+static inline void base64_decode_block(char *out, block64 *b) {
+  base64_decode(out, b->chunks[0]);
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_valid<endianness::BIG>(buf, len,
-                                                               utf8_output);
-}
+template <bool base64_url, bool ignore_garbage, typename chartype>
+full_result
+compress_decode_base64(char *dst, const chartype *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  (void)options;
+  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
+                                        : tables::base64::to_base64_value;
+  size_t equallocation =
+      srclen; // location of the first padding character if any
+  size_t equalsigns = 0;
+  // skip trailing spaces
+  while (!ignore_garbage && srclen > 0 &&
+         scalar::base64::is_eight_byte(src[srclen - 1]) &&
+         to_base64[uint8_t(src[srclen - 1])] == 64) {
+    srclen--;
+  }
+  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
+    equallocation = srclen - 1;
+    srclen--;
+    equalsigns = 1;
+    // skip trailing spaces
+    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+           to_base64[uint8_t(src[srclen - 1])] == 64) {
+      srclen--;
+    }
+    if (srclen > 0 && src[srclen - 1] == '=') {
+      equallocation = srclen - 1;
+      srclen--;
+      equalsigns = 2;
+    }
+  }
+  if (srclen == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, 0, 0};
+  }
+  const chartype *const srcinit = src;
+  const char *const dstinit = dst;
+  const chartype *const srcend = src + srclen;
+
+  // figure out why block_size == 2 is sometimes best???
+  constexpr size_t block_size = 6;
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const chartype *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b;
+      load_block(&b, src);
+      src += 64;
+      uint64_t error = 0;
+      uint64_t badcharmask =
+          to_base64_mask<base64_url, ignore_garbage>(&b, &error);
+      if (!ignore_garbage && error) {
+        src -= 64;
+        size_t error_offset = _tzcnt_u64(error);
+        return {error_code::INVALID_BASE64_CHARACTER,
+                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+      }
+      if (badcharmask != 0) {
+        // optimization opportunity: check for simple masks like those made of
+        // continuous 1s followed by continuous 0s. And masks containing a
+        // single bad character.
+        bufferptr += compress_block(&b, badcharmask, bufferptr);
+      } else if (bufferptr != buffer) {
+        copy_block(&b, bufferptr);
+        bufferptr += 64;
+      } else {
+        base64_decode_block(dst, &b);
+        dst += 48;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 1); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
+  }
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf32_to_latin1::convert(buf, len, latin1_output);
-}
+  int last_block_len = (int)(srcend - src);
+  if (last_block_len != 0) {
+    __mmask64 input_mask = ((__mmask64)1 << last_block_len) - 1;
+    block64 b;
+    load_block_partial(&b, src, input_mask);
+    uint64_t error = 0;
+    uint64_t badcharmask =
+        to_base64_mask<base64_url, ignore_garbage>(&b, &error, input_mask);
+    if (!ignore_garbage && error) {
+      size_t error_offset = _tzcnt_u64(error);
+      return {error_code::INVALID_BASE64_CHARACTER,
+              size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+    }
+    src += last_block_len;
+    bufferptr += compress_block(&b, badcharmask, bufferptr);
+  }
 
-simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf32_to_latin1::convert_with_errors(buf, len, latin1_output);
-}
+  char *buffer_start = buffer;
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    base64_decode_block(dst, buffer_start);
+    dst += 48;
+  }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return scalar::utf32_to_latin1::convert_valid(buf, len, latin1_output);
-}
+  if ((bufferptr - buffer_start) != 0) {
+    size_t rem = (bufferptr - buffer_start);
+    int idx = rem % 4;
+    __mmask64 mask = ((__mmask64)1 << rem) - 1;
+    __m512i input = _mm512_maskz_loadu_epi8(mask, buffer_start);
+    size_t output_len = (rem / 4) * 3;
+    __mmask64 output_mask = mask >> (rem - output_len);
+    const __m512i merge_ab_and_bc =
+        _mm512_maddubs_epi16(input, _mm512_set1_epi32(0x01400140));
+    const __m512i merged =
+        _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
+    const __m512i pack = _mm512_set_epi8(
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
+        52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
+        28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
+        5, 6, 0, 1, 2);
+    const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert(buf, len, utf8_output);
-}
+    if (!ignore_garbage &&
+        last_chunk_options == last_chunk_handling_options::strict &&
+        (idx != 1) && ((idx + equalsigns) & 3) != 0) {
+      // The partial chunk was at src - idx
+      _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+      dst += output_len;
+      return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+              size_t(dst - dstinit)};
+    } else if (!ignore_garbage &&
+               last_chunk_options ==
+                   last_chunk_handling_options::stop_before_partial &&
+               (idx != 1) && ((idx + equalsigns) & 3) != 0) {
+      // Rewind src to before partial chunk
+      _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+      dst += output_len;
+      src -= idx;
+    } else {
+      if (idx == 2) {
+        if (!ignore_garbage &&
+            last_chunk_options == last_chunk_handling_options::strict) {
+          uint32_t triple = (uint32_t(bufferptr[-2]) << 3 * 6) +
+                            (uint32_t(bufferptr[-1]) << 2 * 6);
+          if (triple & 0xffff) {
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                    size_t(dst - dstinit)};
+          }
+        }
+        output_mask = (output_mask << 1) | 1;
+        output_len += 1;
+        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+        dst += output_len;
+      } else if (idx == 3) {
+        if (!ignore_garbage &&
+            last_chunk_options == last_chunk_handling_options::strict) {
+          uint32_t triple = (uint32_t(bufferptr[-3]) << 3 * 6) +
+                            (uint32_t(bufferptr[-2]) << 2 * 6) +
+                            (uint32_t(bufferptr[-1]) << 1 * 6);
+          if (triple & 0xff) {
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                    size_t(dst - dstinit)};
+          }
+        }
+        output_mask = (output_mask << 2) | 3;
+        output_len += 2;
+        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+        dst += output_len;
+      } else if (!ignore_garbage && idx == 1) {
+        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+        dst += output_len;
+        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      } else {
+        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+        dst += output_len;
+      }
+    }
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert_with_errors(buf, len, utf8_output);
-}
+    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
+        equalsigns > 0) {
+      size_t output_count = size_t(dst - dstinit);
+      if ((output_count % 3 == 0) ||
+          ((output_count % 3) + 1 + equalsigns != 4)) {
+        return {INVALID_BASE64_CHARACTER, equallocation, output_count};
+      }
+    }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert_valid(buf, len, utf8_output);
-}
+    return {SUCCESS, srclen, size_t(dst - dstinit)};
+  }
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert<endianness::LITTLE>(buf, len,
-                                                             utf16_output);
+  if (!ignore_garbage && equalsigns > 0) {
+    if (last_chunk_options == last_chunk_handling_options::strict) {
+      return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+              size_t(dst - dstinit)};
+    }
+    if (last_chunk_options ==
+        last_chunk_handling_options::stop_before_partial) {
+      return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
+    }
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
 }
+/* end file src/icelake/icelake_base64.inl.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert<endianness::BIG>(buf, len,
-                                                          utf16_output);
-}
+#include <cstdint>
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf16_output);
-}
+} // namespace
+} // namespace icelake
+} // namespace simdutf
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
-      buf, len, utf16_output);
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf32.h */
+#include <limits>
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_valid<endianness::LITTLE>(
-      buf, len, utf16_output);
-}
+namespace simdutf {
+namespace icelake {
+namespace {
+namespace utf32 {
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_valid<endianness::BIG>(buf, len,
-                                                                utf16_output);
-}
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert<endianness::LITTLE>(buf, len,
-                                                             utf32_output);
-}
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert<endianness::BIG>(buf, len,
-                                                          utf32_output);
-}
+  const char32_t *start = input;
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf32_output);
-}
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-      buf, len, utf32_output);
-}
+  const size_t N = vector_u32::ELEMENTS;
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_valid<endianness::LITTLE>(
-      buf, len, utf32_output);
-}
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_valid<endianness::BIG>(buf, len,
-                                                                utf32_output);
-}
+  size_t counter = 0;
 
-void implementation::change_endianness_utf16(const char16_t *input,
-                                             size_t length,
-                                             char16_t *output) const noexcept {
-  scalar::utf16::change_endianness_utf16(input, length, output);
-}
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
 
-simdutf_warn_unused size_t implementation::count_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::count_code_points<endianness::LITTLE>(input, length);
-}
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
 
-simdutf_warn_unused size_t implementation::count_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::count_code_points<endianness::BIG>(input, length);
-}
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-simdutf_warn_unused size_t
-implementation::count_utf8(const char *input, size_t length) const noexcept {
-  return scalar::utf8::count_code_points(input, length);
-}
+        input += 4 * N;
+      }
 
-simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
-    const char *buf, size_t len) const noexcept {
-  return scalar::utf8::count_code_points(buf, len);
-}
+      counter += acc.sum();
+    }
+  }
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return scalar::utf16::latin1_length_from_utf16(length);
-}
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return length;
-}
+        input += N;
+      }
 
-simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
-    const char *input, size_t length) const noexcept {
-  size_t answer = length;
-  size_t i = 0;
-  auto pop = [](uint64_t v) {
-    return (size_t)(((v >> 7) & UINT64_C(0x0101010101010101)) *
-                        UINT64_C(0x0101010101010101) >>
-                    56);
-  };
-  for (; i + 32 <= length; i += 32) {
-    uint64_t v;
-    memcpy(&v, input + i, 8);
-    answer += pop(v);
-    memcpy(&v, input + i + 8, sizeof(v));
-    answer += pop(v);
-    memcpy(&v, input + i + 16, sizeof(v));
-    answer += pop(v);
-    memcpy(&v, input + i + 24, sizeof(v));
-    answer += pop(v);
-  }
-  for (; i + 8 <= length; i += 8) {
-    uint64_t v;
-    memcpy(&v, input + i, sizeof(v));
-    answer += pop(v);
+      counter += acc.sum();
+    }
   }
-  for (; i + 1 <= length; i += 1) {
-    answer += static_cast<uint8_t>(input[i]) >> 7;
+
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
   }
-  return answer;
-}
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(input,
-                                                                   length);
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
 }
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
-}
+} // namespace utf32
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdutf
+/* end file src/generic/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf32_length_from_utf16<endianness::LITTLE>(input,
-                                                                    length);
-}
+namespace simdutf {
+namespace icelake {
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
-}
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused int
+implementation::detect_encodings(const char *input,
+                                 size_t length) const noexcept {
+  // If there is a BOM, then we trust it.
+  auto bom_encoding = simdutf::BOM::check_bom(input, length);
+  if (bom_encoding != encoding_type::unspecified) {
+    return bom_encoding;
+  }
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf16_length_from_latin1(length);
-}
+  int out = 0;
+  uint32_t utf16_err = (length % 2);
+  uint32_t utf32_err = (length % 4);
+  uint32_t ends_with_high = 0;
+  avx512_utf8_checker checker{};
+  const __m512i offset = _mm512_set1_epi32((uint32_t)0xffff2000);
+  __m512i currentmax = _mm512_setzero_si512();
+  __m512i currentoffsetmax = _mm512_setzero_si512();
+  const char *ptr = input;
+  const char *end = ptr + length;
+  for (; end - ptr >= 64; ptr += 64) {
+    // utf8 checks
+    const __m512i data = _mm512_loadu_si512((const __m512i *)ptr);
+    checker.check_next_input(data);
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return scalar::utf8::utf16_length_from_utf8(input, length);
-}
+    // utf16le_checks
+    __m512i diff = _mm512_sub_epi16(data, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    __mmask32 highsurrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+    __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+    utf16_err |= (((highsurrogates << 1) | ends_with_high) != lowsurrogates);
+    ends_with_high = ((highsurrogates & 0x80000000) != 0);
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  return scalar::utf32::utf8_length_from_utf32(input, length);
-}
+    // utf32le checks
+    currentoffsetmax =
+        _mm512_max_epu32(_mm512_add_epi32(data, offset), currentoffsetmax);
+    currentmax = _mm512_max_epu32(data, currentmax);
+  }
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  return scalar::utf32::utf16_length_from_utf32(input, length);
-}
+  // last block with 0 <= len < 64
+  __mmask64 read_mask = (__mmask64(1) << (end - ptr)) - 1;
+  const __m512i data = _mm512_maskz_loadu_epi8(read_mask, (const __m512i *)ptr);
+  checker.check_next_input(data);
 
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf32_length_from_latin1(length);
-}
+  __m512i diff = _mm512_sub_epi16(data, _mm512_set1_epi16(uint16_t(0xD800)));
+  __mmask32 surrogates =
+      _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+  __mmask32 highsurrogates =
+      _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+  __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+  utf16_err |= (((highsurrogates << 1) | ends_with_high) != lowsurrogates);
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return scalar::utf8::count_code_points(input, length);
-}
+  currentoffsetmax =
+      _mm512_max_epu32(_mm512_add_epi32(data, offset), currentoffsetmax);
+  currentmax = _mm512_max_epu32(data, currentmax);
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
+  const __m512i standardmax = _mm512_set1_epi32((uint32_t)0x10ffff);
+  const __m512i standardoffsetmax = _mm512_set1_epi32((uint32_t)0xfffff7ff);
+  __m512i is_zero =
+      _mm512_xor_si512(_mm512_max_epu32(currentmax, standardmax), standardmax);
+  utf32_err |= (_mm512_test_epi8_mask(is_zero, is_zero) != 0);
+  is_zero = _mm512_xor_si512(
+      _mm512_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
+  utf32_err |= (_mm512_test_epi8_mask(is_zero, is_zero) != 0);
+  checker.check_eof();
+  bool is_valid_utf8 = !checker.errors();
+  if (is_valid_utf8) {
+    out |= encoding_type::UTF8;
+  }
+  if (utf16_err == 0) {
+    out |= encoding_type::UTF16_LE;
+  }
+  if (utf32_err == 0) {
+    out |= encoding_type::UTF32_LE;
+  }
+  return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
-    }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
-    }
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return true;
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation};
-    }
-    return {SUCCESS, 0};
+  avx512_utf8_checker checker{};
+  const char *ptr = buf;
+  const char *end = ptr + len;
+  for (; end - ptr >= 64; ptr += 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    checker.check_next_input(utf8);
   }
-  result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation};
-    }
+  if (end != ptr) {
+    const __m512i utf8 = _mm512_maskz_loadu_epi8(
+        ~UINT64_C(0) >> (64 - (end - ptr)), (const __m512i *)ptr);
+    checker.check_next_input(utf8);
   }
-  return r;
+  checker.check_eof();
+  return !checker.errors();
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
-    }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
-    }
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused result implementation::validate_utf8_with_errors(
+    const char *buf, size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return result(error_code::SUCCESS, len);
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+  avx512_utf8_checker checker{};
+  const char *ptr = buf;
+  const char *end = ptr + len;
+  size_t count{0};
+  for (; end - ptr >= 64; ptr += 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
+    checker.check_next_input(utf8);
+    if (checker.errors()) {
+      if (count != 0) {
+        count--;
+      } // Sometimes the error is only detected in the next chunk
+      result res = scalar::utf8::rewind_and_validate_with_errors(
+          reinterpret_cast<const char *>(buf),
+          reinterpret_cast<const char *>(buf + count), len - count);
+      res.count += count;
+      return res;
     }
-    return {SUCCESS, 0, 0};
+    count += 64;
   }
-  full_result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.output_count % 3 == 0) ||
-        ((r.output_count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, r.output_count};
-    }
+  if (end != ptr) {
+    const __m512i utf8 = _mm512_maskz_loadu_epi8(
+        ~UINT64_C(0) >> (64 - (end - ptr)), (const __m512i *)ptr);
+    checker.check_next_input(utf8);
   }
-  return r;
+  checker.check_eof();
+  if (checker.errors()) {
+    if (count != 0) {
+      count--;
+    } // Sometimes the error is only detected in the next chunk
+    result res = scalar::utf8::rewind_and_validate_with_errors(
+        reinterpret_cast<const char *>(buf),
+        reinterpret_cast<const char *>(buf + count), len - count);
+    res.count += count;
+    return res;
+  }
+  return result(error_code::SUCCESS, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
+#if SIMDUTF_FEATURE_ASCII
+simdutf_warn_unused bool
+implementation::validate_ascii(const char *buf, size_t len) const noexcept {
+  return icelake::validate_ascii(buf, len);
 }
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
+simdutf_warn_unused result implementation::validate_ascii_with_errors(
+    const char *buf, size_t len) const noexcept {
+  const char *buf_orig = buf;
+  const char *end = buf + len;
+  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
+  for (; end - buf >= 64; buf += 64) {
+    const __m512i input = _mm512_loadu_si512((const __m512i *)buf);
+    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
+    if (notascii) {
+      return result(error_code::TOO_LARGE,
+                    buf - buf_orig + _tzcnt_u64(notascii));
     }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
+  }
+  if (end != buf) {
+    const __m512i input = _mm512_maskz_loadu_epi8(
+        ~UINT64_C(0) >> (64 - (end - buf)), (const __m512i *)buf);
+    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
+    if (notascii) {
+      return result(error_code::TOO_LARGE,
+                    buf - buf_orig + _tzcnt_u64(notascii));
     }
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0};
+  return result(error_code::SUCCESS, len);
+}
+#endif // SIMDUTF_FEATURE_ASCII
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf16le(const char16_t *buf,
+                                 size_t len) const noexcept {
+  const char16_t *end = buf + len;
+
+  for (; end - buf >= 32;) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        return false;
       }
-      return {INVALID_BASE64_CHARACTER, equallocation};
+      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
+      if (ends_with_high) {
+        buf += 31; // advance only by 31 code units so that we start with the
+                   // high surrogate on the next round.
+      } else {
+        buf += 32;
+      }
+    } else {
+      buf += 32;
     }
-    return {SUCCESS, 0};
   }
-  result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation};
+  if (buf < end) {
+    __m512i in =
+        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        return false;
+      }
     }
   }
-  return r;
+  return true;
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool
+implementation::validate_utf16be(const char16_t *buf,
+                                 size_t len) const noexcept {
+  const char16_t *end = buf + len;
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  for (; end - buf >= 32;) {
+    __m512i in =
+        _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)buf), byteflip);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        return false;
+      }
+      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
+      if (ends_with_high) {
+        buf += 31; // advance only by 31 code units so that we start with the
+                   // high surrogate on the next round.
+      } else {
+        buf += 32;
+      }
+    } else {
+      buf += 32;
     }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
+  }
+  if (buf < end) {
+    __m512i in = _mm512_shuffle_epi8(
+        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf),
+        byteflip);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        return false;
+      }
     }
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0, 0};
+  return true;
+}
+
+simdutf_warn_unused result implementation::validate_utf16le_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  const char16_t *start_buf = buf;
+  const char16_t *end = buf + len;
+  for (; end - buf >= 32;) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
+        uint32_t extra_high =
+            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
+        return result(error_code::SURROGATE,
+                      (buf - start_buf) +
+                          (extra_low < extra_high ? extra_low : extra_high));
       }
-      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
+      if (ends_with_high) {
+        buf += 31; // advance only by 31 code units so that we start with the
+                   // high surrogate on the next round.
+      } else {
+        buf += 32;
+      }
+    } else {
+      buf += 32;
     }
-    return {SUCCESS, 0, 0};
   }
-  full_result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.output_count % 3 == 0) ||
-        ((r.output_count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, r.output_count};
+  if (buf < end) {
+    __m512i in =
+        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
+        uint32_t extra_high =
+            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
+        return result(error_code::SURROGATE,
+                      (buf - start_buf) +
+                          (extra_low < extra_high ? extra_low : extra_high));
+      }
     }
   }
-  return r;
+  return result(error_code::SUCCESS, len);
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
+simdutf_warn_unused result implementation::validate_utf16be_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  const char16_t *start_buf = buf;
+  const char16_t *end = buf + len;
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  for (; end - buf >= 32;) {
+    __m512i in =
+        _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)buf), byteflip);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
+        uint32_t extra_high =
+            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
+        return result(error_code::SURROGATE,
+                      (buf - start_buf) +
+                          (extra_low < extra_high ? extra_low : extra_high));
+      }
+      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
+      if (ends_with_high) {
+        buf += 31; // advance only by 31 code units so that we start with the
+                   // high surrogate on the next round.
+      } else {
+        buf += 32;
+      }
+    } else {
+      buf += 32;
+    }
+  }
+  if (buf < end) {
+    __m512i in = _mm512_shuffle_epi8(
+        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf),
+        byteflip);
+    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
+    __mmask32 surrogates =
+        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
+    if (surrogates) {
+      __mmask32 highsurrogates =
+          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
+      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
+      // high must be followed by low
+      if ((highsurrogates << 1) != lowsurrogates) {
+        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
+        uint32_t extra_high =
+            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
+        return result(error_code::SURROGATE,
+                      (buf - start_buf) +
+                          (extra_low < extra_high ? extra_low : extra_high));
+      }
+    }
+  }
+  return result(error_code::SUCCESS, len);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
-size_t implementation::binary_to_base64(const char *input, size_t length,
-                                        char *output,
-                                        base64_options options) const noexcept {
-  return scalar::base64::tail_encode_base64(output, input, length, options);
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
+  return icelake::validate_utf32(buf, len);
 }
-} // namespace fallback
-} // namespace simdutf
-
-/* begin file src/simdutf/fallback/end.h */
-/* end file src/simdutf/fallback/end.h */
-/* end file src/fallback/implementation.cpp */
-#endif
-#if SIMDUTF_IMPLEMENTATION_ICELAKE
-/* begin file src/icelake/implementation.cpp */
-#include <tuple>
-#include <utility>
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused result implementation::validate_utf32_with_errors(
+    const char32_t *buf, size_t len) const noexcept {
+  const char32_t *buf_orig = buf;
+  if (len >= 16) {
+    const char32_t *end = buf + len - 16;
+    while (buf <= end) {
+      __m512i utf32 = _mm512_loadu_si512((const __m512i *)buf);
+      __mmask16 outside_range = _mm512_cmp_epu32_mask(
+          utf32, _mm512_set1_epi32(0x10ffff), _MM_CMPINT_GT);
 
-/* begin file src/simdutf/icelake/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "icelake"
-// #define SIMDUTF_IMPLEMENTATION icelake
-
-#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-// nothing needed.
-#else
-SIMDUTF_TARGET_ICELAKE
-#endif
+      __m512i utf32_off =
+          _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));
 
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-// clang-format off
-SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
-// clang-format on
-#endif // end of workaround
-/* end file src/simdutf/icelake/begin.h */
-namespace simdutf {
-namespace icelake {
-namespace {
-#ifndef SIMDUTF_ICELAKE_H
-  #error "icelake.h must be included"
-#endif
-/* begin file src/icelake/icelake_utf8_common.inl.cpp */
-// Common procedures for both validating and non-validating conversions from
-// UTF-8.
-enum block_processing_mode { SIMDUTF_FULL, SIMDUTF_TAIL };
+      __mmask16 surrogate_range = _mm512_cmp_epu32_mask(
+          utf32_off, _mm512_set1_epi32(0xfffff7ff), _MM_CMPINT_GT);
+      if ((outside_range | surrogate_range)) {
+        auto outside_idx = _tzcnt_u32(outside_range);
+        auto surrogate_idx = _tzcnt_u32(surrogate_range);
 
-using utf8_to_utf16_result = std::pair<const char *, char16_t *>;
-using utf8_to_utf32_result = std::pair<const char *, uint32_t *>;
+        if (outside_idx < surrogate_idx) {
+          return result(error_code::TOO_LARGE, buf - buf_orig + outside_idx);
+        }
 
-/*
-    process_block_utf8_to_utf16 converts up to 64 bytes from 'in' from UTF-8
-    to UTF-16. When tail = SIMDUTF_FULL, then the full input buffer (64 bytes)
-    might be used. When tail = SIMDUTF_TAIL, we take into account 'gap' which
-    indicates how many input bytes are relevant.
+        return result(error_code::SURROGATE, buf - buf_orig + surrogate_idx);
+      }
 
-    Returns true when the result is correct, otherwise it returns false.
+      buf += 16;
+    }
+  }
+  if (len > 0) {
+    __m512i utf32 = _mm512_maskz_loadu_epi32(
+        __mmask16((1U << (buf_orig + len - buf)) - 1), (const __m512i *)buf);
+    __mmask16 outside_range = _mm512_cmp_epu32_mask(
+        utf32, _mm512_set1_epi32(0x10ffff), _MM_CMPINT_GT);
+    __m512i utf32_off = _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));
 
-    The provided in and out pointers are advanced according to how many input
-    bytes have been processed, upon success.
-*/
-template <block_processing_mode tail, endianness big_endian>
-simdutf_really_inline bool
-process_block_utf8_to_utf16(const char *&in, char16_t *&out, size_t gap) {
-  // constants
-  __m512i mask_identity = _mm512_set_epi8(
-      63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46,
-      45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28,
-      27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9,
-      8, 7, 6, 5, 4, 3, 2, 1, 0);
-  __m512i mask_c0c0c0c0 = _mm512_set1_epi32(0xc0c0c0c0);
-  __m512i mask_80808080 = _mm512_set1_epi32(0x80808080);
-  __m512i mask_f0f0f0f0 = _mm512_set1_epi32(0xf0f0f0f0);
-  __m512i mask_dfdfdfdf_tail = _mm512_set_epi64(
-      0xffffdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf,
-      0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf,
-      0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf);
-  __m512i mask_c2c2c2c2 = _mm512_set1_epi32(0xc2c2c2c2);
-  __m512i mask_ffffffff = _mm512_set1_epi32(0xffffffff);
-  __m512i mask_d7c0d7c0 = _mm512_set1_epi32(0xd7c0d7c0);
-  __m512i mask_dc00dc00 = _mm512_set1_epi32(0xdc00dc00);
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  // Note that 'tail' is a compile-time constant !
-  __mmask64 b =
-      (tail == SIMDUTF_FULL) ? 0xFFFFFFFFFFFFFFFF : (uint64_t(1) << gap) - 1;
-  __m512i input = (tail == SIMDUTF_FULL) ? _mm512_loadu_si512(in)
-                                         : _mm512_maskz_loadu_epi8(b, in);
-  __mmask64 m1 = (tail == SIMDUTF_FULL)
-                     ? _mm512_cmplt_epu8_mask(input, mask_80808080)
-                     : _mm512_mask_cmplt_epu8_mask(b, input, mask_80808080);
-  if (_ktestc_mask64_u8(m1,
-                        b)) { // NOT(m1) AND b -- if all zeroes, then all ASCII
-                              // alternatively, we could do 'if (m1 == b) { '
-    if (tail == SIMDUTF_FULL) {
-      in += 64; // consumed 64 bytes
-      // we convert a full 64-byte block, writing 128 bytes.
-      __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
-      if (big_endian) {
-        input1 = _mm512_shuffle_epi8(input1, byteflip);
-      }
-      _mm512_storeu_si512(out, input1);
-      out += 32;
-      __m512i input2 =
-          _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
-      if (big_endian) {
-        input2 = _mm512_shuffle_epi8(input2, byteflip);
-      }
-      _mm512_storeu_si512(out, input2);
-      out += 32;
-      return true; // we are done
-    } else {
-      in += gap;
-      if (gap <= 32) {
-        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
-        if (big_endian) {
-          input1 = _mm512_shuffle_epi8(input1, byteflip);
-        }
-        _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << (gap)) - 1),
-                                 input1);
-        out += gap;
-      } else {
-        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
-        if (big_endian) {
-          input1 = _mm512_shuffle_epi8(input1, byteflip);
-        }
-        _mm512_storeu_si512(out, input1);
-        out += 32;
-        __m512i input2 =
-            _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
-        if (big_endian) {
-          input2 = _mm512_shuffle_epi8(input2, byteflip);
-        }
-        _mm512_mask_storeu_epi16(
-            out, __mmask32((uint32_t(1) << (gap - 32)) - 1), input2);
-        out += gap - 32;
+    __mmask16 surrogate_range = _mm512_cmp_epu32_mask(
+        utf32_off, _mm512_set1_epi32(0xfffff7ff), _MM_CMPINT_GT);
+    if ((outside_range | surrogate_range)) {
+      auto outside_idx = _tzcnt_u32(outside_range);
+      auto surrogate_idx = _tzcnt_u32(surrogate_range);
+
+      if (outside_idx < surrogate_idx) {
+        return result(error_code::TOO_LARGE, buf - buf_orig + outside_idx);
       }
-      return true; // we are done
+
+      return result(error_code::SURROGATE, buf - buf_orig + surrogate_idx);
     }
   }
-  // classify characters further
-  __mmask64 m234 = _mm512_cmp_epu8_mask(
-      mask_c0c0c0c0, input,
-      _MM_CMPINT_LE); // 0xc0 <= input, 2, 3, or 4 leading byte
-  __mmask64 m34 =
-      _mm512_cmp_epu8_mask(mask_dfdfdfdf_tail, input,
-                           _MM_CMPINT_LT); // 0xdf < input,  3 or 4 leading byte
 
-  __mmask64 milltwobytes = _mm512_mask_cmp_epu8_mask(
-      m234, input, mask_c2c2c2c2,
-      _MM_CMPINT_LT); // 0xc0 <= input < 0xc2 (illegal two byte sequence)
-                      // Overlong 2-byte sequence
-  if (_ktestz_mask64_u8(milltwobytes, milltwobytes) == 0) {
-    // Overlong 2-byte sequence
-    return false;
+  return result(error_code::SUCCESS, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
+    const char *buf, size_t len, char *utf8_output) const noexcept {
+  return icelake::latin1_to_utf8_avx512_start(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return icelake_convert_latin1_to_utf16<endianness::LITTLE>(buf, len,
+                                                             utf16_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return icelake_convert_latin1_to_utf16<endianness::BIG>(buf, len,
+                                                          utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  avx512_convert_latin1_to_utf32(buf, len, utf32_output);
+  return len;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake::utf8_to_latin1_avx512(buf, len, latin1_output);
+}
+
+simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  // First, try to convert as much as possible using the SIMD implementation.
+  const char *obuf = buf;
+  char *olatin1_output = latin1_output;
+  size_t written = icelake::utf8_to_latin1_avx512(obuf, len, olatin1_output);
+
+  // If we have completely converted the string
+  if (obuf == buf + len) {
+    return {simdutf::SUCCESS, written};
   }
-  if (_ktestz_mask64_u8(m34, m34) == 0) {
-    // We have a 3-byte sequence and/or a 2-byte sequence, or possibly even a
-    // 4-byte sequence!
-    __mmask64 m4 = _mm512_cmp_epu8_mask(
-        input, mask_f0f0f0f0,
-        _MM_CMPINT_NLT); // 0xf0 <= zmm0 (4 byte start bytes)
+  size_t pos = obuf - buf;
+  result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+      pos, buf + pos, len - pos, latin1_output);
+  res.count += pos;
+  return res;
+}
 
-    __mmask64 mask_not_ascii = (tail == SIMDUTF_FULL)
-                                   ? _knot_mask64(m1)
-                                   : _kand_mask64(_knot_mask64(m1), b);
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake::valid_utf8_to_latin1_avx512(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-    __mmask64 mp1 = _kshiftli_mask64(m234, 1);
-    __mmask64 mp2 = _kshiftli_mask64(m34, 2);
-    // We could do it as follows...
-    // if (_kortestz_mask64_u8(m4,m4)) { // compute the bitwise OR of the 64-bit
-    // masks a and b and return 1 if all zeroes but GCC generates better code
-    // when we do:
-    if (m4 == 0) { // compute the bitwise OR of the 64-bit masks a and b and
-                   // return 1 if all zeroes
-      // Fast path with 1,2,3 bytes
-      __mmask64 mc = _kor_mask64(mp1, mp2); // expected continuation bytes
-      __mmask64 m1234 = _kor_mask64(m1, m234);
-      // mismatched continuation bytes:
-      if (tail == SIMDUTF_FULL) {
-        __mmask64 xnormcm1234 = _kxnor_mask64(
-            mc,
-            m1234); // XNOR of mc and m1234 should be all zero if they differ
-        // the presence of a 1 bit indicates that they overlap.
-        // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return
-        // 1 if all zeroes.
-        if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) {
-          return false;
-        }
-      } else {
-        __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
-        if (mc != bxorm1234) {
-          return false;
-        }
-      }
-      // mend: identifying the last bytes of each sequence to be decoded
-      __mmask64 mend = _kshiftri_mask64(m1234, 1);
-      if (tail != SIMDUTF_FULL) {
-        mend = _kor_mask64(mend, (uint64_t(1) << (gap - 1)));
-      }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16_result ret =
+      fast_avx512_convert_utf8_to_utf16<endianness::LITTLE>(buf, len,
+                                                            utf16_output);
+  if (ret.second == nullptr) {
+    return 0;
+  }
+  return ret.second - utf16_output;
+}
 
-      __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
-      __m512i last_and_thirdu16 =
-          _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16_result ret = fast_avx512_convert_utf8_to_utf16<endianness::BIG>(
+      buf, len, utf16_output);
+  if (ret.second == nullptr) {
+    return 0;
+  }
+  return ret.second - utf16_output;
+}
 
-      __m512i nonasciitags = _mm512_maskz_mov_epi8(
-          mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
-      __m512i clearedbytes = _mm512_andnot_si512(
-          nonasciitags, input); // high two bits cleared where not ASCII
-      __m512i lastbytes = _mm512_maskz_permutexvar_epi8(
-          0x5555555555555555, last_and_thirdu16,
-          clearedbytes); // the last byte of each character
+simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::LITTLE>(
+      buf, len, utf16_output);
+}
 
-      __mmask64 mask_before_non_ascii = _kshiftri_mask64(
-          mask_not_ascii, 1); // bytes that precede non-ASCII bytes
-      __m512i indexofsecondlastbytes = _mm512_add_epi16(
-          mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
-      __m512i beforeasciibytes =
-          _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
-      __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(
-          0x5555555555555555, indexofsecondlastbytes,
-          beforeasciibytes); // the second last bytes (of two, three byte seq,
-                             // surrogates)
-      secondlastbytes =
-          _mm512_slli_epi16(secondlastbytes, 6); // shifted into position
+simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::BIG>(
+      buf, len, utf16_output);
+}
 
-      __m512i indexofthirdlastbytes = _mm512_add_epi16(
-          mask_ffffffff,
-          indexofsecondlastbytes); // indices of the second last bytes
-      __m512i thirdlastbyte =
-          _mm512_maskz_mov_epi8(m34,
-                                clearedbytes); // only those that are the third
-                                               // last byte of a sequence
-      __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(
-          0x5555555555555555, indexofthirdlastbytes,
-          thirdlastbyte); // the third last bytes (of three byte sequences, hi
-                          // surrogate)
-      thirdlastbytes =
-          _mm512_slli_epi16(thirdlastbytes, 12); // shifted into position
-      __m512i Wout = _mm512_ternarylogic_epi32(lastbytes, secondlastbytes,
-                                               thirdlastbytes, 254);
-      // the elements of Wout excluding the last element if it happens to be a
-      // high surrogate:
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16_result ret =
+      icelake::valid_utf8_to_fixed_length<endianness::LITTLE, char16_t>(
+          buf, len, utf16_output);
+  size_t saved_bytes = ret.second - utf16_output;
+  const char *end = buf + len;
+  if (ret.first == end) {
+    return saved_bytes;
+  }
 
-      __mmask64 mprocessed =
-          (tail == SIMDUTF_FULL)
-              ? _pdep_u64(0xFFFFFFFF, mend)
-              : _pdep_u64(
-                    0xFFFFFFFF,
-                    _kand_mask64(
-                        mend, b)); // we adjust mend at the end of the output.
+  // Note: AVX512 procedure looks up 4 bytes forward, and
+  //       correctly converts multi-byte chars even if their
+  //       continuation bytes lie outsiede 16-byte window.
+  //       It meas, we have to skip continuation bytes from
+  //       the beginning ret.first, as they were already consumed.
+  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
+    ret.first += 1;
+  }
 
-      // Encodings out of range...
-      {
-        // the location of 3-byte sequence start bytes in the input
-        __mmask64 m3 = m34 & (b ^ m4);
-        // code units in Wout corresponding to 3-byte sequences.
-        __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
-        __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
-        __mmask32 Msmall800 =
-            _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
-        __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
-        __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
-        __mmask32 M3s =
-            _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
-        if (_kor_mask32(Msmall800, M3s)) {
-          return false;
-        }
-      }
-      int64_t nout = _mm_popcnt_u64(mprocessed);
-      in += 64 - _lzcnt_u64(mprocessed);
-      if (big_endian) {
-        Wout = _mm512_shuffle_epi8(Wout, byteflip);
-      }
-      _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
-      out += nout;
-      return true; // ok
+  if (ret.first != end) {
+    const size_t scalar_saved_bytes =
+        scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    //
-    // We have a 4-byte sequence, this is the general case.
-    // Slow!
-    __mmask64 mp3 = _kshiftli_mask64(m4, 3);
-    __mmask64 mc =
-        _kor_mask64(_kor_mask64(mp1, mp2), mp3); // expected continuation bytes
-    __mmask64 m1234 = _kor_mask64(m1, m234);
+    saved_bytes += scalar_saved_bytes;
+  }
 
-    // mend: identifying the last bytes of each sequence to be decoded
-    __mmask64 mend =
-        _kor_mask64(_kshiftri_mask64(_kor_mask64(mp3, m1234), 1), mp3);
-    if (tail != SIMDUTF_FULL) {
-      mend = _kor_mask64(mend, __mmask64(uint64_t(1) << (gap - 1)));
+  return saved_bytes;
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16_result ret =
+      icelake::valid_utf8_to_fixed_length<endianness::BIG, char16_t>(
+          buf, len, utf16_output);
+  size_t saved_bytes = ret.second - utf16_output;
+  const char *end = buf + len;
+  if (ret.first == end) {
+    return saved_bytes;
+  }
+
+  // Note: AVX512 procedure looks up 4 bytes forward, and
+  //       correctly converts multi-byte chars even if their
+  //       continuation bytes lie outsiede 16-byte window.
+  //       It meas, we have to skip continuation bytes from
+  //       the beginning ret.first, as they were already consumed.
+  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
+    ret.first += 1;
+  }
+
+  if (ret.first != end) {
+    const size_t scalar_saved_bytes =
+        scalar::utf8_to_utf16::convert_valid<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
-    __m512i last_and_thirdu16 =
-        _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));
+    saved_bytes += scalar_saved_bytes;
+  }
 
-    __m512i nonasciitags = _mm512_maskz_mov_epi8(
-        mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
-    __m512i clearedbytes = _mm512_andnot_si512(
-        nonasciitags, input); // high two bits cleared where not ASCII
-    __m512i lastbytes = _mm512_maskz_permutexvar_epi8(
-        0x5555555555555555, last_and_thirdu16,
-        clearedbytes); // the last byte of each character
+  return saved_bytes;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    __mmask64 mask_before_non_ascii = _kshiftri_mask64(
-        mask_not_ascii, 1); // bytes that precede non-ASCII bytes
-    __m512i indexofsecondlastbytes = _mm512_add_epi16(
-        mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
-    __m512i beforeasciibytes =
-        _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
-    __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(
-        0x5555555555555555, indexofsecondlastbytes,
-        beforeasciibytes); // the second last bytes (of two, three byte seq,
-                           // surrogates)
-    secondlastbytes =
-        _mm512_slli_epi16(secondlastbytes, 6); // shifted into position
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_out) const noexcept {
+  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
+  utf8_to_utf32_result ret =
+      icelake::validating_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(
+          buf, len, utf32_output);
+  if (ret.second == nullptr)
+    return 0;
 
-    __m512i indexofthirdlastbytes = _mm512_add_epi16(
-        mask_ffffffff,
-        indexofsecondlastbytes); // indices of the second last bytes
-    __m512i thirdlastbyte = _mm512_maskz_mov_epi8(
-        m34,
-        clearedbytes); // only those that are the third last byte of a sequence
-    __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(
-        0x5555555555555555, indexofthirdlastbytes,
-        thirdlastbyte); // the third last bytes (of three byte sequences, hi
-                        // surrogate)
-    thirdlastbytes =
-        _mm512_slli_epi16(thirdlastbytes, 12); // shifted into position
-    __m512i thirdsecondandlastbytes = _mm512_ternarylogic_epi32(
-        lastbytes, secondlastbytes, thirdlastbytes, 254);
-    uint64_t Mlo_uint64 = _pext_u64(mp3, mend);
-    __mmask32 Mlo = __mmask32(Mlo_uint64);
-    __mmask32 Mhi = __mmask32(Mlo_uint64 >> 1);
-    __m512i lo_surr_mask = _mm512_maskz_mov_epi16(
-        Mlo,
-        mask_dc00dc00); // lo surr: 1101110000000000, other:  0000000000000000
-    __m512i shifted4_thirdsecondandlastbytes =
-        _mm512_srli_epi16(thirdsecondandlastbytes,
-                          4); // hi surr: 00000WVUTSRQPNML  vuts = WVUTS - 1
-    __m512i tagged_lo_surrogates = _mm512_or_si512(
-        thirdsecondandlastbytes,
-        lo_surr_mask); // lo surr: 110111KJHGFEDCBA, other:  unchanged
-    __m512i Wout = _mm512_mask_add_epi16(
-        tagged_lo_surrogates, Mhi, shifted4_thirdsecondandlastbytes,
-        mask_d7c0d7c0); // hi sur: 110110vutsRQPNML, other:  unchanged
-    // the elements of Wout excluding the last element if it happens to be a
-    // high surrogate:
-    __mmask32 Mout = ~(Mhi & 0x80000000);
-    __mmask64 mprocessed =
-        (tail == SIMDUTF_FULL)
-            ? _pdep_u64(Mout, mend)
-            : _pdep_u64(
-                  Mout,
-                  _kand_mask64(mend,
-                               b)); // we adjust mend at the end of the output.
+  size_t saved_bytes = ret.second - utf32_output;
+  const char *end = buf + len;
+  if (ret.first == end) {
+    return saved_bytes;
+  }
 
-    // mismatched continuation bytes:
-    if (tail == SIMDUTF_FULL) {
-      __mmask64 xnormcm1234 = _kxnor_mask64(
-          mc, m1234); // XNOR of mc and m1234 should be all zero if they differ
-      // the presence of a 1 bit indicates that they overlap.
-      // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return 1
-      // if all zeroes.
-      if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) {
-        return false;
-      }
-    } else {
-      __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
-      if (mc != bxorm1234) {
-        return false;
-      }
+  // Note: the AVX512 procedure looks up 4 bytes forward, and
+  //       correctly converts multi-byte chars even if their
+  //       continuation bytes lie outside 16-byte window.
+  //       It means, we have to skip continuation bytes from
+  //       the beginning ret.first, as they were already consumed.
+  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
+    ret.first += 1;
+  }
+  if (ret.first != end) {
+    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert(
+        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
-    // Encodings out of range...
-    {
-      // the location of 3-byte sequence start bytes in the input
-      __mmask64 m3 = m34 & (b ^ m4);
-      // code units in Wout corresponding to 3-byte sequences.
-      __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
-      __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
-      __mmask32 Msmall800 =
-          _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
-      __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
-      __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
-      __mmask32 M3s =
-          _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
-      __m512i mask_04000400 = _mm512_set1_epi32(0x04000400);
-      __mmask32 M4s =
-          _mm512_mask_cmpge_epu16_mask(Mhi, Moutminusd800, mask_04000400);
-      if (!_kortestz_mask32_u8(M4s, _kor_mask32(Msmall800, M3s))) {
-        return false;
+    saved_bytes += scalar_saved_bytes;
+  }
+
+  return saved_bytes;
+}
+
+simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
+    const char *buf, size_t len, char32_t *utf32) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    return {error_code::SUCCESS, 0};
+  }
+  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32);
+  auto ret = icelake::validating_utf8_to_fixed_length_with_constant_checks<
+      endianness::LITTLE, uint32_t>(buf, len, utf32_output);
+
+  if (!std::get<2>(ret)) {
+    size_t pos = std::get<0>(ret) - buf;
+    // We might have an error that occurs right before  pos.
+    // This is only a concern if buf[pos] is not a continuation byte.
+    if ((buf[pos] & 0xc0) != 0x80 && pos >= 64) {
+      pos -= 1;
+    } else if ((buf[pos] & 0xc0) == 0x80 && pos >= 64) {
+      // We must check whether we are the fourth continuation byte
+      bool c1 = (buf[pos - 1] & 0xc0) == 0x80;
+      bool c2 = (buf[pos - 2] & 0xc0) == 0x80;
+      bool c3 = (buf[pos - 3] & 0xc0) == 0x80;
+      if (c1 && c2 && c3) {
+        return {simdutf::TOO_LONG, pos};
       }
     }
-    in += 64 - _lzcnt_u64(mprocessed);
-    int64_t nout = _mm_popcnt_u64(mprocessed);
-    if (big_endian) {
-      Wout = _mm512_shuffle_epi8(Wout, byteflip);
-    }
-    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
-    out += nout;
-    return true; // ok
+    // todo: we reset the output to utf32 instead of using std::get<2.(ret) as
+    // you'd expect. that is because
+    // validating_utf8_to_fixed_length_with_constant_checks may have processed
+    // data beyond the error.
+    result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+        pos, buf + pos, len - pos, utf32);
+    res.count += pos;
+    return res;
   }
-  // Fast path 2: all ASCII or 2 byte
-  __mmask64 continuation_or_ascii = (tail == SIMDUTF_FULL)
-                                        ? _knot_mask64(m234)
-                                        : _kand_mask64(_knot_mask64(m234), b);
-  // on top of -0xc0 we subtract -2 which we get back later of the
-  // continuation byte tags
-  __m512i leading2byte = _mm512_maskz_sub_epi8(m234, input, mask_c2c2c2c2);
-  __mmask64 leading = tail == (tail == SIMDUTF_FULL)
-                          ? _kor_mask64(m1, m234)
-                          : _kand_mask64(_kor_mask64(m1, m234),
-                                         b); // first bytes of each sequence
-  if (tail == SIMDUTF_FULL) {
-    __mmask64 xnor234leading =
-        _kxnor_mask64(_kshiftli_mask64(m234, 1), leading);
-    if (!_kortestz_mask64_u8(xnor234leading, xnor234leading)) {
-      return false;
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  const char *end = buf + len;
+  if (std::get<0>(ret) == end) {
+    return {simdutf::SUCCESS, saved_bytes};
+  }
+
+  // Note: the AVX512 procedure looks up 4 bytes forward, and
+  //       correctly converts multi-byte chars even if their
+  //       continuation bytes lie outside 16-byte window.
+  //       It means, we have to skip continuation bytes from
+  //       the beginning ret.first, as they were already consumed.
+  while (std::get<0>(ret) != end and
+         ((uint8_t(*std::get<0>(ret)) & 0xc0) == 0x80)) {
+    std::get<0>(ret) += 1;
+  }
+
+  if (std::get<0>(ret) != end) {
+    auto scalar_result = scalar::utf8_to_utf32::convert_with_errors(
+        std::get<0>(ret), len - (std::get<0>(ret) - buf),
+        reinterpret_cast<char32_t *>(utf32_output) + saved_bytes);
+    if (scalar_result.error != simdutf::SUCCESS) {
+      scalar_result.count += (std::get<0>(ret) - buf);
+    } else {
+      scalar_result.count += saved_bytes;
     }
-  } else {
-    __mmask64 bxorleading = _kxor_mask64(b, leading);
-    if (_kshiftli_mask64(m234, 1) != bxorleading) {
-      return false;
+    return scalar_result;
+  }
+
+  return {simdutf::SUCCESS, size_t(std::get<1>(ret) - utf32_output)};
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_out) const noexcept {
+  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
+  utf8_to_utf32_result ret =
+      icelake::valid_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(
+          buf, len, utf32_output);
+  size_t saved_bytes = ret.second - utf32_output;
+  const char *end = buf + len;
+  if (ret.first == end) {
+    return saved_bytes;
+  }
+
+  // Note: AVX512 procedure looks up 4 bytes forward, and
+  //       correctly converts multi-byte chars even if their
+  //       continuation bytes lie outsiede 16-byte window.
+  //       It meas, we have to skip continuation bytes from
+  //       the beginning ret.first, as they were already consumed.
+  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
+    ret.first += 1;
+  }
+
+  if (ret.first != end) {
+    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert_valid(
+        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
+    saved_bytes += scalar_saved_bytes;
   }
-  //
-  if (tail == SIMDUTF_FULL) {
-    // In the two-byte/ASCII scenario, we are easily latency bound, so we want
-    // to increment the input buffer as quickly as possible.
-    // We process 32 bytes unless the byte at index 32 is a continuation byte,
-    // in which case we include it as well for a total of 33 bytes.
-    // Note that if x is an ASCII byte, then the following is false:
-    // int8_t(x) <= int8_t(0xc0) under two's complement.
-    in += 32;
-    if (int8_t(*in) <= int8_t(0xc0))
-      in++;
-    // The alternative is to do
-    // in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
-    // but it requires loading the input, doing the mask computation, and
-    // converting back the mask to a general register. It just takes too long,
-    // leaving the processor likely to be idle.
-  } else {
-    in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
+
+  return saved_bytes;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf16_to_latin1<endianness::LITTLE>(buf, len,
+                                                             latin1_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf16_to_latin1<endianness::BIG>(buf, len,
+                                                          latin1_output);
+}
+
+simdutf_warn_unused result
+implementation::convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
+             buf, len, latin1_output)
+      .first;
+}
+
+simdutf_warn_unused result
+implementation::convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf16_to_latin1_with_errors<endianness::BIG>(
+             buf, len, latin1_output)
+      .first;
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement custom function
+  return convert_utf16be_to_latin1(buf, len, latin1_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement custom function
+  return convert_utf16le_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  size_t outlen;
+  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(
+      buf, len, (unsigned char *)utf8_output, &outlen);
+  if (inlen != len) {
+    return 0;
   }
-  __m512i lead = _mm512_maskz_compress_epi8(
-      leading, leading2byte); // will contain zero for ascii, and the data
-  lead = _mm512_cvtepu8_epi16(
-      _mm512_castsi512_si256(lead)); // ... zero extended into code units
-  __m512i follow = _mm512_maskz_compress_epi8(
-      continuation_or_ascii, input); // the last bytes of each sequence
-  follow = _mm512_cvtepu8_epi16(
-      _mm512_castsi512_si256(follow)); // ... zero extended into code units
-  lead = _mm512_slli_epi16(lead, 6);   // shifted into position
-  __m512i final = _mm512_add_epi16(follow, lead); // combining lead and follow
+  return outlen;
+}
 
-  if (big_endian) {
-    final = _mm512_shuffle_epi8(final, byteflip);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  size_t outlen;
+  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(
+      buf, len, (unsigned char *)utf8_output, &outlen);
+  if (inlen != len) {
+    return 0;
   }
-  if (tail == SIMDUTF_FULL) {
-    // Next part is UTF-16 specific and can be generalized to UTF-32.
-    int nout = _mm_popcnt_u32(uint32_t(leading));
-    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
-    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
-  } else {
-    int nout = int(_mm_popcnt_u64(_pdep_u64(0xFFFFFFFF, leading)));
-    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
-    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
+  return outlen;
+}
+
+simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  size_t outlen;
+  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(
+      buf, len, (unsigned char *)utf8_output, &outlen);
+  if (inlen != len) {
+    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
+        buf + inlen, len - inlen, utf8_output + outlen);
+    res.count += inlen;
+    return res;
+  }
+  return {simdutf::SUCCESS, outlen};
+}
+
+simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  size_t outlen;
+  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(
+      buf, len, (unsigned char *)utf8_output, &outlen);
+  if (inlen != len) {
+    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
+        buf + inlen, len - inlen, utf8_output + outlen);
+    res.count += inlen;
+    return res;
   }
+  return {simdutf::SUCCESS, outlen};
+}
 
-  return true; // we are fine.
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16le_to_utf8(buf, len, utf8_output);
 }
 
-/*
-    utf32_to_utf16_masked converts `count` lower UTF-32 code units
-    from input `utf32` into UTF-16. It differs from utf32_to_utf16
-    in that it 'masks' the writes.
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16be_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    Returns how many 16-bit code units were stored.
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf32_to_latin1(buf, len, latin1_output);
+}
 
-    byteflip is used for flipping 16-bit code units, and it should be
-        __m512i byteflip = _mm512_setr_epi64(
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809
-        );
-    We pass it to the (always inlined) function to encourage the compiler to
-    keep the value in a (constant) register.
-*/
-template <endianness big_endian>
-simdutf_really_inline size_t utf32_to_utf16_masked(const __m512i byteflip,
-                                                   __m512i utf32,
-                                                   unsigned int count,
-                                                   char16_t *output) {
+simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf32_to_latin1_with_errors(buf, len, latin1_output)
+      .first;
+}
 
-  const __mmask16 valid = uint16_t((1 << count) - 1);
-  // 1. check if we have any surrogate pairs
-  const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
-  const __mmask16 sp_mask =
-      _mm512_mask_cmpgt_epu32_mask(valid, utf32, v_0000_ffff);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return icelake_convert_utf32_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-  if (sp_mask == 0) {
-    if (big_endian) {
-      _mm256_mask_storeu_epi16(
-          (__m256i *)output, valid,
-          _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32),
-                              _mm512_castsi512_si256(byteflip)));
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char32_t *, char *> ret =
+      avx512_convert_utf32_to_utf8(buf, len, utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
+simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      icelake::avx512_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
+  if (ret.first.count != len) {
+    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
+        buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
     } else {
-      _mm256_mask_storeu_epi16((__m256i *)output, valid,
-                               _mm512_cvtepi32_epi16(utf32));
+      ret.second += scalar_res.count;
     }
-    return count;
   }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  {
-    // build surrogate pair code units in 32-bit lanes
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf32_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-    //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
-    const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
-    const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      avx512_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  return saved_bytes;
+}
 
-    //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
-    const __m512i t1 = _mm512_slli_epi32(t0, 6);
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      avx512_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  return saved_bytes;
+}
 
-    //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1
-    //    to t0
-    //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
-    const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
-    const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);
+simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      avx512_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(
+          buf, len, utf16_output);
+  if (ret.first.error) {
+    return ret.first;
+  }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-    //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1
-    //    to t0
-    //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
-    const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
-    const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
-    const __m512i t3 =
-        _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
-    const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
-    __m512i t5 = _mm512_ror_epi32(t4, 16);
-    // Here we want to trim all of the upper 16-bit code units from the 2-byte
-    // characters represented as 4-byte values. We can compute it from
-    // sp_mask or the following... It can be more optimized!
-    const __mmask32 nonzero = _kor_mask32(
-        0xaaaaaaaa, _mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
-    const __mmask32 nonzero_masked =
-        _kand_mask32(nonzero, __mmask32((uint64_t(1) << (2 * count)) - 1));
-    if (big_endian) {
-      t5 = _mm512_shuffle_epi8(t5, byteflip);
-    }
-    // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
-    // (zen4)
-    __m512i compressed = _mm512_maskz_compress_epi16(nonzero_masked, t5);
-    _mm512_mask_storeu_epi16(
-        output,
-        (1 << (count + static_cast<unsigned int>(count_ones(sp_mask)))) - 1,
-        compressed);
-    //_mm512_mask_compressstoreu_epi16(output, nonzero_masked, t5);
+simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      avx512_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len,
+                                                                 utf16_output);
+  if (ret.first.error) {
+    return ret.first;
   }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  return count + static_cast<unsigned int>(count_ones(sp_mask));
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16le(buf, len, utf16_output);
 }
 
-/*
-    utf32_to_utf16 converts `count` lower UTF-32 code units
-    from input `utf32` into UTF-16. It may overflow.
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16be(buf, len, utf16_output);
+}
 
-    Returns how many 16-bit code units were stored.
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
+                                                          utf32_output);
+  if (!std::get<2>(ret)) {
+    return 0;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    byteflip is used for flipping 16-bit code units, and it should be
-        __m512i byteflip = _mm512_setr_epi64(
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809,
-            0x0607040502030001,
-            0x0e0f0c0d0a0b0809
-        );
-    We pass it to the (always inlined) function to encourage the compiler to
-    keep the value in a (constant) register.
-*/
-template <endianness big_endian>
-simdutf_really_inline size_t utf32_to_utf16(const __m512i byteflip,
-                                            __m512i utf32, unsigned int count,
-                                            char16_t *output) {
-  // check if we have any surrogate pairs
-  const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
-  const __mmask16 sp_mask = _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
+  if (!std::get<2>(ret)) {
+    return 0;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::BIG>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-  if (sp_mask == 0) {
-    // technically, it should be _mm256_storeu_epi16
-    if (big_endian) {
-      _mm256_storeu_si256(
-          (__m256i *)output,
-          _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32),
-                              _mm512_castsi512_si256(byteflip)));
+simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
+                                                          utf32_output);
+  if (!std::get<2>(ret)) {
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    scalar_res.count += (std::get<0>(ret) - buf);
+    return scalar_res;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_res.error) {
+      scalar_res.count += (std::get<0>(ret) - buf);
+      return scalar_res;
     } else {
-      _mm256_storeu_si256((__m256i *)output, _mm512_cvtepi32_epi16(utf32));
+      scalar_res.count += saved_bytes;
+      return scalar_res;
     }
-    return count;
   }
+  return simdutf::result(simdutf::SUCCESS, saved_bytes);
+}
 
-  {
-    // build surrogate pair code units in 32-bit lanes
-
-    //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
-    const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
-    const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);
-
-    //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
-    const __m512i t1 = _mm512_slli_epi32(t0, 6);
-
-    //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1
-    //    to t0
-    //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
-    const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
-    const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);
-
-    //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1
-    //    to t0
-    //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
-    const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
-    const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
-    const __m512i t3 =
-        _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
-    const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
-    __m512i t5 = _mm512_ror_epi32(t4, 16);
-    const __mmask32 nonzero = _kor_mask32(
-        0xaaaaaaaa, _mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
-    if (big_endian) {
-      t5 = _mm512_shuffle_epi8(t5, byteflip);
+simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
+  if (!std::get<2>(ret)) {
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    scalar_res.count += (std::get<0>(ret) - buf);
+    return scalar_res;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_res.error) {
+      scalar_res.count += (std::get<0>(ret) - buf);
+      return scalar_res;
+    } else {
+      scalar_res.count += saved_bytes;
+      return scalar_res;
     }
-    // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability
-    // (zen4)
-    __m512i compressed = _mm512_maskz_compress_epi16(nonzero, t5);
-    _mm512_mask_storeu_epi16(
-        output,
-        (1 << (count + static_cast<unsigned int>(count_ones(sp_mask)))) - 1,
-        compressed);
-    //_mm512_mask_compressstoreu_epi16(output, nonzero, t5);
   }
-
-  return count + static_cast<unsigned int>(count_ones(sp_mask));
+  return simdutf::result(simdutf::SUCCESS, saved_bytes);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
+                                                          utf32_output);
+  if (!std::get<2>(ret)) {
+    return 0;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
 }
 
-/**
- * Store the last N bytes of previous followed by 512-N bytes from input.
- */
-template <int N> __m512i prev(__m512i input, __m512i previous) {
-  static_assert(N <= 32, "N must be no larger than 32");
-  const __m512i movemask =
-      _mm512_setr_epi32(28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11);
-  const __m512i rotated = _mm512_permutex2var_epi32(input, movemask, previous);
-#if SIMDUTF_GCC8 || SIMDUTF_GCC9
-  constexpr int shift = 16 - N; // workaround for GCC8,9
-  return _mm512_alignr_epi8(input, rotated, shift);
-#else
-  return _mm512_alignr_epi8(input, rotated, 16 - N);
-#endif // SIMDUTF_GCC8 || SIMDUTF_GCC9
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::tuple<const char16_t *, char32_t *, bool> ret =
+      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
+  if (!std::get<2>(ret)) {
+    return 0;
+  }
+  size_t saved_bytes = std::get<1>(ret) - utf32_output;
+  if (std::get<0>(ret) != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::BIG>(
+            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-template <unsigned idx0, unsigned idx1, unsigned idx2, unsigned idx3>
-__m512i shuffle_epi128(__m512i v) {
-  static_assert((idx0 >= 0 && idx0 <= 3), "idx0 must be in range 0..3");
-  static_assert((idx1 >= 0 && idx1 <= 3), "idx1 must be in range 0..3");
-  static_assert((idx2 >= 0 && idx2 <= 3), "idx2 must be in range 0..3");
-  static_assert((idx3 >= 0 && idx3 <= 3), "idx3 must be in range 0..3");
-
-  constexpr unsigned shuffle = idx0 | (idx1 << 2) | (idx2 << 4) | (idx3 << 6);
-  return _mm512_shuffle_i32x4(v, v, shuffle);
+#if SIMDUTF_FEATURE_UTF16
+void implementation::change_endianness_utf16(const char16_t *input,
+                                             size_t length,
+                                             char16_t *output) const noexcept {
+  size_t pos = 0;
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  while (pos + 32 <= length) {
+    __m512i utf16 = _mm512_loadu_si512((const __m512i *)(input + pos));
+    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
+    _mm512_storeu_si512(output + pos, utf16);
+    pos += 32;
+  }
+  if (pos < length) {
+    __mmask32 m((1U << (length - pos)) - 1);
+    __m512i utf16 = _mm512_maskz_loadu_epi16(m, (const __m512i *)(input + pos));
+    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
+    _mm512_mask_storeu_epi16(output + pos, m, utf16);
+  }
 }
 
-template <unsigned idx> constexpr __m512i broadcast_epi128(__m512i v) {
-  return shuffle_epi128<idx, idx, idx, idx>(v);
-}
+simdutf_warn_unused size_t implementation::count_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  const char16_t *ptr = input;
+  size_t count{0};
 
-/**
- * Current unused.
- */
-template <int N> __m512i rotate_by_N_epi8(const __m512i input) {
+  if (length >= 32) {
+    const char16_t *end = input + length - 32;
+
+    const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
+    const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);
 
-  // lanes order: 1, 2, 3, 0 => 0b00_11_10_01
-  const __m512i permuted = _mm512_shuffle_i32x4(input, input, 0x39);
+    while (ptr <= end) {
+      __m512i utf16 = _mm512_loadu_si512((const __m512i *)ptr);
+      ptr += 32;
+      uint64_t not_high_surrogate =
+          static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) |
+                                _mm512_cmplt_epu16_mask(utf16, low));
+      count += count_ones(not_high_surrogate);
+    }
+  }
 
-  return _mm512_alignr_epi8(permuted, input, N);
+  return count + scalar::utf16::count_code_points<endianness::LITTLE>(
+                     ptr, length - (ptr - input));
 }
 
-/*
-    expanded_utf8_to_utf32 converts expanded UTF-8 characters (`utf8`)
-    stored at separate 32-bit lanes.
-
-    For each lane we have also a character class (`char_class), given in form
-    0x8080800N, where N is 4 highest bits from the leading byte; 0x80 resets
-    corresponding bytes during pshufb.
-*/
-simdutf_really_inline __m512i expanded_utf8_to_utf32(__m512i char_class,
-                                                     __m512i utf8) {
-  /*
-      Input:
-      - utf8: bytes stored at separate 32-bit code units
-      - valid: which code units have valid UTF-8 characters
+simdutf_warn_unused size_t implementation::count_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  const char16_t *ptr = input;
+  size_t count{0};
+  if (length >= 32) {
 
-      Bit layout of single word. We show 4 cases for each possible
-      UTF-8 character encoding. The `?` denotes bits we must not
-      assume their value.
+    const char16_t *end = input + length - 32;
 
-      |10dd.dddd|10cc.cccc|10bb.bbbb|1111.0aaa| 4-byte char
-      |????.????|10cc.cccc|10bb.bbbb|1110.aaaa| 3-byte char
-      |????.????|????.????|10bb.bbbb|110a.aaaa| 2-byte char
-      |????.????|????.????|????.????|0aaa.aaaa| ASCII char
-        byte 3    byte 2    byte 1     byte 0
-  */
+    const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
+    const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);
 
-  /* 1. Reset control bits of continuation bytes and the MSB
-        of the leading byte; this makes all bytes unsigned (and
-        does not alter ASCII char).
+    const __m512i byteflip = _mm512_setr_epi64(
+        0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+        0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+        0x0607040502030001, 0x0e0f0c0d0a0b0809);
+    while (ptr <= end) {
+      __m512i utf16 =
+          _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)ptr), byteflip);
+      ptr += 32;
+      uint64_t not_high_surrogate =
+          static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) |
+                                _mm512_cmplt_epu16_mask(utf16, low));
+      count += count_ones(not_high_surrogate);
+    }
+  }
 
-      |00dd.dddd|00cc.cccc|00bb.bbbb|0111.0aaa| 4-byte char
-      |00??.????|00cc.cccc|00bb.bbbb|0110.aaaa| 3-byte char
-      |00??.????|00??.????|00bb.bbbb|010a.aaaa| 2-byte char
-      |00??.????|00??.????|00??.????|0aaa.aaaa| ASCII char
-       ^^        ^^        ^^        ^
-  */
-  __m512i values;
-  const __m512i v_3f3f_3f7f = _mm512_set1_epi32(0x3f3f3f7f);
-  values = _mm512_and_si512(utf8, v_3f3f_3f7f);
+  return count + scalar::utf16::count_code_points<endianness::BIG>(
+                     ptr, length - (ptr - input));
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-  /* 2. Swap and join fields A-B and C-D
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused size_t
+implementation::count_utf8(const char *input, size_t length) const noexcept {
+  const uint8_t *str = reinterpret_cast<const uint8_t *>(input);
+  size_t answer =
+      length / sizeof(__m512i) *
+      sizeof(__m512i); // Number of 512-bit chunks that fits into the length.
+  size_t i = 0;
+  __m512i unrolled_popcount{0};
 
-      |0000.cccc|ccdd.dddd|0001.110a|aabb.bbbb| 4-byte char
-      |0000.cccc|cc??.????|0001.10aa|aabb.bbbb| 3-byte char
-      |0000.????|????.????|0001.0aaa|aabb.bbbb| 2-byte char
-      |0000.????|????.????|000a.aaaa|aa??.????| ASCII char */
-  const __m512i v_0140_0140 = _mm512_set1_epi32(0x01400140);
-  values = _mm512_maddubs_epi16(values, v_0140_0140);
+  const __m512i continuation = _mm512_set1_epi8(char(0b10111111));
 
-  /* 3. Swap and join fields AB & CD
+  while (i + sizeof(__m512i) <= length) {
+    size_t iterations = (length - i) / sizeof(__m512i);
 
-      |0000.0001|110a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char
-      |0000.0001|10aa.aabb|bbbb.cccc|cc??.????| 3-byte char
-      |0000.0001|0aaa.aabb|bbbb.????|????.????| 2-byte char
-      |0000.000a|aaaa.aa??|????.????|????.????| ASCII char */
-  const __m512i v_0001_1000 = _mm512_set1_epi32(0x00011000);
-  values = _mm512_madd_epi16(values, v_0001_1000);
+    size_t max_i = i + iterations * sizeof(__m512i) - sizeof(__m512i);
+    for (; i + 8 * sizeof(__m512i) <= max_i; i += 8 * sizeof(__m512i)) {
+      __m512i input1 = _mm512_loadu_si512((const __m512i *)(str + i));
+      __m512i input2 =
+          _mm512_loadu_si512((const __m512i *)(str + i + sizeof(__m512i)));
+      __m512i input3 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 2 * sizeof(__m512i)));
+      __m512i input4 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 3 * sizeof(__m512i)));
+      __m512i input5 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 4 * sizeof(__m512i)));
+      __m512i input6 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 5 * sizeof(__m512i)));
+      __m512i input7 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 6 * sizeof(__m512i)));
+      __m512i input8 =
+          _mm512_loadu_si512((const __m512i *)(str + i + 7 * sizeof(__m512i)));
 
-  /* 4. Shift left the values by variable amounts to reset highest UTF-8 bits
-      |aaab.bbbb|bccc.cccd|dddd.d000|0000.0000| 4-byte char -- by 11
-      |aaaa.bbbb|bbcc.cccc|????.??00|0000.0000| 3-byte char -- by 10
-      |aaaa.abbb|bbb?.????|????.???0|0000.0000| 2-byte char -- by 9
-      |aaaa.aaa?|????.????|????.????|?000.0000| ASCII char -- by 7 */
-  {
-    /** pshufb
+      __mmask64 mask1 = _mm512_cmple_epi8_mask(input1, continuation);
+      __mmask64 mask2 = _mm512_cmple_epi8_mask(input2, continuation);
+      __mmask64 mask3 = _mm512_cmple_epi8_mask(input3, continuation);
+      __mmask64 mask4 = _mm512_cmple_epi8_mask(input4, continuation);
+      __mmask64 mask5 = _mm512_cmple_epi8_mask(input5, continuation);
+      __mmask64 mask6 = _mm512_cmple_epi8_mask(input6, continuation);
+      __mmask64 mask7 = _mm512_cmple_epi8_mask(input7, continuation);
+      __mmask64 mask8 = _mm512_cmple_epi8_mask(input8, continuation);
 
-    continuation = 0
-    ascii    = 7
-    _2_bytes = 9
-    _3_bytes = 10
-    _4_bytes = 11
+      __m512i mask_register = _mm512_set_epi64(mask8, mask7, mask6, mask5,
+                                               mask4, mask3, mask2, mask1);
 
-    shift_left_v3 = 4 * [
-        ascii, # 0000
-        ascii, # 0001
-        ascii, # 0010
-        ascii, # 0011
-        ascii, # 0100
-        ascii, # 0101
-        ascii, # 0110
-        ascii, # 0111
-        continuation, # 1000
-        continuation, # 1001
-        continuation, # 1010
-        continuation, # 1011
-        _2_bytes, # 1100
-        _2_bytes, # 1101
-        _3_bytes, # 1110
-        _4_bytes, # 1111
-    ] */
-    const __m512i shift_left_v3 = _mm512_setr_epi64(
-        0x0707070707070707, 0x0b0a090900000000, 0x0707070707070707,
-        0x0b0a090900000000, 0x0707070707070707, 0x0b0a090900000000,
-        0x0707070707070707, 0x0b0a090900000000);
+      unrolled_popcount = _mm512_add_epi64(unrolled_popcount,
+                                           _mm512_popcnt_epi64(mask_register));
+    }
 
-    const __m512i shift = _mm512_shuffle_epi8(shift_left_v3, char_class);
-    values = _mm512_sllv_epi32(values, shift);
+    for (; i <= max_i; i += sizeof(__m512i)) {
+      __m512i more_input = _mm512_loadu_si512((const __m512i *)(str + i));
+      uint64_t continuation_bitmask = static_cast<uint64_t>(
+          _mm512_cmple_epi8_mask(more_input, continuation));
+      answer -= count_ones(continuation_bitmask);
+    }
   }
 
-  /* 5. Shift right the values by variable amounts to reset lowest bits
-      |0000.0000|000a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char -- by 11
-      |0000.0000|0000.0000|aaaa.bbbb|bbcc.cccc| 3-byte char -- by 16
-      |0000.0000|0000.0000|0000.0aaa|aabb.bbbb| 2-byte char -- by 21
-      |0000.0000|0000.0000|0000.0000|0aaa.aaaa| ASCII char -- by 25 */
-  {
-    // 4 * [25, 25, 25, 25, 25, 25, 25, 25, 0, 0, 0, 0, 21, 21, 16, 11]
-    const __m512i shift_right = _mm512_setr_epi64(
-        0x1919191919191919, 0x0b10151500000000, 0x1919191919191919,
-        0x0b10151500000000, 0x1919191919191919, 0x0b10151500000000,
-        0x1919191919191919, 0x0b10151500000000);
+  answer -= _mm512_reduce_add_epi64(unrolled_popcount);
 
-    const __m512i shift = _mm512_shuffle_epi8(shift_right, char_class);
-    values = _mm512_srlv_epi32(values, shift);
-  }
+  return answer + scalar::utf8::count_code_points(
+                      reinterpret_cast<const char *>(str + i), length - i);
+}
+#endif // SIMDUTF_FEATURE_UTF8
 
-  return values;
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
+    const char *buf, size_t len) const noexcept {
+  return count_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-simdutf_really_inline __m512i expand_and_identify(__m512i lane0, __m512i lane1,
-                                                  int &count) {
-  const __m512i merged = _mm512_mask_mov_epi32(lane0, 0x1000, lane1);
-  const __m512i expand_ver2 = _mm512_setr_epi64(
-      0x0403020103020100, 0x0605040305040302, 0x0807060507060504,
-      0x0a09080709080706, 0x0c0b0a090b0a0908, 0x0e0d0c0b0d0c0b0a,
-      0x000f0e0d0f0e0d0c, 0x0201000f01000f0e);
-  const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);
-  const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);
-  const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);
-  const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);
-  const __mmask16 leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);
-  count = static_cast<int>(count_ones(leading_bytes));
-  return _mm512_mask_compress_epi32(_mm512_setzero_si512(), leading_bytes,
-                                    input);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return icelake_utf8_length_from_utf16<endianness::LITTLE>(input, length);
 }
 
-simdutf_really_inline __m512i expand_utf8_to_utf32(__m512i input) {
-  __m512i char_class = _mm512_srli_epi32(input, 4);
-  /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */
-  const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);
-  const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);
-  char_class =
-      _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea);
-  return expanded_utf8_to_utf32(char_class, input);
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return icelake_utf8_length_from_utf16<endianness::BIG>(input, length);
 }
-/* end file src/icelake/icelake_utf8_common.inl.cpp */
-/* begin file src/icelake/icelake_macros.inl.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-/*
-    This upcoming macro (SIMDUTF_ICELAKE_TRANSCODE16) takes 16 + 4 bytes (of a
-   UTF-8 string) and loads all possible 4-byte substring into an AVX512
-   register.
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return implementation::count_utf16le(input, length);
+}
 
-    For example if we have bytes abcdefgh... we create following 32-bit lanes
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return implementation::count_utf16be(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    [abcd|bcde|cdef|defg|efgh|...]
-     ^                          ^
-     byte 0 of reg              byte 63 of reg
-*/
-/** pshufb
-        # lane{0,1,2} have got bytes: [  0,  1,  2,  3,  4,  5,  6,  8,  9, 10,
-   11, 12, 13, 14, 15] # lane3 has got bytes:        [ 16, 17, 18, 19,  4,  5,
-   6,  8,  9, 10, 11, 12, 13, 14, 15]
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
+    const char *input, size_t length) const noexcept {
+  const uint8_t *str = reinterpret_cast<const uint8_t *>(input);
+  size_t answer = length / sizeof(__m512i) * sizeof(__m512i);
+  size_t i = 0;
+  if (answer >= 2048) { // long strings optimization
+    unsigned char v_0xFF = 0xff;
+    __m512i eight_64bits = _mm512_setzero_si512();
+    while (i + sizeof(__m512i) <= length) {
+      __m512i runner = _mm512_setzero_si512();
+      size_t iterations = (length - i) / sizeof(__m512i);
+      if (iterations > 255) {
+        iterations = 255;
+      }
+      size_t max_i = i + iterations * sizeof(__m512i) - sizeof(__m512i);
+      for (; i + 4 * sizeof(__m512i) <= max_i; i += 4 * sizeof(__m512i)) {
+        // Load four __m512i vectors
+        __m512i input1 = _mm512_loadu_si512((const __m512i *)(str + i));
+        __m512i input2 =
+            _mm512_loadu_si512((const __m512i *)(str + i + sizeof(__m512i)));
+        __m512i input3 = _mm512_loadu_si512(
+            (const __m512i *)(str + i + 2 * sizeof(__m512i)));
+        __m512i input4 = _mm512_loadu_si512(
+            (const __m512i *)(str + i + 3 * sizeof(__m512i)));
 
-        expand_ver2 = [
-            # lane 0:
-            0, 1, 2, 3,
-            1, 2, 3, 4,
-            2, 3, 4, 5,
-            3, 4, 5, 6,
+        // Generate four masks
+        __mmask64 mask1 =
+            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input1);
+        __mmask64 mask2 =
+            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input2);
+        __mmask64 mask3 =
+            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input3);
+        __mmask64 mask4 =
+            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input4);
+        // Apply the masks and subtract from the runner
+        __m512i not_ascii1 =
+            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask1, v_0xFF);
+        __m512i not_ascii2 =
+            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask2, v_0xFF);
+        __m512i not_ascii3 =
+            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask3, v_0xFF);
+        __m512i not_ascii4 =
+            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask4, v_0xFF);
 
-            # lane 1:
-            4, 5, 6, 7,
-            5, 6, 7, 8,
-            6, 7, 8, 9,
-            7, 8, 9, 10,
+        runner = _mm512_sub_epi8(runner, not_ascii1);
+        runner = _mm512_sub_epi8(runner, not_ascii2);
+        runner = _mm512_sub_epi8(runner, not_ascii3);
+        runner = _mm512_sub_epi8(runner, not_ascii4);
+      }
 
-            # lane 2:
-             8,  9, 10, 11,
-             9, 10, 11, 12,
-            10, 11, 12, 13,
-            11, 12, 13, 14,
+      for (; i <= max_i; i += sizeof(__m512i)) {
+        __m512i more_input = _mm512_loadu_si512((const __m512i *)(str + i));
 
-            # lane 3 order: 13, 14, 15, 16 14, 15, 16, 17, 15, 16, 17, 18, 16,
-   17, 18, 19 12, 13, 14, 15, 13, 14, 15,  0, 14, 15,  0,  1, 15,  0,  1,  2,
-        ]
-*/
+        __mmask64 mask =
+            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), more_input);
+        __m512i not_ascii =
+            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask, v_0xFF);
+        runner = _mm512_sub_epi8(runner, not_ascii);
+      }
 
-#define SIMDUTF_ICELAKE_TRANSCODE16(LANE0, LANE1, MASKED)                      \
-  {                                                                            \
-    const __m512i merged = _mm512_mask_mov_epi32(LANE0, 0x1000, LANE1);        \
-    const __m512i expand_ver2 = _mm512_setr_epi64(                             \
-        0x0403020103020100, 0x0605040305040302, 0x0807060507060504,            \
-        0x0a09080709080706, 0x0c0b0a090b0a0908, 0x0e0d0c0b0d0c0b0a,            \
-        0x000f0e0d0f0e0d0c, 0x0201000f01000f0e);                               \
-    const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);            \
-                                                                               \
-    __mmask16 leading_bytes;                                                   \
-    const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);                       \
-    const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);                   \
-    const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);                       \
-    leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);                 \
-                                                                               \
-    __m512i char_class;                                                        \
-    char_class = _mm512_srli_epi32(input, 4);                                  \
-    /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */                     \
-    const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);                       \
-    const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);                 \
-    char_class =                                                               \
-        _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea); \
-                                                                               \
-    const int valid_count = static_cast<int>(count_ones(leading_bytes));       \
-    const __m512i utf32 = expanded_utf8_to_utf32(char_class, input);           \
-                                                                               \
-    const __m512i out = _mm512_mask_compress_epi32(_mm512_setzero_si512(),     \
-                                                   leading_bytes, utf32);      \
-                                                                               \
-    if (UTF32) {                                                               \
-      if (MASKED) {                                                            \
-        const __mmask16 valid = uint16_t((1 << valid_count) - 1);              \
-        _mm512_mask_storeu_epi32((__m512i *)output, valid, out);               \
-      } else {                                                                 \
-        _mm512_storeu_si512((__m512i *)output, out);                           \
-      }                                                                        \
-      output += valid_count;                                                   \
-    } else {                                                                   \
-      if (MASKED) {                                                            \
-        output += utf32_to_utf16_masked<big_endian>(                           \
-            byteflip, out, valid_count, reinterpret_cast<char16_t *>(output)); \
-      } else {                                                                 \
-        output += utf32_to_utf16<big_endian>(                                  \
-            byteflip, out, valid_count, reinterpret_cast<char16_t *>(output)); \
-      }                                                                        \
-    }                                                                          \
-  }
+      eight_64bits = _mm512_add_epi64(
+          eight_64bits, _mm512_sad_epu8(runner, _mm512_setzero_si512()));
+    }
 
-#define SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(INPUT, VALID_COUNT, MASKED)       \
-  {                                                                            \
-    if (UTF32) {                                                               \
-      if (MASKED) {                                                            \
-        const __mmask16 valid_mask = uint16_t((1 << VALID_COUNT) - 1);         \
-        _mm512_mask_storeu_epi32((__m512i *)output, valid_mask, INPUT);        \
-      } else {                                                                 \
-        _mm512_storeu_si512((__m512i *)output, INPUT);                         \
-      }                                                                        \
-      output += VALID_COUNT;                                                   \
-    } else {                                                                   \
-      if (MASKED) {                                                            \
-        output += utf32_to_utf16_masked<big_endian>(                           \
-            byteflip, INPUT, VALID_COUNT,                                      \
-            reinterpret_cast<char16_t *>(output));                             \
-      } else {                                                                 \
-        output +=                                                              \
-            utf32_to_utf16<big_endian>(byteflip, INPUT, VALID_COUNT,           \
-                                       reinterpret_cast<char16_t *>(output));  \
-      }                                                                        \
-    }                                                                          \
+    answer += _mm512_reduce_add_epi64(eight_64bits);
+  } else if (answer > 0) {
+    for (; i + sizeof(__m512i) <= length; i += sizeof(__m512i)) {
+      __m512i latin = _mm512_loadu_si512((const __m512i *)(str + i));
+      uint64_t non_ascii = _mm512_movepi8_mask(latin);
+      answer += count_ones(non_ascii);
+    }
   }
+  return answer + scalar::latin1::utf8_length_from_latin1(
+                      reinterpret_cast<const char *>(str + i), length - i);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-#define SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)                       \
-  if (UTF32) {                                                                 \
-    const __m128i t0 = _mm512_castsi512_si128(utf8);                           \
-    const __m128i t1 = _mm512_extracti32x4_epi32(utf8, 1);                     \
-    const __m128i t2 = _mm512_extracti32x4_epi32(utf8, 2);                     \
-    const __m128i t3 = _mm512_extracti32x4_epi32(utf8, 3);                     \
-    _mm512_storeu_si512((__m512i *)(output + 0 * 16),                          \
-                        _mm512_cvtepu8_epi32(t0));                             \
-    _mm512_storeu_si512((__m512i *)(output + 1 * 16),                          \
-                        _mm512_cvtepu8_epi32(t1));                             \
-    _mm512_storeu_si512((__m512i *)(output + 2 * 16),                          \
-                        _mm512_cvtepu8_epi32(t2));                             \
-    _mm512_storeu_si512((__m512i *)(output + 3 * 16),                          \
-                        _mm512_cvtepu8_epi32(t3));                             \
-  } else {                                                                     \
-    const __m256i h0 = _mm512_castsi512_si256(utf8);                           \
-    const __m256i h1 = _mm512_extracti64x4_epi64(utf8, 1);                     \
-    if (big_endian) {                                                          \
-      _mm512_storeu_si512(                                                     \
-          (__m512i *)(output + 0 * 16),                                        \
-          _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h0), byteflip));            \
-      _mm512_storeu_si512(                                                     \
-          (__m512i *)(output + 2 * 16),                                        \
-          _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h1), byteflip));            \
-    } else {                                                                   \
-      _mm512_storeu_si512((__m512i *)(output + 0 * 16),                        \
-                          _mm512_cvtepu8_epi16(h0));                           \
-      _mm512_storeu_si512((__m512i *)(output + 2 * 16),                        \
-                          _mm512_cvtepu8_epi16(h1));                           \
-    }                                                                          \
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  size_t pos = 0;
+  size_t count = 0;
+  // This algorithm could no doubt be improved!
+  for (; pos + 64 <= length; pos += 64) {
+    __m512i utf8 = _mm512_loadu_si512((const __m512i *)(input + pos));
+    uint64_t utf8_continuation_mask =
+        _mm512_cmplt_epi8_mask(utf8, _mm512_set1_epi8(-65 + 1));
+    // We count one word for anything that is not a continuation (so
+    // leading bytes).
+    count += 64 - count_ones(utf8_continuation_mask);
+    uint64_t utf8_4byte =
+        _mm512_cmpge_epu8_mask(utf8, _mm512_set1_epi8(int8_t(240)));
+    count += count_ones(utf8_4byte);
   }
-/* end file src/icelake/icelake_macros.inl.cpp */
-/* begin file src/icelake/icelake_from_valid_utf8.inl.cpp */
-// file included directly
-
-// File contains conversion procedure from VALID UTF-8 strings.
+  return count +
+         scalar::utf8::utf16_length_from_utf8(input + pos, length - pos);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-/*
-    valid_utf8_to_fixed_length converts a valid UTF-8 string into UTF-32.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  return utf32::utf8_length_from_utf32(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-    The `OUTPUT` template type decides what to do with UTF-32: store
-    it directly or convert into UTF-16 (with AVX512).
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  const char32_t *ptr = input;
+  size_t count{0};
 
-    Input:
-    - str           - valid UTF-8 string
-    - len           - string length
-    - out_buffer    - output buffer
+  if (length >= 16) {
+    const char32_t *end = input + length - 16;
 
-    Result:
-    - pair.first    - the first unprocessed input byte
-    - pair.second   - the first unprocessed output word
-*/
-template <endianness big_endian, typename OUTPUT>
-std::pair<const char *, OUTPUT *>
-valid_utf8_to_fixed_length(const char *str, size_t len, OUTPUT *dwords) {
-  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
-  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
-  static_assert(
-      UTF32 or UTF16,
-      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
-  static_assert(!(UTF32 and big_endian),
-                "we do not currently support big-endian UTF-32");
+    const __m512i v_0000_ffff = _mm512_set1_epi32((uint32_t)0x0000ffff);
 
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  const char *ptr = str;
-  const char *end = ptr + len;
+    while (ptr <= end) {
+      __m512i utf32 = _mm512_loadu_si512((const __m512i *)ptr);
+      ptr += 16;
+      __mmask16 surrogates_bitmask =
+          _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);
 
-  OUTPUT *output = dwords;
-  /**
-   * In the main loop, we consume 64 bytes per iteration,
-   * but we access 64 + 4 bytes.
-   * We check for ptr + 64 + 64 <= end because
-   * we want to be do maskless writes without overruns.
-   */
-  while (end - ptr >= 64 + 4) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
-    const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
-    if (ascii == 0) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
-      continue;
+      count += 16 + count_ones(surrogates_bitmask);
     }
+  }
 
-    const __m512i lane0 = broadcast_epi128<0>(utf8);
-    const __m512i lane1 = broadcast_epi128<1>(utf8);
-    int valid_count0;
-    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-    const __m512i lane2 = broadcast_epi128<2>(utf8);
-    int valid_count1;
-    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-    if (valid_count0 + valid_count1 <= 16) {
-      vec0 = _mm512_mask_expand_epi32(
-          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-      valid_count0 += valid_count1;
-      vec0 = expand_utf8_to_utf32(vec0);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
+  return count +
+         scalar::utf32::utf16_length_from_utf32(ptr, length - (ptr - input));
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return implementation::count_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
     } else {
-      vec0 = expand_utf8_to_utf32(vec0);
-      vec1 = expand_utf8_to_utf32(vec1);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
     }
-    const __m512i lane3 = broadcast_epi128<3>(utf8);
-    int valid_count2;
-    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
-    uint32_t tmp1;
-    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
-    const __m512i lane4 = _mm512_set1_epi32(tmp1);
-    int valid_count3;
-    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
-    if (valid_count2 + valid_count3 <= 16) {
-      vec2 = _mm512_mask_expand_epi32(
-          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
-      valid_count2 += valid_count3;
-      vec2 = expand_utf8_to_utf32(vec2);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
     } else {
-      vec2 = expand_utf8_to_utf32(vec2);
-      vec3 = expand_utf8_to_utf32(vec3);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
     }
-    ptr += 4 * 16;
   }
+}
 
-  if (end - ptr >= 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
-    const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
-    if (ascii == 0) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
     } else {
-      const __m512i lane0 = broadcast_epi128<0>(utf8);
-      const __m512i lane1 = broadcast_epi128<1>(utf8);
-      int valid_count0;
-      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-      const __m512i lane2 = broadcast_epi128<2>(utf8);
-      int valid_count1;
-      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-      if (valid_count0 + valid_count1 <= 16) {
-        vec0 = _mm512_mask_expand_epi32(
-            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-        valid_count0 += valid_count1;
-        vec0 = expand_utf8_to_utf32(vec0);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      } else {
-        vec0 = expand_utf8_to_utf32(vec0);
-        vec1 = expand_utf8_to_utf32(vec1);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
-      }
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
+}
 
-      const __m512i lane3 = broadcast_epi128<3>(utf8);
-      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
+}
 
-      ptr += 3 * 16;
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
     }
   }
-  return {ptr, output};
 }
 
-using utf8_to_utf16_result = std::pair<const char *, char16_t *>;
-/* end file src/icelake/icelake_from_valid_utf8.inl.cpp */
-/* begin file src/icelake/icelake_utf8_validation.inl.cpp */
-// file included directly
+size_t implementation::binary_to_base64(const char *input, size_t length,
+                                        char *output,
+                                        base64_options options) const noexcept {
+  if (options & base64_url) {
+    return encode_base64<true>(output, input, length, options);
+  } else {
+    return encode_base64<false>(output, input, length, options);
+  }
+}
+#endif // SIMDUTF_FEATURE_BASE64
 
-simdutf_really_inline __m512i check_special_cases(__m512i input,
-                                                  const __m512i prev1) {
-  __m512i mask1 = _mm512_setr_epi64(0x0202020202020202, 0x4915012180808080,
-                                    0x0202020202020202, 0x4915012180808080,
-                                    0x0202020202020202, 0x4915012180808080,
-                                    0x0202020202020202, 0x4915012180808080);
-  const __m512i v_0f = _mm512_set1_epi8(0x0f);
-  __m512i index1 = _mm512_and_si512(_mm512_srli_epi16(prev1, 4), v_0f);
+} // namespace icelake
+} // namespace simdutf
 
-  __m512i byte_1_high = _mm512_shuffle_epi8(mask1, index1);
-  __m512i mask2 = _mm512_setr_epi64(0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
-                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
-                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb,
-                                    0xcbcbcb8b8383a3e7, 0xcbcbdbcbcbcbcbcb);
-  __m512i index2 = _mm512_and_si512(prev1, v_0f);
+/* begin file src/simdutf/icelake/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
 
-  __m512i byte_1_low = _mm512_shuffle_epi8(mask2, index2);
-  __m512i mask3 =
-      _mm512_setr_epi64(0x101010101010101, 0x1010101babaaee6, 0x101010101010101,
-                        0x1010101babaaee6, 0x101010101010101, 0x1010101babaaee6,
-                        0x101010101010101, 0x1010101babaaee6);
-  __m512i index3 = _mm512_and_si512(_mm512_srli_epi16(input, 4), v_0f);
-  __m512i byte_2_high = _mm512_shuffle_epi8(mask3, index3);
-  return _mm512_ternarylogic_epi64(byte_1_high, byte_1_low, byte_2_high, 128);
-}
 
-simdutf_really_inline __m512i check_multibyte_lengths(const __m512i input,
-                                                      const __m512i prev_input,
-                                                      const __m512i sc) {
-  __m512i prev2 = prev<2>(input, prev_input);
-  __m512i prev3 = prev<3>(input, prev_input);
-  __m512i is_third_byte = _mm512_subs_epu8(
-      prev2, _mm512_set1_epi8(0b11100000u - 1)); // Only 111_____ will be > 0
-  __m512i is_fourth_byte = _mm512_subs_epu8(
-      prev3, _mm512_set1_epi8(0b11110000u - 1)); // Only 1111____ will be > 0
-  __m512i is_third_or_fourth_byte =
-      _mm512_or_si512(is_third_byte, is_fourth_byte);
-  const __m512i v_7f = _mm512_set1_epi8(char(0x7f));
-  is_third_or_fourth_byte = _mm512_adds_epu8(v_7f, is_third_or_fourth_byte);
-  // We want to compute (is_third_or_fourth_byte AND v80) XOR sc.
-  const __m512i v_80 = _mm512_set1_epi8(char(0x80));
-  return _mm512_ternarylogic_epi32(is_third_or_fourth_byte, v_80, sc,
-                                   0b1101010);
-  //__m512i is_third_or_fourth_byte_mask =
-  //_mm512_and_si512(is_third_or_fourth_byte, v_80); return
-  // _mm512_xor_si512(is_third_or_fourth_byte_mask, sc);
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+SIMDUTF_POP_DISABLE_WARNINGS
+#endif // end of workaround
+/* end file src/simdutf/icelake/end.h */
+/* end file src/icelake/implementation.cpp */
+#endif
+#if SIMDUTF_IMPLEMENTATION_HASWELL
+/* begin file src/haswell/implementation.cpp */
+/* begin file src/simdutf/haswell/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "haswell"
+// #define SIMDUTF_IMPLEMENTATION haswell
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
+
+#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
+// nothing needed.
+#else
+SIMDUTF_TARGET_HASWELL
+#endif
+
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+// clang-format off
+SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
+// clang-format on
+#endif // end of workaround
+/* end file src/simdutf/haswell/begin.h */
+
+namespace simdutf {
+namespace haswell {
+namespace {
+#ifndef SIMDUTF_HASWELL_H
+  #error "haswell.h must be included"
+#endif
+using namespace simd;
+
+#if SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    SIMDUTF_FEATURE_UTF8
+simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
+  return input.reduce_or().is_ascii();
 }
-//
-// Return nonzero if there are incomplete multibyte characters at the end of the
-// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
-//
-simdutf_really_inline __m512i is_incomplete(const __m512i input) {
-  // If the previous input's last 3 bytes match this, they're too short (they
-  // ended at EOF):
-  // ... 1111____ 111_____ 11______
-  __m512i max_value = _mm512_setr_epi64(0xffffffffffffffff, 0xffffffffffffffff,
-                                        0xffffffffffffffff, 0xffffffffffffffff,
-                                        0xffffffffffffffff, 0xffffffffffffffff,
-                                        0xffffffffffffffff, 0xbfdfefffffffffff);
-  return _mm512_subs_epu8(input, max_value);
+#endif // SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_really_inline simd8<bool>
+must_be_2_3_continuation(const simd8<uint8_t> prev2,
+                         const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte =
+      prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be > 0x80
+  simd8<uint8_t> is_fourth_byte =
+      prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be > 0x80
+  return simd8<bool>(is_third_byte | is_fourth_byte);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-struct avx512_utf8_checker {
-  // If this is nonzero, there has been a UTF-8 error.
-  __m512i error{};
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+namespace utf16 {
+/* begin file src/haswell/avx2_validate_utf16.cpp */
+template <endianness big_endian>
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> &in0,
+                                       const simd16<uint16_t> &in1) {
+  if (big_endian) {
+    // we want lower bytes
+    const auto mask = simd16<uint16_t>(0x00ff);
+    const auto t0 = in0 & mask;
+    const auto t1 = in1 & mask;
 
-  // The last input we received
-  __m512i prev_input_block{};
-  // Whether the last input we received was incomplete (used for ASCII fast
-  // path)
-  __m512i prev_incomplete{};
+    return simd16<uint16_t>::pack(t0, t1);
+  } else {
+    const auto t0 = in0.shr<8>();
+    const auto t1 = in1.shr<8>();
 
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const __m512i input,
-                                              const __m512i prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    __m512i prev1 = prev<1>(input, prev_input);
-    __m512i sc = check_special_cases(input, prev1);
-    this->error = _mm512_or_si512(
-        check_multibyte_lengths(input, prev_input, sc), this->error);
+    return simd16<uint16_t>::pack(t0, t1);
   }
+}
+/* end file src/haswell/avx2_validate_utf16.cpp */
+}
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-  // The only problem that can happen at EOF is that a multibyte character is
-  // too short or a byte value too large in the last bytes: check_special_cases
-  // only checks for bytes too large in the first of two bytes.
-  simdutf_really_inline void check_eof() {
-    // If the previous block had incomplete UTF-8 characters at the end, an
-    // ASCII block can't possibly finish them.
-    this->error = _mm512_or_si512(this->error, this->prev_incomplete);
-  }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_latin1_to_utf8.cpp */
+std::pair<const char *, char *>
+avx2_convert_latin1_to_utf8(const char *latin1_input, size_t len,
+                            char *utf8_output) {
+  const char *end = latin1_input + len;
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
+  const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
+  const size_t safety_margin = 12;
 
-  // returns true if ASCII.
-  simdutf_really_inline bool check_next_input(const __m512i input) {
-    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
-    const __mmask64 ascii = _mm512_test_epi8_mask(input, v_80);
-    if (ascii == 0) {
-      this->error = _mm512_or_si512(this->error, this->prev_incomplete);
-      return true;
-    } else {
-      this->check_utf8_bytes(input, this->prev_input_block);
-      this->prev_incomplete = is_incomplete(input);
-      this->prev_input_block = input;
-      return false;
+  while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) {
+    __m128i in8 = _mm_loadu_si128((__m128i *)latin1_input);
+    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
+    const __m128i v_80 = _mm_set1_epi8((char)0x80);
+    if (_mm_testz_si128(in8, v_80)) { // ASCII fast path!!!!
+      // 1. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, in8);
+      // 2. adjust pointers
+      latin1_input += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
+    }
+    // We proceed only with the first 16 bytes.
+    const __m256i in = _mm256_cvtepu8_epi16((in8));
+
+    // 1. prepare 2-byte values
+    // input 16-bit word : [0000|0000|aabb|bbbb] x 8
+    // expected output   : [1100|00aa|10bb|bbbb] x 8
+    const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+    const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
+
+    // t0 = [0000|00aa|bbbb|bb00]
+    const __m256i t0 = _mm256_slli_epi16(in, 2);
+    // t1 = [0000|00aa|0000|0000]
+    const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+    // t2 = [0000|0000|00bb|bbbb]
+    const __m256i t2 = _mm256_and_si256(in, v_003f);
+    // t3 = [000a|aaaa|00bb|bbbb]
+    const __m256i t3 = _mm256_or_si256(t1, t2);
+    // t4 = [1100|00aa|10bb|bbbb]
+    const __m256i t4 = _mm256_or_si256(t3, v_c080);
+
+    // 2. merge ASCII and 2-byte codewords
+
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
+
+    const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask);
+
+    // 3. prepare bitmask for 8-bit lookup
+    const uint32_t M0 = one_byte_bitmask & 0x55555555;
+    const uint32_t M1 = M0 >> 7;
+    const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+    // 4. pack the bytes
+
+    const uint8_t *row =
+        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+    const uint8_t *row_2 =
+        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 16)]
+                                                            [0];
+
+    const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+    const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+
+    const __m256i utf8_packed = _mm256_shuffle_epi8(
+        utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+    // 5. store bytes
+    _mm_storeu_si128((__m128i *)utf8_output,
+                     _mm256_castsi256_si128(utf8_packed));
+    utf8_output += row[0];
+    _mm_storeu_si128((__m128i *)utf8_output,
+                     _mm256_extractf128_si256(utf8_packed, 1));
+    utf8_output += row_2[0];
+
+    // 6. adjust pointers
+    latin1_input += 16;
+    continue;
+
+  } // while
+  return std::make_pair(latin1_input, utf8_output);
+}
+/* end file src/haswell/avx2_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_latin1_to_utf16.cpp */
+template <endianness big_endian>
+std::pair<const char *, char16_t *>
+avx2_convert_latin1_to_utf16(const char *latin1_input, size_t len,
+                             char16_t *utf16_output) {
+  size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16
+
+  size_t i = 0;
+  for (; i < rounded_len; i += 16) {
+    // Load 16 bytes from the address (input + i) into a xmm register
+    const __m128i latin1 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(latin1_input + i));
+
+    // Zero extend each byte in `in` to word
+    __m256i utf16 = _mm256_cvtepu8_epi16(latin1);
+
+    if (big_endian) {
+      const __m128i swap128 =
+          _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+      const __m256i swap = _mm256_set_m128i(swap128, swap128);
+      utf16 = _mm256_shuffle_epi8(utf16, swap);
     }
+
+    // Store the contents of xmm1 into the address pointed by (output + i)
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16);
   }
-  // do not forget to call check_eof!
-  simdutf_really_inline bool errors() const {
-    return _mm512_test_epi8_mask(this->error, this->error) != 0;
+
+  return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len);
+}
+/* end file src/haswell/avx2_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_latin1_to_utf32.cpp */
+std::pair<const char *, char32_t *>
+avx2_convert_latin1_to_utf32(const char *buf, size_t len,
+                             char32_t *utf32_output) {
+  size_t rounded_len = ((len | 7) ^ 7); // Round down to nearest multiple of 8
+
+  for (size_t i = 0; i < rounded_len; i += 8) {
+    // Load 8 Latin1 characters into a 64-bit register
+    __m128i in = _mm_loadl_epi64((__m128i *)&buf[i]);
+
+    // Zero extend each set of 8 Latin1 characters to 8 32-bit integers using
+    // vpmovzxbd
+    __m256i out = _mm256_cvtepu8_epi32(in);
+
+    // Store the results back to memory
+    _mm256_storeu_si256((__m256i *)&utf32_output[i], out);
   }
-}; // struct avx512_utf8_checker
-/* end file src/icelake/icelake_utf8_validation.inl.cpp */
-/* begin file src/icelake/icelake_from_utf8.inl.cpp */
-// file included directly
 
-// File contains conversion procedure from possibly invalid UTF-8 strings.
+  // return pointers pointing to where we left off
+  return std::make_pair(buf + rounded_len, utf32_output + rounded_len);
+}
+/* end file src/haswell/avx2_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-/**
- * Attempts to convert up to len 1-byte code units from in (in UTF-8 format) to
- * out.
- * Returns the position of the input and output after the processing is
- * completed. Upon error, the output is set to null.
- */
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/haswell/avx2_convert_utf8_to_utf16.cpp */
+// depends on "tables/utf8_to_utf16_tables.h"
 
+// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
 template <endianness big_endian>
-utf8_to_utf16_result
-fast_avx512_convert_utf8_to_utf16(const char *in, size_t len, char16_t *out) {
-  const char *const final_in = in + len;
-  bool result = true;
-  while (result) {
-    if (final_in - in >= 64) {
-      result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(
-          in, out, final_in - in);
-    } else if (in < final_in) {
-      result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(
-          in, out, final_in - in);
-    } else {
-      break;
+size_t convert_masked_utf8_to_utf16(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char16_t *&utf16_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  // We first try a few fast paths.
+  const __m128i swap =
+      _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+  const __m128i in = _mm_loadu_si128((__m128i *)input);
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process the data in chunks of 12 bytes.
+    __m256i ascii = _mm256_cvtepu8_epi16(in);
+    if (big_endian) {
+      const __m256i swap256 = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      ascii = _mm256_shuffle_epi8(ascii, swap256);
     }
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii);
+    utf16_output += 12; // We wrote 12 16-bit characters.
+    return 12;          // We consumed 12 bytes.
   }
-  if (!result) {
-    out = nullptr;
+  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte
+    // UTF-16 code units. There is probably a more efficient sequence, but the
+    // following might do.
+    const __m128i sh =
+        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
+    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
+    if (big_endian)
+      composed = _mm_shuffle_epi8(composed, swap);
+    _mm_storeu_si128((__m128i *)utf16_output, composed);
+    utf16_output += 8; // We wrote 16 bytes, 8 code points.
+    return 16;
+  }
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte
+    // UTF-16 code units. There is probably a more efficient sequence, but the
+    // following might do.
+    const __m128i sh =
+        _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii =
+        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
+    const __m128i middlebyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    const __m128i highbyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
+    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
+    if (big_endian)
+      composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
+    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
+    utf16_output += 4;
+    return 12;
   }
-  return std::make_pair(in, out);
-}
 
-template <endianness big_endian>
-simdutf::result fast_avx512_convert_utf8_to_utf16_with_errors(const char *in,
-                                                              size_t len,
-                                                              char16_t *out) {
-  const char *const init_in = in;
-  const char16_t *const init_out = out;
-  const char *const final_in = in + len;
-  bool result = true;
-  while (result) {
-    if (final_in - in >= 64) {
-      result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(
-          in, out, final_in - in);
-    } else if (in < final_in) {
-      result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(
-          in, out, final_in - in);
-    } else {
-      break;
+  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
+      [input_utf8_end_of_code_point_mask][1];
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
+    // processors where pdep/pext is fast, we might be able to use a small
+    // lookup table.
+    const __m128i sh = _mm_loadu_si128(
+        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
+    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
+    if (big_endian)
+      composed = _mm_shuffle_epi8(composed, swap);
+    _mm_storeu_si128((__m128i *)utf16_output, composed);
+    utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential
+                       // overflow of 4 bytes.
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    const __m128i sh = _mm_loadu_si128(
+        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii =
+        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
+    const __m128i middlebyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    const __m128i highbyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
+    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
+    if (big_endian)
+      composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
+    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
+    utf16_output += 4; // Here we overflow by 8 bytes.
+  } else if (idx < 209) {
+    // TWO (2) input code-code units
+    //////////////
+    // There might be garbage inputs where a leading byte mascarades as a
+    // four-byte leading byte (by being followed by 3 continuation byte), but is
+    // not greater than 0xf0. This could trigger a buffer overflow if we only
+    // counted leading bytes of the form 0xf0 as generating surrogate pairs,
+    // without further UTF-8 validation. Thus we must be careful to ensure that
+    // only leading bytes at least as large as 0xf0 generate surrogate pairs. We
+    // do as at the cost of an extra mask.
+    /////////////
+    const __m128i sh = _mm_loadu_si128(
+        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
+    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
+    // correct for spurious high bit
+    const __m128i correct =
+        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
+    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
+    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
+    // We deliberately carry the leading four bits in highbyte if they are
+    // present, we remove them later when computing hightenbits.
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000));
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
+    // When we need to generate a surrogate pair (leading byte > 0xF0), then
+    // the corresponding 32-bit value in 'composed'  will be greater than
+    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
+    // location of the surrogate pairs.
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
+                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
+    const __m128i composedminus =
+        _mm_sub_epi32(composed, _mm_set1_epi32(0x10000));
+    const __m128i lowtenbits =
+        _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff));
+    // Notice the 0x3ff mask:
+    const __m128i hightenbits =
+        _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff));
+    const __m128i lowtenbitsadd =
+        _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00));
+    const __m128i hightenbitsadd =
+        _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800));
+    const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16);
+    __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted);
+    uint32_t basic_buffer[4];
+    uint32_t basic_buffer_swap[4];
+    if (big_endian) {
+      _mm_storeu_si128((__m128i *)basic_buffer_swap,
+                       _mm_shuffle_epi8(composed, swap));
+      surrogates = _mm_shuffle_epi8(surrogates, swap);
     }
-  }
-  if (!result) {
-    size_t pos = size_t(in - init_in);
-    if (pos < len && (init_in[pos] & 0xc0) == 0x80 && pos >= 64) {
-      // We must check whether we are the fourth continuation byte
-      bool c1 = (init_in[pos - 1] & 0xc0) == 0x80;
-      bool c2 = (init_in[pos - 2] & 0xc0) == 0x80;
-      bool c3 = (init_in[pos - 3] & 0xc0) == 0x80;
-      if (c1 && c2 && c3) {
-        return {simdutf::TOO_LONG, pos};
+    _mm_storeu_si128((__m128i *)basic_buffer, composed);
+    uint32_t surrogate_buffer[4];
+    _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates);
+    for (size_t i = 0; i < 3; i++) {
+      if (basic_buffer[i] > 0x3c00000) {
+        utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff);
+        utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16);
+        utf16_output += 2;
+      } else {
+        utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i])
+                                     : uint16_t(basic_buffer[i]);
+        utf16_output++;
       }
     }
-    // rewind_and_convert_with_errors will seek a potential error from in
-    // onward, with the ability to go back up to in - init_in bytes, and read
-    // final_in - in bytes forward.
-    simdutf::result res =
-        scalar::utf8_to_utf16::rewind_and_convert_with_errors<big_endian>(
-            in - init_in, in, final_in - in, out);
-    res.count += (in - init_in);
-    return res;
   } else {
-    return simdutf::result(error_code::SUCCESS, out - init_out);
+    // here we know that there is an error but we do not handle errors
   }
+  return consumed;
 }
+/* end file src/haswell/avx2_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-template <endianness big_endian, typename OUTPUT>
-// todo: replace with the utf-8 to utf-16 routine adapted to utf-32. This code
-// is legacy.
-std::pair<const char *, OUTPUT *>
-validating_utf8_to_fixed_length(const char *str, size_t len, OUTPUT *dwords) {
-  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
-  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
-  static_assert(
-      UTF32 or UTF16,
-      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
-  static_assert(!(UTF32 and big_endian),
-                "we do not currently support big-endian UTF-32");
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/haswell/avx2_convert_utf8_to_utf32.cpp */
+// depends on "tables/utf8_to_utf16_tables.h"
 
-  const char *ptr = str;
-  const char *end = ptr + len;
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  OUTPUT *output = dwords;
-  avx512_utf8_checker checker{};
-  /**
-   * In the main loop, we consume 64 bytes per iteration,
-   * but we access 64 + 4 bytes.
-   * We use masked writes to avoid overruns, see
-   * https://github.com/simdutf/simdutf/issues/471
-   */
-  while (end - ptr >= 64 + 4) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    if (checker.check_next_input(utf8)) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
-      continue;
-    }
-    const __m512i lane0 = broadcast_epi128<0>(utf8);
-    const __m512i lane1 = broadcast_epi128<1>(utf8);
-    int valid_count0;
-    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-    const __m512i lane2 = broadcast_epi128<2>(utf8);
-    int valid_count1;
-    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-    if (valid_count0 + valid_count1 <= 16) {
-      vec0 = _mm512_mask_expand_epi32(
-          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-      valid_count0 += valid_count1;
-      vec0 = expand_utf8_to_utf32(vec0);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-    } else {
-      vec0 = expand_utf8_to_utf32(vec0);
-      vec1 = expand_utf8_to_utf32(vec1);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
-    }
-    const __m512i lane3 = broadcast_epi128<3>(utf8);
-    int valid_count2;
-    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
-    uint32_t tmp1;
-    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
-    const __m512i lane4 = _mm512_set1_epi32(tmp1);
-    int valid_count3;
-    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
-    if (valid_count2 + valid_count3 <= 16) {
-      vec2 = _mm512_mask_expand_epi32(
-          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
-      valid_count2 += valid_count3;
-      vec2 = expand_utf8_to_utf32(vec2);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
-    } else {
-      vec2 = expand_utf8_to_utf32(vec2);
-      vec3 = expand_utf8_to_utf32(vec3);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
-    }
-    ptr += 4 * 16;
+// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
+size_t convert_masked_utf8_to_utf32(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char32_t *&utf32_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  // We first try a few fast paths.
+  const __m128i in = _mm_loadu_si128((__m128i *)input);
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process the data in chunks of 12 bytes.
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
+                        _mm256_cvtepu8_epi32(in));
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8),
+                        _mm256_cvtepu8_epi32(_mm_srli_si128(in, 8)));
+    utf32_output += 12; // We wrote 12 32-bit characters.
+    return 12;          // We consumed 12 bytes.
   }
-  const char *validatedptr = ptr; // validated up to ptr
-
-  // For the final pass, we validate 64 bytes, but we only transcode
-  // 3*16 bytes, so we may end up double-validating 16 bytes.
-  if (end - ptr >= 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    if (checker.check_next_input(utf8)) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
-    } else {
-      const __m512i lane0 = broadcast_epi128<0>(utf8);
-      const __m512i lane1 = broadcast_epi128<1>(utf8);
-      int valid_count0;
-      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-      const __m512i lane2 = broadcast_epi128<2>(utf8);
-      int valid_count1;
-      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-      if (valid_count0 + valid_count1 <= 16) {
-        vec0 = _mm512_mask_expand_epi32(
-            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-        valid_count0 += valid_count1;
-        vec0 = expand_utf8_to_utf32(vec0);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      } else {
-        vec0 = expand_utf8_to_utf32(vec0);
-        vec1 = expand_utf8_to_utf32(vec1);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
-      }
-
-      const __m512i lane3 = broadcast_epi128<3>(utf8);
-      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
-
-      ptr += 3 * 16;
-    }
-    validatedptr += 4 * 16;
+  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte
+    // UTF-32 code units. There is probably a more efficient sequence, but the
+    // following might do.
+    const __m128i sh =
+        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
+    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
+    _mm256_storeu_si256((__m256i *)utf32_output,
+                        _mm256_cvtepu16_epi32(composed));
+    utf32_output += 8; // We wrote 16 bytes, 8 code points.
+    return 16;
   }
-  if (end != validatedptr) {
-    const __m512i utf8 =
-        _mm512_maskz_loadu_epi8(~UINT64_C(0) >> (64 - (end - validatedptr)),
-                                (const __m512i *)validatedptr);
-    checker.check_next_input(utf8);
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte
+    // UTF-32 code units. There is probably a more efficient sequence, but the
+    // following might do.
+    const __m128i sh =
+        _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii =
+        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
+    const __m128i middlebyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    const __m128i highbyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
+    _mm_storeu_si128((__m128i *)utf32_output, composed);
+    utf32_output += 4;
+    return 12;
   }
-  checker.check_eof();
-  if (checker.errors()) {
-    return {ptr, nullptr}; // We found an error.
+  /// We do not have a fast path available, so we fallback.
+
+  const uint8_t idx =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
+    // processors where pdep/pext is fast, we might be able to use a small
+    // lookup table.
+    const __m128i sh =
+        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
+    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
+    _mm256_storeu_si256((__m256i *)utf32_output,
+                        _mm256_cvtepu16_epi32(composed));
+    utf32_output += 6; // We wrote 24 bytes, 6 code points. There is a potential
+    // overflow of 32 - 24 = 8 bytes.
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    const __m128i sh =
+        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii =
+        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
+    const __m128i middlebyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    const __m128i highbyte =
+        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
+    _mm_storeu_si128((__m128i *)utf32_output, composed);
+    utf32_output += 4;
+  } else if (idx < 209) {
+    // TWO (2) input code-code units
+    const __m128i sh =
+        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
+    const __m128i perm = _mm_shuffle_epi8(in, sh);
+    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
+    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
+    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
+    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
+    // correct for spurious high bit
+    const __m128i correct =
+        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
+    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
+    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
+    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000));
+    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
+    const __m128i composed =
+        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
+                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
+    _mm_storeu_si128((__m128i *)utf32_output, composed);
+    utf32_output +=
+        3; // We wrote 3 * 4 bytes, there is a potential overflow of 4 bytes.
+  } else {
+    // here we know that there is an error but we do not handle errors
   }
-  return {ptr, output};
+  return consumed;
 }
+/* end file src/haswell/avx2_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-// Like validating_utf8_to_fixed_length but returns as soon as an error is
-// identified todo: replace with the utf-8 to utf-16 routine adapted to utf-32.
-// This code is legacy.
-template <endianness big_endian, typename OUTPUT>
-std::tuple<const char *, OUTPUT *, bool>
-validating_utf8_to_fixed_length_with_constant_checks(const char *str,
-                                                     size_t len,
-                                                     OUTPUT *dwords) {
-  constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
-  constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
-  static_assert(
-      UTF32 or UTF16,
-      "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
-  static_assert(!(UTF32 and big_endian),
-                "we do not currently support big-endian UTF-32");
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_utf16_to_latin1.cpp */
+template <endianness big_endian>
+std::pair<const char16_t *, char *>
+avx2_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                             char *latin1_output) {
+  const char16_t *end = buf + len;
+  while (end - buf >= 32) {
+    // Load 16 UTF-16 characters into 256-bit AVX2 register
+    __m256i in0 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+    __m256i in1 =
+        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + 16));
 
-  const char *ptr = str;
-  const char *end = ptr + len;
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  OUTPUT *output = dwords;
-  avx512_utf8_checker checker{};
-  /**
-   * In the main loop, we consume 64 bytes per iteration,
-   * but we access 64 + 4 bytes.
-   */
-  while (end - ptr >= 4 + 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    bool ascii = checker.check_next_input(utf8);
-    if (checker.errors()) {
-      return {ptr, output, false}; // We found an error.
-    }
-    if (ascii) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
-      continue;
-    }
-    const __m512i lane0 = broadcast_epi128<0>(utf8);
-    const __m512i lane1 = broadcast_epi128<1>(utf8);
-    int valid_count0;
-    __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-    const __m512i lane2 = broadcast_epi128<2>(utf8);
-    int valid_count1;
-    __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-    if (valid_count0 + valid_count1 <= 16) {
-      vec0 = _mm512_mask_expand_epi32(
-          vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-      valid_count0 += valid_count1;
-      vec0 = expand_utf8_to_utf32(vec0);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-    } else {
-      vec0 = expand_utf8_to_utf32(vec0);
-      vec1 = expand_utf8_to_utf32(vec1);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+    if (!match_system(big_endian)) {
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in0 = _mm256_shuffle_epi8(in0, swap);
+      in1 = _mm256_shuffle_epi8(in1, swap);
     }
-    const __m512i lane3 = broadcast_epi128<3>(utf8);
-    int valid_count2;
-    __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
-    uint32_t tmp1;
-    ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
-    const __m512i lane4 = _mm512_set1_epi32(tmp1);
-    int valid_count3;
-    __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
-    if (valid_count2 + valid_count3 <= 16) {
-      vec2 = _mm512_mask_expand_epi32(
-          vec2, __mmask16(((1 << valid_count3) - 1) << valid_count2), vec3);
-      valid_count2 += valid_count3;
-      vec2 = expand_utf8_to_utf32(vec2);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
+
+    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
+    if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) {
+      // Pack 16-bit characters into 8-bit and store in latin1_output
+      const __m256i packed = _mm256_packus_epi16(in0, in1);
+
+      const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000);
+
+      _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result);
+      // Adjust pointers for the next iteration
+      buf += 32;
+      latin1_output += 32;
     } else {
-      vec2 = expand_utf8_to_utf32(vec2);
-      vec3 = expand_utf8_to_utf32(vec3);
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, true)
-      SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, true)
+      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
     }
-    ptr += 4 * 16;
-  }
-  const char *validatedptr = ptr; // validated up to ptr
+  } // while
+  return std::make_pair(buf, latin1_output);
+}
 
-  // For the final pass, we validate 64 bytes, but we only transcode
-  // 3*16 bytes, so we may end up double-validating 16 bytes.
-  if (end - ptr >= 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    bool ascii = checker.check_next_input(utf8);
-    if (checker.errors()) {
-      return {ptr, output, false}; // We found an error.
+template <endianness big_endian>
+std::pair<result, char *>
+avx2_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
+                                         char *latin1_output) {
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+  while (end - buf >= 16) {
+    __m256i in = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+
+    if (!match_system(big_endian)) {
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in = _mm256_shuffle_epi8(in, swap);
     }
-    if (ascii) {
-      SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
-      output += 64;
-      ptr += 64;
+
+    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
+    if (_mm256_testz_si256(in, high_byte_mask)) {
+      __m128i lo = _mm256_extractf128_si256(in, 0);
+      __m128i hi = _mm256_extractf128_si256(in, 1);
+      __m128i latin1_packed_lo = _mm_packus_epi16(lo, lo);
+      __m128i latin1_packed_hi = _mm_packus_epi16(hi, hi);
+      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output),
+                       latin1_packed_lo);
+      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output + 8),
+                       latin1_packed_hi);
+      buf += 16;
+      latin1_output += 16;
     } else {
-      const __m512i lane0 = broadcast_epi128<0>(utf8);
-      const __m512i lane1 = broadcast_epi128<1>(utf8);
-      int valid_count0;
-      __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
-      const __m512i lane2 = broadcast_epi128<2>(utf8);
-      int valid_count1;
-      __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
-      if (valid_count0 + valid_count1 <= 16) {
-        vec0 = _mm512_mask_expand_epi32(
-            vec0, __mmask16(((1 << valid_count1) - 1) << valid_count0), vec1);
-        valid_count0 += valid_count1;
-        vec0 = expand_utf8_to_utf32(vec0);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-      } else {
-        vec0 = expand_utf8_to_utf32(vec0);
-        vec1 = expand_utf8_to_utf32(vec1);
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
-        SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
+      // Fallback to scalar code for handling errors
+      for (int k = 0; k < 16; k++) {
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if (word <= 0xff) {
+          *latin1_output++ = char(word);
+        } else {
+          return std::make_pair(
+              result{error_code::TOO_LARGE, (size_t)(buf - start + k)},
+              latin1_output);
+        }
       }
-
-      const __m512i lane3 = broadcast_epi128<3>(utf8);
-      SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)
-
-      ptr += 3 * 16;
+      buf += 16;
     }
-    validatedptr += 4 * 16;
-  }
-  if (end != validatedptr) {
-    const __m512i utf8 =
-        _mm512_maskz_loadu_epi8(~UINT64_C(0) >> (64 - (end - validatedptr)),
-                                (const __m512i *)validatedptr);
-    checker.check_next_input(utf8);
-  }
-  checker.check_eof();
-  if (checker.errors()) {
-    return {ptr, output, false}; // We found an error.
-  }
-  return {ptr, output, true};
+  } // while
+  return std::make_pair(result{error_code::SUCCESS, (size_t)(buf - start)},
+                        latin1_output);
 }
-/* end file src/icelake/icelake_from_utf8.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf8_to_latin1.inl.cpp */
-// file included directly
+/* end file src/haswell/avx2_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-// File contains conversion procedure from possibly invalid UTF-8 strings.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/haswell/avx2_convert_utf16_to_utf8.cpp */
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
 
-template <bool is_remaining>
-simdutf_really_inline size_t process_block_from_utf8_to_latin1(
-    const char *buf, size_t len, char *latin_output, __m512i minus64,
-    __m512i one, __mmask64 *next_leading_ptr, __mmask64 *next_bit6_ptr) {
-  __mmask64 load_mask =
-      is_remaining ? _bzhi_u64(~0ULL, (unsigned int)len) : ~0ULL;
-  __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)buf);
-  __mmask64 nonascii = _mm512_movepi8_mask(input);
-  if (nonascii == 0) {
-    if (*next_leading_ptr) { // If we ended with a leading byte, it is an error.
-      return 0;              // Indicates error
-    }
-    is_remaining
-        ? _mm512_mask_storeu_epi8((__m512i *)latin_output, load_mask, input)
-        : _mm512_storeu_si512((__m512i *)latin_output, input);
-    return len;
-  }
+    We consider three cases:
+    1. an input register contains no surrogates and each value
+       is in range 0x0000 .. 0x07ff.
+    2. an input register contains no surrogates and values are
+       is in range 0x0000 .. 0xffff.
+    3. an input register contains surrogates --- i.e. codepoints
+       can have 16 or 32 bits.
 
-  const __mmask64 leading = _mm512_cmpge_epu8_mask(input, minus64);
+    Ad 1.
 
-  __m512i highbits = _mm512_xor_si512(input, _mm512_set1_epi8(-62));
-  __mmask64 invalid_leading_bytes =
-      _mm512_mask_cmpgt_epu8_mask(leading, highbits, one);
+    When values are less than 0x0800, it means that a 16-bit code unit
+    can be converted into: 1) single UTF8 byte (when it is an ASCII
+    char) or 2) two UTF8 bytes.
 
-  if (invalid_leading_bytes) {
-    return 0; // Indicates error
-  }
+    For this case we do only some shuffle to obtain these 2-byte
+    codes and finally compress the whole SSE register with a single
+    shuffle.
 
-  __mmask64 leading_shift = (leading << 1) | *next_leading_ptr;
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
 
-  if ((nonascii ^ leading) != leading_shift) {
-    return 0; // Indicates error
-  }
+    Ad 2.
 
-  const __mmask64 bit6 = _mm512_cmpeq_epi8_mask(highbits, one);
-  input =
-      _mm512_mask_sub_epi8(input, (bit6 << 1) | *next_bit6_ptr, input, minus64);
+    When values fit in 16-bit code units, but are above 0x07ff, then
+    a single word may produce one, two or three UTF8 bytes.
 
-  __mmask64 retain = ~leading & load_mask;
-  __m512i output = _mm512_maskz_compress_epi8(retain, input);
-  int64_t written_out = count_ones(retain);
-  if (written_out == 0) {
-    return 0; // Indicates error
-  }
-  *next_bit6_ptr = bit6 >> 63;
-  *next_leading_ptr = leading >> 63;
+    We prepare data for all these three cases in two registers.
+    The first register contains lower two UTF8 bytes (used in all
+    cases), while the second one contains just the third byte for
+    the three-UTF8-bytes case.
 
-  __mmask64 store_mask = ~UINT64_C(0) >> (64 - written_out);
+    Finally these two registers are interleaved forming eight-element
+    array of 32-bit values. The array spans two SSE registers.
+    The bytes from the registers are compressed using two shuffles.
 
-  _mm512_mask_storeu_epi8((__m512i *)latin_output, store_mask, output);
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
 
-  return written_out;
-}
 
-size_t utf8_to_latin1_avx512(const char *&inbuf, size_t len,
-                             char *&inlatin_output) {
-  const char *buf = inbuf;
-  char *latin_output = inlatin_output;
-  char *start = latin_output;
-  size_t pos = 0;
-  __m512i minus64 = _mm512_set1_epi8(-64); // 11111111111 ... 1100 0000
-  __m512i one = _mm512_set1_epi8(1);
-  __mmask64 next_leading = 0;
-  __mmask64 next_bit6 = 0;
+    To summarize:
+    - We need two 256-entry tables that have 8704 bytes in total.
+*/
 
-  while (pos + 64 <= len) {
-    size_t written = process_block_from_utf8_to_latin1<false>(
-        buf + pos, 64, latin_output, minus64, one, &next_leading, &next_bit6);
-    if (written == 0) {
-      inlatin_output = latin_output;
-      inbuf = buf + pos - next_leading;
-      return 0; // Indicates error at pos or after, or just before pos (too
-                // short error)
-    }
-    latin_output += written;
-    pos += 64;
-  }
+/*
+  Returns a pair: the first unprocessed byte from buf and utf8_output
+  A scalar routing should carry on the conversion of the tail.
+*/
+template <endianness big_endian>
+std::pair<const char16_t *, char *>
+avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) {
+  const char16_t *end = buf + len;
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-  if (pos < len) {
-    size_t remaining = len - pos;
-    size_t written = process_block_from_utf8_to_latin1<true>(
-        buf + pos, remaining, latin_output, minus64, one, &next_leading,
-        &next_bit6);
-    if (written == 0) {
-      inbuf = buf + pos - next_leading;
-      inlatin_output = latin_output;
-      return 0; // Indicates error at pos or after, or just before pos (too
-                // short error)
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+    if (big_endian) {
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in = _mm256_shuffle_epi8(in, swap);
     }
-    latin_output += written;
-  }
-  if (next_leading) {
-    inbuf = buf + len - next_leading;
-    inlatin_output = latin_output;
-    return 0; // Indicates error at end of buffer
-  }
-  inlatin_output = latin_output;
-  inbuf += len;
-  return size_t(latin_output - start);
-}
-/* end file src/icelake/icelake_convert_utf8_to_latin1.inl.cpp */
-/* begin file src/icelake/icelake_convert_valid_utf8_to_latin1.inl.cpp */
-// file included directly
+    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
+    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
+    if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
+      // 1. pack the bytes
+      const __m128i utf8_packed = _mm_packus_epi16(
+          _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1));
+      // 2. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
+      // 3. adjust pointers
+      buf += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
+    }
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-// File contains conversion procedure from valid UTF-8 strings.
+    // no bits set above 11th bit
+    const __m256i one_or_two_bytes_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
+    const uint32_t one_or_two_bytes_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
+    if (one_or_two_bytes_bitmask == 0xffffffff) {
 
-template <bool is_remaining>
-simdutf_really_inline size_t process_valid_block_from_utf8_to_latin1(
-    const char *buf, size_t len, char *latin_output, __m512i minus64,
-    __m512i one, __mmask64 *next_leading_ptr, __mmask64 *next_bit6_ptr) {
-  __mmask64 load_mask =
-      is_remaining ? _bzhi_u64(~0ULL, (unsigned int)len) : ~0ULL;
-  __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)buf);
-  __mmask64 nonascii = _mm512_movepi8_mask(input);
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
 
-  if (nonascii == 0) {
-    is_remaining
-        ? _mm512_mask_storeu_epi8((__m512i *)latin_output, load_mask, input)
-        : _mm512_storeu_si512((__m512i *)latin_output, input);
-    return len;
-  }
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const __m256i t0 = _mm256_slli_epi16(in, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const __m256i t2 = _mm256_and_si256(in, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const __m256i t3 = _mm256_or_si256(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const __m256i t4 = _mm256_or_si256(t3, v_c080);
 
-  __mmask64 leading = _mm512_cmpge_epu8_mask(input, minus64);
+      // 2. merge ASCII and 2-byte codewords
+      const __m256i utf8_unpacked =
+          _mm256_blendv_epi8(t4, in, one_byte_bytemask);
 
-  __m512i highbits = _mm512_xor_si512(input, _mm512_set1_epi8(-62));
+      // 3. prepare bitmask for 8-bit lookup
+      const uint32_t M0 = one_byte_bitmask & 0x55555555;
+      const uint32_t M1 = M0 >> 7;
+      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+      // 4. pack the bytes
 
-  *next_leading_ptr = leading >> 63;
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+      const uint8_t *row_2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
+                                                                       16)][0];
 
-  __mmask64 bit6 = _mm512_cmpeq_epi8_mask(highbits, one);
-  input =
-      _mm512_mask_sub_epi8(input, (bit6 << 1) | *next_bit6_ptr, input, minus64);
-  *next_bit6_ptr = bit6 >> 63;
+      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
 
-  __mmask64 retain = ~leading & load_mask;
-  __m512i output = _mm512_maskz_compress_epi8(retain, input);
-  int64_t written_out = count_ones(retain);
-  if (written_out == 0) {
-    return 0; // Indicates error
-  }
-  __mmask64 store_mask = ~UINT64_C(0) >> (64 - written_out);
-  // Optimization opportunity: sometimes, masked writes are not needed.
-  _mm512_mask_storeu_epi8((__m512i *)latin_output, store_mask, output);
-  return written_out;
-}
+      const __m256i utf8_packed = _mm256_shuffle_epi8(
+          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+      // 5. store bytes
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_castsi256_si128(utf8_packed));
+      utf8_output += row[0];
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_extractf128_si256(utf8_packed, 1));
+      utf8_output += row_2[0];
+
+      // 6. adjust pointers
+      buf += 16;
+      continue;
+    }
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const __m256i surrogates_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+
+    // bitmask = 0x0000 if there are no surrogates
+    //         = 0xc000 if the last word is a surrogate
+    const uint32_t surrogates_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x00000000) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+      const __m256i dup_even = _mm256_setr_epi16(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+
+      /* In this branch we handle three cases:
+         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
+
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
+
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
+
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
+
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+
+      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
+      const __m256i s0 = _mm256_srli_epi16(in, 4);
+      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
+      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
+      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
+      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
+      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
+      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
+                                             simdutf_vec(0b0100000000000000));
+      const __m256i s4 = _mm256_xor_si256(s3, m0);
+#undef simdutf_vec
 
-size_t valid_utf8_to_latin1_avx512(const char *buf, size_t len,
-                                   char *latin_output) {
-  char *start = latin_output;
-  size_t pos = 0;
-  __m512i minus64 = _mm512_set1_epi8(-64); // 11111111111 ... 1100 0000
-  __m512i one = _mm512_set1_epi8(1);
-  __mmask64 next_leading = 0;
-  __mmask64 next_bit6 = 0;
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-  while (pos + 64 <= len) {
-    size_t written = process_valid_block_from_utf8_to_latin1<false>(
-        buf + pos, 64, latin_output, minus64, one, &next_leading, &next_bit6);
-    latin_output += written;
-    pos += 64;
-  }
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-  if (pos < len) {
-    size_t remaining = len - pos;
-    size_t written = process_valid_block_from_utf8_to_latin1<true>(
-        buf + pos, remaining, latin_output, minus64, one, &next_leading,
-        &next_bit6);
-    latin_output += written;
-  }
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-  return (size_t)(latin_output - start);
-}
-/* end file src/icelake/icelake_convert_valid_utf8_to_latin1.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf16_to_latin1.inl.cpp */
-// file included directly
-template <endianness big_endian>
-size_t icelake_convert_utf16_to_latin1(const char16_t *buf, size_t len,
-                                       char *latin1_output) {
-  const char16_t *end = buf + len;
-  __m512i v_0xFF = _mm512_set1_epi16(0xff);
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  __m512i shufmask = _mm512_set_epi8(
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
-      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
-  while (end - buf >= 32) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
-    if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
-    }
-    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
-      return 0;
-    }
-    _mm256_storeu_si256(
-        (__m256i *)latin1_output,
-        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
-    latin1_output += 32;
-    buf += 32;
-  }
-  if (buf < end) {
-    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
-    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
-    if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
-    }
-    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
-      return 0;
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xF800) != 0xD800) {
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word =
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(nullptr, utf8_output);
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf8_output++ = char((value >> 18) | 0b11110000);
+          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((value & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
     }
-    _mm256_mask_storeu_epi8(
-        latin1_output, mask,
-        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
-  }
-  return len;
+  } // while
+  return std::make_pair(buf, utf8_output);
 }
 
+/*
+  Returns a pair: a result struct and utf8_output.
+  If there is an error, the count field of the result is the position of the
+  error. Otherwise, it is the position of the first unprocessed byte in buf
+  (even if finished). A scalar routing should carry on the conversion of the
+  tail if needed.
+*/
 template <endianness big_endian>
 std::pair<result, char *>
-icelake_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
-                                            char *latin1_output) {
-  const char16_t *end = buf + len;
+avx2_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
+                                       char *utf8_output) {
   const char16_t *start = buf;
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  __m512i v_0xFF = _mm512_set1_epi16(0xff);
-  __m512i shufmask = _mm512_set_epi8(
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
-      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
-  while (end - buf >= 32) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+  const char16_t *end = buf + len;
+
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
     if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
-    }
-    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
-      uint16_t word;
-      while ((word = (big_endian ? scalar::utf16::swap_bytes(uint16_t(*buf))
-                                 : uint16_t(*buf))) <= 0xff) {
-        *latin1_output++ = uint8_t(word);
-        buf++;
-      }
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            latin1_output);
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in = _mm256_shuffle_epi8(in, swap);
     }
-    _mm256_storeu_si256(
-        (__m256i *)latin1_output,
-        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
-    latin1_output += 32;
-    buf += 32;
-  }
-  if (buf < end) {
-    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
-    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
-    if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
+    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
+    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
+    if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
+      // 1. pack the bytes
+      const __m128i utf8_packed = _mm_packus_epi16(
+          _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1));
+      // 2. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
+      // 3. adjust pointers
+      buf += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
     }
-    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-      uint16_t word;
-      while ((word = (big_endian ? scalar::utf16::swap_bytes(uint16_t(*buf))
-                                 : uint16_t(*buf))) <= 0xff) {
-        *latin1_output++ = uint8_t(word);
-        buf++;
-      }
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            latin1_output);
-    }
-    _mm256_mask_storeu_epi8(
-        latin1_output, mask,
-        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
-  }
-  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
-}
-/* end file src/icelake/icelake_convert_utf16_to_latin1.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */
-// file included directly
+    // no bits set above 11th bit
+    const __m256i one_or_two_bytes_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
+    const uint32_t one_or_two_bytes_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
+    if (one_or_two_bytes_bitmask == 0xffffffff) {
 
-/**
- * This function converts the input (inbuf, inlen), assumed to be valid
- * UTF16 (little endian) into UTF-8 (to outbuf). The number of code units
- * written is written to 'outlen' and the function reports the number of input
- * word consumed.
- */
-template <endianness big_endian>
-size_t utf16_to_utf8_avx512i(const char16_t *inbuf, size_t inlen,
-                             unsigned char *outbuf, size_t *outlen) {
-  __m512i in;
-  __mmask32 inmask = _cvtu32_mask32(0x7fffffff);
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  const char16_t *const inbuf_orig = inbuf;
-  const unsigned char *const outbuf_orig = outbuf;
-  int adjust = 0;
-  int carry = 0;
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
 
-  while (inlen >= 32) {
-    in = _mm512_loadu_si512(inbuf);
-    if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
-    }
-    inlen -= 31;
-  lastiteration:
-    inbuf += 31;
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const __m256i t0 = _mm256_slli_epi16(in, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const __m256i t2 = _mm256_and_si256(in, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const __m256i t3 = _mm256_or_si256(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const __m256i t4 = _mm256_or_si256(t3, v_c080);
 
-  failiteration:
-    const __mmask32 is234byte = _mm512_mask_cmp_epu16_mask(
-        inmask, in, _mm512_set1_epi16(0x0080), _MM_CMPINT_NLT);
+      // 2. merge ASCII and 2-byte codewords
+      const __m256i utf8_unpacked =
+          _mm256_blendv_epi8(t4, in, one_byte_bytemask);
 
-    if (_ktestz_mask32_u8(inmask, is234byte)) {
-      // fast path for ASCII only
-      _mm512_mask_cvtepi16_storeu_epi8(outbuf, inmask, in);
-      outbuf += 31;
-      carry = 0;
+      // 3. prepare bitmask for 8-bit lookup
+      const uint32_t M0 = one_byte_bitmask & 0x55555555;
+      const uint32_t M1 = M0 >> 7;
+      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+      // 4. pack the bytes
 
-      if (inlen < 32) {
-        goto tail;
-      } else {
-        continue;
-      }
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+      const uint8_t *row_2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
+                                                                       16)][0];
+
+      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+
+      const __m256i utf8_packed = _mm256_shuffle_epi8(
+          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+      // 5. store bytes
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_castsi256_si128(utf8_packed));
+      utf8_output += row[0];
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_extractf128_si256(utf8_packed, 1));
+      utf8_output += row_2[0];
+
+      // 6. adjust pointers
+      buf += 16;
+      continue;
     }
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const __m256i surrogates_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
 
-    const __mmask32 is12byte =
-        _mm512_cmp_epu16_mask(in, _mm512_set1_epi16(0x0800), _MM_CMPINT_LT);
+    // bitmask = 0x0000 if there are no surrogates
+    //         = 0xc000 if the last word is a surrogate
+    const uint32_t surrogates_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x00000000) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+      const __m256i dup_even = _mm256_setr_epi16(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
 
-    if (_ktestc_mask32_u8(is12byte, inmask)) {
-      // fast path for 1 and 2 byte only
+      /* In this branch we handle three cases:
+         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
 
-      const __m512i twobytes = _mm512_ternarylogic_epi32(
-          _mm512_slli_epi16(in, 8), _mm512_srli_epi16(in, 6),
-          _mm512_set1_epi16(0x3f3f), 0xa8); // (A|B)&C
-      in = _mm512_mask_add_epi16(in, is234byte, twobytes,
-                                 _mm512_set1_epi16(int16_t(0x80c0)));
-      const __m512i cmpmask =
-          _mm512_mask_blend_epi16(inmask, _mm512_set1_epi16(int16_t(0xffff)),
-                                  _mm512_set1_epi16(0x0800));
-      const __mmask64 smoosh =
-          _mm512_cmp_epu8_mask(in, cmpmask, _MM_CMPINT_NLT);
-      const __m512i out = _mm512_maskz_compress_epi8(smoosh, in);
-      _mm512_mask_storeu_epi8(outbuf,
-                              _cvtu64_mask64(_pext_u64(_cvtmask64_u64(smoosh),
-                                                       _cvtmask64_u64(smoosh))),
-                              out);
-      outbuf += 31 + _mm_popcnt_u32(_cvtmask32_u32(is234byte));
-      carry = 0;
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
 
-      if (inlen < 32) {
-        goto tail;
-      } else {
-        continue;
-      }
-    }
-    __m512i lo = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
-    __m512i hi = _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1));
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
 
-    __m512i taglo = _mm512_set1_epi32(0x8080e000);
-    __m512i taghi = taglo;
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
 
-    const __m512i fc00masked =
-        _mm512_and_epi32(in, _mm512_set1_epi16(int16_t(0xfc00)));
-    const __mmask32 hisurr = _mm512_mask_cmp_epu16_mask(
-        inmask, fc00masked, _mm512_set1_epi16(int16_t(0xd800)), _MM_CMPINT_EQ);
-    const __mmask32 losurr = _mm512_cmp_epu16_mask(
-        fc00masked, _mm512_set1_epi16(int16_t(0xdc00)), _MM_CMPINT_EQ);
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
 
-    int carryout = 0;
-    if (!_kortestz_mask32_u8(hisurr, losurr)) {
-      // handle surrogates
+      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
+      const __m256i s0 = _mm256_srli_epi16(in, 4);
+      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
+      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
+      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
+      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
+      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
+      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
+                                             simdutf_vec(0b0100000000000000));
+      const __m256i s4 = _mm256_xor_si256(s3, m0);
+#undef simdutf_vec
 
-      __m512i los = _mm512_alignr_epi32(hi, lo, 1);
-      __m512i his = _mm512_alignr_epi32(lo, hi, 1);
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-      const __mmask32 hisurrhi = _kshiftri_mask32(hisurr, 16);
-      taglo = _mm512_mask_mov_epi32(taglo, __mmask16(hisurr),
-                                    _mm512_set1_epi32(0x808080f0));
-      taghi = _mm512_mask_mov_epi32(taghi, __mmask16(hisurrhi),
-                                    _mm512_set1_epi32(0x808080f0));
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-      lo = _mm512_mask_slli_epi32(lo, __mmask16(hisurr), lo, 10);
-      hi = _mm512_mask_slli_epi32(hi, __mmask16(hisurrhi), hi, 10);
-      los = _mm512_add_epi32(los, _mm512_set1_epi32(0xfca02400));
-      his = _mm512_add_epi32(his, _mm512_set1_epi32(0xfca02400));
-      lo = _mm512_mask_add_epi32(lo, __mmask16(hisurr), lo, los);
-      hi = _mm512_mask_add_epi32(hi, __mmask16(hisurrhi), hi, his);
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-      carryout = _cvtu32_mask32(_kshiftri_mask32(hisurr, 30));
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
 
-      const uint32_t h = _cvtmask32_u32(hisurr);
-      const uint32_t l = _cvtmask32_u32(losurr);
-      // check for mismatched surrogates
-      if ((h + h + carry) ^ l) {
-        const uint32_t lonohi = l & ~(h + h + carry);
-        const uint32_t hinolo = h & ~(l >> 1);
-        inlen = _tzcnt_u32(hinolo | lonohi);
-        inmask = __mmask32(0x7fffffff & ((1U << inlen) - 1));
-        in = _mm512_maskz_mov_epi16(inmask, in);
-        adjust = (int)inlen - 31;
-        inlen = 0;
-        goto failiteration;
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xF800) != 0xD800) {
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word =
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k - 1),
+                utf8_output);
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf8_output++ = char((value >> 18) | 0b11110000);
+          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((value & 0b111111) | 0b10000000);
+        }
       }
+      buf += k;
     }
+  } // while
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
+}
+/* end file src/haswell/avx2_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    hi = _mm512_maskz_mov_epi32(_cvtu32_mask16(0x7fff), hi);
-    carry = carryout;
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/haswell/avx2_convert_utf16_to_utf32.cpp */
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
 
-    __m512i mslo =
-        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), lo);
+    We consider three cases:
+    1. an input register contains no surrogates and each value
+       is in range 0x0000 .. 0x07ff.
+    2. an input register contains no surrogates and values are
+       in range 0x0000 .. 0xffff.
+    3. an input register contains surrogates --- i.e. codepoints
+       can have 16 or 32 bits.
 
-    __m512i mshi =
-        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), hi);
+    Ad 1.
 
-    const __mmask32 outmask = __mmask32(_kandn_mask64(losurr, inmask));
-    const __mmask64 outmhi = _kshiftri_mask64(outmask, 16);
+    When values are less than 0x0800, it means that a 16-bit code unit
+    can be converted into: 1) single UTF8 byte (when it is an ASCII
+    char) or 2) two UTF8 bytes.
 
-    const __mmask32 is1byte = __mmask32(_knot_mask64(is234byte));
-    const __mmask64 is1bhi = _kshiftri_mask64(is1byte, 16);
-    const __mmask64 is12bhi = _kshiftri_mask64(is12byte, 16);
+    For this case we do only some shuffle to obtain these 2-byte
+    codes and finally compress the whole SSE register with a single
+    shuffle.
 
-    taglo = _mm512_mask_mov_epi32(taglo, __mmask16(is12byte),
-                                  _mm512_set1_epi32(0x80c00000));
-    taghi = _mm512_mask_mov_epi32(taghi, __mmask16(is12bhi),
-                                  _mm512_set1_epi32(0x80c00000));
-    __m512i magiclo = _mm512_mask_blend_epi32(__mmask16(outmask),
-                                              _mm512_set1_epi32(0xffffffff),
-                                              _mm512_set1_epi32(0x00010101));
-    __m512i magichi = _mm512_mask_blend_epi32(__mmask16(outmhi),
-                                              _mm512_set1_epi32(0xffffffff),
-                                              _mm512_set1_epi32(0x00010101));
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
 
-    magiclo = _mm512_mask_blend_epi32(__mmask16(outmask),
-                                      _mm512_set1_epi32(0xffffffff),
-                                      _mm512_set1_epi32(0x00010101));
-    magichi = _mm512_mask_blend_epi32(__mmask16(outmhi),
-                                      _mm512_set1_epi32(0xffffffff),
-                                      _mm512_set1_epi32(0x00010101));
+    Ad 2.
 
-    mslo = _mm512_ternarylogic_epi32(mslo, _mm512_set1_epi32(0x3f3f3f3f), taglo,
-                                     0xea); // A&B|C
-    mshi = _mm512_ternarylogic_epi32(mshi, _mm512_set1_epi32(0x3f3f3f3f), taghi,
-                                     0xea);
-    mslo = _mm512_mask_slli_epi32(mslo, __mmask16(is1byte), lo, 24);
+    When values fit in 16-bit code units, but are above 0x07ff, then
+    a single word may produce one, two or three UTF8 bytes.
 
-    mshi = _mm512_mask_slli_epi32(mshi, __mmask16(is1bhi), hi, 24);
+    We prepare data for all these three cases in two registers.
+    The first register contains lower two UTF8 bytes (used in all
+    cases), while the second one contains just the third byte for
+    the three-UTF8-bytes case.
 
-    const __mmask64 wantlo =
-        _mm512_cmp_epu8_mask(mslo, magiclo, _MM_CMPINT_NLT);
-    const __mmask64 wanthi =
-        _mm512_cmp_epu8_mask(mshi, magichi, _MM_CMPINT_NLT);
-    const __m512i outlo = _mm512_maskz_compress_epi8(wantlo, mslo);
-    const __m512i outhi = _mm512_maskz_compress_epi8(wanthi, mshi);
-    const uint64_t wantlo_uint64 = _cvtmask64_u64(wantlo);
-    const uint64_t wanthi_uint64 = _cvtmask64_u64(wanthi);
+    Finally these two registers are interleaved forming eight-element
+    array of 32-bit values. The array spans two SSE registers.
+    The bytes from the registers are compressed using two shuffles.
 
-    uint64_t advlo = _mm_popcnt_u64(wantlo_uint64);
-    uint64_t advhi = _mm_popcnt_u64(wanthi_uint64);
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
 
-    _mm512_mask_storeu_epi8(
-        outbuf, _cvtu64_mask64(_pext_u64(wantlo_uint64, wantlo_uint64)), outlo);
-    _mm512_mask_storeu_epi8(
-        outbuf + advlo, _cvtu64_mask64(_pext_u64(wanthi_uint64, wanthi_uint64)),
-        outhi);
-    outbuf += advlo + advhi;
-  }
-  outbuf += -adjust;
 
-tail:
-  if (inlen != 0) {
-    // We must have inlen < 31.
-    inmask = _cvtu32_mask32((1U << inlen) - 1);
-    in = _mm512_maskz_loadu_epi16(inmask, inbuf);
-    if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
-    }
-    adjust = (int)inlen - 31;
-    inlen = 0;
-    goto lastiteration;
-  }
-  *outlen = (outbuf - outbuf_orig) + adjust;
-  return ((inbuf - inbuf_orig) + adjust);
-}
-/* end file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
-// file included directly
+    To summarize:
+    - We need two 256-entry tables that have 8704 bytes in total.
+*/
 
 /*
   Returns a pair: the first unprocessed byte from buf and utf32_output
   A scalar routing should carry on the conversion of the tail.
 */
 template <endianness big_endian>
-std::tuple<const char16_t *, char32_t *, bool>
-convert_utf16_to_utf32(const char16_t *buf, size_t len,
-                       char32_t *utf32_output) {
+std::pair<const char16_t *, char32_t *>
+avx2_convert_utf16_to_utf32(const char16_t *buf, size_t len,
+                            char32_t *utf32_output) {
   const char16_t *end = buf + len;
-  const __m512i v_fc00 = _mm512_set1_epi16((uint16_t)0xfc00);
-  const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);
-  const __m512i v_dc00 = _mm512_set1_epi16((uint16_t)0xdc00);
-  __mmask32 carry{0};
-  const __m512i byteflip = _mm512_setr_epi64(
-      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-      0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  while (std::distance(buf, end) >= 32) {
-    // Always safe because buf + 32 <= end so that end - buf >= 32 bytes:
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+
+  while (end - buf >= 16) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
     if (big_endian) {
-      in = _mm512_shuffle_epi8(in, byteflip);
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in = _mm256_shuffle_epi8(in, swap);
     }
 
-    // H - bitmask for high surrogates
-    const __mmask32 H =
-        _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_d800);
-    // H - bitmask for low surrogates
-    const __mmask32 L =
-        _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_dc00);
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const __m256i surrogates_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
 
-    if ((H | L)) {
+    // bitmask = 0x0000 if there are no surrogates
+    //         = 0xc000 if the last word is a surrogate
+    const uint32_t surrogates_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x00000000) {
+      // case: we extend all sixteen 16-bit code units to sixteen 32-bit code
+      // units
+      _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
+                          _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
+      _mm256_storeu_si256(
+          reinterpret_cast<__m256i *>(utf32_output + 8),
+          _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in, 1)));
+      utf32_output += 16;
+      buf += 16;
       // surrogate pair(s) in a register
-      const __mmask32 V =
-          (L ^
-           (carry | (H << 1))); // A high surrogate must be followed by low one
-                                // and a low one must be preceded by a high one.
-                                // If valid, V should be equal to 0
-
-      if (V == 0) {
-        // valid case
-        /*
-            Input surrogate pair:
-            |1101.11aa.aaaa.aaaa|1101.10bb.bbbb.bbbb|
-                low surrogate      high surrogate
-        */
-        /*  1. Expand all code units to 32-bit code units
-            in
-           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
-        */
-        const __m512i first = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
-        const __m512i second =
-            _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1));
-
-        /*  2. Shift by one 16-bit word to align low surrogates with high
-           surrogates in
-           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
-            shifted
-           |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
-        */
-        const __m512i shifted_first = _mm512_alignr_epi32(second, first, 1);
-        const __m512i shifted_second =
-            _mm512_alignr_epi32(_mm512_setzero_si512(), second, 1);
-
-        /*  3. Align all high surrogates in first and second by shifting to the
-           left by 10 bits
-            |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
-        */
-        const __m512i aligned_first =
-            _mm512_mask_slli_epi32(first, (__mmask16)H, first, 10);
-        const __m512i aligned_second =
-            _mm512_mask_slli_epi32(second, (__mmask16)(H >> 16), second, 10);
-
-        /*  4. Remove surrogate prefixes and add offset 0x10000 by adding in,
-           shifted and constant in
-           |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
-            shifted
-           |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
-            constant|1111.1100.1010.0000.0010.0100.0000.0000|1111.1100.1010.0000.0010.0100.0000.0000|
-        */
-        const __m512i constant = _mm512_set1_epi32((uint32_t)0xfca02400);
-        const __m512i added_first = _mm512_mask_add_epi32(
-            aligned_first, (__mmask16)H, aligned_first, shifted_first);
-        const __m512i utf32_first = _mm512_mask_add_epi32(
-            added_first, (__mmask16)H, added_first, constant);
-
-        const __m512i added_second =
-            _mm512_mask_add_epi32(aligned_second, (__mmask16)(H >> 16),
-                                  aligned_second, shifted_second);
-        const __m512i utf32_second = _mm512_mask_add_epi32(
-            added_second, (__mmask16)(H >> 16), added_second, constant);
-
-        //  5. Store all valid UTF-32 code units (low surrogate positions and
-        //  32nd word are invalid)
-        const __mmask32 valid = ~L & 0x7fffffff;
-        // We deliberately do a _mm512_maskz_compress_epi32 followed by
-        // storeu_epi32 to ease performance portability to Zen 4.
-        const __m512i compressed_first =
-            _mm512_maskz_compress_epi32((__mmask16)(valid), utf32_first);
-        const size_t howmany1 = count_ones((uint16_t)(valid));
-        _mm512_storeu_si512((__m512i *)utf32_output, compressed_first);
-        utf32_output += howmany1;
-        const __m512i compressed_second =
-            _mm512_maskz_compress_epi32((__mmask16)(valid >> 16), utf32_second);
-        const size_t howmany2 = count_ones((uint16_t)(valid >> 16));
-        // The following could be unsafe in some cases?
-        //_mm512_storeu_epi32((__m512i *) utf32_output, compressed_second);
-        _mm512_mask_storeu_epi32((__m512i *)utf32_output,
-                                 __mmask16((1 << howmany2) - 1),
-                                 compressed_second);
-        utf32_output += howmany2;
-        // Only process 31 code units, but keep track if the 31st word is a high
-        // surrogate as a carry
-        buf += 31;
-        carry = (H >> 30) & 0x1;
-      } else {
-        // invalid case
-        return std::make_tuple(buf + carry, utf32_output, false);
-      }
     } else {
-      // no surrogates
-      // extend all thirty-two 16-bit code units to thirty-two 32-bit code units
-      _mm512_storeu_si512((__m512i *)(utf32_output),
-                          _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in)));
-      _mm512_storeu_si512(
-          (__m512i *)(utf32_output) + 1,
-          _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1)));
-      utf32_output += 32;
-      buf += 32;
-      carry = 0;
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xF800) != 0xD800) {
+          // No surrogate pair
+          *utf32_output++ = char32_t(word);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word =
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(nullptr, utf32_output);
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf32_output++ = char32_t(value);
+        }
+      }
+      buf += k;
     }
   } // while
-  return std::make_tuple(buf + carry, utf32_output, true);
+  return std::make_pair(buf, utf32_output);
 }
-/* end file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf32_to_latin1.inl.cpp */
-// file included directly
-size_t icelake_convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                                       char *latin1_output) {
-  const char32_t *end = buf + len;
-  __m512i v_0xFF = _mm512_set1_epi32(0xff);
-  __m512i shufmask = _mm512_set_epi8(
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
-      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
+
+/*
+  Returns a pair: a result struct and utf8_output.
+  If there is an error, the count field of the result is the position of the
+  error. Otherwise, it is the position of the first unprocessed byte in buf
+  (even if finished). A scalar routing should carry on the conversion of the
+  tail if needed.
+*/
+template <endianness big_endian>
+std::pair<result, char32_t *>
+avx2_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
+                                        char32_t *utf32_output) {
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+
   while (end - buf >= 16) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
-    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
-      return 0;
-    }
-    _mm_storeu_si128(
-        (__m128i *)latin1_output,
-        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
-    latin1_output += 16;
-    buf += 16;
-  }
-  if (buf < end) {
-    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
-    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
-    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
-      return 0;
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+    if (big_endian) {
+      const __m256i swap = _mm256_setr_epi8(
+          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
+          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
+      in = _mm256_shuffle_epi8(in, swap);
     }
-    _mm_mask_storeu_epi8(
-        latin1_output, mask,
-        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
-  }
-  return len;
-}
 
-std::pair<result, char *>
-icelake_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                            char *latin1_output) {
-  const char32_t *end = buf + len;
-  const char32_t *start = buf;
-  __m512i v_0xFF = _mm512_set1_epi32(0xff);
-  __m512i shufmask = _mm512_set_epi8(
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
-      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
-  while (end - buf >= 16) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
-    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
-      while (uint32_t(*buf) <= 0xff) {
-        *latin1_output++ = uint8_t(*buf++);
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const __m256i surrogates_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+
+    // bitmask = 0x0000 if there are no surrogates
+    //         = 0xc000 if the last word is a surrogate
+    const uint32_t surrogates_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x00000000) {
+      // case: we extend all sixteen 16-bit code units to sixteen 32-bit code
+      // units
+      _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
+                          _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
+      _mm256_storeu_si256(
+          reinterpret_cast<__m256i *>(utf32_output + 8),
+          _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in, 1)));
+      utf32_output += 16;
+      buf += 16;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
       }
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            latin1_output);
-    }
-    _mm_storeu_si128(
-        (__m128i *)latin1_output,
-        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
-    latin1_output += 16;
-    buf += 16;
-  }
-  if (buf < end) {
-    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
-    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
-    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
-      while (uint32_t(*buf) <= 0xff) {
-        *latin1_output++ = uint8_t(*buf++);
+      for (; k < forward; k++) {
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
+        if ((word & 0xF800) != 0xD800) {
+          // No surrogate pair
+          *utf32_output++ = char32_t(word);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word =
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k - 1),
+                utf32_output);
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf32_output++ = char32_t(value);
+        }
       }
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            latin1_output);
-    }
-    _mm_mask_storeu_epi8(
-        latin1_output, mask,
-        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
-  }
-  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
-}
-/* end file src/icelake/icelake_convert_utf32_to_latin1.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
-// file included directly
-
-// Todo: currently, this is just the haswell code, optimize for icelake kernel.
-std::pair<const char32_t *, char *>
-avx512_convert_utf32_to_utf8(const char32_t *buf, size_t len,
-                             char *utf8_output) {
-  const char32_t *end = buf + len;
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
-  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
-  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
-  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
-  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
-  __m256i running_max = _mm256_setzero_si256();
-  __m256i forbidden_bytemask = _mm256_setzero_si256();
+      buf += k;
+    }
+  } // while
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf32_output);
+}
+/* end file src/haswell/avx2_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_utf32_to_latin1.cpp */
+std::pair<const char32_t *, char *>
+avx2_convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                             char *latin1_output) {
+  const size_t rounded_len =
+      len & ~0x1F; // Round down to nearest multiple of 32
 
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
-    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);
+  const __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00);
 
-    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
-    // saturation
-    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
-                                        _mm256_and_si256(nextin, v_7fffffff));
-    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
+  for (size_t i = 0; i < rounded_len; i += 4 * 8) {
+    __m256i a = _mm256_loadu_si256((__m256i *)(buf + 0 * 8));
+    __m256i b = _mm256_loadu_si256((__m256i *)(buf + 1 * 8));
+    __m256i c = _mm256_loadu_si256((__m256i *)(buf + 2 * 8));
+    __m256i d = _mm256_loadu_si256((__m256i *)(buf + 3 * 8));
 
-    // Try to apply UTF-16 => UTF-8 routine on 256 bits
-    // (haswell/avx2_convert_utf16_to_utf8.cpp)
+    const __m256i check_combined =
+        _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d));
 
-    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
-      // 1. pack the bytes
-      const __m128i utf8_packed = _mm_packus_epi16(
-          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
-      // 2. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
-      // 3. adjust pointers
-      buf += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
+    if (!_mm256_testz_si256(check_combined, high_bytes_mask)) {
+      return std::make_pair(nullptr, latin1_output);
     }
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-    // no bits set above 11th bit
-    const __m256i one_or_two_bytes_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
-    const uint32_t one_or_two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
-    if (one_or_two_bytes_bitmask == 0xffffffff) {
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
+    b = _mm256_slli_epi32(b, 1 * 8);
+    c = _mm256_slli_epi32(c, 2 * 8);
+    d = _mm256_slli_epi32(d, 3 * 8);
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const __m256i t3 = _mm256_or_si256(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const __m256i t4 = _mm256_or_si256(t3, v_c080);
+    // clang-format off
 
-      // 2. merge ASCII and 2-byte codewords
-      const __m256i utf8_unpacked =
-          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
+    // a  = [.. .. .. a7|.. .. .. a6|.. .. .. a5|.. .. .. a4||.. .. .. a3|.. .. .. a2|.. .. .. a1|.. .. .. a0]
+    // b  = [.. .. b7 ..|.. .. b6 ..|.. .. b5 ..|.. .. b4 ..||.. .. b3 ..|.. .. b2 ..|.. .. b1 ..|.. .. b0 ..]
+    // c  = [.. c7 .. ..|.. c6 .. ..|.. c5 .. ..|.. c4 .. ..||.. c3 .. ..|.. c2 .. ..|.. c1 .. ..|.. c0 .. ..]
+    // d  = [d7 .. .. ..|d6 .. .. ..|d5 .. .. ..|d4 .. .. ..||d3 .. .. ..|d2 .. .. ..|d1 .. .. ..|d0 .. .. ..]
 
-      // 3. prepare bitmask for 8-bit lookup
-      const uint32_t M0 = one_byte_bitmask & 0x55555555;
-      const uint32_t M1 = M0 >> 7;
-      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-      // 4. pack the bytes
+    // t0 = [d7 c7 b7 a7|d6 c6 b6 a6|d5 c5 b5 a5|d4 c4 b4 a4||d3 c3 b3 a3|d2 c2 b2 a2|d1 c1 b1 a1|d0 c0 b0 a0]
+    const __m256i t0 =
+        _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d));
 
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-      const uint8_t *row_2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
-                                                                       16)][0];
+    // shuffle bytes within 128-bit lanes
+    // t1 = [d7 d6 d5 d4|c7 c6 c5 c4|b7 b6 b5 b4|a7 a6 a5 a4||d3 d2 d1 d0|c3 c2 c1 c0|b3 b2 b1 b0|a3 a2 a1 a0]
+    const __m256i shuffle_bytes =
+        _mm256_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
+                         0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
 
-      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+    const __m256i t1 = _mm256_shuffle_epi8(t0, shuffle_bytes);
 
-      const __m256i utf8_packed = _mm256_shuffle_epi8(
-          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-      // 5. store bytes
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_castsi256_si128(utf8_packed));
-      utf8_output += row[0];
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_extractf128_si256(utf8_packed, 1));
-      utf8_output += row_2[0];
+    // reshuffle dwords
+    // t2 = [d7 d6 d5 d4|d3 d2 d1 d0|c7 c6 c5 c4|c3 c2 c1 c0||b7 b6 b5 b4|b3 b2 b1 b0|a7 a6 a5 a4|a3 a2 a1 a0]
+    const __m256i shuffle_dwords = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7);
+    const __m256i t2 = _mm256_permutevar8x32_epi32(t1, shuffle_dwords);
+// clang format on
 
-      // 6. adjust pointers
-      buf += 16;
-      continue;
-    }
-    // Must check for overflow in packing
-    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
-        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
-    if (saturation_bitmask == 0xffffffff) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
-      forbidden_bytemask = _mm256_or_si256(
-          forbidden_bytemask,
-          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));
+    _mm256_storeu_si256((__m256i *)latin1_output, t2);
 
-      const __m256i dup_even = _mm256_setr_epi16(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+    latin1_output += 32;
+    buf += 32;
+  }
 
-      /* In this branch we handle three cases:
-        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+  return std::make_pair(buf, latin1_output);
+}
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+std::pair<result, char *>
+avx2_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                         char *latin1_output) {
+  const size_t rounded_len =
+      len & ~0x1F; // Round down to nearest multiple of 32
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+  const char32_t *start = buf;
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+  const __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00);
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+  for (size_t i = 0; i < rounded_len; i += 4 * 8) {
+    __m256i a = _mm256_loadu_si256((__m256i *)(buf + 0 * 8));
+    __m256i b = _mm256_loadu_si256((__m256i *)(buf + 1 * 8));
+    __m256i c = _mm256_loadu_si256((__m256i *)(buf + 2 * 8));
+    __m256i d = _mm256_loadu_si256((__m256i *)(buf + 3 * 8));
 
-      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
-      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
-      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
-      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
-      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
-      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
-      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
-      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
-                                             simdutf_vec(0b0100000000000000));
-      const __m256i s4 = _mm256_xor_si256(s3, m0);
-#undef simdutf_vec
+    const __m256i check_combined =
+        _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d));
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
+    if (!_mm256_testz_si256(check_combined, high_bytes_mask)) {
+      // Fallback to scalar code for handling errors
+      for (int k = 0; k < 4 * 8; k++) {
+        char32_t codepoint = buf[k];
+        if (codepoint <= 0xFF) {
+          *latin1_output++ = static_cast<char>(codepoint);
+        } else {
+          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
+                                latin1_output);
+        }
+      }
+    }
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
+    b = _mm256_slli_epi32(b, 1 * 8);
+    c = _mm256_slli_epi32(c, 2 * 8);
+    d = _mm256_slli_epi32(d, 3 * 8);
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
+    const __m256i t0 =
+        _mm256_or_si256(_mm256_or_si256(a, b), _mm256_or_si256(c, d));
 
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+    const __m256i shuffle_bytes =
+        _mm256_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15,
+                         0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
 
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+    const __m256i t1 = _mm256_shuffle_epi8(t0, shuffle_bytes);
 
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-    } else {
-      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
-      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
-      // wasteful to use scalar code, but being efficient with SIMD may require
-      // large, non-trivial tables?
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr, utf8_output);
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else { // 4-byte
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr, utf8_output);
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+    const __m256i shuffle_dwords = _mm256_setr_epi32(0, 4, 1, 5, 2, 6, 3, 7);
+    const __m256i t2 = _mm256_permutevar8x32_epi32(t1, shuffle_dwords);
 
-  // check for invalid input
-  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(
-          _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
-    return std::make_pair(nullptr, utf8_output);
-  }
+    _mm256_storeu_si256((__m256i *)latin1_output, t2);
 
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
-    return std::make_pair(nullptr, utf8_output);
+    latin1_output += 32;
+    buf += 32;
   }
 
-  return std::make_pair(buf, utf8_output);
+  return std::make_pair(result(error_code::SUCCESS, buf - start),
+                        latin1_output);
 }
+/* end file src/haswell/avx2_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-// Todo: currently, this is just the haswell code, optimize for icelake kernel.
-std::pair<result, char *>
-avx512_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
-                                         char *utf8_output) {
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/haswell/avx2_convert_utf32_to_utf8.cpp */
+std::pair<const char32_t *, char *>
+avx2_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) {
   const char32_t *end = buf + len;
-  const char32_t *start = buf;
-
   const __m256i v_0000 = _mm256_setzero_si256();
   const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
   const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
   const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
   const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
   const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
-  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
+  __m256i running_max = _mm256_setzero_si256();
+  __m256i forbidden_bytemask = _mm256_setzero_si256();
 
   const size_t safety_margin =
       12; // to avoid overruns, see issue
@@ -25692,14 +33833,7 @@ avx512_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
   while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
     __m256i in = _mm256_loadu_si256((__m256i *)buf);
     __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
-    // Check for too large input
-    const __m256i max_input =
-        _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
-    if (static_cast<uint32_t>(_mm256_movemask_epi8(
-            _mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            utf8_output);
-    }
+    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);
 
     // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
     // saturation
@@ -25790,16 +33924,10 @@ avx512_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
     if (saturation_bitmask == 0xffffffff) {
       // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-
-      // Check for illegal surrogate code units
       const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
-      const __m256i forbidden_bytemask =
-          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
-      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
-          0x0) {
-        return std::make_pair(result(error_code::SURROGATE, buf - start),
-                              utf8_output);
-      }
+      forbidden_bytemask = _mm256_or_si256(
+          forbidden_bytemask,
+          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));
 
       const __m256i dup_even = _mm256_setr_epi16(
           0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
@@ -25853,2720 +33981,2805 @@ avx512_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
       const __m256i s4 = _mm256_xor_si256(s3, m0);
 #undef simdutf_vec
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
-
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
-
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
-
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
-
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
-
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-    } else {
-      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
-      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
-      // wasteful to use scalar code, but being efficient with SIMD may require
-      // large, non-trivial tables?
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k), utf8_output);
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else { // 4-byte
-          if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k), utf8_output);
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
-
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
-}
-/* end file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
-/* begin file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
-// file included directly
-
-// Todo: currently, this is just the haswell code, optimize for icelake kernel.
-template <endianness big_endian>
-std::pair<const char32_t *, char16_t *>
-avx512_convert_utf32_to_utf16(const char32_t *buf, size_t len,
-                              char16_t *utf16_output) {
-  const char32_t *end = buf + len;
-
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-  __m256i forbidden_bytemask = _mm256_setzero_si256();
-
-  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-
-    const __m256i v_00000000 = _mm256_setzero_si256();
-    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
-
-    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
-    const __m256i saturation_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
-
-    if (saturation_bitmask == 0xffffffff) {
-      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
-      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
-      forbidden_bytemask = _mm256_or_si256(
-          forbidden_bytemask,
-          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800));
-
-      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
-                                              _mm256_extractf128_si256(in, 1));
-      if (big_endian) {
-        const __m128i swap =
-            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
-      }
-      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
-      utf16_output += 8;
-      buf += 8;
-    } else {
-      size_t forward = 7;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          *utf16_output++ =
-              big_endian
-                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
-                  : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (big_endian) {
-            high_surrogate =
-                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
-            low_surrogate =
-                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
-        }
-      }
-      buf += k;
-    }
-  }
-
-  // check for invalid input
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
-    return std::make_pair(nullptr, utf16_output);
-  }
-
-  return std::make_pair(buf, utf16_output);
-}
-
-// Todo: currently, this is just the haswell code, optimize for icelake kernel.
-template <endianness big_endian>
-std::pair<result, char16_t *>
-avx512_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
-                                          char16_t *utf16_output) {
-  const char32_t *start = buf;
-  const char32_t *end = buf + len;
-
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-    const __m256i v_00000000 = _mm256_setzero_si256();
-    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
-    const __m256i saturation_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
 
-    if (saturation_bitmask == 0xffffffff) {
-      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
-      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
-      const __m256i forbidden_bytemask =
-          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800);
-      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
-          0x0) {
-        return std::make_pair(result(error_code::SURROGATE, buf - start),
-                              utf16_output);
-      }
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
 
-      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
-                                              _mm256_extractf128_si256(in, 1));
-      if (big_endian) {
-        const __m128i swap =
-            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
-      }
-      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
-      utf16_output += 8;
-      buf += 8;
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
     } else {
-      size_t forward = 7;
+      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
+      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
+      // wasteful to use scalar code, but being efficient with SIMD may require
+      // large, non-trivial tables?
+      size_t forward = 15;
       size_t k = 0;
       if (size_t(end - buf) < forward + 1) {
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
         uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
+        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
           if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k), utf16_output);
+            return std::make_pair(nullptr, utf8_output);
           }
-          *utf16_output++ =
-              big_endian
-                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
-                  : char16_t(word);
-        } else {
-          // will generate a surrogate pair
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else { // 4-byte
           if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k), utf16_output);
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (big_endian) {
-            high_surrogate =
-                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
-            low_surrogate =
-                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
+            return std::make_pair(nullptr, utf8_output);
           }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
         }
       }
       buf += k;
     }
-  }
-
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
-}
-/* end file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
-/* begin file src/icelake/icelake_ascii_validation.inl.cpp */
-// file included directly
-
-bool validate_ascii(const char *buf, size_t len) {
-  const char *end = buf + len;
-  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
-  __m512i running_or = _mm512_setzero_si512();
-  for (; end - buf >= 64; buf += 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)buf);
-    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
-                                           0xf8); // running_or | (utf8 & ascii)
-  }
-  if (buf < end) {
-    const __m512i utf8 = _mm512_maskz_loadu_epi8(
-        (uint64_t(1) << (end - buf)) - 1, (const __m512i *)buf);
-    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
-                                           0xf8); // running_or | (utf8 & ascii)
-  }
-  return (_mm512_test_epi8_mask(running_or, running_or) == 0);
-}
-/* end file src/icelake/icelake_ascii_validation.inl.cpp */
-/* begin file src/icelake/icelake_utf32_validation.inl.cpp */
-// file included directly
-
-bool validate_utf32(const char32_t *buf, size_t len) {
-  if (len == 0) {
-    return true;
-  }
-  const char32_t *end = buf + len;
-
-  const __m512i offset = _mm512_set1_epi32((uint32_t)0xffff2000);
-  __m512i currentmax = _mm512_setzero_si512();
-  __m512i currentoffsetmax = _mm512_setzero_si512();
+  } // while
 
-  while (buf < end - 16) {
-    __m512i utf32 = _mm512_loadu_si512((const __m512i *)buf);
-    buf += 16;
-    currentoffsetmax =
-        _mm512_max_epu32(_mm512_add_epi32(utf32, offset), currentoffsetmax);
-    currentmax = _mm512_max_epu32(utf32, currentmax);
+  // check for invalid input
+  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
+  if (static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(
+          _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
+    return std::make_pair(nullptr, utf8_output);
   }
 
-  __m512i utf32 =
-      _mm512_maskz_loadu_epi32(__mmask16((1 << (end - buf)) - 1), buf);
-  currentoffsetmax =
-      _mm512_max_epu32(_mm512_add_epi32(utf32, offset), currentoffsetmax);
-  currentmax = _mm512_max_epu32(utf32, currentmax);
-
-  const __m512i standardmax = _mm512_set1_epi32((uint32_t)0x10ffff);
-  const __m512i standardoffsetmax = _mm512_set1_epi32((uint32_t)0xfffff7ff);
-  __m512i is_zero =
-      _mm512_xor_si512(_mm512_max_epu32(currentmax, standardmax), standardmax);
-  if (_mm512_test_epi8_mask(is_zero, is_zero) != 0) {
-    return false;
-  }
-  is_zero = _mm512_xor_si512(
-      _mm512_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
-  if (_mm512_test_epi8_mask(is_zero, is_zero) != 0) {
-    return false;
+  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
+    return std::make_pair(nullptr, utf8_output);
   }
 
-  return true;
+  return std::make_pair(buf, utf8_output);
 }
-/* end file src/icelake/icelake_utf32_validation.inl.cpp */
-/* begin file src/icelake/icelake_convert_latin1_to_utf8.inl.cpp */
-// file included directly
 
-static inline size_t latin1_to_utf8_avx512_vec(__m512i input, size_t input_len,
-                                               char *utf8_output,
-                                               int mask_output) {
-  __mmask64 nonascii = _mm512_movepi8_mask(input);
-  size_t output_size = input_len + (size_t)count_ones(nonascii);
+std::pair<result, char *>
+avx2_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
+                                       char *utf8_output) {
+  const char32_t *end = buf + len;
+  const char32_t *start = buf;
 
-  // Mask to denote whether the byte is a leading byte that is not ascii
-  __mmask64 sixth = _mm512_cmpge_epu8_mask(
-      input, _mm512_set1_epi8(-64)); // binary representation of -64: 1100 0000
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
+  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
+  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
+  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
+  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
+  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
 
-  const uint64_t alternate_bits = UINT64_C(0x5555555555555555);
-  uint64_t ascii = ~nonascii;
-  // the bits in ascii are inverted and zeros are interspersed in between them
-  uint64_t maskA = ~_pdep_u64(ascii, alternate_bits);
-  uint64_t maskB = ~_pdep_u64(ascii >> 32, alternate_bits);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-  // interleave bytes from top and bottom halves (abcd...ABCD -> aAbBcCdD)
-  __m512i input_interleaved = _mm512_permutexvar_epi8(
-      _mm512_set_epi32(0x3f1f3e1e, 0x3d1d3c1c, 0x3b1b3a1a, 0x39193818,
-                       0x37173616, 0x35153414, 0x33133212, 0x31113010,
-                       0x2f0f2e0e, 0x2d0d2c0c, 0x2b0b2a0a, 0x29092808,
-                       0x27072606, 0x25052404, 0x23032202, 0x21012000),
-      input);
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
+    // Check for too large input
+    const __m256i max_input =
+        _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
+    if (static_cast<uint32_t>(_mm256_movemask_epi8(
+            _mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            utf8_output);
+    }
 
-  // double size of each byte, and insert the leading byte 1100 0010
+    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
+    // saturation
+    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
+                                        _mm256_and_si256(nextin, v_7fffffff));
+    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
 
-  /*
-  upscale the bytes to 16-bit value, adding the 0b11000000 leading byte in the
-  process. We adjust for the bytes that have their two most significant bits.
-  This takes care of the first 32 bytes, assuming we interleaved the bytes. */
-  __m512i outputA =
-      _mm512_shldi_epi16(input_interleaved, _mm512_set1_epi8(-62), 8);
-  outputA = _mm512_mask_add_epi16(
-      outputA, (__mmask32)sixth, outputA,
-      _mm512_set1_epi16(1 - 0x4000)); // 1- 0x4000 = 1100 0000 0000 0001????
+    // Try to apply UTF-16 => UTF-8 routine on 256 bits
+    // (haswell/avx2_convert_utf16_to_utf8.cpp)
 
-  // in the second 32-bit half, set first or second option based on whether
-  // original input is leading byte (second case) or not (first case)
-  __m512i leadingB =
-      _mm512_mask_blend_epi16((__mmask32)(sixth >> 32),
-                              _mm512_set1_epi16(0x00c2),  // 0000 0000 1101 0010
-                              _mm512_set1_epi16(0x40c3)); // 0100 0000 1100 0011
-  __m512i outputB = _mm512_ternarylogic_epi32(
-      input_interleaved, leadingB, _mm512_set1_epi16((short)0xff00),
-      (240 & 170) ^ 204); // (input_interleaved & 0xff00) ^ leadingB
+    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
+      // 1. pack the bytes
+      const __m128i utf8_packed = _mm_packus_epi16(
+          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
+      // 2. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
+      // 3. adjust pointers
+      buf += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
+    }
+    // no bits set above 7th bit
+    const __m256i one_byte_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
+    const uint32_t one_byte_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
 
-  // prune redundant bytes
-  outputA = _mm512_maskz_compress_epi8(maskA, outputA);
-  outputB = _mm512_maskz_compress_epi8(maskB, outputB);
+    // no bits set above 11th bit
+    const __m256i one_or_two_bytes_bytemask =
+        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
+    const uint32_t one_or_two_bytes_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
+    if (one_or_two_bytes_bitmask == 0xffffffff) {
+      // 1. prepare 2-byte values
+      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+      // expected output   : [110a|aaaa|10bb|bbbb] x 8
+      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
+      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
 
-  size_t output_sizeA = (size_t)count_ones((uint32_t)nonascii) + 32;
+      // t0 = [000a|aaaa|bbbb|bb00]
+      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
+      // t1 = [000a|aaaa|0000|0000]
+      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
+      // t2 = [0000|0000|00bb|bbbb]
+      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
+      // t3 = [000a|aaaa|00bb|bbbb]
+      const __m256i t3 = _mm256_or_si256(t1, t2);
+      // t4 = [110a|aaaa|10bb|bbbb]
+      const __m256i t4 = _mm256_or_si256(t3, v_c080);
 
-  if (mask_output) {
-    if (input_len > 32) { // is the second half of the input vector used?
-      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_sizeA);
-      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
-      utf8_output += output_sizeA;
-      write_mask = _bzhi_u64(~0ULL, (unsigned int)(output_size - output_sizeA));
-      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputB);
-    } else {
-      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_size);
-      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
-    }
-  } else {
-    _mm512_storeu_si512(utf8_output, outputA);
-    utf8_output += output_sizeA;
-    _mm512_storeu_si512(utf8_output, outputB);
-  }
-  return output_size;
-}
+      // 2. merge ASCII and 2-byte codewords
+      const __m256i utf8_unpacked =
+          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
 
-static inline size_t latin1_to_utf8_avx512_branch(__m512i input,
-                                                  char *utf8_output) {
-  __mmask64 nonascii = _mm512_movepi8_mask(input);
-  if (nonascii) {
-    return latin1_to_utf8_avx512_vec(input, 64, utf8_output, 0);
-  } else {
-    _mm512_storeu_si512(utf8_output, input);
-    return 64;
-  }
-}
+      // 3. prepare bitmask for 8-bit lookup
+      const uint32_t M0 = one_byte_bitmask & 0x55555555;
+      const uint32_t M1 = M0 >> 7;
+      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
+      // 4. pack the bytes
 
-size_t latin1_to_utf8_avx512_start(const char *buf, size_t len,
-                                   char *utf8_output) {
-  char *start = utf8_output;
-  size_t pos = 0;
-  // if there's at least 128 bytes remaining, we don't need to mask the output
-  for (; pos + 128 <= len; pos += 64) {
-    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
-    utf8_output += latin1_to_utf8_avx512_branch(input, utf8_output);
-  }
-  // in the last 128 bytes, the first 64 may require masking the output
-  if (pos + 64 <= len) {
-    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
-    utf8_output += latin1_to_utf8_avx512_vec(input, 64, utf8_output, 1);
-    pos += 64;
-  }
-  // with the last 64 bytes, the input also needs to be masked
-  if (pos < len) {
-    __mmask64 load_mask = _bzhi_u64(~0ULL, (unsigned int)(len - pos));
-    __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)(buf + pos));
-    utf8_output += latin1_to_utf8_avx512_vec(input, len - pos, utf8_output, 1);
-  }
-  return (size_t)(utf8_output - start);
-}
-/* end file src/icelake/icelake_convert_latin1_to_utf8.inl.cpp */
-/* begin file src/icelake/icelake_convert_latin1_to_utf16.inl.cpp */
-// file included directly
-template <endianness big_endian>
-size_t icelake_convert_latin1_to_utf16(const char *latin1_input, size_t len,
-                                       char16_t *utf16_output) {
-  size_t rounded_len = len & ~0x1F; // Round down to nearest multiple of 32
+      const uint8_t *row =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
+      const uint8_t *row_2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
+                                                                       16)][0];
 
-  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
-                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  for (size_t i = 0; i < rounded_len; i += 32) {
-    // Load 32 Latin1 characters into a 256-bit register
-    __m256i in = _mm256_loadu_si256((__m256i *)&latin1_input[i]);
-    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
-    __m512i out = _mm512_cvtepu8_epi16(in);
-    if (big_endian) {
-      out = _mm512_shuffle_epi8(out, byteflip);
-    }
-    // Store the results back to memory
-    _mm512_storeu_si512((__m512i *)&utf16_output[i], out);
-  }
-  if (rounded_len != len) {
-    uint32_t mask = uint32_t(1 << (len - rounded_len)) - 1;
-    __m256i in = _mm256_maskz_loadu_epi8(mask, latin1_input + rounded_len);
+      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
+      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
 
-    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
-    __m512i out = _mm512_cvtepu8_epi16(in);
-    if (big_endian) {
-      out = _mm512_shuffle_epi8(out, byteflip);
-    }
-    // Store the results back to memory
-    _mm512_mask_storeu_epi16(utf16_output + rounded_len, mask, out);
-  }
+      const __m256i utf8_packed = _mm256_shuffle_epi8(
+          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
+      // 5. store bytes
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_castsi256_si128(utf8_packed));
+      utf8_output += row[0];
+      _mm_storeu_si128((__m128i *)utf8_output,
+                       _mm256_extractf128_si256(utf8_packed, 1));
+      utf8_output += row_2[0];
 
-  return len;
-}
-/* end file src/icelake/icelake_convert_latin1_to_utf16.inl.cpp */
-/* begin file src/icelake/icelake_convert_latin1_to_utf32.inl.cpp */
-std::pair<const char *, char32_t *>
-avx512_convert_latin1_to_utf32(const char *buf, size_t len,
-                               char32_t *utf32_output) {
-  size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16
+      // 6. adjust pointers
+      buf += 16;
+      continue;
+    }
+    // Must check for overflow in packing
+    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
+        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
+    const uint32_t saturation_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+    if (saturation_bitmask == 0xffffffff) {
+      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
 
-  for (size_t i = 0; i < rounded_len; i += 16) {
-    // Load 16 Latin1 characters into a 128-bit register
-    __m128i in = _mm_loadu_si128((__m128i *)&buf[i]);
+      // Check for illegal surrogate code units
+      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
+      const __m256i forbidden_bytemask =
+          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
+      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
+          0x0) {
+        return std::make_pair(result(error_code::SURROGATE, buf - start),
+                              utf8_output);
+      }
 
-    // Zero extend each set of 8 Latin1 characters to 16 32-bit integers using
-    // vpmovzxbd
-    __m512i out = _mm512_cvtepu8_epi32(in);
+      const __m256i dup_even = _mm256_setr_epi16(
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
+          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
 
-    // Store the results back to memory
-    _mm512_storeu_si512((__m512i *)&utf32_output[i], out);
-  }
+      /* In this branch we handle three cases:
+        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+        single UFT-8 byte
+        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+        UTF-8 bytes
+        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+        three UTF-8 bytes
 
-  // Return pointers pointing to where we left off
-  return std::make_pair(buf + rounded_len, utf32_output + rounded_len);
-}
-/* end file src/icelake/icelake_convert_latin1_to_utf32.inl.cpp */
-/* begin file src/icelake/icelake_base64.inl.cpp */
-// file included directly
-/**
- * References and further reading:
- *
- * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
- * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
- * https://arxiv.org/abs/1910.05109
- *
- * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
- * Instructions, ACM Transactions on the Web 12 (3), 2018.
- * https://arxiv.org/abs/1704.00605
- *
- * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
- * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
- * Request for Comments: 4648.
- *
- * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
- * http://www.alfredklomp.com/programming/sse-base64/. (2014).
- *
- * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
- * acceleration. https://github.com/aklomp/base64. (2014).
- *
- * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
- * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
- *
- * Nick Kopp. 2013. Base64 Encoding on a GPU.
- * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
- */
+        We expand the input word (16-bit) into two code units (32-bit), thus
+        we have room for four bytes. However, we need five distinct bit
+        layouts. Note that the last byte in cases #2 and #3 is the same.
 
-struct block64 {
-  __m512i chunks[1];
-};
+        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+        in register t2.
 
-template <bool base64_url>
-size_t encode_base64(char *dst, const char *src, size_t srclen,
-                     base64_options options) {
-  // credit: Wojciech Muła
-  const uint8_t *input = (const uint8_t *)src;
+        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+        either byte 1 for case #2 or byte 2 for case #3. Note that they
+        differ by exactly one bit.
 
-  uint8_t *out = (uint8_t *)dst;
-  static const char *lookup_tbl =
-      base64_url
-          ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
-          : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+        Finally from these two code units we build proper UTF-8 sequence, taking
+        into account the case (i.e, the number of bytes to write).
+      */
+      /**
+       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+       * t2 => [0ccc|cccc] [10cc|cccc]
+       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+       */
+#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
+      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
+      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
+      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
 
-  const __m512i shuffle_input = _mm512_setr_epi32(
-      0x01020001, 0x04050304, 0x07080607, 0x0a0b090a, 0x0d0e0c0d, 0x10110f10,
-      0x13141213, 0x16171516, 0x191a1819, 0x1c1d1b1c, 0x1f201e1f, 0x22232122,
-      0x25262425, 0x28292728, 0x2b2c2a2b, 0x2e2f2d2e);
-  const __m512i lookup =
-      _mm512_loadu_si512(reinterpret_cast<const __m512i *>(lookup_tbl));
-  const __m512i multi_shifts = _mm512_set1_epi64(UINT64_C(0x3036242a1016040a));
-  size_t size = srclen;
-  __mmask64 input_mask = 0xffffffffffff; // (1 << 48) - 1
-  while (size >= 48) {
-    const __m512i v = _mm512_maskz_loadu_epi8(
-        input_mask, reinterpret_cast<const __m512i *>(input));
-    const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
-    const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
-    const __m512i result = _mm512_permutexvar_epi8(indices, lookup);
-    _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), result);
-    out += 64;
-    input += 48;
-    size -= 48;
-  }
-  input_mask = ((__mmask64)1 << size) - 1;
-  const __m512i v = _mm512_maskz_loadu_epi8(
-      input_mask, reinterpret_cast<const __m512i *>(input));
-  const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
-  const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
-  bool padding_needed =
-      (((options & base64_url) == 0) ^
-       ((options & base64_reverse_padding) == base64_reverse_padding));
-  size_t padding_amount = ((size % 3) > 0) ? (3 - (size % 3)) : 0;
-  size_t output_len = ((size + 2) / 3) * 4;
-  size_t non_padded_output_len = output_len - padding_amount;
-  if (!padding_needed) {
-    output_len = non_padded_output_len;
-  }
-  __mmask64 output_mask = output_len == 64 ? (__mmask64)UINT64_MAX
-                                           : ((__mmask64)1 << output_len) - 1;
-  __m512i result = _mm512_mask_permutexvar_epi8(
-      _mm512_set1_epi8('='), ((__mmask64)1 << non_padded_output_len) - 1,
-      indices, lookup);
-  _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out), output_mask,
-                          result);
-  return (size_t)(out - (uint8_t *)dst) + output_len;
-}
+      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
+      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
+      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
+      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
+      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
+      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
+      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
+      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
+      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
+                                             simdutf_vec(0b0100000000000000));
+      const __m256i s4 = _mm256_xor_si256(s3, m0);
+#undef simdutf_vec
 
-template <bool base64_url, bool ignore_garbage>
-static inline uint64_t to_base64_mask(block64 *b, uint64_t *error,
-                                      uint64_t input_mask = UINT64_MAX) {
-  __m512i input = b->chunks[0];
-  const __m512i ascii_space_tbl = _mm512_set_epi8(
-      0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10,
-      9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0,
-      0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32);
-  __m512i lookup0;
-  if (base64_url) {
-    lookup0 = _mm512_set_epi8(
-        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
-        52, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128, -128,
-        -128, -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128,
-        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1,
-        -128, -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -1);
-  } else {
-    lookup0 = _mm512_set_epi8(
-        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
-        52, 63, -128, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128,
-        -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128, -128,
-        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1, -128,
-        -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -128);
-  }
-  __m512i lookup1;
-  if (base64_url) {
-    lookup1 = _mm512_set_epi8(
-        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
-        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
-        63, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15,
-        14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
-  } else {
-    lookup1 = _mm512_set_epi8(
-        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
-        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
-        -128, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16,
-        15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
-  }
+      // 4. expand code units 16-bit => 32-bit
+      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
+      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
 
-  const __m512i translated = _mm512_permutex2var_epi8(lookup0, input, lookup1);
-  const __m512i combined = _mm512_or_si512(translated, input);
-  const __mmask64 mask = _mm512_movepi8_mask(combined) & input_mask;
-  if (!ignore_garbage && mask) {
-    const __mmask64 spaces =
-        _mm512_cmpeq_epi8_mask(_mm512_shuffle_epi8(ascii_space_tbl, input),
-                               input) &
-        input_mask;
-    *error = (mask ^ spaces);
-  }
-  b->chunks[0] = translated;
+      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
+                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
+      // Due to the wider registers, the following path is less likely to be
+      // useful.
+      /*if(mask == 0) {
+        // We only have three-byte code units. Use fast path.
+        const __m256i shuffle =
+      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
+      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
+      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
+      _mm256_shuffle_epi8(out1, shuffle);
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
+        utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
+        _mm_storeu_si128((__m128i*)utf8_output,
+      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
+        continue;
+      }*/
+      const uint8_t mask0 = uint8_t(mask);
+      const uint8_t *row0 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
+      const __m128i utf8_0 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
 
-  return mask | (~input_mask);
-}
+      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
+      const uint8_t *row1 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
+      const __m128i utf8_1 =
+          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
 
-static inline void copy_block(block64 *b, char *output) {
-  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), b->chunks[0]);
-}
+      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
+      const uint8_t *row2 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
+      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
+      const __m128i utf8_2 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
 
-static inline uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
-  uint64_t nmask = ~mask;
-  __m512i c = _mm512_maskz_compress_epi8(nmask, b->chunks[0]);
-  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), c);
-  return _mm_popcnt_u64(nmask);
-}
+      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
+      const uint8_t *row3 =
+          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
+      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
+      const __m128i utf8_3 =
+          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
 
-// The caller of this function is responsible to ensure that there are 64 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char *src) {
-  b->chunks[0] = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
-}
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
+      utf8_output += row0[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
+      utf8_output += row1[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
+      utf8_output += row2[0];
+      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
+      utf8_output += row3[0];
+      buf += 16;
+    } else {
+      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
+      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
+      // wasteful to use scalar code, but being efficient with SIMD may require
+      // large, non-trivial tables?
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k), utf8_output);
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else { // 4-byte
+          if (word > 0x10FFFF) {
+            return std::make_pair(
+                result(error_code::TOO_LARGE, buf - start + k), utf8_output);
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
+    }
+  } // while
 
-static inline void load_block_partial(block64 *b, const char *src,
-                                      __mmask64 input_mask) {
-  b->chunks[0] = _mm512_maskz_loadu_epi8(
-      input_mask, reinterpret_cast<const __m512i *>(src));
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
 }
+/* end file src/haswell/avx2_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-// The caller of this function is responsible to ensure that there are 128 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char16_t *src) {
-  __m512i m1 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
-  __m512i m2 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src + 32));
-  __m512i p = _mm512_packus_epi16(m1, m2);
-  b->chunks[0] =
-      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
-}
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/haswell/avx2_convert_utf32_to_utf16.cpp */
+template <endianness big_endian>
+std::pair<const char32_t *, char16_t *>
+avx2_convert_utf32_to_utf16(const char32_t *buf, size_t len,
+                            char16_t *utf16_output) {
+  const char32_t *end = buf + len;
 
-static inline void load_block_partial(block64 *b, const char16_t *src,
-                                      __mmask64 input_mask) {
-  __m512i m1 = _mm512_maskz_loadu_epi16((__mmask32)input_mask,
-                                        reinterpret_cast<const __m512i *>(src));
-  __m512i m2 =
-      _mm512_maskz_loadu_epi16((__mmask32)(input_mask >> 32),
-                               reinterpret_cast<const __m512i *>(src + 32));
-  __m512i p = _mm512_packus_epi16(m1, m2);
-  b->chunks[0] =
-      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
-}
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+  __m256i forbidden_bytemask = _mm256_setzero_si256();
 
-static inline void base64_decode(char *out, __m512i str) {
-  const __m512i merge_ab_and_bc =
-      _mm512_maddubs_epi16(str, _mm512_set1_epi32(0x01400140));
-  const __m512i merged =
-      _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
-  const __m512i pack = _mm512_set_epi8(
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
-      52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
-      28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
-      5, 6, 0, 1, 2);
-  const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
-  _mm512_mask_storeu_epi8(
-      (__m512i *)out, 0xffffffffffff,
-      shuffled); // mask would be 0xffffffffffff since we write 48 bytes.
-}
-// decode 64 bytes and output 48 bytes
-static inline void base64_decode_block(char *out, const char *src) {
-  base64_decode(out,
-                _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src)));
-}
-static inline void base64_decode_block(char *out, block64 *b) {
-  base64_decode(out, b->chunks[0]);
-}
+  const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
+  const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
 
-template <bool base64_url, bool ignore_garbage, typename chartype>
-full_result
-compress_decode_base64(char *dst, const chartype *src, size_t srclen,
-                       base64_options options,
-                       last_chunk_handling_options last_chunk_options) {
-  (void)options;
-  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
-                                        : tables::base64::to_base64_value;
-  size_t equallocation =
-      srclen; // location of the first padding character if any
-  size_t equalsigns = 0;
-  // skip trailing spaces
-  while (!ignore_garbage && srclen > 0 &&
-         scalar::base64::is_eight_byte(src[srclen - 1]) &&
-         to_base64[uint8_t(src[srclen - 1])] == 64) {
-    srclen--;
-  }
-  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
-    equallocation = srclen - 1;
-    srclen--;
-    equalsigns = 1;
-    // skip trailing spaces
-    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
-           to_base64[uint8_t(src[srclen - 1])] == 64) {
-      srclen--;
-    }
-    if (srclen > 0 && src[srclen - 1] == '=') {
-      equallocation = srclen - 1;
-      srclen--;
-      equalsigns = 2;
-    }
-  }
-  if (srclen == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation, 0};
-    }
-    return {SUCCESS, 0, 0};
-  }
-  const chartype *const srcinit = src;
-  const char *const dstinit = dst;
-  const chartype *const srcend = src + srclen;
+  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
+    const __m256i in = _mm256_loadu_si256((__m256i *)buf);
 
-  // figure out why block_size == 2 is sometimes best???
-  constexpr size_t block_size = 6;
-  char buffer[block_size * 64];
-  char *bufferptr = buffer;
-  if (srclen >= 64) {
-    const chartype *const srcend64 = src + srclen - 64;
-    while (src <= srcend64) {
-      block64 b;
-      load_block(&b, src);
-      src += 64;
-      uint64_t error = 0;
-      uint64_t badcharmask =
-          to_base64_mask<base64_url, ignore_garbage>(&b, &error);
-      if (!ignore_garbage && error) {
-        src -= 64;
-        size_t error_offset = _tzcnt_u64(error);
-        return {error_code::INVALID_BASE64_CHARACTER,
-                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+    if (simdutf_likely(_mm256_testz_si256(in, v_ffff0000))) {
+      // no bits set above 16th bit <=> can pack to UTF16
+      // without surrogate pairs
+      forbidden_bytemask = _mm256_or_si256(
+          forbidden_bytemask,
+          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800));
+
+      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
+                                              _mm256_extractf128_si256(in, 1));
+      if (big_endian) {
+        const __m128i swap =
+            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
       }
-      if (badcharmask != 0) {
-        // optimization opportunity: check for simple masks like those made of
-        // continuous 1s followed by continuous 0s. And masks containing a
-        // single bad character.
-        bufferptr += compress_block(&b, badcharmask, bufferptr);
-      } else if (bufferptr != buffer) {
-        copy_block(&b, bufferptr);
-        bufferptr += 64;
-      } else {
-        base64_decode_block(dst, &b);
-        dst += 48;
+      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
+      utf16_output += 8;
+      buf += 8;
+    } else {
+      size_t forward = 7;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
       }
-      if (bufferptr >= (block_size - 1) * 64 + buffer) {
-        for (size_t i = 0; i < (block_size - 1); i++) {
-          base64_decode_block(dst, buffer + i * 64);
-          dst += 48;
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFF0000) == 0) {
+          // will not generate a surrogate pair
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(nullptr, utf16_output);
+          }
+          *utf16_output++ =
+              big_endian
+                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
+                  : char16_t(word);
+        } else {
+          // will generate a surrogate pair
+          if (word > 0x10FFFF) {
+            return std::make_pair(nullptr, utf16_output);
+          }
+          word -= 0x10000;
+          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+          if (big_endian) {
+            high_surrogate =
+                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
+            low_surrogate =
+                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
+          }
+          *utf16_output++ = char16_t(high_surrogate);
+          *utf16_output++ = char16_t(low_surrogate);
         }
-        std::memcpy(buffer, buffer + (block_size - 1) * 64,
-                    64); // 64 might be too much
-        bufferptr -= (block_size - 1) * 64;
       }
+      buf += k;
     }
   }
 
-  int last_block_len = (int)(srcend - src);
-  if (last_block_len != 0) {
-    __mmask64 input_mask = ((__mmask64)1 << last_block_len) - 1;
-    block64 b;
-    load_block_partial(&b, src, input_mask);
-    uint64_t error = 0;
-    uint64_t badcharmask =
-        to_base64_mask<base64_url, ignore_garbage>(&b, &error, input_mask);
-    if (!ignore_garbage && error) {
-      size_t error_offset = _tzcnt_u64(error);
-      return {error_code::INVALID_BASE64_CHARACTER,
-              size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
-    }
-    src += last_block_len;
-    bufferptr += compress_block(&b, badcharmask, bufferptr);
+  // check for invalid input
+  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
+    return std::make_pair(nullptr, utf16_output);
   }
 
-  char *buffer_start = buffer;
-  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
-    base64_decode_block(dst, buffer_start);
-    dst += 48;
-  }
+  return std::make_pair(buf, utf16_output);
+}
 
-  if ((bufferptr - buffer_start) != 0) {
-    size_t rem = (bufferptr - buffer_start);
-    int idx = rem % 4;
-    __mmask64 mask = ((__mmask64)1 << rem) - 1;
-    __m512i input = _mm512_maskz_loadu_epi8(mask, buffer_start);
-    size_t output_len = (rem / 4) * 3;
-    __mmask64 output_mask = mask >> (rem - output_len);
-    const __m512i merge_ab_and_bc =
-        _mm512_maddubs_epi16(input, _mm512_set1_epi32(0x01400140));
-    const __m512i merged =
-        _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
-    const __m512i pack = _mm512_set_epi8(
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
-        52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
-        28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
-        5, 6, 0, 1, 2);
-    const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
+template <endianness big_endian>
+std::pair<result, char16_t *>
+avx2_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
+                                        char16_t *utf16_output) {
+  const char32_t *start = buf;
+  const char32_t *end = buf + len;
 
-    if (!ignore_garbage &&
-        last_chunk_options == last_chunk_handling_options::strict &&
-        (idx != 1) && ((idx + equalsigns) & 3) != 0) {
-      // The partial chunk was at src - idx
-      _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-      dst += output_len;
-      return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
-              size_t(dst - dstinit)};
-    } else if (!ignore_garbage &&
-               last_chunk_options ==
-                   last_chunk_handling_options::stop_before_partial &&
-               (idx != 1) && ((idx + equalsigns) & 3) != 0) {
-      // Rewind src to before partial chunk
-      _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-      dst += output_len;
-      src -= idx;
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+
+  const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
+  const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
+  const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
+
+  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
+    const __m256i in = _mm256_loadu_si256((__m256i *)buf);
+
+    if (simdutf_likely(_mm256_testz_si256(in, v_ffff0000))) {
+      // no bits set above 16th bit <=> can pack to UTF16 without surrogate
+      // pairs
+      const __m256i forbidden_bytemask =
+          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800);
+      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
+          0x0) {
+        return std::make_pair(result(error_code::SURROGATE, buf - start),
+                              utf16_output);
+      }
+
+      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
+                                              _mm256_extractf128_si256(in, 1));
+      if (big_endian) {
+        const __m128i swap =
+            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
+      }
+      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
+      utf16_output += 8;
+      buf += 8;
     } else {
-      if (idx == 2) {
-        if (!ignore_garbage &&
-            last_chunk_options == last_chunk_handling_options::strict) {
-          uint32_t triple = (uint32_t(bufferptr[-2]) << 3 * 6) +
-                            (uint32_t(bufferptr[-1]) << 2 * 6);
-          if (triple & 0xffff) {
-            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-            dst += output_len;
-            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
-                    size_t(dst - dstinit)};
+      size_t forward = 7;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFF0000) == 0) {
+          // will not generate a surrogate pair
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return std::make_pair(
+                result(error_code::SURROGATE, buf - start + k), utf16_output);
           }
-        }
-        output_mask = (output_mask << 1) | 1;
-        output_len += 1;
-        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-        dst += output_len;
-      } else if (idx == 3) {
-        if (!ignore_garbage &&
-            last_chunk_options == last_chunk_handling_options::strict) {
-          uint32_t triple = (uint32_t(bufferptr[-3]) << 3 * 6) +
-                            (uint32_t(bufferptr[-2]) << 2 * 6) +
-                            (uint32_t(bufferptr[-1]) << 1 * 6);
-          if (triple & 0xff) {
-            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-            dst += output_len;
-            return {BASE64_EXTRA_BITS, size_t(src - srcinit),
-                    size_t(dst - dstinit)};
+          *utf16_output++ =
+              big_endian
+                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
+                  : char16_t(word);
+        } else {
+          // will generate a surrogate pair
+          if (word > 0x10FFFF) {
+            return std::make_pair(
+                result(error_code::TOO_LARGE, buf - start + k), utf16_output);
+          }
+          word -= 0x10000;
+          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+          if (big_endian) {
+            high_surrogate =
+                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
+            low_surrogate =
+                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
           }
+          *utf16_output++ = char16_t(high_surrogate);
+          *utf16_output++ = char16_t(low_surrogate);
         }
-        output_mask = (output_mask << 2) | 3;
-        output_len += 2;
-        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-        dst += output_len;
-      } else if (!ignore_garbage && idx == 1) {
-        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-        dst += output_len;
-        return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      } else {
-        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
-        dst += output_len;
       }
+      buf += k;
     }
+  }
 
-    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
-        equalsigns > 0) {
-      size_t output_count = size_t(dst - dstinit);
-      if ((output_count % 3 == 0) ||
-          ((output_count % 3) + 1 + equalsigns != 4)) {
-        return {INVALID_BASE64_CHARACTER, equallocation, output_count};
-      }
-    }
+  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
+}
+/* end file src/haswell/avx2_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    return {SUCCESS, srclen, size_t(dst - dstinit)};
-  }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/haswell/avx2_convert_utf8_to_latin1.cpp */
+// depends on "tables/utf8_to_utf16_tables.h"
 
-  if (!ignore_garbage && equalsigns > 0) {
-    if (last_chunk_options == last_chunk_handling_options::strict) {
-      return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
-              size_t(dst - dstinit)};
-    }
-    if (last_chunk_options ==
-        last_chunk_handling_options::stop_before_partial) {
-      return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
-    }
-    if ((size_t(dst - dstinit) % 3 == 0) ||
-        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
-    }
+// Convert up to 12 bytes from utf8 to latin1 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
+size_t convert_masked_utf8_to_latin1(const char *input,
+                                     uint64_t utf8_end_of_code_point_mask,
+                                     char *&latin1_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  const __m128i in = _mm_loadu_si128((__m128i *)input);
+
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask &
+      0xfff; // we are only processing 12 bytes in case it is not all ASCII
+
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process the data in chunks of 12 bytes.
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in);
+    latin1_output += 12; // We wrote 12 characters.
+    return 12;           // We consumed 1 bytes.
   }
-  return {SUCCESS, srclen, size_t(dst - dstinit)};
+  /// We do not have a fast path available, so we fallback.
+  const uint8_t idx =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
+  // this indicates an invalid input:
+  if (idx >= 64) {
+    return consumed;
+  }
+  // Here we should have (idx < 64), if not, there is a bug in the validation or
+  // elsewhere. SIX (6) input code-code units this is a relatively easy scenario
+  // we process SIX (6) input code-code units. The max length in bytes of six
+  // code code units spanning between 1 and 2 bytes each is 12 bytes. On
+  // processors where pdep/pext is fast, we might be able to use a small lookup
+  // table.
+  const __m128i sh =
+      _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
+  const __m128i perm = _mm_shuffle_epi8(in, sh);
+  const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
+  const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
+  __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
+  const __m128i latin1_packed = _mm_packus_epi16(composed, composed);
+  // writing 8 bytes even though we only care about the first 6 bytes.
+  // performance note: it would be faster to use _mm_storeu_si128, we should
+  // investigate.
+  _mm_storel_epi64((__m128i *)latin1_output, latin1_packed);
+  latin1_output += 6; // We wrote 6 bytes.
+  return consumed;
 }
-/* end file src/icelake/icelake_base64.inl.cpp */
+/* end file src/haswell/avx2_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-#include <cstdint>
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/haswell/avx2_base64.cpp */
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
 
-} // namespace
-} // namespace icelake
-} // namespace simdutf
+template <bool base64_url>
+simdutf_really_inline __m256i lookup_pshufb_improved(const __m256i input) {
+  // credit: Wojciech Muła
+  __m256i result = _mm256_subs_epu8(input, _mm256_set1_epi8(51));
+  const __m256i less = _mm256_cmpgt_epi8(_mm256_set1_epi8(26), input);
+  result =
+      _mm256_or_si256(result, _mm256_and_si256(less, _mm256_set1_epi8(13)));
+  __m256i shift_LUT;
+  if (base64_url) {
+    shift_LUT = _mm256_setr_epi8(
+        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0,
 
-namespace simdutf {
-namespace icelake {
+        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0);
+  } else {
+    shift_LUT = _mm256_setr_epi8(
+        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0,
 
-simdutf_warn_unused int
-implementation::detect_encodings(const char *input,
-                                 size_t length) const noexcept {
-  // If there is a BOM, then we trust it.
-  auto bom_encoding = simdutf::BOM::check_bom(input, length);
-  if (bom_encoding != encoding_type::unspecified) {
-    return bom_encoding;
+        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0);
   }
 
-  int out = 0;
-  uint32_t utf16_err = (length % 2);
-  uint32_t utf32_err = (length % 4);
-  uint32_t ends_with_high = 0;
-  avx512_utf8_checker checker{};
-  const __m512i offset = _mm512_set1_epi32((uint32_t)0xffff2000);
-  __m512i currentmax = _mm512_setzero_si512();
-  __m512i currentoffsetmax = _mm512_setzero_si512();
-  const char *ptr = input;
-  const char *end = ptr + length;
-  for (; end - ptr >= 64; ptr += 64) {
-    // utf8 checks
-    const __m512i data = _mm512_loadu_si512((const __m512i *)ptr);
-    checker.check_next_input(data);
+  result = _mm256_shuffle_epi8(shift_LUT, result);
+  return _mm256_add_epi8(result, input);
+}
 
-    // utf16le_checks
-    __m512i diff = _mm512_sub_epi16(data, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    __mmask32 highsurrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-    __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-    utf16_err |= (((highsurrogates << 1) | ends_with_high) != lowsurrogates);
-    ends_with_high = ((highsurrogates & 0x80000000) != 0);
+template <bool isbase64url>
+size_t encode_base64(char *dst, const char *src, size_t srclen,
+                     base64_options options) {
+  // credit: Wojciech Muła
+  const uint8_t *input = (const uint8_t *)src;
 
-    // utf32le checks
-    currentoffsetmax =
-        _mm512_max_epu32(_mm512_add_epi32(data, offset), currentoffsetmax);
-    currentmax = _mm512_max_epu32(data, currentmax);
-  }
+  uint8_t *out = (uint8_t *)dst;
+  const __m256i shuf =
+      _mm256_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1,
 
-  // last block with 0 <= len < 64
-  __mmask64 read_mask = (__mmask64(1) << (end - ptr)) - 1;
-  const __m512i data = _mm512_maskz_loadu_epi8(read_mask, (const __m512i *)ptr);
-  checker.check_next_input(data);
+                      10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1);
+  size_t i = 0;
+  for (; i + 100 <= srclen; i += 96) {
+    const __m128i lo0 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 0));
+    const __m128i hi0 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 1));
+    const __m128i lo1 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 2));
+    const __m128i hi1 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 3));
+    const __m128i lo2 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 4));
+    const __m128i hi2 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 5));
+    const __m128i lo3 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 6));
+    const __m128i hi3 = _mm_loadu_si128(
+        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 7));
 
-  __m512i diff = _mm512_sub_epi16(data, _mm512_set1_epi16(uint16_t(0xD800)));
-  __mmask32 surrogates =
-      _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-  __mmask32 highsurrogates =
-      _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-  __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-  utf16_err |= (((highsurrogates << 1) | ends_with_high) != lowsurrogates);
+    __m256i in0 = _mm256_shuffle_epi8(_mm256_set_m128i(hi0, lo0), shuf);
+    __m256i in1 = _mm256_shuffle_epi8(_mm256_set_m128i(hi1, lo1), shuf);
+    __m256i in2 = _mm256_shuffle_epi8(_mm256_set_m128i(hi2, lo2), shuf);
+    __m256i in3 = _mm256_shuffle_epi8(_mm256_set_m128i(hi3, lo3), shuf);
 
-  currentoffsetmax =
-      _mm512_max_epu32(_mm512_add_epi32(data, offset), currentoffsetmax);
-  currentmax = _mm512_max_epu32(data, currentmax);
+    const __m256i t0_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x0fc0fc00));
+    const __m256i t0_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x0fc0fc00));
+    const __m256i t0_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x0fc0fc00));
+    const __m256i t0_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x0fc0fc00));
 
-  const __m512i standardmax = _mm512_set1_epi32((uint32_t)0x10ffff);
-  const __m512i standardoffsetmax = _mm512_set1_epi32((uint32_t)0xfffff7ff);
-  __m512i is_zero =
-      _mm512_xor_si512(_mm512_max_epu32(currentmax, standardmax), standardmax);
-  utf32_err |= (_mm512_test_epi8_mask(is_zero, is_zero) != 0);
-  is_zero = _mm512_xor_si512(
-      _mm512_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
-  utf32_err |= (_mm512_test_epi8_mask(is_zero, is_zero) != 0);
-  checker.check_eof();
-  bool is_valid_utf8 = !checker.errors();
-  if (is_valid_utf8) {
-    out |= encoding_type::UTF8;
-  }
-  if (utf16_err == 0) {
-    out |= encoding_type::UTF16_LE;
-  }
-  if (utf32_err == 0) {
-    out |= encoding_type::UTF32_LE;
-  }
-  return out;
-}
+    const __m256i t1_0 =
+        _mm256_mulhi_epu16(t0_0, _mm256_set1_epi32(0x04000040));
+    const __m256i t1_1 =
+        _mm256_mulhi_epu16(t0_1, _mm256_set1_epi32(0x04000040));
+    const __m256i t1_2 =
+        _mm256_mulhi_epu16(t0_2, _mm256_set1_epi32(0x04000040));
+    const __m256i t1_3 =
+        _mm256_mulhi_epu16(t0_3, _mm256_set1_epi32(0x04000040));
 
-simdutf_warn_unused bool
-implementation::validate_utf8(const char *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return true;
-  }
-  avx512_utf8_checker checker{};
-  const char *ptr = buf;
-  const char *end = ptr + len;
-  for (; end - ptr >= 64; ptr += 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    checker.check_next_input(utf8);
-  }
-  if (end != ptr) {
-    const __m512i utf8 = _mm512_maskz_loadu_epi8(
-        ~UINT64_C(0) >> (64 - (end - ptr)), (const __m512i *)ptr);
-    checker.check_next_input(utf8);
-  }
-  checker.check_eof();
-  return !checker.errors();
-}
+    const __m256i t2_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x003f03f0));
+    const __m256i t2_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x003f03f0));
+    const __m256i t2_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x003f03f0));
+    const __m256i t2_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x003f03f0));
 
-simdutf_warn_unused result implementation::validate_utf8_with_errors(
-    const char *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return result(error_code::SUCCESS, len);
-  }
-  avx512_utf8_checker checker{};
-  const char *ptr = buf;
-  const char *end = ptr + len;
-  size_t count{0};
-  for (; end - ptr >= 64; ptr += 64) {
-    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)ptr);
-    checker.check_next_input(utf8);
-    if (checker.errors()) {
-      if (count != 0) {
-        count--;
-      } // Sometimes the error is only detected in the next chunk
-      result res = scalar::utf8::rewind_and_validate_with_errors(
-          reinterpret_cast<const char *>(buf),
-          reinterpret_cast<const char *>(buf + count), len - count);
-      res.count += count;
-      return res;
-    }
-    count += 64;
-  }
-  if (end != ptr) {
-    const __m512i utf8 = _mm512_maskz_loadu_epi8(
-        ~UINT64_C(0) >> (64 - (end - ptr)), (const __m512i *)ptr);
-    checker.check_next_input(utf8);
-  }
-  checker.check_eof();
-  if (checker.errors()) {
-    if (count != 0) {
-      count--;
-    } // Sometimes the error is only detected in the next chunk
-    result res = scalar::utf8::rewind_and_validate_with_errors(
-        reinterpret_cast<const char *>(buf),
-        reinterpret_cast<const char *>(buf + count), len - count);
-    res.count += count;
-    return res;
-  }
-  return result(error_code::SUCCESS, len);
-}
+    const __m256i t3_0 =
+        _mm256_mullo_epi16(t2_0, _mm256_set1_epi32(0x01000010));
+    const __m256i t3_1 =
+        _mm256_mullo_epi16(t2_1, _mm256_set1_epi32(0x01000010));
+    const __m256i t3_2 =
+        _mm256_mullo_epi16(t2_2, _mm256_set1_epi32(0x01000010));
+    const __m256i t3_3 =
+        _mm256_mullo_epi16(t2_3, _mm256_set1_epi32(0x01000010));
 
-simdutf_warn_unused bool
-implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return icelake::validate_ascii(buf, len);
-}
+    const __m256i input0 = _mm256_or_si256(t1_0, t3_0);
+    const __m256i input1 = _mm256_or_si256(t1_1, t3_1);
+    const __m256i input2 = _mm256_or_si256(t1_2, t3_2);
+    const __m256i input3 = _mm256_or_si256(t1_3, t3_3);
 
-simdutf_warn_unused result implementation::validate_ascii_with_errors(
-    const char *buf, size_t len) const noexcept {
-  const char *buf_orig = buf;
-  const char *end = buf + len;
-  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
-  for (; end - buf >= 64; buf += 64) {
-    const __m512i input = _mm512_loadu_si512((const __m512i *)buf);
-    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
-    if (notascii) {
-      return result(error_code::TOO_LARGE,
-                    buf - buf_orig + _tzcnt_u64(notascii));
-    }
-  }
-  if (end != buf) {
-    const __m512i input = _mm512_maskz_loadu_epi8(
-        ~UINT64_C(0) >> (64 - (end - buf)), (const __m512i *)buf);
-    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
-    if (notascii) {
-      return result(error_code::TOO_LARGE,
-                    buf - buf_orig + _tzcnt_u64(notascii));
-    }
-  }
-  return result(error_code::SUCCESS, len);
-}
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
+                        lookup_pshufb_improved<isbase64url>(input0));
+    out += 32;
 
-simdutf_warn_unused bool
-implementation::validate_utf16le(const char16_t *buf,
-                                 size_t len) const noexcept {
-  const char16_t *end = buf + len;
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
+                        lookup_pshufb_improved<isbase64url>(input1));
+    out += 32;
 
-  for (; end - buf >= 32;) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        return false;
-      }
-      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
-      if (ends_with_high) {
-        buf += 31; // advance only by 31 code units so that we start with the
-                   // high surrogate on the next round.
-      } else {
-        buf += 32;
-      }
-    } else {
-      buf += 32;
-    }
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
+                        lookup_pshufb_improved<isbase64url>(input2));
+    out += 32;
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
+                        lookup_pshufb_improved<isbase64url>(input3));
+    out += 32;
   }
-  if (buf < end) {
-    __m512i in =
-        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        return false;
-      }
-    }
+  for (; i + 28 <= srclen; i += 24) {
+    // lo = [xxxx|DDDC|CCBB|BAAA]
+    // hi = [xxxx|HHHG|GGFF|FEEE]
+    const __m128i lo =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(input + i));
+    const __m128i hi =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(input + i + 4 * 3));
+
+    // bytes from groups A, B and C are needed in separate 32-bit lanes
+    // in = [0HHH|0GGG|0FFF|0EEE[0DDD|0CCC|0BBB|0AAA]
+    __m256i in = _mm256_shuffle_epi8(_mm256_set_m128i(hi, lo), shuf);
+
+    // this part is well commented in encode.sse.cpp
+
+    const __m256i t0 = _mm256_and_si256(in, _mm256_set1_epi32(0x0fc0fc00));
+    const __m256i t1 = _mm256_mulhi_epu16(t0, _mm256_set1_epi32(0x04000040));
+    const __m256i t2 = _mm256_and_si256(in, _mm256_set1_epi32(0x003f03f0));
+    const __m256i t3 = _mm256_mullo_epi16(t2, _mm256_set1_epi32(0x01000010));
+    const __m256i indices = _mm256_or_si256(t1, t3);
+
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
+                        lookup_pshufb_improved<isbase64url>(indices));
+    out += 32;
   }
-  return true;
+  return i / 3 * 4 + scalar::base64::tail_encode_base64((char *)out, src + i,
+                                                        srclen - i, options);
 }
 
-simdutf_warn_unused bool
-implementation::validate_utf16be(const char16_t *buf,
-                                 size_t len) const noexcept {
-  const char16_t *end = buf + len;
-  const __m512i byteflip = _mm512_setr_epi64(
-      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-      0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  for (; end - buf >= 32;) {
-    __m512i in =
-        _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)buf), byteflip);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        return false;
-      }
-      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
-      if (ends_with_high) {
-        buf += 31; // advance only by 31 code units so that we start with the
-                   // high surrogate on the next round.
-      } else {
-        buf += 32;
-      }
-    } else {
-      buf += 32;
-    }
-  }
-  if (buf < end) {
-    __m512i in = _mm512_shuffle_epi8(
-        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf),
-        byteflip);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        return false;
-      }
-    }
+static inline void compress(__m128i data, uint16_t mask, char *output) {
+  if (mask == 0) {
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data);
+    return;
   }
-  return true;
+  // this particular implementation was inspired by work done by @animetosho
+  // we do it in two steps, first 8 bytes and then second 8 bytes
+  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+  // next line just loads the 64-bit values thintable_epi8[mask1] and
+  // thintable_epi8[mask2] into a 128-bit register, using only
+  // two instructions on most compilers.
+
+  __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2],
+                                    tables::base64::thintable_epi8[mask1]);
+  // we increment by 0x08 the second half of the mask
+  shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+  // this is the version "nearly pruned"
+  __m128i pruned = _mm_shuffle_epi8(data, shufmask);
+  // we still need to put the two halves together.
+  // we compute the popcount of the first half:
+  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
+  // then load the corresponding mask, what it does is to write
+  // only the first pop1 bytes from the first 8 bytes, and then
+  // it fills in with the bytes from the second 8 bytes + some filling
+  // at the end.
+  __m128i compactmask = _mm_loadu_si128(reinterpret_cast<const __m128i *>(
+      tables::base64::pshufb_combine_table + pop1 * 8));
+  __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+
+  _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
 }
 
-simdutf_warn_unused result implementation::validate_utf16le_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  const char16_t *start_buf = buf;
-  const char16_t *end = buf + len;
-  for (; end - buf >= 32;) {
-    __m512i in = _mm512_loadu_si512((__m512i *)buf);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
-        uint32_t extra_high =
-            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
-        return result(error_code::SURROGATE,
-                      (buf - start_buf) +
-                          (extra_low < extra_high ? extra_low : extra_high));
-      }
-      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
-      if (ends_with_high) {
-        buf += 31; // advance only by 31 code units so that we start with the
-                   // high surrogate on the next round.
-      } else {
-        buf += 32;
-      }
-    } else {
-      buf += 32;
-    }
-  }
-  if (buf < end) {
-    __m512i in =
-        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
-        uint32_t extra_high =
-            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
-        return result(error_code::SURROGATE,
-                      (buf - start_buf) +
-                          (extra_low < extra_high ? extra_low : extra_high));
-      }
-    }
+// --- decoding -----------------------------------------------
+
+template <typename = void>
+simdutf_really_inline void compress(__m256i data, uint32_t mask, char *output) {
+  if (mask == 0) {
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), data);
+    return;
   }
-  return result(error_code::SUCCESS, len);
+  compress(_mm256_castsi256_si128(data), uint16_t(mask), output);
+  compress(_mm256_extracti128_si256(data, 1), uint16_t(mask >> 16),
+           output + count_ones(~mask & 0xFFFF));
 }
 
-simdutf_warn_unused result implementation::validate_utf16be_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  const char16_t *start_buf = buf;
-  const char16_t *end = buf + len;
-  const __m512i byteflip = _mm512_setr_epi64(
-      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-      0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  for (; end - buf >= 32;) {
-    __m512i in =
-        _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)buf), byteflip);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
-        uint32_t extra_high =
-            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
-        return result(error_code::SURROGATE,
-                      (buf - start_buf) +
-                          (extra_low < extra_high ? extra_low : extra_high));
-      }
-      bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
-      if (ends_with_high) {
-        buf += 31; // advance only by 31 code units so that we start with the
-                   // high surrogate on the next round.
-      } else {
-        buf += 32;
-      }
-    } else {
-      buf += 32;
-    }
-  }
-  if (buf < end) {
-    __m512i in = _mm512_shuffle_epi8(
-        _mm512_maskz_loadu_epi16((1U << (end - buf)) - 1, (__m512i *)buf),
-        byteflip);
-    __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
-    __mmask32 surrogates =
-        _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
-    if (surrogates) {
-      __mmask32 highsurrogates =
-          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
-      __mmask32 lowsurrogates = surrogates ^ highsurrogates;
-      // high must be followed by low
-      if ((highsurrogates << 1) != lowsurrogates) {
-        uint32_t extra_low = _tzcnt_u32(lowsurrogates & ~(highsurrogates << 1));
-        uint32_t extra_high =
-            _tzcnt_u32(highsurrogates & ~(lowsurrogates >> 1));
-        return result(error_code::SURROGATE,
-                      (buf - start_buf) +
-                          (extra_low < extra_high ? extra_low : extra_high));
-      }
-    }
-  }
-  return result(error_code::SUCCESS, len);
+template <typename = void>
+simdutf_really_inline void base64_decode(char *out, __m256i str) {
+  // credit: aqrit
+  const __m256i pack_shuffle =
+      _mm256_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1,
+                       2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1);
+  const __m256i t0 = _mm256_maddubs_epi16(str, _mm256_set1_epi32(0x01400140));
+  const __m256i t1 = _mm256_madd_epi16(t0, _mm256_set1_epi32(0x00011000));
+  const __m256i t2 = _mm256_shuffle_epi8(t1, pack_shuffle);
+
+  // Store the output:
+  _mm_storeu_si128((__m128i *)out, _mm256_castsi256_si128(t2));
+  _mm_storeu_si128((__m128i *)(out + 12), _mm256_extracti128_si256(t2, 1));
 }
 
-simdutf_warn_unused bool
-implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
-  return icelake::validate_utf32(buf, len);
+template <typename = void>
+simdutf_really_inline void base64_decode_block(char *out, const char *src) {
+  base64_decode(out,
+                _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src)));
+  base64_decode(out + 24, _mm256_loadu_si256(
+                              reinterpret_cast<const __m256i *>(src + 32)));
 }
 
-simdutf_warn_unused result implementation::validate_utf32_with_errors(
-    const char32_t *buf, size_t len) const noexcept {
-  const char32_t *buf_orig = buf;
-  if (len >= 16) {
-    const char32_t *end = buf + len - 16;
-    while (buf <= end) {
-      __m512i utf32 = _mm512_loadu_si512((const __m512i *)buf);
-      __mmask16 outside_range = _mm512_cmp_epu32_mask(
-          utf32, _mm512_set1_epi32(0x10ffff), _MM_CMPINT_GT);
+template <typename = void>
+simdutf_really_inline void base64_decode_block_safe(char *out,
+                                                    const char *src) {
+  base64_decode(out,
+                _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src)));
+  char buffer[32]; // We enforce safety with a buffer.
+  base64_decode(
+      buffer, _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32)));
+  std::memcpy(out + 24, buffer, 24);
+}
 
-      __m512i utf32_off =
-          _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));
+// --- decoding - base64 class --------------------------------
 
-      __mmask16 surrogate_range = _mm512_cmp_epu32_mask(
-          utf32_off, _mm512_set1_epi32(0xfffff7ff), _MM_CMPINT_GT);
-      if ((outside_range | surrogate_range)) {
-        auto outside_idx = _tzcnt_u32(outside_range);
-        auto surrogate_idx = _tzcnt_u32(surrogate_range);
+class block64 {
+  __m256i chunks[2];
 
-        if (outside_idx < surrogate_idx) {
-          return result(error_code::TOO_LARGE, buf - buf_orig + outside_idx);
-        }
+public:
+  // The caller of this function is responsible to ensure that there are 64
+  // bytes available from reading at src.
+  simdutf_really_inline block64(const char *src) {
+    chunks[0] = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src));
+    chunks[1] = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32));
+  }
+
+  // The caller of this function is responsible to ensure that there are 128
+  // bytes available from reading at src.
+  simdutf_really_inline block64(const char16_t *src) {
+    const auto m1 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src));
+    const auto m2 =
+        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 16));
+    const auto m3 =
+        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32));
+    const auto m4 =
+        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 48));
+
+    const auto m1p = _mm256_permute2x128_si256(m1, m2, 0x20);
+    const auto m2p = _mm256_permute2x128_si256(m1, m2, 0x31);
+    const auto m3p = _mm256_permute2x128_si256(m3, m4, 0x20);
+    const auto m4p = _mm256_permute2x128_si256(m3, m4, 0x31);
+
+    chunks[0] = _mm256_packus_epi16(m1p, m2p);
+    chunks[1] = _mm256_packus_epi16(m3p, m4p);
+  }
+
+  simdutf_really_inline void copy_block(char *output) {
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), chunks[0]);
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(output + 32), chunks[1]);
+  }
+
+  // decode 64 bytes and output 48 bytes
+  simdutf_really_inline void base64_decode_block(char *out) {
+    base64_decode(out, chunks[0]);
+    base64_decode(out + 24, chunks[1]);
+  }
+
+  simdutf_really_inline void base64_decode_block_safe(char *out) {
+    base64_decode(out, chunks[0]);
+    char buffer[32]; // We enforce safety with a buffer.
+    base64_decode(buffer, chunks[1]);
+    std::memcpy(out + 24, buffer, 24);
+  }
+
+  template <bool base64_url, bool ignore_garbage>
+  simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) {
+    uint32_t err0 = 0;
+    uint32_t err1 = 0;
+    uint64_t m0 = to_base64_mask<base64_url, ignore_garbage>(&chunks[0], &err0);
+    uint64_t m1 = to_base64_mask<base64_url, ignore_garbage>(&chunks[1], &err1);
+    if (!ignore_garbage) {
+      *error = err0 | ((uint64_t)err1 << 32);
+    }
+    return m0 | (m1 << 32);
+  }
+
+  template <bool base64_url, bool ignore_garbage>
+  simdutf_really_inline uint32_t to_base64_mask(__m256i *src, uint32_t *error) {
+    const __m256i ascii_space_tbl =
+        _mm256_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa,
+                         0x0, 0xc, 0xd, 0x0, 0x0, 0x20, 0x0, 0x0, 0x0, 0x0, 0x0,
+                         0x0, 0x0, 0x0, 0x9, 0xa, 0x0, 0xc, 0xd, 0x0, 0x0);
+    // credit: aqrit
+    __m256i delta_asso;
+    if (base64_url) {
+      delta_asso = _mm256_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0,
+                                    0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF, 0x1, 0x1,
+                                    0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0, 0x0,
+                                    0x0, 0x0, 0xF, 0x0, 0xF);
+    } else {
+      delta_asso = _mm256_setr_epi8(
+          0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00,
+          0x00, 0x00, 0x0F, 0x00, 0x0F, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+          0x01, 0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F);
+    }
+
+    __m256i delta_values;
+    if (base64_url) {
+      delta_values = _mm256_setr_epi8(
+          0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9),
+          uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), uint8_t(0xE0),
+          uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF),
+          uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3),
+          uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9));
+    } else {
+      delta_values = _mm256_setr_epi8(
+          int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), int8_t(0x04),
+          int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), int8_t(0x00),
+          int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9),
+          int8_t(0xB9), int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13),
+          int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9),
+          int8_t(0x00), int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF),
+          int8_t(0xB9), int8_t(0xB9));
+    }
+
+    __m256i check_asso;
+    if (base64_url) {
+      check_asso = _mm256_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1,
+                                    0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6, 0xD, 0x1,
+                                    0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x3,
+                                    0x7, 0xB, 0xE, 0xB, 0x6);
+    } else {
+      check_asso = _mm256_setr_epi8(
+          0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03,
+          0x07, 0x0B, 0x0B, 0x0B, 0x0F, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01,
+          0x01, 0x01, 0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F);
+    }
+    __m256i check_values;
+    if (base64_url) {
+      check_values = _mm256_setr_epi8(
+          uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
+          uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), uint8_t(0xA6),
+          uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, uint8_t(0x80),
+          0x0, uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
+          uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6),
+          uint8_t(0xA6), uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0,
+          uint8_t(0x80), 0x0, uint8_t(0x80));
+    } else {
+      check_values = _mm256_setr_epi8(
+          int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0xCF),
+          int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), int8_t(0x86),
+          int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), int8_t(0x91),
+          int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80),
+          int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5),
+          int8_t(0x86), int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80),
+          int8_t(0x91), int8_t(0x80));
+    }
+    const __m256i shifted = _mm256_srli_epi32(*src, 3);
+    const __m256i delta_hash =
+        _mm256_avg_epu8(_mm256_shuffle_epi8(delta_asso, *src), shifted);
+    const __m256i check_hash =
+        _mm256_avg_epu8(_mm256_shuffle_epi8(check_asso, *src), shifted);
+    const __m256i out =
+        _mm256_adds_epi8(_mm256_shuffle_epi8(delta_values, delta_hash), *src);
+    const __m256i chk =
+        _mm256_adds_epi8(_mm256_shuffle_epi8(check_values, check_hash), *src);
+    const int mask = _mm256_movemask_epi8(chk);
+    if (!ignore_garbage && mask) {
+      __m256i ascii_space =
+          _mm256_cmpeq_epi8(_mm256_shuffle_epi8(ascii_space_tbl, *src), *src);
+      *error = (mask ^ _mm256_movemask_epi8(ascii_space));
+    }
+    *src = out;
+    return (uint32_t)mask;
+  }
+
+  simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) {
+    if (is_power_of_two(mask)) {
+      return compress_block_single(mask, output);
+    }
+
+    uint64_t nmask = ~mask;
+    compress(chunks[0], uint32_t(mask), output);
+    compress(chunks[1], uint32_t(mask >> 32),
+             output + count_ones(nmask & 0xFFFFFFFF));
+    return count_ones(nmask);
+  }
+
+  simdutf_really_inline size_t compress_block_single(uint64_t mask,
+                                                     char *output) {
+    const size_t pos64 = trailing_zeroes(mask);
+    const int8_t pos = pos64 & 0xf;
+    switch (pos64 >> 4) {
+    case 0b00: {
+      const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0);
+      const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(lane0, sh);
+
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed);
+      _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), lane1);
+      _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]);
+    } break;
+    case 0b01: {
+      const __m128i lane0 = _mm256_extracti128_si256(chunks[0], 0);
+      const __m128i lane1 = _mm256_extracti128_si256(chunks[0], 1);
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), lane0);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(lane1, sh);
+
+      _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed);
+      _mm256_storeu_si256((__m256i *)(output + 2 * 16 - 1), chunks[1]);
+    } break;
+    case 0b10: {
+      const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0);
+      const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1);
+
+      _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(lane2, sh);
+
+      _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed);
+      _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), lane3);
+    } break;
+    case 0b11: {
+      const __m128i lane2 = _mm256_extracti128_si256(chunks[1], 0);
+      const __m128i lane3 = _mm256_extracti128_si256(chunks[1], 1);
+
+      _mm256_storeu_si256((__m256i *)(output + 0 * 16), chunks[0]);
+      _mm_storeu_si128((__m128i *)(output + 2 * 16), lane2);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(lane3, sh);
+
+      _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed);
+    } break;
+    }
+
+    return 63;
+  }
+};
+/* end file src/haswell/avx2_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
 
-        return result(error_code::SURROGATE, buf - buf_orig + surrogate_idx);
-      }
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
 
-      buf += 16;
-    }
-  }
-  if (len > 0) {
-    __m512i utf32 = _mm512_maskz_loadu_epi32(
-        __mmask16((1U << (buf_orig + len - buf)) - 1), (const __m512i *)buf);
-    __mmask16 outside_range = _mm512_cmp_epu32_mask(
-        utf32, _mm512_set1_epi32(0x10ffff), _MM_CMPINT_GT);
-    __m512i utf32_off = _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));
+/* begin file src/generic/buf_block_reader.h */
+namespace simdutf {
+namespace haswell {
+namespace {
 
-    __mmask16 surrogate_range = _mm512_cmp_epu32_mask(
-        utf32_off, _mm512_set1_epi32(0xfffff7ff), _MM_CMPINT_GT);
-    if ((outside_range | surrogate_range)) {
-      auto outside_idx = _tzcnt_u32(outside_range);
-      auto surrogate_idx = _tzcnt_u32(surrogate_range);
+// Walks through a buffer in block-sized increments, loading the last part with
+// spaces
+template <size_t STEP_SIZE> struct buf_block_reader {
+public:
+  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdutf_really_inline size_t block_index();
+  simdutf_really_inline bool has_full_block() const;
+  simdutf_really_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0
+   * (in which case this function fills the buffer with spaces and returns 0. In
+   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
+   * block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
+  simdutf_really_inline void advance();
 
-      if (outside_idx < surrogate_idx) {
-        return result(error_code::TOO_LARGE, buf - buf_orig + outside_idx);
-      }
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
 
-      return result(error_code::SURROGATE, buf - buf_orig + surrogate_idx);
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text_64(const uint8_t *text) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  in.store(reinterpret_cast<uint8_t *>(buf));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') {
+      buf[i] = '_';
     }
   }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
 
-  return result(error_code::SUCCESS, len);
+simdutf_unused static char *format_mask(uint64_t mask) {
+  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
+  for (size_t i = 0; i < 64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
-    const char *buf, size_t len, char *utf8_output) const noexcept {
-  return icelake::latin1_to_utf8_avx512_start(buf, len, utf8_output);
+template <size_t STEP_SIZE>
+simdutf_really_inline
+buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
+    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
+      idx{0} {}
+
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
+  return idx;
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return icelake_convert_latin1_to_utf16<endianness::LITTLE>(buf, len,
-                                                             utf16_output);
+template <size_t STEP_SIZE>
+simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return icelake_convert_latin1_to_utf16<endianness::BIG>(buf, len,
-                                                          utf16_output);
+template <size_t STEP_SIZE>
+simdutf_really_inline const uint8_t *
+buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
 }
 
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char *, char32_t *> ret =
-      avx512_convert_latin1_to_utf32(buf, len, utf32_output);
-  if (ret.first == nullptr) {
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t
+buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if (len == idx) {
     return 0;
-  }
-  size_t converted_chars = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_converted_chars == 0) {
-      return 0;
-    }
-    converted_chars += scalar_converted_chars;
-  }
-  return converted_chars;
+  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20,
+              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
+                          // to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake::utf8_to_latin1_avx512(buf, len, latin1_output);
+template <size_t STEP_SIZE>
+simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
 }
 
-simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  // First, try to convert as much as possible using the SIMD implementation.
-  const char *obuf = buf;
-  char *olatin1_output = latin1_output;
-  size_t written = icelake::utf8_to_latin1_avx512(obuf, len, olatin1_output);
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/buf_block_reader.h */
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_validation {
 
-  // If we have completely converted the string
-  if (obuf == buf + len) {
-    return {simdutf::SUCCESS, written};
-  }
-  size_t pos = obuf - buf;
-  result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-      pos, buf + pos, len - pos, latin1_output);
-  res.count += pos;
-  return res;
-}
+using namespace simd;
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake::valid_utf8_to_latin1_avx512(buf, len, latin1_output);
-}
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
+
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16_result ret =
-      fast_avx512_convert_utf8_to_utf16<endianness::LITTLE>(buf, len,
-                                                            utf16_output);
-  if (ret.second == nullptr) {
-    return 0;
-  }
-  return ret.second - utf16_output;
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
 }
-
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16_result ret = fast_avx512_convert_utf8_to_utf16<endianness::BIG>(
-      buf, len, utf16_output);
-  if (ret.second == nullptr) {
-    return 0;
-  }
-  return ret.second - utf16_output;
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::LITTLE>(
-      buf, len, utf16_output);
+//
+// Return nonzero if there are incomplete multibyte characters at the end of the
+// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
+//
+simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  // If the previous input's last 3 bytes match this, they're too short (they
+  // ended at EOF):
+  // ... 1111____ 111_____ 11______
+  static const uint8_t max_array[32] = {255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        0b11110000u - 1,
+                                        0b11100000u - 1,
+                                        0b11000000u - 1};
+  const simd8<uint8_t> max_value(
+      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
+  return input.gt_bits(max_value);
 }
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::BIG>(
-      buf, len, utf16_output);
-}
+struct utf8_checker {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+  // The last input we received
+  simd8<uint8_t> prev_input_block;
+  // Whether the last input we received was incomplete (used for ASCII fast
+  // path)
+  simd8<uint8_t> prev_incomplete;
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16_result ret =
-      icelake::valid_utf8_to_fixed_length<endianness::LITTLE, char16_t>(
-          buf, len, utf16_output);
-  size_t saved_bytes = ret.second - utf16_output;
-  const char *end = buf + len;
-  if (ret.first == end) {
-    return saved_bytes;
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
 
-  // Note: AVX512 procedure looks up 4 bytes forward, and
-  //       correctly converts multi-byte chars even if their
-  //       continuation bytes lie outsiede 16-byte window.
-  //       It meas, we have to skip continuation bytes from
-  //       the beginning ret.first, as they were already consumed.
-  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
-    ret.first += 1;
+  // The only problem that can happen at EOF is that a multibyte character is
+  // too short or a byte value too large in the last bytes: check_special_cases
+  // only checks for bytes too large in the first of two bytes.
+  simdutf_really_inline void check_eof() {
+    // If the previous block had incomplete UTF-8 characters at the end, an
+    // ASCII block can't possibly finish them.
+    this->error |= this->prev_incomplete;
   }
 
-  if (ret.first != end) {
-    const size_t scalar_saved_bytes =
-        scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
+    if (simdutf_likely(is_ascii(input))) {
+      this->error |= this->prev_incomplete;
+    } else {
+      // you might think that a for-loop would work, but under Visual Studio, it
+      // is not good enough.
+      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                    "We support either two or four chunks per 64-byte block.");
+      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+      }
+      this->prev_incomplete =
+          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
+      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-
-  return saved_bytes;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16_result ret =
-      icelake::valid_utf8_to_fixed_length<endianness::BIG, char16_t>(
-          buf, len, utf16_output);
-  size_t saved_bytes = ret.second - utf16_output;
-  const char *end = buf + len;
-  if (ret.first == end) {
-    return saved_bytes;
-  }
-
-  // Note: AVX512 procedure looks up 4 bytes forward, and
-  //       correctly converts multi-byte chars even if their
-  //       continuation bytes lie outsiede 16-byte window.
-  //       It meas, we have to skip continuation bytes from
-  //       the beginning ret.first, as they were already consumed.
-  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
-    ret.first += 1;
   }
 
-  if (ret.first != end) {
-    const size_t scalar_saved_bytes =
-        scalar::utf8_to_utf16::convert_valid<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+  // do not forget to call check_eof!
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
 
-  return saved_bytes;
-}
+}; // struct utf8_checker
+} // namespace utf8_validation
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_out) const noexcept {
-  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
-  utf8_to_utf32_result ret =
-      icelake::validating_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(
-          buf, len, utf32_output);
-  if (ret.second == nullptr)
-    return 0;
+using utf8_validation::utf8_checker;
 
-  size_t saved_bytes = ret.second - utf32_output;
-  const char *end = buf + len;
-  if (ret.first == end) {
-    return saved_bytes;
-  }
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
+/* begin file src/generic/utf8_validation/utf8_validator.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_validation {
 
-  // Note: the AVX512 procedure looks up 4 bytes forward, and
-  //       correctly converts multi-byte chars even if their
-  //       continuation bytes lie outside 16-byte window.
-  //       It means, we have to skip continuation bytes from
-  //       the beginning ret.first, as they were already consumed.
-  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
-    ret.first += 1;
-  }
-  if (ret.first != end) {
-    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert(
-        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template <class checker>
+bool generic_validate_utf8(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    c.check_next_input(in);
+    reader.advance();
   }
-
-  return saved_bytes;
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
+  reader.advance();
+  c.check_eof();
+  return !c.errors();
 }
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
-    const char *buf, size_t len, char32_t *utf32) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    return {error_code::SUCCESS, 0};
-  }
-  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32);
-  auto ret = icelake::validating_utf8_to_fixed_length_with_constant_checks<
-      endianness::LITTLE, uint32_t>(buf, len, utf32_output);
+bool generic_validate_utf8(const char *input, size_t length) {
+  return generic_validate_utf8<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
+}
 
-  if (!std::get<2>(ret)) {
-    size_t pos = std::get<0>(ret) - buf;
-    // We might have an error that occurs right before  pos.
-    // This is only a concern if buf[pos] is not a continuation byte.
-    if ((buf[pos] & 0xc0) != 0x80 && pos >= 64) {
-      pos -= 1;
-    } else if ((buf[pos] & 0xc0) == 0x80 && pos >= 64) {
-      // We must check whether we are the fourth continuation byte
-      bool c1 = (buf[pos - 1] & 0xc0) == 0x80;
-      bool c2 = (buf[pos - 2] & 0xc0) == 0x80;
-      bool c3 = (buf[pos - 3] & 0xc0) == 0x80;
-      if (c1 && c2 && c3) {
-        return {simdutf::TOO_LONG, pos};
-      }
+/**
+ * Validates that the string is actual UTF-8 and stops on errors.
+ */
+template <class checker>
+result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
+  size_t count{0};
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    c.check_next_input(in);
+    if (c.errors()) {
+      if (count != 0) {
+        count--;
+      } // Sometimes the error is only detected in the next chunk
+      result res = scalar::utf8::rewind_and_validate_with_errors(
+          reinterpret_cast<const char *>(input),
+          reinterpret_cast<const char *>(input + count), length - count);
+      res.count += count;
+      return res;
     }
-    // todo: we reset the output to utf32 instead of using std::get<2.(ret) as
-    // you'd expect. that is because
-    // validating_utf8_to_fixed_length_with_constant_checks may have processed
-    // data beyond the error.
-    result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-        pos, buf + pos, len - pos, utf32);
-    res.count += pos;
-    return res;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  const char *end = buf + len;
-  if (std::get<0>(ret) == end) {
-    return {simdutf::SUCCESS, saved_bytes};
-  }
-
-  // Note: the AVX512 procedure looks up 4 bytes forward, and
-  //       correctly converts multi-byte chars even if their
-  //       continuation bytes lie outside 16-byte window.
-  //       It means, we have to skip continuation bytes from
-  //       the beginning ret.first, as they were already consumed.
-  while (std::get<0>(ret) != end and
-         ((uint8_t(*std::get<0>(ret)) & 0xc0) == 0x80)) {
-    std::get<0>(ret) += 1;
+    reader.advance();
+    count += 64;
   }
-
-  if (std::get<0>(ret) != end) {
-    auto scalar_result = scalar::utf8_to_utf32::convert_with_errors(
-        std::get<0>(ret), len - (std::get<0>(ret) - buf),
-        reinterpret_cast<char32_t *>(utf32_output) + saved_bytes);
-    if (scalar_result.error != simdutf::SUCCESS) {
-      scalar_result.count += (std::get<0>(ret) - buf);
-    } else {
-      scalar_result.count += saved_bytes;
-    }
-    return scalar_result;
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
+  reader.advance();
+  c.check_eof();
+  if (c.errors()) {
+    if (count != 0) {
+      count--;
+    } // Sometimes the error is only detected in the next chunk
+    result res = scalar::utf8::rewind_and_validate_with_errors(
+        reinterpret_cast<const char *>(input),
+        reinterpret_cast<const char *>(input) + count, length - count);
+    res.count += count;
+    return res;
+  } else {
+    return result(error_code::SUCCESS, length);
   }
+}
 
-  return {simdutf::SUCCESS, size_t(std::get<1>(ret) - utf32_output)};
+result generic_validate_utf8_with_errors(const char *input, size_t length) {
+  return generic_validate_utf8_with_errors<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_out) const noexcept {
-  uint32_t *utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
-  utf8_to_utf32_result ret =
-      icelake::valid_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(
-          buf, len, utf32_output);
-  size_t saved_bytes = ret.second - utf32_output;
-  const char *end = buf + len;
-  if (ret.first == end) {
-    return saved_bytes;
-  }
+} // namespace utf8_validation
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-  // Note: AVX512 procedure looks up 4 bytes forward, and
-  //       correctly converts multi-byte chars even if their
-  //       continuation bytes lie outsiede 16-byte window.
-  //       It meas, we have to skip continuation bytes from
-  //       the beginning ret.first, as they were already consumed.
-  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
-    ret.first += 1;
-  }
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace ascii_validation {
 
-  if (ret.first != end) {
-    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert_valid(
-        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+  uint8_t blocks[64]{};
+  simd::simd8x64<uint8_t> running_or(blocks);
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    running_or |= in;
+    reader.advance();
   }
-
-  return saved_bytes;
-}
-
-simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf16_to_latin1<endianness::LITTLE>(buf, len,
-                                                             latin1_output);
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  running_or |= in;
+  return running_or.is_ascii();
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf16_to_latin1<endianness::BIG>(buf, len,
-                                                          latin1_output);
-}
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+  size_t count{0};
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    if (!in.is_ascii()) {
+      result res = scalar::ascii::validate_with_errors(
+          reinterpret_cast<const char *>(input + count), length - count);
+      return result(res.error, count + res.count);
+    }
+    reader.advance();
 
-simdutf_warn_unused result
-implementation::convert_utf16le_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
-             buf, len, latin1_output)
-      .first;
+    count += 64;
+  }
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  if (!in.is_ascii()) {
+    result res = scalar::ascii::validate_with_errors(
+        reinterpret_cast<const char *>(input + count), length - count);
+    return result(res.error, count + res.count);
+  } else {
+    return result(error_code::SUCCESS, length);
+  }
 }
 
-simdutf_warn_unused result
-implementation::convert_utf16be_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf16_to_latin1_with_errors<endianness::BIG>(
-             buf, len, latin1_output)
-      .first;
-}
+} // namespace ascii_validation
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement custom function
-  return convert_utf16be_to_latin1(buf, len, latin1_output);
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  // transcoding from UTF-8 to UTF-16
+/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_utf16 {
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement custom function
-  return convert_utf16le_to_latin1(buf, len, latin1_output);
-}
+using namespace simd;
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  size_t outlen;
-  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(
-      buf, len, (unsigned char *)utf8_output, &outlen);
-  if (inlen != len) {
-    return 0;
+template <endianness endian>
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char16_t *utf16_output) noexcept {
+  // The implementation is not specific to haswell and should be moved to the
+  // generic directory.
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    // this loop could be unrolled further. For example, we could process the
+    // mask far more than 64 bytes.
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf16<endian>(utf16_output);
+      utf16_output += 64;
+      pos += 64;
+    } else {
+      // Slow path. We hope that the compiler will recognize that this is a slow
+      // path. Anything that is not a continuation mask is a 'leading byte',
+      // that is, the start of a new code point.
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      // The *start* of code points is not so useful, rather, we want the *end*
+      // of code points.
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
+      size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times when using solely
+      // the slow/regular path, and at least four times if there are fast paths.
+      while (pos < max_starting_point) {
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        //
+        // Thus we may allow convert_masked_utf8_to_utf16 to process
+        // more bytes at a time under a fast-path mode where 16 bytes
+        // are consumed at once (e.g., when encountering ASCII).
+        size_t consumed = convert_masked_utf8_to_utf16<endian>(
+            input + pos, utf8_end_of_code_point_mask, utf16_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
+    }
   }
-  return outlen;
+  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
+      input + pos, size - pos, utf16_output);
+  return utf16_output - start;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  size_t outlen;
-  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(
-      buf, len, (unsigned char *)utf8_output, &outlen);
-  if (inlen != len) {
-    return 0;
-  }
-  return outlen;
-}
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_utf16 {
+using namespace simd;
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  size_t outlen;
-  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(
-      buf, len, (unsigned char *)utf8_output, &outlen);
-  if (inlen != len) {
-    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
-        buf + inlen, len - inlen, utf8_output + outlen);
-    res.count += inlen;
-    return res;
-  }
-  return {simdutf::SUCCESS, outlen};
-}
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  size_t outlen;
-  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(
-      buf, len, (unsigned char *)utf8_output, &outlen);
-  if (inlen != len) {
-    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
-        buf + inlen, len - inlen, utf8_output + outlen);
-    res.count += inlen;
-    return res;
-  }
-  return {simdutf::SUCCESS, outlen};
-}
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16le_to_utf8(buf, len, utf8_output);
-}
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16be_to_utf8(buf, len, utf8_output);
-}
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf32_to_latin1(buf, len, latin1_output);
-}
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
 
-simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf32_to_latin1_with_errors(buf, len, latin1_output)
-      .first;
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
 }
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return icelake_convert_utf32_to_latin1(buf, len, latin1_output);
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char32_t *, char *> ret =
-      avx512_convert_utf32_to_utf8(buf, len, utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      icelake::avx512_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
-  if (ret.first.count != len) {
-    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
-        buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf32_to_utf8(buf, len, utf8_output);
-}
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      avx512_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+  template <endianness endian>
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
-
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      avx512_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
       return 0;
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
-
-simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      avx512_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(
-          buf, len, utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+    if (pos < size) {
+      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
+          in + pos, size - pos, utf16_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      utf16_output += howmany;
     }
+    return utf16_output - start;
   }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      avx512_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len,
-                                                                 utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+  template <endianness endian>
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res =
+              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+                  pos, in + pos, size - pos, utf16_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf16_output += res.count;
+      }
     }
+    return result(error_code::SUCCESS, utf16_output - start);
   }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16le(buf, len, utf16_output);
-}
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
+  }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16be(buf, len, utf16_output);
-}
+}; // struct utf8_checker
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
-                                                          utf32_output);
-  if (!std::get<2>(ret)) {
-    return 0;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  // transcoding from UTF-8 to UTF-32
+/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_utf32 {
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
-  if (!std::get<2>(ret)) {
-    return 0;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::BIG>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+using namespace simd;
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
-                                                          utf32_output);
-  if (!std::get<2>(ret)) {
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    scalar_res.count += (std::get<0>(ret) - buf);
-    return scalar_res;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_res.error) {
-      scalar_res.count += (std::get<0>(ret) - buf);
-      return scalar_res;
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char32_t *utf32_output) noexcept {
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf32(utf32_output);
+      utf32_output += 64;
+      pos += 64;
     } else {
-      scalar_res.count += saved_bytes;
-      return scalar_res;
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      size_t max_starting_point = (pos + 64) - 12;
+      while (pos < max_starting_point) {
+        size_t consumed = convert_masked_utf8_to_utf32(
+            input + pos, utf8_end_of_code_point_mask, utf32_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
     }
   }
-  return simdutf::result(simdutf::SUCCESS, saved_bytes);
+  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
+                                                       utf32_output);
+  return utf32_output - start;
 }
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
-  if (!std::get<2>(ret)) {
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    scalar_res.count += (std::get<0>(ret) - buf);
-    return scalar_res;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_res.error) {
-      scalar_res.count += (std::get<0>(ret) - buf);
-      return scalar_res;
-    } else {
-      scalar_res.count += saved_bytes;
-      return scalar_res;
-    }
-  }
-  return simdutf::result(simdutf::SUCCESS, saved_bytes);
-}
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_utf32 {
+using namespace simd;
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
-                                                          utf32_output);
-  if (!std::get<2>(ret)) {
-    return 0;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
 
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::tuple<const char16_t *, char32_t *, bool> ret =
-      icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
-  if (!std::get<2>(ret)) {
-    return 0;
-  }
-  size_t saved_bytes = std::get<1>(ret) - utf32_output;
-  if (std::get<0>(ret) != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::BIG>(
-            std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-void implementation::change_endianness_utf16(const char16_t *input,
-                                             size_t length,
-                                             char16_t *output) const noexcept {
-  size_t pos = 0;
-  const __m512i byteflip = _mm512_setr_epi64(
-      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-      0x0607040502030001, 0x0e0f0c0d0a0b0809);
-  while (pos + 32 <= length) {
-    __m512i utf16 = _mm512_loadu_si512((const __m512i *)(input + pos));
-    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
-    _mm512_storeu_si512(output + pos, utf16);
-    pos += 32;
-  }
-  if (pos < length) {
-    __mmask32 m((1U << (length - pos)) - 1);
-    __m512i utf16 = _mm512_maskz_loadu_epi16(m, (const __m512i *)(input + pos));
-    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
-    _mm512_mask_storeu_epi16(output + pos, m, utf16);
-  }
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-simdutf_warn_unused size_t implementation::count_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  const char16_t *ptr = input;
-  size_t count{0};
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
 
-  if (length >= 32) {
-    const char16_t *end = input + length - 32;
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
+  }
 
-    const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
-    const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 16 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // we have an error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      return 0;
+    }
+    if (pos < size) {
+      size_t howmany =
+          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      utf32_output += howmany;
+    }
+    return utf32_output - start;
+  }
 
-    while (ptr <= end) {
-      __m512i utf16 = _mm512_loadu_si512((const __m512i *)ptr);
-      ptr += 32;
-      uint64_t not_high_surrogate =
-          static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) |
-                                _mm512_cmplt_epu16_mask(utf16, low));
-      count += count_ones(not_high_surrogate);
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, utf32_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf32_output += res.count;
+      }
     }
+    return result(error_code::SUCCESS, utf32_output - start);
   }
 
-  return count + scalar::utf16::count_code_points<endianness::LITTLE>(
-                     ptr, length - (ptr - input));
-}
-
-simdutf_warn_unused size_t implementation::count_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  const char16_t *ptr = input;
-  size_t count{0};
-  if (length >= 32) {
-
-    const char16_t *end = input + length - 32;
-
-    const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
-    const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);
-
-    const __m512i byteflip = _mm512_setr_epi64(
-        0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-        0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-        0x0607040502030001, 0x0e0f0c0d0a0b0809);
-    while (ptr <= end) {
-      __m512i utf16 =
-          _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i *)ptr), byteflip);
-      ptr += 32;
-      uint64_t not_high_surrogate =
-          static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) |
-                                _mm512_cmplt_epu16_mask(utf16, low));
-      count += count_ones(not_high_surrogate);
-    }
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
 
-  return count + scalar::utf16::count_code_points<endianness::BIG>(
-                     ptr, length - (ptr - input));
-}
-
-simdutf_warn_unused size_t
-implementation::count_utf8(const char *input, size_t length) const noexcept {
-  const uint8_t *str = reinterpret_cast<const uint8_t *>(input);
-  size_t answer =
-      length / sizeof(__m512i) *
-      sizeof(__m512i); // Number of 512-bit chunks that fits into the length.
-  size_t i = 0;
-  __m512i unrolled_popcount{0};
-
-  const __m512i continuation = _mm512_set1_epi8(char(0b10111111));
-
-  while (i + sizeof(__m512i) <= length) {
-    size_t iterations = (length - i) / sizeof(__m512i);
-
-    size_t max_i = i + iterations * sizeof(__m512i) - sizeof(__m512i);
-    for (; i + 8 * sizeof(__m512i) <= max_i; i += 8 * sizeof(__m512i)) {
-      __m512i input1 = _mm512_loadu_si512((const __m512i *)(str + i));
-      __m512i input2 =
-          _mm512_loadu_si512((const __m512i *)(str + i + sizeof(__m512i)));
-      __m512i input3 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 2 * sizeof(__m512i)));
-      __m512i input4 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 3 * sizeof(__m512i)));
-      __m512i input5 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 4 * sizeof(__m512i)));
-      __m512i input6 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 5 * sizeof(__m512i)));
-      __m512i input7 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 6 * sizeof(__m512i)));
-      __m512i input8 =
-          _mm512_loadu_si512((const __m512i *)(str + i + 7 * sizeof(__m512i)));
-
-      __mmask64 mask1 = _mm512_cmple_epi8_mask(input1, continuation);
-      __mmask64 mask2 = _mm512_cmple_epi8_mask(input2, continuation);
-      __mmask64 mask3 = _mm512_cmple_epi8_mask(input3, continuation);
-      __mmask64 mask4 = _mm512_cmple_epi8_mask(input4, continuation);
-      __mmask64 mask5 = _mm512_cmple_epi8_mask(input5, continuation);
-      __mmask64 mask6 = _mm512_cmple_epi8_mask(input6, continuation);
-      __mmask64 mask7 = _mm512_cmple_epi8_mask(input7, continuation);
-      __mmask64 mask8 = _mm512_cmple_epi8_mask(input8, continuation);
+}; // struct utf8_checker
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+/* begin file src/generic/utf32.h */
+#include <limits>
 
-      __m512i mask_register = _mm512_set_epi64(mask8, mask7, mask6, mask5,
-                                               mask4, mask3, mask2, mask1);
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf32 {
 
-      unrolled_popcount = _mm512_add_epi64(unrolled_popcount,
-                                           _mm512_popcnt_epi64(mask_register));
-    }
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
 
-    for (; i <= max_i; i += sizeof(__m512i)) {
-      __m512i more_input = _mm512_loadu_si512((const __m512i *)(str + i));
-      uint64_t continuation_bitmask = static_cast<uint64_t>(
-          _mm512_cmple_epi8_mask(more_input, continuation));
-      answer -= count_ones(continuation_bitmask);
-    }
-  }
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
 
-  answer -= _mm512_reduce_add_epi64(unrolled_popcount);
+  const char32_t *start = input;
 
-  return answer + scalar::utf8::count_code_points(
-                      reinterpret_cast<const char *>(str + i), length - i);
-}
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
 
-simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
-    const char *buf, size_t len) const noexcept {
-  return count_utf8(buf, len);
-}
+  const size_t N = vector_u32::ELEMENTS;
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return scalar::utf16::latin1_length_from_utf16(length);
-}
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return scalar::utf32::latin1_length_from_utf32(length);
-}
+  size_t counter = 0;
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  const char16_t *ptr = input;
-  size_t count{0};
-  if (length >= 32) {
-    const char16_t *end = input + length - 32;
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
 
-    const __m512i v_007f = _mm512_set1_epi16((uint16_t)0x007f);
-    const __m512i v_07ff = _mm512_set1_epi16((uint16_t)0x07ff);
-    const __m512i v_dfff = _mm512_set1_epi16((uint16_t)0xdfff);
-    const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
 
-    while (ptr <= end) {
-      __m512i utf16 = _mm512_loadu_si512((const __m512i *)ptr);
-      ptr += 32;
-      __mmask32 ascii_bitmask = _mm512_cmple_epu16_mask(utf16, v_007f);
-      __mmask32 two_bytes_bitmask =
-          _mm512_mask_cmple_epu16_mask(~ascii_bitmask, utf16, v_07ff);
-      __mmask32 not_one_two_bytes = ~(ascii_bitmask | two_bytes_bitmask);
-      __mmask32 surrogates_bitmask =
-          _mm512_mask_cmple_epu16_mask(not_one_two_bytes, utf16, v_dfff) &
-          _mm512_mask_cmpge_epu16_mask(not_one_two_bytes, utf16, v_d800);
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-      size_t ascii_count = count_ones(ascii_bitmask);
-      size_t two_bytes_count = count_ones(two_bytes_bitmask);
-      size_t surrogate_bytes_count = count_ones(surrogates_bitmask);
-      size_t three_bytes_count =
-          32 - ascii_count - two_bytes_count - surrogate_bytes_count;
+        input += 4 * N;
+      }
 
-      count += ascii_count + 2 * two_bytes_count + 3 * three_bytes_count +
-               2 * surrogate_bytes_count;
+      counter += acc.sum();
     }
   }
 
-  return count + scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(
-                     ptr, length - (ptr - input));
-}
-
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  const char16_t *ptr = input;
-  size_t count{0};
-
-  if (length >= 32) {
-    const char16_t *end = input + length - 32;
-
-    const __m512i v_007f = _mm512_set1_epi16((uint16_t)0x007f);
-    const __m512i v_07ff = _mm512_set1_epi16((uint16_t)0x07ff);
-    const __m512i v_dfff = _mm512_set1_epi16((uint16_t)0xdfff);
-    const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-    const __m512i byteflip = _mm512_setr_epi64(
-        0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
-        0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
-        0x0607040502030001, 0x0e0f0c0d0a0b0809);
-    while (ptr <= end) {
-      __m512i utf16 = _mm512_loadu_si512((const __m512i *)ptr);
-      utf16 = _mm512_shuffle_epi8(utf16, byteflip);
-      ptr += 32;
-      __mmask32 ascii_bitmask = _mm512_cmple_epu16_mask(utf16, v_007f);
-      __mmask32 two_bytes_bitmask =
-          _mm512_mask_cmple_epu16_mask(~ascii_bitmask, utf16, v_07ff);
-      __mmask32 not_one_two_bytes = ~(ascii_bitmask | two_bytes_bitmask);
-      __mmask32 surrogates_bitmask =
-          _mm512_mask_cmple_epu16_mask(not_one_two_bytes, utf16, v_dfff) &
-          _mm512_mask_cmpge_epu16_mask(not_one_two_bytes, utf16, v_d800);
+        input += N;
+      }
 
-      size_t ascii_count = count_ones(ascii_bitmask);
-      size_t two_bytes_count = count_ones(two_bytes_bitmask);
-      size_t surrogate_bytes_count = count_ones(surrogates_bitmask);
-      size_t three_bytes_count =
-          32 - ascii_count - two_bytes_count - surrogate_bytes_count;
-      count += ascii_count + 2 * two_bytes_count + 3 * three_bytes_count +
-               2 * surrogate_bytes_count;
+      counter += acc.sum();
     }
   }
 
-  return count + scalar::utf16::utf8_length_from_utf16<endianness::BIG>(
-                     ptr, length - (ptr - input));
-}
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
+  }
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return implementation::count_utf16le(input, length);
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
 }
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return implementation::count_utf16be(input, length);
-}
+} // namespace utf32
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf16_length_from_latin1(length);
-}
+// other functions
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/generic/utf8.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8 {
 
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf32_length_from_latin1(length);
+using namespace simd;
+
+simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.gt(-65);
+    count += count_ones(utf8_continuation_mask);
+  }
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
-simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
-    const char *input, size_t length) const noexcept {
-  const uint8_t *str = reinterpret_cast<const uint8_t *>(input);
-  size_t answer = length / sizeof(__m512i) * sizeof(__m512i);
-  size_t i = 0;
-  if (answer >= 2048) { // long strings optimization
-    unsigned char v_0xFF = 0xff;
-    __m512i eight_64bits = _mm512_setzero_si512();
-    while (i + sizeof(__m512i) <= length) {
-      __m512i runner = _mm512_setzero_si512();
-      size_t iterations = (length - i) / sizeof(__m512i);
-      if (iterations > 255) {
-        iterations = 255;
-      }
-      size_t max_i = i + iterations * sizeof(__m512i) - sizeof(__m512i);
-      for (; i + 4 * sizeof(__m512i) <= max_i; i += 4 * sizeof(__m512i)) {
-        // Load four __m512i vectors
-        __m512i input1 = _mm512_loadu_si512((const __m512i *)(str + i));
-        __m512i input2 =
-            _mm512_loadu_si512((const __m512i *)(str + i + sizeof(__m512i)));
-        __m512i input3 = _mm512_loadu_si512(
-            (const __m512i *)(str + i + 2 * sizeof(__m512i)));
-        __m512i input4 = _mm512_loadu_si512(
-            (const __m512i *)(str + i + 3 * sizeof(__m512i)));
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
 
-        // Generate four masks
-        __mmask64 mask1 =
-            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input1);
-        __mmask64 mask2 =
-            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input2);
-        __mmask64 mask3 =
-            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input3);
-        __mmask64 mask4 =
-            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), input4);
-        // Apply the masks and subtract from the runner
-        __m512i not_ascii1 =
-            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask1, v_0xFF);
-        __m512i not_ascii2 =
-            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask2, v_0xFF);
-        __m512i not_ascii3 =
-            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask3, v_0xFF);
-        __m512i not_ascii4 =
-            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask4, v_0xFF);
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
 
-        runner = _mm512_sub_epi8(runner, not_ascii1);
-        runner = _mm512_sub_epi8(runner, not_ascii2);
-        runner = _mm512_sub_epi8(runner, not_ascii3);
-        runner = _mm512_sub_epi8(runner, not_ascii4);
-      }
+  size_t pos = 0;
+  size_t count = 0;
 
-      for (; i <= max_i; i += sizeof(__m512i)) {
-        __m512i more_input = _mm512_loadu_si512((const __m512i *)(str + i));
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
 
-        __mmask64 mask =
-            _mm512_cmpgt_epi8_mask(_mm512_setzero_si512(), more_input);
-        __m512i not_ascii =
-            _mm512_mask_set1_epi8(_mm512_setzero_si512(), mask, v_0xFF);
-        runner = _mm512_sub_epi8(runner, not_ascii);
-      }
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
 
-      eight_64bits = _mm512_add_epi64(
-          eight_64bits, _mm512_sad_epu8(runner, _mm512_setzero_si512()));
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
     }
+  }
 
-    answer += _mm512_reduce_add_epi64(eight_64bits);
-  } else if (answer > 0) {
-    for (; i + sizeof(__m512i) <= length; i += sizeof(__m512i)) {
-      __m512i latin = _mm512_loadu_si512((const __m512i *)(str + i));
-      uint64_t non_ascii = _mm512_movepi8_mask(latin);
-      answer += count_ones(non_ascii);
-    }
+  if (iterations > 0) {
+    count += local.sum_bytes();
   }
-  return answer + scalar::latin1::utf8_length_from_latin1(
-                      reinterpret_cast<const char *>(str + i), length - i);
+
+  count += counters.sum();
+
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
+#endif
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
-    const char *input, size_t length) const noexcept {
+simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
+                                                    size_t size) {
   size_t pos = 0;
   size_t count = 0;
   // This algorithm could no doubt be improved!
-  for (; pos + 64 <= length; pos += 64) {
-    __m512i utf8 = _mm512_loadu_si512((const __m512i *)(input + pos));
-    uint64_t utf8_continuation_mask =
-        _mm512_cmplt_epi8_mask(utf8, _mm512_set1_epi8(-65 + 1));
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
     // We count one word for anything that is not a continuation (so
     // leading bytes).
     count += 64 - count_ones(utf8_continuation_mask);
-    uint64_t utf8_4byte =
-        _mm512_cmpge_epu8_mask(utf8, _mm512_set1_epi8(int8_t(240)));
+    int64_t utf8_4byte = input.gteq_unsigned(240);
     count += count_ones(utf8_4byte);
   }
-  return count +
-         scalar::utf8::utf16_length_from_utf8(input + pos, length - pos);
+  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
 }
+} // namespace utf8
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const char32_t *ptr = input;
-  size_t count{0};
-
-  if (length >= 16) {
-    const char32_t *end = input + length - 16;
-
-    const __m512i v_0000_007f = _mm512_set1_epi32((uint32_t)0x7f);
-    const __m512i v_0000_07ff = _mm512_set1_epi32((uint32_t)0x7ff);
-    const __m512i v_0000_ffff = _mm512_set1_epi32((uint32_t)0x0000ffff);
-
-    while (ptr <= end) {
-      __m512i utf32 = _mm512_loadu_si512((const __m512i *)ptr);
-      ptr += 16;
-      __mmask16 ascii_bitmask = _mm512_cmple_epu32_mask(utf32, v_0000_007f);
-      __mmask16 two_bytes_bitmask = _mm512_mask_cmple_epu32_mask(
-          _knot_mask16(ascii_bitmask), utf32, v_0000_07ff);
-      __mmask16 three_bytes_bitmask = _mm512_mask_cmple_epu32_mask(
-          _knot_mask16(_mm512_kor(ascii_bitmask, two_bytes_bitmask)), utf32,
-          v_0000_ffff);
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/generic/utf16.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf16 {
 
-      size_t ascii_count = count_ones(ascii_bitmask);
-      size_t two_bytes_count = count_ones(two_bytes_bitmask);
-      size_t three_bytes_count = count_ones(three_bytes_bitmask);
-      size_t four_bytes_count =
-          16 - ascii_count - two_bytes_count - three_bytes_count;
-      count += ascii_count + 2 * two_bytes_count + 3 * three_bytes_count +
-               4 * four_bytes_count;
+template <endianness big_endian>
+simdutf_really_inline size_t count_code_points(const char16_t *in,
+                                               size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
     }
+    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
+    count += count_ones(not_pair) / 2;
   }
-
   return count +
-         scalar::utf32::utf8_length_from_utf32(ptr, length - (ptr - input));
+         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
 }
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const char32_t *ptr = input;
-  size_t count{0};
-
-  if (length >= 16) {
-    const char32_t *end = input + length - 16;
-
-    const __m512i v_0000_ffff = _mm512_set1_epi32((uint32_t)0x0000ffff);
-
-    while (ptr <= end) {
-      __m512i utf32 = _mm512_loadu_si512((const __m512i *)ptr);
-      ptr += 16;
-      __mmask16 surrogates_bitmask =
-          _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);
-
-      count += 16 + count_ones(surrogates_bitmask);
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
+                                                    size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  // This algorithm could no doubt be improved!
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
     }
-  }
+    uint64_t ascii_mask = input.lteq(0x7F);
+    uint64_t twobyte_mask = input.lteq(0x7FF);
+    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
 
-  return count +
-         scalar::utf32::utf16_length_from_utf32(ptr, length - (ptr - input));
+    size_t ascii_count = count_ones(ascii_mask) / 2;
+    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
+    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
+    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
+    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
+             ascii_count;
+  }
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return implementation::count_utf8(input, length);
-}
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
+  const auto one = vector_u16::splat(1);
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
+  auto v_count = vector_u16::zero();
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
     }
-  }
-}
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
-}
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
 
-size_t implementation::binary_to_base64(const char *input, size_t length,
-                                        char *output,
-                                        base64_options options) const noexcept {
-  if (options & base64_url) {
-    return encode_base64<true>(output, input, length, options);
-  } else {
-    return encode_base64<false>(output, input, length, options);
-  }
-}
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
 
-} // namespace icelake
-} // namespace simdutf
+    /*
+        Explanation how the counting works.
 
-/* begin file src/simdutf/icelake/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
 
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
 
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-SIMDUTF_POP_DISABLE_WARNINGS
-#endif // end of workaround
-/* end file src/simdutf/icelake/end.h */
-/* end file src/icelake/implementation.cpp */
-#endif
-#if SIMDUTF_IMPLEMENTATION_HASWELL
-/* begin file src/haswell/implementation.cpp */
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
 
-/* begin file src/simdutf/haswell/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "haswell"
-// #define SIMDUTF_IMPLEMENTATION haswell
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
 
-#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
-// nothing needed.
-#else
-SIMDUTF_TARGET_HASWELL
-#endif
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
+    }
+  }
 
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-// clang-format off
-SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
-// clang-format on
-#endif // end of workaround
-/* end file src/simdutf/haswell/begin.h */
-namespace simdutf {
-namespace haswell {
-namespace {
-#ifndef SIMDUTF_HASWELL_H
-  #error "haswell.h must be included"
-#endif
-using namespace simd;
+  if (iteration > 0) {
+    count += v_count.sum();
+  }
 
-simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
-  return input.reduce_or().is_ascii();
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte =
-      prev1.saturating_sub(0b11000000u - 1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte =
-      prev2.saturating_sub(0b11100000u - 1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte =
-      prev3.saturating_sub(0b11110000u - 1); // Only 1111____ will be > 0
-  // Caller requires a bool (all 1's). All values resulting from the subtraction
-  // will be <= 64, so signed comparison is fine.
-  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) >
-         int8_t(0);
+template <endianness big_endian>
+simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
+                                                     size_t size) {
+  return count_code_points<big_endian>(in, size);
 }
 
-simdutf_really_inline simd8<bool>
-must_be_2_3_continuation(const simd8<uint8_t> prev2,
-                         const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte =
-      prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be > 0x80
-  simd8<uint8_t> is_fourth_byte =
-      prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be > 0x80
-  return simd8<bool>(is_third_byte | is_fourth_byte);
+simdutf_really_inline void
+change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
+  size_t pos = 0;
+
+  while (pos < size / 32 * 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    input.swap_bytes();
+    input.store(reinterpret_cast<uint16_t *>(output));
+    pos += 32;
+    output += 32;
+  }
+
+  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
 }
 
-/* begin file src/haswell/avx2_validate_utf16.cpp */
+} // namespace utf16
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf16.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf16 {
 /*
+    UTF-16 validation
+    --------------------------------------------------
+
     In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
 
     In a vectorized algorithm we want to examine the most significant
@@ -28586,7 +36799,7 @@ must_be_2_3_continuation(const simd8<uint8_t> prev2,
     - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
     - there must not be sole low surrogate nor high surrogate
 
-    We're going to build three bitmasks based on the 3rd nibble:
+    We are going to build three bitmasks based on the 3rd nibble:
     - V = valid word,
     - L = low surrogate (0xd800 .. 0xdbff)
     - H = high surrogate (0xdc00 .. 0xdfff)
@@ -28606,96 +36819,12 @@ must_be_2_3_continuation(const simd8<uint8_t> prev2,
                                   the last bit can be zero, we just consume 7
    code units and recheck this word in the next iteration
 */
-
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
-template <endianness big_endian>
-const char16_t *avx2_validate_utf16(const char16_t *input, size_t size) {
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
-
-    const auto in = simd16<uint16_t>::pack(t0, t1);
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-    if (surrogates_bitmask == 0x0) {
-      input += simd16<uint16_t>::ELEMENTS * 2;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint32_t V = ~surrogates_bitmask;
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint32_t H = vH.to_bitmask();
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint32_t L = ~H & surrogates_bitmask;
-
-      const uint32_t a =
-          L & (H >> 1); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint32_t b =
-          a << 1; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint32_t c = V | a | b; // Combine all the masks into the final one.
-
-      if (c == 0xffffffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += simd16<uint16_t>::ELEMENTS * 2;
-      } else if (c == 0x7fffffff) {
-        // The 31 lower code units of the input register contains valid UTF-16.
-        // The 31 word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += simd16<uint16_t>::ELEMENTS * 2 - 1;
-      } else {
-        return nullptr;
-      }
-    }
-  }
-
-  return input;
-}
-
 template <endianness big_endian>
-const result avx2_validate_utf16_with_errors(const char16_t *input,
-                                             size_t size) {
+const result validate_utf16_with_errors(const char16_t *input, size_t size) {
   if (simdutf_unlikely(size == 0)) {
     return result(error_code::SUCCESS, 0);
   }
+
   const char16_t *start = input;
   const char16_t *end = input + size;
 
@@ -28704,2378 +36833,3299 @@ const result avx2_validate_utf16_with_errors(const char16_t *input,
   const auto v_fc = simd8<uint8_t>::splat(0xfc);
   const auto v_dc = simd8<uint8_t>::splat(0xdc);
 
-  while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
+  while (input + simd16<uint16_t>::SIZE * 2 < end) {
     // 0. Load data: since the validation takes into account only higher
     //    byte of each word, we compress the two vectors into one which
     //    consists only the higher bytes.
     auto in0 = simd16<uint16_t>(input);
-    auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
 
-    const auto in = simd16<uint16_t>::pack(t0, t1);
+    // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16
+    // and yields a single vector having only higher bytes of characters.
+    const auto in = utf16_gather_high_bytes<big_endian>(in0, in1);
 
     // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
     const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-    if (surrogates_bitmask == 0x0) {
-      input += simd16<uint16_t>::ELEMENTS * 2;
+    const uint16_t surrogates_bitmask =
+        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
+    if (surrogates_bitmask == 0x0000) {
+      input += 16;
     } else {
       // 2. We have some surrogates that have to be distinguished:
       //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
       //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
       //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher byte)
 
       // V - non-surrogate code units
       //     V = not surrogates_wordmask
-      const uint32_t V = ~surrogates_bitmask;
+      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
 
       // H - word-mask for high surrogates: the six highest bits are 0b1101'11
       const auto vH = (in & v_fc) == v_dc;
-      const uint32_t H = vH.to_bitmask();
+      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
 
       // L - word mask for low surrogates
       //     L = not H and surrogates_wordmask
-      const uint32_t L = ~H & surrogates_bitmask;
-
-      const uint32_t a =
-          L & (H >> 1); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint32_t b =
-          a << 1; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint32_t c = V | a | b; // Combine all the masks into the final one.
-
-      if (c == 0xffffffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += simd16<uint16_t>::ELEMENTS * 2;
-      } else if (c == 0x7fffffff) {
-        // The 31 lower code units of the input register contains valid UTF-16.
-        // The 31 word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += simd16<uint16_t>::ELEMENTS * 2 - 1;
-      } else {
-        return result(error_code::SURROGATE, input - start);
-      }
-    }
-  }
-
-  return result(error_code::SUCCESS, input - start);
-}
-/* end file src/haswell/avx2_validate_utf16.cpp */
-/* begin file src/haswell/avx2_validate_utf32le.cpp */
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
-const char32_t *avx2_validate_utf32le(const char32_t *input, size_t size) {
-  const char32_t *end = input + size;
-
-  const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
-  const __m256i offset = _mm256_set1_epi32(0xffff2000);
-  const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
-  __m256i currentmax = _mm256_setzero_si256();
-  __m256i currentoffsetmax = _mm256_setzero_si256();
-
-  while (input + 8 < end) {
-    const __m256i in = _mm256_loadu_si256((__m256i *)input);
-    currentmax = _mm256_max_epu32(in, currentmax);
-    currentoffsetmax =
-        _mm256_max_epu32(_mm256_add_epi32(in, offset), currentoffsetmax);
-    input += 8;
-  }
-  __m256i is_zero =
-      _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
-  if (_mm256_testz_si256(is_zero, is_zero) == 0) {
-    return nullptr;
-  }
-
-  is_zero = _mm256_xor_si256(
-      _mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
-  if (_mm256_testz_si256(is_zero, is_zero) == 0) {
-    return nullptr;
-  }
-
-  return input;
-}
-
-const result avx2_validate_utf32le_with_errors(const char32_t *input,
-                                               size_t size) {
-  const char32_t *start = input;
-  const char32_t *end = input + size;
-
-  const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
-  const __m256i offset = _mm256_set1_epi32(0xffff2000);
-  const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
-  __m256i currentmax = _mm256_setzero_si256();
-  __m256i currentoffsetmax = _mm256_setzero_si256();
-
-  while (input + 8 < end) {
-    const __m256i in = _mm256_loadu_si256((__m256i *)input);
-    currentmax = _mm256_max_epu32(in, currentmax);
-    currentoffsetmax =
-        _mm256_max_epu32(_mm256_add_epi32(in, offset), currentoffsetmax);
-
-    __m256i is_zero = _mm256_xor_si256(
-        _mm256_max_epu32(currentmax, standardmax), standardmax);
-    if (_mm256_testz_si256(is_zero, is_zero) == 0) {
-      return result(error_code::TOO_LARGE, input - start);
-    }
-
-    is_zero =
-        _mm256_xor_si256(_mm256_max_epu32(currentoffsetmax, standardoffsetmax),
-                         standardoffsetmax);
-    if (_mm256_testz_si256(is_zero, is_zero) == 0) {
-      return result(error_code::SURROGATE, input - start);
-    }
-    input += 8;
-  }
-
-  return result(error_code::SUCCESS, input - start);
-}
-/* end file src/haswell/avx2_validate_utf32le.cpp */
-
-/* begin file src/haswell/avx2_convert_latin1_to_utf8.cpp */
-std::pair<const char *, char *>
-avx2_convert_latin1_to_utf8(const char *latin1_input, size_t len,
-                            char *utf8_output) {
-  const char *end = latin1_input + len;
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
-  const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
-  const size_t safety_margin = 12;
-
-  while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) {
-    __m128i in8 = _mm_loadu_si128((__m128i *)latin1_input);
-    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
-    const __m128i v_80 = _mm_set1_epi8((char)0x80);
-    if (_mm_testz_si128(in8, v_80)) { // ASCII fast path!!!!
-      // 1. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, in8);
-      // 2. adjust pointers
-      latin1_input += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
-    }
-    // We proceed only with the first 16 bytes.
-    const __m256i in = _mm256_cvtepu8_epi16((in8));
-
-    // 1. prepare 2-byte values
-    // input 16-bit word : [0000|0000|aabb|bbbb] x 8
-    // expected output   : [1100|00aa|10bb|bbbb] x 8
-    const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-    const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
-
-    // t0 = [0000|00aa|bbbb|bb00]
-    const __m256i t0 = _mm256_slli_epi16(in, 2);
-    // t1 = [0000|00aa|0000|0000]
-    const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-    // t2 = [0000|0000|00bb|bbbb]
-    const __m256i t2 = _mm256_and_si256(in, v_003f);
-    // t3 = [000a|aaaa|00bb|bbbb]
-    const __m256i t3 = _mm256_or_si256(t1, t2);
-    // t4 = [1100|00aa|10bb|bbbb]
-    const __m256i t4 = _mm256_or_si256(t3, v_c080);
-
-    // 2. merge ASCII and 2-byte codewords
-
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
-
-    const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask);
-
-    // 3. prepare bitmask for 8-bit lookup
-    const uint32_t M0 = one_byte_bitmask & 0x55555555;
-    const uint32_t M1 = M0 >> 7;
-    const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-    // 4. pack the bytes
-
-    const uint8_t *row =
-        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-    const uint8_t *row_2 =
-        &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >> 16)]
-                                                            [0];
-
-    const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-    const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
-
-    const __m256i utf8_packed = _mm256_shuffle_epi8(
-        utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-    // 5. store bytes
-    _mm_storeu_si128((__m128i *)utf8_output,
-                     _mm256_castsi256_si128(utf8_packed));
-    utf8_output += row[0];
-    _mm_storeu_si128((__m128i *)utf8_output,
-                     _mm256_extractf128_si256(utf8_packed, 1));
-    utf8_output += row_2[0];
-
-    // 6. adjust pointers
-    latin1_input += 16;
-    continue;
-
-  } // while
-  return std::make_pair(latin1_input, utf8_output);
-}
-/* end file src/haswell/avx2_convert_latin1_to_utf8.cpp */
-/* begin file src/haswell/avx2_convert_latin1_to_utf16.cpp */
-template <endianness big_endian>
-std::pair<const char *, char16_t *>
-avx2_convert_latin1_to_utf16(const char *latin1_input, size_t len,
-                             char16_t *utf16_output) {
-  size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 32
-
-  size_t i = 0;
-  for (; i < rounded_len; i += 16) {
-    // Load 16 bytes from the address (input + i) into a xmm register
-    __m128i xmm0 =
-        _mm_loadu_si128(reinterpret_cast<const __m128i *>(latin1_input + i));
-
-    // Zero extend each byte in xmm0 to word and put it in another xmm register
-    __m128i xmm1 = _mm_cvtepu8_epi16(xmm0);
-
-    // Shift xmm0 to the right by 8 bytes
-    xmm0 = _mm_srli_si128(xmm0, 8);
-
-    // Zero extend each byte in the shifted xmm0 to word in xmm0
-    xmm0 = _mm_cvtepu8_epi16(xmm0);
-
-    if (big_endian) {
-      const __m128i swap =
-          _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-      xmm0 = _mm_shuffle_epi8(xmm0, swap);
-      xmm1 = _mm_shuffle_epi8(xmm1, swap);
-    }
-
-    // Store the contents of xmm1 into the address pointed by (output + i)
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + i), xmm1);
-
-    // Store the contents of xmm0 into the address pointed by (output + i + 8)
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + i + 8), xmm0);
-  }
-
-  return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len);
-}
-/* end file src/haswell/avx2_convert_latin1_to_utf16.cpp */
-/* begin file src/haswell/avx2_convert_latin1_to_utf32.cpp */
-std::pair<const char *, char32_t *>
-avx2_convert_latin1_to_utf32(const char *buf, size_t len,
-                             char32_t *utf32_output) {
-  size_t rounded_len = ((len | 7) ^ 7); // Round down to nearest multiple of 8
-
-  for (size_t i = 0; i < rounded_len; i += 8) {
-    // Load 8 Latin1 characters into a 64-bit register
-    __m128i in = _mm_loadl_epi64((__m128i *)&buf[i]);
-
-    // Zero extend each set of 8 Latin1 characters to 8 32-bit integers using
-    // vpmovzxbd
-    __m256i out = _mm256_cvtepu8_epi32(in);
-
-    // Store the results back to memory
-    _mm256_storeu_si256((__m256i *)&utf32_output[i], out);
-  }
-
-  // return pointers pointing to where we left off
-  return std::make_pair(buf + rounded_len, utf32_output + rounded_len);
-}
-/* end file src/haswell/avx2_convert_latin1_to_utf32.cpp */
-
-/* begin file src/haswell/avx2_convert_utf8_to_utf16.cpp */
-// depends on "tables/utf8_to_utf16_tables.h"
-
-// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
-// end of the code points. Only the least significant 12 bits of the mask
-// are accessed.
-// It returns how many bytes were consumed (up to 12).
-template <endianness big_endian>
-size_t convert_masked_utf8_to_utf16(const char *input,
-                                    uint64_t utf8_end_of_code_point_mask,
-                                    char16_t *&utf16_output) {
-  // we use an approach where we try to process up to 12 input bytes.
-  // Why 12 input bytes and not 16? Because we are concerned with the size of
-  // the lookup tables. Also 12 is nicely divisible by two and three.
-  //
-  //
-  // Optimization note: our main path below is load-latency dependent. Thus it
-  // is maybe beneficial to have fast paths that depend on branch prediction but
-  // have less latency. This results in more instructions but, potentially, also
-  // higher speeds.
-  //
-  // We first try a few fast paths.
-  const __m128i swap =
-      _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-  const __m128i in = _mm_loadu_si128((__m128i *)input);
-  const uint16_t input_utf8_end_of_code_point_mask =
-      utf8_end_of_code_point_mask & 0xfff;
-  if (utf8_end_of_code_point_mask == 0xfff) {
-    // We process the data in chunks of 12 bytes.
-    __m256i ascii = _mm256_cvtepu8_epi16(in);
-    if (big_endian) {
-      const __m256i swap256 = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      ascii = _mm256_shuffle_epi8(ascii, swap256);
-    }
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii);
-    utf16_output += 12; // We wrote 12 16-bit characters.
-    return 12;          // We consumed 12 bytes.
-  }
-  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
-    // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte
-    // UTF-16 code units. There is probably a more efficient sequence, but the
-    // following might do.
-    const __m128i sh =
-        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
-    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
-    if (big_endian)
-      composed = _mm_shuffle_epi8(composed, swap);
-    _mm_storeu_si128((__m128i *)utf16_output, composed);
-    utf16_output += 8; // We wrote 16 bytes, 8 code points.
-    return 16;
-  }
-  if (input_utf8_end_of_code_point_mask == 0x924) {
-    // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte
-    // UTF-16 code units. There is probably a more efficient sequence, but the
-    // following might do.
-    const __m128i sh =
-        _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii =
-        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
-    const __m128i middlebyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    const __m128i highbyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
-    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
-    if (big_endian)
-      composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
-    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
-    utf16_output += 4;
-    return 12;
-  }
+      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
 
-  const uint8_t idx = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][0];
-  const uint8_t consumed = simdutf::tables::utf8_to_utf16::utf8bigindex
-      [input_utf8_end_of_code_point_mask][1];
-  if (idx < 64) {
-    // SIX (6) input code-code units
-    // this is a relatively easy scenario
-    // we process SIX (6) input code-code units. The max length in bytes of six
-    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
-    // processors where pdep/pext is fast, we might be able to use a small
-    // lookup table.
-    const __m128i sh = _mm_loadu_si128(
-        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
-    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
-    if (big_endian)
-      composed = _mm_shuffle_epi8(composed, swap);
-    _mm_storeu_si128((__m128i *)utf16_output, composed);
-    utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential
-                       // overflow of 4 bytes.
-  } else if (idx < 145) {
-    // FOUR (4) input code-code units
-    const __m128i sh = _mm_loadu_si128(
-        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii =
-        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
-    const __m128i middlebyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    const __m128i highbyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
-    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
-    if (big_endian)
-      composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
-    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
-    utf16_output += 4; // Here we overflow by 8 bytes.
-  } else if (idx < 209) {
-    // TWO (2) input code-code units
-    //////////////
-    // There might be garbage inputs where a leading byte mascarades as a
-    // four-byte leading byte (by being followed by 3 continuation byte), but is
-    // not greater than 0xf0. This could trigger a buffer overflow if we only
-    // counted leading bytes of the form 0xf0 as generating surrogate pairs,
-    // without further UTF-8 validation. Thus we must be careful to ensure that
-    // only leading bytes at least as large as 0xf0 generate surrogate pairs. We
-    // do as at the cost of an extra mask.
-    /////////////
-    const __m128i sh = _mm_loadu_si128(
-        (const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
-    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
-    // correct for spurious high bit
-    const __m128i correct =
-        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
-    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
-    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
-    // We deliberately carry the leading four bits in highbyte if they are
-    // present, we remove them later when computing hightenbits.
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000));
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
-    // When we need to generate a surrogate pair (leading byte > 0xF0), then
-    // the corresponding 32-bit value in 'composed'  will be greater than
-    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
-    // location of the surrogate pairs.
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
-                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
-    const __m128i composedminus =
-        _mm_sub_epi32(composed, _mm_set1_epi32(0x10000));
-    const __m128i lowtenbits =
-        _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff));
-    // Notice the 0x3ff mask:
-    const __m128i hightenbits =
-        _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff));
-    const __m128i lowtenbitsadd =
-        _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00));
-    const __m128i hightenbitsadd =
-        _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800));
-    const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16);
-    __m128i surrogates = _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted);
-    uint32_t basic_buffer[4];
-    uint32_t basic_buffer_swap[4];
-    if (big_endian) {
-      _mm_storeu_si128((__m128i *)basic_buffer_swap,
-                       _mm_shuffle_epi8(composed, swap));
-      surrogates = _mm_shuffle_epi8(surrogates, swap);
-    }
-    _mm_storeu_si128((__m128i *)basic_buffer, composed);
-    uint32_t surrogate_buffer[4];
-    _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates);
-    for (size_t i = 0; i < 3; i++) {
-      if (basic_buffer[i] > 0x3c00000) {
-        utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff);
-        utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16);
-        utf16_output += 2;
+      const uint16_t a = static_cast<uint16_t>(
+          L & (H >> 1)); // A low surrogate must be followed by high one.
+                         // (A low surrogate placed in the 7th register's word
+                         // is an exception we handle.)
+      const uint16_t b = static_cast<uint16_t>(
+          a << 1); // Just mark that the opinput - startite fact is hold,
+                   // thanks to that we have only two masks for valid case.
+      const uint16_t c = static_cast<uint16_t>(
+          V | a | b); // Combine all the masks into the final one.
+
+      if (c == 0xffff) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0x7fff) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
       } else {
-        utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i])
-                                     : uint16_t(basic_buffer[i]);
-        utf16_output++;
+        return result(error_code::SURROGATE, input - start);
       }
     }
-  } else {
-    // here we know that there is an error but we do not handle errors
   }
-  return consumed;
+
+  return result(error_code::SUCCESS, input - start);
 }
-/* end file src/haswell/avx2_convert_utf8_to_utf16.cpp */
-/* begin file src/haswell/avx2_convert_utf8_to_utf32.cpp */
-// depends on "tables/utf8_to_utf16_tables.h"
 
-// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
-// end of the code points. Only the least significant 12 bits of the mask
-// are accessed.
-// It returns how many bytes were consumed (up to 12).
-size_t convert_masked_utf8_to_utf32(const char *input,
-                                    uint64_t utf8_end_of_code_point_mask,
-                                    char32_t *&utf32_output) {
-  // we use an approach where we try to process up to 12 input bytes.
-  // Why 12 input bytes and not 16? Because we are concerned with the size of
-  // the lookup tables. Also 12 is nicely divisible by two and three.
+} // namespace utf16
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/validate_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  // transcoding from UTF-8 to Latin 1
+/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_latin1 {
+using namespace simd;
+
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // For UTF-8 to Latin 1, we can allow any ASCII character, and any
+  // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or
+  // 0b11000010 and nothing else.
   //
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  constexpr const uint8_t FORBIDDEN = 0xff;
+
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      FORBIDDEN,
+      // 1110____ ________ <three byte lead in byte 1>
+      FORBIDDEN,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      FORBIDDEN);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
+
+              // ____0100 ________
+              FORBIDDEN,
+              // ____0101 ________
+              FORBIDDEN,
+              // ____011_ ________
+              FORBIDDEN, FORBIDDEN,
+
+              // ____1___ ________
+              FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN,
+              // ____1101 ________
+              FORBIDDEN, FORBIDDEN, FORBIDDEN);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+
+  validating_transcoder() : error(uint8_t(0)) {}
   //
-  // Optimization note: our main path below is load-latency dependent. Thus it
-  // is maybe beneficial to have fast paths that depend on branch prediction but
-  // have less latency. This results in more instructions but, potentially, also
-  // higher speeds.
+  // Check whether the current bytes are valid UTF-8.
   //
-  // We first try a few fast paths.
-  const __m128i in = _mm_loadu_si128((__m128i *)input);
-  const uint16_t input_utf8_end_of_code_point_mask =
-      utf8_end_of_code_point_mask & 0xfff;
-  if (utf8_end_of_code_point_mask == 0xfff) {
-    // We process the data in chunks of 12 bytes.
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
-                        _mm256_cvtepu8_epi32(in));
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8),
-                        _mm256_cvtepu8_epi32(_mm_srli_si128(in, 8)));
-    utf32_output += 12; // We wrote 12 32-bit characters.
-    return 12;          // We consumed 12 bytes.
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    this->error |= check_special_cases(input, prev1);
   }
-  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
-    // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte
-    // UTF-32 code units. There is probably a more efficient sequence, but the
-    // following might do.
-    const __m128i sh =
-        _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
-    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
-    _mm256_storeu_si256((__m256i *)utf32_output,
-                        _mm256_cvtepu16_epi32(composed));
-    utf32_output += 8; // We wrote 16 bytes, 8 code points.
-    return 16;
+
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char *latin1_output) {
+    size_t pos = 0;
+    char *start{latin1_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 16 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 16; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) >
+                       -65); // twos complement of -65 is 1011 1111 ...
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask =
+            input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                               // this case, we also have ASCII to account for.
+        if (utf8_continuation_mask & 1) {
+          return 0; // error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      return 0;
+    }
+    if (pos < size) {
+      size_t howmany =
+          scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      latin1_output += howmany;
+    }
+    return latin1_output - start;
   }
-  if (input_utf8_end_of_code_point_mask == 0x924) {
-    // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte
-    // UTF-32 code units. There is probably a more efficient sequence, but the
-    // following might do.
-    const __m128i sh =
-        _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii =
-        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
-    const __m128i middlebyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    const __m128i highbyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
-    _mm_storeu_si128((__m128i *)utf32_output, composed);
-    utf32_output += 4;
-    return 12;
+
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char *latin1_output) {
+    size_t pos = 0;
+    char *start{latin1_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        if (errors()) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, latin1_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        latin1_output += res.count;
+      }
+    }
+    return result(error_code::SUCCESS, latin1_output - start);
   }
-  /// We do not have a fast path available, so we fallback.
 
-  const uint8_t idx =
-      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
-  const uint8_t consumed =
-      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
-  if (idx < 64) {
-    // SIX (6) input code-code units
-    // this is a relatively easy scenario
-    // we process SIX (6) input code-code units. The max length in bytes of six
-    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
-    // processors where pdep/pext is fast, we might be able to use a small
-    // lookup table.
-    const __m128i sh =
-        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
-    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
-    _mm256_storeu_si256((__m256i *)utf32_output,
-                        _mm256_cvtepu16_epi32(composed));
-    utf32_output += 6; // We wrote 24 bytes, 6 code points. There is a potential
-    // overflow of 32 - 24 = 8 bytes.
-  } else if (idx < 145) {
-    // FOUR (4) input code-code units
-    const __m128i sh =
-        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii =
-        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
-    const __m128i middlebyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    const __m128i highbyte =
-        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
-    _mm_storeu_si128((__m128i *)utf32_output, composed);
-    utf32_output += 4;
-  } else if (idx < 209) {
-    // TWO (2) input code-code units
-    const __m128i sh =
-        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
-    const __m128i perm = _mm_shuffle_epi8(in, sh);
-    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
-    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
-    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
-    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
-    // correct for spurious high bit
-    const __m128i correct =
-        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
-    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
-    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
-    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000));
-    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
-    const __m128i composed =
-        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
-                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
-    _mm_storeu_si128((__m128i *)utf32_output, composed);
-    utf32_output +=
-        3; // We wrote 3 * 4 bytes, there is a potential overflow of 4 bytes.
-  } else {
-    // here we know that there is an error but we do not handle errors
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
+  }
+
+}; // struct utf8_checker
+} // namespace utf8_to_latin1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+/* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf8_to_latin1 {
+using namespace simd;
+
+simdutf_really_inline size_t convert_valid(const char *in, size_t size,
+                                           char *latin1_output) {
+  size_t pos = 0;
+  char *start{latin1_output};
+  // In the worst case, we have the haswell kernel which can cause an overflow
+  // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the last
+  // 16 bytes, and if the data is valid, then it is entirely safe because 16
+  // UTF-8 bytes generate much more than 8 bytes. However, you cannot generally
+  // assume that you have valid UTF-8 input, so we are going to go back from the
+  // end counting 8 leading bytes, to give us a good margin.
+  size_t leading_byte = 0;
+  size_t margin = size;
+  for (; margin > 0 && leading_byte < 8; margin--) {
+    leading_byte += (int8_t(in[margin - 1]) >
+                     -65); // twos complement of -65 is 1011 1111 ...
+  }
+  // If the input is long enough, then we have that margin-1 is the eight last
+  // leading byte.
+  const size_t safety_margin = size - margin + 1; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    if (input.is_ascii()) {
+      input.store((int8_t *)latin1_output);
+      latin1_output += 64;
+      pos += 64;
+    } else {
+      // you might think that a for-loop would work, but under Visual Studio, it
+      // is not good enough.
+      uint64_t utf8_continuation_mask =
+          input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                             // this case, we also have ASCII to account for.
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
+      size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times.
+      while (pos < max_starting_point) {
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        size_t consumed = convert_masked_utf8_to_latin1(
+            in + pos, utf8_end_of_code_point_mask, latin1_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
+    }
   }
-  return consumed;
+  if (pos < size) {
+    size_t howmany = scalar::utf8_to_latin1::convert_valid(in + pos, size - pos,
+                                                           latin1_output);
+    latin1_output += howmany;
+  }
+  return latin1_output - start;
 }
-/* end file src/haswell/avx2_convert_utf8_to_utf32.cpp */
 
-/* begin file src/haswell/avx2_convert_utf16_to_latin1.cpp */
-template <endianness big_endian>
-std::pair<const char16_t *, char *>
-avx2_convert_utf16_to_latin1(const char16_t *buf, size_t len,
-                             char *latin1_output) {
-  const char16_t *end = buf + len;
-  while (end - buf >= 16) {
-    // Load 16 UTF-16 characters into 256-bit AVX2 register
-    __m256i in = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+} // namespace utf8_to_latin1
+} // namespace
+} // namespace haswell
+} // namespace simdutf
+  // namespace simdutf
+/* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-    if (!match_system(big_endian)) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
-    }
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf32.h */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace utf32 {
 
-    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
-    if (_mm256_testz_si256(in, high_byte_mask)) {
-      // Pack 16-bit characters into 8-bit and store in latin1_output
-      __m128i lo = _mm256_extractf128_si256(in, 0);
-      __m128i hi = _mm256_extractf128_si256(in, 1);
-      __m128i latin1_packed_lo = _mm_packus_epi16(lo, lo);
-      __m128i latin1_packed_hi = _mm_packus_epi16(hi, hi);
-      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output),
-                       latin1_packed_lo);
-      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output + 8),
-                       latin1_packed_hi);
-      // Adjust pointers for next iteration
-      buf += 16;
-      latin1_output += 16;
-    } else {
-      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
-    }
-  } // while
-  return std::make_pair(buf, latin1_output);
-}
+simdutf_really_inline bool validate(const char32_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return true;
+  }
 
-template <endianness big_endian>
-std::pair<result, char *>
-avx2_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
-                                         char *latin1_output) {
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
-  while (end - buf >= 16) {
-    __m256i in = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+  const char32_t *end = input + size;
 
-    if (!match_system(big_endian)) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
-    }
+  using vector_u32 = simd32<uint32_t>;
 
-    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
-    if (_mm256_testz_si256(in, high_byte_mask)) {
-      __m128i lo = _mm256_extractf128_si256(in, 0);
-      __m128i hi = _mm256_extractf128_si256(in, 1);
-      __m128i latin1_packed_lo = _mm_packus_epi16(lo, lo);
-      __m128i latin1_packed_hi = _mm_packus_epi16(hi, hi);
-      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output),
-                       latin1_packed_lo);
-      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output + 8),
-                       latin1_packed_hi);
-      buf += 16;
-      latin1_output += 16;
-    } else {
-      // Fallback to scalar code for handling errors
-      for (int k = 0; k < 16; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
-        if (word <= 0xff) {
-          *latin1_output++ = char(word);
-        } else {
-          return std::make_pair(
-              result{error_code::TOO_LARGE, (size_t)(buf - start + k)},
-              latin1_output);
-        }
-      }
-      buf += 16;
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+  auto currentmax = vector_u32::zero();
+  auto currentoffsetmax = vector_u32::zero();
+
+  constexpr size_t N = vector_u32::ELEMENTS;
+
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
     }
-  } // while
-  return std::make_pair(result{error_code::SUCCESS, (size_t)(buf - start)},
-                        latin1_output);
+
+    currentmax = max(currentmax, in);
+    currentoffsetmax = max(currentoffsetmax, in + offset);
+    input += N;
+  }
+
+  const auto too_large = currentmax > standardmax;
+  if (too_large.any()) {
+    return false;
+  }
+
+  const auto surrogate = currentoffsetmax > standardoffsetmax;
+  if (surrogate.any()) {
+    return false;
+  }
+
+  return scalar::utf32::validate(input, end - input);
 }
-/* end file src/haswell/avx2_convert_utf16_to_latin1.cpp */
-/* begin file src/haswell/avx2_convert_utf16_to_utf8.cpp */
-/*
-    The vectorized algorithm works on single SSE register i.e., it
-    loads eight 16-bit code units.
 
-    We consider three cases:
-    1. an input register contains no surrogates and each value
-       is in range 0x0000 .. 0x07ff.
-    2. an input register contains no surrogates and values are
-       is in range 0x0000 .. 0xffff.
-    3. an input register contains surrogates --- i.e. codepoints
-       can have 16 or 32 bits.
+simdutf_really_inline result validate_with_errors(const char32_t *input,
+                                                  size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return result(error_code::SUCCESS, 0);
+  }
 
-    Ad 1.
+  const char32_t *start = input;
+  const char32_t *end = input + size;
 
-    When values are less than 0x0800, it means that a 16-bit code unit
-    can be converted into: 1) single UTF8 byte (when it is an ASCII
-    char) or 2) two UTF8 bytes.
+  using vector_u32 = simd32<uint32_t>;
 
-    For this case we do only some shuffle to obtain these 2-byte
-    codes and finally compress the whole SSE register with a single
-    shuffle.
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+  constexpr size_t N = vector_u32::ELEMENTS;
 
-    Ad 2.
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
+    }
 
-    When values fit in 16-bit code units, but are above 0x07ff, then
-    a single word may produce one, two or three UTF8 bytes.
+    const auto too_large = in > standardmax;
+    const auto surrogate = (in + offset) > standardoffsetmax;
 
-    We prepare data for all these three cases in two registers.
-    The first register contains lower two UTF8 bytes (used in all
-    cases), while the second one contains just the third byte for
-    the three-UTF8-bytes case.
+    const auto combined = too_large | surrogate;
+    if (simdutf_unlikely(combined.any())) {
+      const size_t consumed = input - start;
+      auto sr = scalar::utf32::validate_with_errors(input, end - input);
+      sr.count += consumed;
 
-    Finally these two registers are interleaved forming eight-element
-    array of 32-bit values. The array spans two SSE registers.
-    The bytes from the registers are compressed using two shuffles.
+      return sr;
+    }
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+    input += N;
+  }
 
+  const size_t consumed = input - start;
+  auto sr = scalar::utf32::validate_with_errors(input, end - input);
+  sr.count += consumed;
 
-    To summarize:
-    - We need two 256-entry tables that have 8704 bytes in total.
-*/
+  return sr;
+}
+
+} // namespace utf32
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/validate_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/generic/base64.h */
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
+namespace simdutf {
+namespace haswell {
+namespace {
+namespace base64 {
 
 /*
-  Returns a pair: the first unprocessed byte from buf and utf8_output
-  A scalar routing should carry on the conversion of the tail.
-*/
-template <endianness big_endian>
-std::pair<const char16_t *, char *>
-avx2_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) {
-  const char16_t *end = buf + len;
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
-  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
-  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+    The following template function implements API for Base64 decoding.
 
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    if (big_endian) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
+    An implementation is responsible for providing the `block64` type and
+    associated methods that perform actual conversion. Please refer
+    to any vectorized implementation to learn the API of these procedures.
+*/
+template <bool base64_url, bool ignore_garbage, typename chartype>
+full_result
+compress_decode_base64(char *dst, const chartype *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
+                                        : tables::base64::to_base64_value;
+  size_t equallocation =
+      srclen; // location of the first padding character if any
+  // skip trailing spaces
+  while (!ignore_garbage && srclen > 0 &&
+         scalar::base64::is_eight_byte(src[srclen - 1]) &&
+         to_base64[uint8_t(src[srclen - 1])] == 64) {
+    srclen--;
+  }
+  size_t equalsigns = 0;
+  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
+    equallocation = srclen - 1;
+    srclen--;
+    equalsigns = 1;
+    // skip trailing spaces
+    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+           to_base64[uint8_t(src[srclen - 1])] == 64) {
+      srclen--;
     }
-    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
-    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
-    if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
-      // 1. pack the bytes
-      const __m128i utf8_packed = _mm_packus_epi16(
-          _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1));
-      // 2. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
-      // 3. adjust pointers
-      buf += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
+    if (srclen > 0 && src[srclen - 1] == '=') {
+      equallocation = srclen - 1;
+      srclen--;
+      equalsigns = 2;
     }
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
+  }
+  if (srclen == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, 0, 0};
+  }
+  char *end_of_safe_64byte_zone =
+      (srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 : dst;
 
-    // no bits set above 11th bit
-    const __m256i one_or_two_bytes_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
-    const uint32_t one_or_two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
-    if (one_or_two_bytes_bitmask == 0xffffffff) {
+  const chartype *const srcinit = src;
+  const char *const dstinit = dst;
+  const chartype *const srcend = src + srclen;
 
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
+  constexpr size_t block_size = 6;
+  static_assert(block_size >= 2, "block_size must be at least two");
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const chartype *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b(src);
+      src += 64;
+      uint64_t error = 0;
+      const uint64_t badcharmask =
+          b.to_base64_mask<base64_url, ignore_garbage>(&error);
+      if (!ignore_garbage && error) {
+        src -= 64;
+        const size_t error_offset = trailing_zeroes(error);
+        return {error_code::INVALID_BASE64_CHARACTER,
+                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+      }
+      if (badcharmask != 0) {
+        bufferptr += b.compress_block(badcharmask, bufferptr);
+      } else if (bufferptr != buffer) {
+        b.copy_block(bufferptr);
+        bufferptr += 64;
+      } else {
+        if (dst >= end_of_safe_64byte_zone) {
+          b.base64_decode_block_safe(dst);
+        } else {
+          b.base64_decode_block(dst);
+        }
+        dst += 48;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 2); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        if (dst >= end_of_safe_64byte_zone) {
+          base64_decode_block_safe(dst, buffer + (block_size - 2) * 64);
+        } else {
+          base64_decode_block(dst, buffer + (block_size - 2) * 64);
+        }
+        dst += 48;
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
+  }
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const __m256i t0 = _mm256_slli_epi16(in, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const __m256i t2 = _mm256_and_si256(in, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const __m256i t3 = _mm256_or_si256(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const __m256i t4 = _mm256_or_si256(t3, v_c080);
+  char *buffer_start = buffer;
+  // Optimization note: if this is almost full, then it is worth our
+  // time, otherwise, we should just decode directly.
+  int last_block = (int)((bufferptr - buffer_start) % 64);
+  if (last_block != 0 && srcend - src + last_block >= 64) {
 
-      // 2. merge ASCII and 2-byte codewords
-      const __m256i utf8_unpacked =
-          _mm256_blendv_epi8(t4, in, one_byte_bytemask);
+    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
+      uint8_t val = to_base64[uint8_t(*src)];
+      *bufferptr = char(val);
+      if (!ignore_garbage &&
+          (!scalar::base64::is_eight_byte(*src) || val > 64)) {
+        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      }
+      bufferptr += (val <= 63);
+      src++;
+    }
+  }
 
-      // 3. prepare bitmask for 8-bit lookup
-      const uint32_t M0 = one_byte_bitmask & 0x55555555;
-      const uint32_t M1 = M0 >> 7;
-      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-      // 4. pack the bytes
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    if (dst >= end_of_safe_64byte_zone) {
+      base64_decode_block_safe(dst, buffer_start);
+    } else {
+      base64_decode_block(dst, buffer_start);
+    }
+    dst += 48;
+  }
+  if ((bufferptr - buffer_start) % 64 != 0) {
+    while (buffer_start + 4 < bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
 
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-      const uint8_t *row_2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
-                                                                       16)][0];
+      dst += 3;
+      buffer_start += 4;
+    }
+    if (buffer_start + 4 <= bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
 
-      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+      dst += 3;
+      buffer_start += 4;
+    }
+    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
+    // backtrack
+    int leftover = int(bufferptr - buffer_start);
+    while (leftover > 0) {
+      if (!ignore_garbage) {
+        while (to_base64[uint8_t(*(src - 1))] == 64) {
+          src--;
+        }
+      } else {
+        while (to_base64[uint8_t(*(src - 1))] >= 64) {
+          src--;
+        }
+      }
+      src--;
+      leftover--;
+    }
+  }
+  if (src < srcend + equalsigns) {
+    full_result r = scalar::base64::base64_tail_decode(
+        dst, src, srcend - src, equalsigns, options, last_chunk_options);
+    r.input_count += size_t(src - srcinit);
+    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
+        r.error == error_code::BASE64_EXTRA_BITS) {
+      return r;
+    } else {
+      r.output_count += size_t(dst - dstinit);
+    }
+    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
+        r.error == error_code::SUCCESS && equalsigns > 0) {
+      // additional checks
+      if ((r.output_count % 3 == 0) ||
+          ((r.output_count % 3) + 1 + equalsigns != 4)) {
+        r.error = error_code::INVALID_BASE64_CHARACTER;
+        r.input_count = equallocation;
+      }
+    }
+    return r;
+  }
+  if (!ignore_garbage && equalsigns > 0) {
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
+}
 
-      const __m256i utf8_packed = _mm256_shuffle_epi8(
-          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-      // 5. store bytes
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_castsi256_si128(utf8_packed));
-      utf8_output += row[0];
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_extractf128_si256(utf8_packed, 1));
-      utf8_output += row_2[0];
+} // namespace base64
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdutf
+/* end file src/generic/base64.h */
+#endif // SIMDUTF_FEATURE_BASE64
 
-      // 6. adjust pointers
-      buf += 16;
-      continue;
-    }
-    // 1. Check if there are any surrogate word in the input chunk.
-    //    We have also deal with situation when there is a surrogate word
-    //    at the end of a chunk.
-    const __m256i surrogates_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+namespace simdutf {
+namespace haswell {
 
-    // bitmask = 0x0000 if there are no surrogates
-    //         = 0xc000 if the last word is a surrogate
-    const uint32_t surrogates_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (surrogates_bitmask == 0x00000000) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-      const __m256i dup_even = _mm256_setr_epi16(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#if SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused int
+implementation::detect_encodings(const char *input,
+                                 size_t length) const noexcept {
+  // If there is a BOM, then we trust it.
+  auto bom_encoding = simdutf::BOM::check_bom(input, length);
+  if (bom_encoding != encoding_type::unspecified) {
+    return bom_encoding;
+  }
 
-      /* In this branch we handle three cases:
-         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+  int out = 0;
+  uint32_t utf16_err = (length % 2);
+  uint32_t utf32_err = (length % 4);
+  uint32_t ends_with_high = 0;
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+  const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
+  const __m256i offset = _mm256_set1_epi32(0xffff2000);
+  const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
+  __m256i currentmax = _mm256_setzero_si256();
+  __m256i currentoffsetmax = _mm256_setzero_si256();
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+  utf8_checker c{};
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    // utf8 checks
+    c.check_next_input(in);
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+    // utf16le checks
+    auto in0 = simd16<uint16_t>(in.chunks[0]);
+    auto in1 = simd16<uint16_t>(in.chunks[1]);
+    const auto t0 = in0.shr<8>();
+    const auto t1 = in1.shr<8>();
+    const auto in2 = simd16<uint16_t>::pack(t0, t1);
+    const auto surrogates_wordmask = (in2 & v_f8) == v_d8;
+    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
+    const auto vL = (in2 & v_fc) == v_dc;
+    const uint32_t L = vL.to_bitmask();
+    const uint32_t H = L ^ surrogates_bitmask;
+    utf16_err |= (((H << 1) | ends_with_high) != L);
+    ends_with_high = (H & 0x80000000) != 0;
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+    // utf32le checks
+    currentmax = _mm256_max_epu32(in.chunks[0], currentmax);
+    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[0], offset),
+                                        currentoffsetmax);
+    currentmax = _mm256_max_epu32(in.chunks[1], currentmax);
+    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[1], offset),
+                                        currentoffsetmax);
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+    reader.advance();
+  }
 
-      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
-      const __m256i s0 = _mm256_srli_epi16(in, 4);
-      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
-      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
-      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
-      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
-      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
-      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
-                                             simdutf_vec(0b0100000000000000));
-      const __m256i s4 = _mm256_xor_si256(s3, m0);
-#undef simdutf_vec
+  uint8_t block[64]{};
+  size_t idx = reader.block_index();
+  std::memcpy(block, &input[idx], length - idx);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
+  // utf16le last block check
+  auto in0 = simd16<uint16_t>(in.chunks[0]);
+  auto in1 = simd16<uint16_t>(in.chunks[1]);
+  const auto t0 = in0.shr<8>();
+  const auto t1 = in1.shr<8>();
+  const auto in2 = simd16<uint16_t>::pack(t0, t1);
+  const auto surrogates_wordmask = (in2 & v_f8) == v_d8;
+  const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
+  const auto vL = (in2 & v_fc) == v_dc;
+  const uint32_t L = vL.to_bitmask();
+  const uint32_t H = L ^ surrogates_bitmask;
+  utf16_err |= (((H << 1) | ends_with_high) != L);
+  // this is required to check for last byte ending in high and end of input
+  // is reached
+  ends_with_high = (H & 0x80000000) != 0;
+  utf16_err |= ends_with_high;
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
+  // utf32le last block check
+  currentmax = _mm256_max_epu32(in.chunks[0], currentmax);
+  currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[0], offset),
+                                      currentoffsetmax);
+  currentmax = _mm256_max_epu32(in.chunks[1], currentmax);
+  currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[1], offset),
+                                      currentoffsetmax);
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
+  reader.advance();
 
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+  c.check_eof();
+  bool is_valid_utf8 = !c.errors();
+  __m256i is_zero =
+      _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
+  utf32_err |= (_mm256_testz_si256(is_zero, is_zero) == 0);
 
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+  is_zero = _mm256_xor_si256(
+      _mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
+  utf32_err |= (_mm256_testz_si256(is_zero, is_zero) == 0);
+  if (is_valid_utf8) {
+    out |= encoding_type::UTF8;
+  }
+  if (utf16_err == 0) {
+    out |= encoding_type::UTF16_LE;
+  }
+  if (utf32_err == 0) {
+    out |= encoding_type::UTF32_LE;
+  }
+  return out;
+}
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-      // surrogate pair(s) in a register
-    } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
-        if ((word & 0xFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xF800) != 0xD800) {
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(nullptr, utf8_output);
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf8_output++ = char((value >> 18) | 0b11110000);
-          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((value & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
-  return std::make_pair(buf, utf8_output);
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return haswell::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-/*
-  Returns a pair: a result struct and utf8_output.
-  If there is an error, the count field of the result is the position of the
-  error. Otherwise, it is the position of the first unprocessed byte in buf
-  (even if finished). A scalar routing should carry on the conversion of the
-  tail if needed.
-*/
-template <endianness big_endian>
-std::pair<result, char *>
-avx2_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
-                                       char *utf8_output) {
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused result implementation::validate_utf8_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return haswell::utf8_validation::generic_validate_utf8_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8
 
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
-  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
-  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+#if SIMDUTF_FEATURE_ASCII
+simdutf_warn_unused bool
+implementation::validate_ascii(const char *buf, size_t len) const noexcept {
+  return haswell::ascii_validation::generic_validate_ascii(buf, len);
+}
 
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    if (big_endian) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
-    }
-    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
-    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
-    if (_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
-      // 1. pack the bytes
-      const __m128i utf8_packed = _mm_packus_epi16(
-          _mm256_castsi256_si128(in), _mm256_extractf128_si256(in, 1));
-      // 2. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
-      // 3. adjust pointers
-      buf += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
-    }
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
+simdutf_warn_unused result implementation::validate_ascii_with_errors(
+    const char *buf, size_t len) const noexcept {
+  return haswell::ascii_validation::generic_validate_ascii_with_errors(buf,
+                                                                       len);
+}
+#endif // SIMDUTF_FEATURE_ASCII
 
-    // no bits set above 11th bit
-    const __m256i one_or_two_bytes_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
-    const uint32_t one_or_two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
-    if (one_or_two_bytes_bitmask == 0xffffffff) {
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf16le(const char16_t *buf,
+                                 size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    // empty input is valid UTF-16. protect the implementation from
+    // handling nullptr
+    return true;
+  }
+  const auto res =
+      haswell::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.is_err()) {
+    return false;
+  }
 
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
+  if (res.count == len) {
+    return true;
+  }
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const __m256i t0 = _mm256_slli_epi16(in, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const __m256i t2 = _mm256_and_si256(in, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const __m256i t3 = _mm256_or_si256(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const __m256i t4 = _mm256_or_si256(t3, v_c080);
+  return scalar::utf16::validate<endianness::LITTLE>(buf + res.count,
+                                                     len - res.count);
+}
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-      // 2. merge ASCII and 2-byte codewords
-      const __m256i utf8_unpacked =
-          _mm256_blendv_epi8(t4, in, one_byte_bytemask);
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool
+implementation::validate_utf16be(const char16_t *buf,
+                                 size_t len) const noexcept {
+  if (simdutf_unlikely(len == 0)) {
+    // empty input is valid UTF-16. protect the implementation from
+    // handling nullptr
+    return true;
+  }
+  const auto res =
+      haswell::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.is_err()) {
+    return false;
+  }
 
-      // 3. prepare bitmask for 8-bit lookup
-      const uint32_t M0 = one_byte_bitmask & 0x55555555;
-      const uint32_t M1 = M0 >> 7;
-      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-      // 4. pack the bytes
+  if (res.count == len) {
+    return true;
+  }
 
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-      const uint8_t *row_2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
-                                                                       16)][0];
+  return scalar::utf16::validate<endianness::BIG>(buf + res.count,
+                                                  len - res.count);
+}
 
-      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+simdutf_warn_unused result implementation::validate_utf16le_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
 
-      const __m256i utf8_packed = _mm256_shuffle_epi8(
-          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-      // 5. store bytes
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_castsi256_si128(utf8_packed));
-      utf8_output += row[0];
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_extractf128_si256(utf8_packed, 1));
-      utf8_output += row_2[0];
+  const result res =
+      haswell::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.count != len) {
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::LITTLE>(
+            buf + res.count, len - res.count);
+    return result(scalar_res.error, res.count + scalar_res.count);
+  } else {
+    return res;
+  }
+}
 
-      // 6. adjust pointers
-      buf += 16;
-      continue;
-    }
-    // 1. Check if there are any surrogate word in the input chunk.
-    //    We have also deal with situation when there is a surrogate word
-    //    at the end of a chunk.
-    const __m256i surrogates_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+simdutf_warn_unused result implementation::validate_utf16be_with_errors(
+    const char16_t *buf, size_t len) const noexcept {
+  const result res =
+      haswell::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.count != len) {
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count,
+                                                             len - res.count);
+    return result(scalar_res.error, res.count + scalar_res.count);
+  } else {
+    return res;
+  }
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-    // bitmask = 0x0000 if there are no surrogates
-    //         = 0xc000 if the last word is a surrogate
-    const uint32_t surrogates_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (surrogates_bitmask == 0x00000000) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-      const __m256i dup_even = _mm256_setr_epi16(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
+  return utf32::validate(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
-      /* In this branch we handle three cases:
-         1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-         2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-         3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+#if SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused result implementation::validate_utf32_with_errors(
+    const char32_t *buf, size_t len) const noexcept {
+  return utf32::validate_with_errors(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
+    const char *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char *, char *> ret =
+      avx2_convert_latin1_to_utf8(buf, len, utf8_output);
+  size_t converted_chars = ret.second - utf8_output;
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
+  }
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char *, char16_t *> ret =
+      avx2_convert_latin1_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t converted_chars = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars =
+        scalar::latin1_to_utf16::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_converted_chars == 0) {
+      return 0;
+    }
+    converted_chars += scalar_converted_chars;
+  }
+  return converted_chars;
+}
 
-      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
-      const __m256i s0 = _mm256_srli_epi16(in, 4);
-      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
-      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
-      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
-      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
-      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
-      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
-                                             simdutf_vec(0b0100000000000000));
-      const __m256i s4 = _mm256_xor_si256(s3, m0);
-#undef simdutf_vec
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char *, char16_t *> ret =
+      avx2_convert_latin1_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t converted_chars = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars =
+        scalar::latin1_to_utf16::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_converted_chars == 0) {
+      return 0;
+    }
+    converted_chars += scalar_converted_chars;
+  }
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char *, char32_t *> ret =
+      avx2_convert_latin1_to_utf32(buf, len, utf32_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t converted_chars = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_converted_chars == 0) {
+      return 0;
+    }
+    converted_chars += scalar_converted_chars;
+  }
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert(buf, len, latin1_output);
+}
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
+simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, latin1_output);
+}
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
+    const char *input, size_t size, char *latin1_output) const noexcept {
+  return utf8_to_latin1::convert_valid(input, size, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
+}
 
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::BIG>(buf, len, utf16_output);
+}
 
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-      // surrogate pair(s) in a register
-    } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
-        if ((word & 0xFF80) == 0) {
-          *utf8_output++ = char(word);
-        } else if ((word & 0xF800) == 0) {
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xF800) != 0xD800) {
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k - 1),
-                utf8_output);
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf8_output++ = char((value >> 18) | 0b11110000);
-          *utf8_output++ = char(((value >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((value >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((value & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
+simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::LITTLE>(buf, len,
+                                                           utf16_output);
 }
-/* end file src/haswell/avx2_convert_utf16_to_utf8.cpp */
-/* begin file src/haswell/avx2_convert_utf16_to_utf32.cpp */
-/*
-    The vectorized algorithm works on single SSE register i.e., it
-    loads eight 16-bit code units.
 
-    We consider three cases:
-    1. an input register contains no surrogates and each value
-       is in range 0x0000 .. 0x07ff.
-    2. an input register contains no surrogates and values are
-       in range 0x0000 .. 0xffff.
-    3. an input register contains surrogates --- i.e. codepoints
-       can have 16 or 32 bits.
+simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
+}
 
-    Ad 1.
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
+    const char *input, size_t size, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,
+                                                          utf16_output);
+}
 
-    When values are less than 0x0800, it means that a 16-bit code unit
-    can be converted into: 1) single UTF8 byte (when it is an ASCII
-    char) or 2) two UTF8 bytes.
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
+    const char *input, size_t size, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
+                                                       utf16_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    For this case we do only some shuffle to obtain these 2-byte
-    codes and finally compress the whole SSE register with a single
-    shuffle.
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert(buf, len, utf32_output);
+}
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, utf32_output);
+}
 
-    Ad 2.
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
+    const char *input, size_t size, char32_t *utf32_output) const noexcept {
+  return utf8_to_utf32::convert_valid(input, size, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-    When values fit in 16-bit code units, but are above 0x07ff, then
-    a single word may produce one, two or three UTF8 bytes.
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      haswell::avx2_convert_utf16_to_latin1<endianness::LITTLE>(buf, len,
+                                                                latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - latin1_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_latin1::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    We prepare data for all these three cases in two registers.
-    The first register contains lower two UTF8 bytes (used in all
-    cases), while the second one contains just the third byte for
-    the three-UTF8-bytes case.
+simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      haswell::avx2_convert_utf16_to_latin1<endianness::BIG>(buf, len,
+                                                             latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - latin1_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_latin1::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    Finally these two registers are interleaved forming eight-element
-    array of 32-bit values. The array spans two SSE registers.
-    The bytes from the registers are compressed using two shuffles.
+simdutf_warn_unused result
+implementation::convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<result, char *> ret =
+      avx2_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
+          buf, len, latin1_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-    We need 256-entry lookup table to get a compression pattern
-    and the number of output bytes in the compressed vector register.
-    Each entry occupies 17 bytes.
+simdutf_warn_unused result
+implementation::convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<result, char *> ret =
+      avx2_convert_utf16_to_latin1_with_errors<endianness::BIG>(buf, len,
+                                                                latin1_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement a custom function
+  return convert_utf16be_to_latin1(buf, len, latin1_output);
+}
 
-    To summarize:
-    - We need two 256-entry tables that have 8704 bytes in total.
-*/
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: implement a custom function
+  return convert_utf16le_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-/*
-  Returns a pair: the first unprocessed byte from buf and utf32_output
-  A scalar routing should carry on the conversion of the tail.
-*/
-template <endianness big_endian>
-std::pair<const char16_t *, char32_t *>
-avx2_convert_utf16_to_utf32(const char16_t *buf, size_t len,
-                            char32_t *utf32_output) {
-  const char16_t *end = buf + len;
-  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
-  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      haswell::avx2_convert_utf16_to_utf8<endianness::LITTLE>(buf, len,
+                                                              utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf8::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-  while (end - buf >= 16) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    if (big_endian) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char16_t *, char *> ret =
+      haswell::avx2_convert_utf16_to_utf8<endianness::BIG>(buf, len,
+                                                           utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf8::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // 1. Check if there are any surrogate word in the input chunk.
-    //    We have also deal with situation when there is a surrogate word
-    //    at the end of a chunk.
-    const __m256i surrogates_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(
+          buf, len, utf8_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-    // bitmask = 0x0000 if there are no surrogates
-    //         = 0xc000 if the last word is a surrogate
-    const uint32_t surrogates_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (surrogates_bitmask == 0x00000000) {
-      // case: we extend all sixteen 16-bit code units to sixteen 32-bit code
-      // units
-      _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
-                          _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
-      _mm256_storeu_si256(
-          reinterpret_cast<__m256i *>(utf32_output + 8),
-          _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in, 1)));
-      utf32_output += 16;
-      buf += 16;
-      // surrogate pair(s) in a register
+simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::BIG>(
+          buf, len, utf8_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
     } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
-        if ((word & 0xF800) != 0xD800) {
-          // No surrogate pair
-          *utf32_output++ = char32_t(word);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(nullptr, utf32_output);
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf32_output++ = char32_t(value);
-        }
-      }
-      buf += k;
+      ret.second += scalar_res.count;
     }
-  } // while
-  return std::make_pair(buf, utf32_output);
+  }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
 }
 
-/*
-  Returns a pair: a result struct and utf8_output.
-  If there is an error, the count field of the result is the position of the
-  error. Otherwise, it is the position of the first unprocessed byte in buf
-  (even if finished). A scalar routing should carry on the conversion of the
-  tail if needed.
-*/
-template <endianness big_endian>
-std::pair<result, char32_t *>
-avx2_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
-                                        char32_t *utf32_output) {
-  const char16_t *start = buf;
-  const char16_t *end = buf + len;
-  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
-  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16le_to_utf8(buf, len, utf8_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
+    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf16be_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  std::pair<const char32_t *, char *> ret =
+      avx2_convert_utf32_to_utf8(buf, len, utf8_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf8_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  std::pair<const char32_t *, char *> ret =
+      avx2_convert_utf32_to_latin1(buf, len, latin1_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - latin1_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-  while (end - buf >= 16) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    if (big_endian) {
-      const __m256i swap = _mm256_setr_epi8(
-          1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14, 17, 16, 19, 18,
-          21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
-      in = _mm256_shuffle_epi8(in, swap);
+simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      avx2_convert_utf32_to_latin1_with_errors(buf, len, latin1_output);
+  if (ret.first.count != len) {
+    result scalar_res = scalar::utf32_to_latin1::convert_with_errors(
+        buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
     }
+  }
+  ret.first.count =
+      ret.second -
+      latin1_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-    // 1. Check if there are any surrogate word in the input chunk.
-    //    We have also deal with situation when there is a surrogate word
-    //    at the end of a chunk.
-    const __m256i surrogates_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return convert_utf32_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-    // bitmask = 0x0000 if there are no surrogates
-    //         = 0xc000 if the last word is a surrogate
-    const uint32_t surrogates_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
-    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
-    // However, it is likely an uncommon occurrence.
-    if (surrogates_bitmask == 0x00000000) {
-      // case: we extend all sixteen 16-bit code units to sixteen 32-bit code
-      // units
-      _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output),
-                          _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
-      _mm256_storeu_si256(
-          reinterpret_cast<__m256i *>(utf32_output + 8),
-          _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in, 1)));
-      utf32_output += 16;
-      buf += 16;
-      // surrogate pair(s) in a register
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char *> ret =
+      haswell::avx2_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
+  if (ret.first.count != len) {
+    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
+        buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
     } else {
-      // Let us do a scalar fallback.
-      // It may seem wasteful to use scalar code, but being efficient with SIMD
-      // in the presence of surrogate pairs may require non-trivial tables.
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
-        if ((word & 0xF800) != 0xD800) {
-          // No surrogate pair
-          *utf32_output++ = char32_t(word);
-        } else {
-          // must be a surrogate pair
-          uint16_t diff = uint16_t(word - 0xD800);
-          uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
-          k++;
-          uint16_t diff2 = uint16_t(next_word - 0xDC00);
-          if ((diff | diff2) > 0x3FF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k - 1),
-                utf32_output);
-          }
-          uint32_t value = (diff << 10) + diff2 + 0x10000;
-          *utf32_output++ = char32_t(value);
-        }
-      }
-      buf += k;
+      ret.second += scalar_res.count;
     }
-  } // while
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf32_output);
+  }
+  ret.first.count =
+      ret.second -
+      utf8_output; // Set count to the number of 8-bit code units written
+  return ret.first;
 }
-/* end file src/haswell/avx2_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-/* begin file src/haswell/avx2_convert_utf32_to_latin1.cpp */
-std::pair<const char32_t *, char *>
-avx2_convert_utf32_to_latin1(const char32_t *buf, size_t len,
-                             char *latin1_output) {
-  const size_t rounded_len =
-      len & ~0x1F; // Round down to nearest multiple of 32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char16_t *, char32_t *> ret =
+      haswell::avx2_convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
+                                                               utf32_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-  __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00);
+simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  std::pair<const char16_t *, char32_t *> ret =
+      haswell::avx2_convert_utf16_to_utf32<endianness::BIG>(buf, len,
+                                                            utf32_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf32_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf16_to_utf32::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-  __m256i shufmask = _mm256_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-                                     -1, 12, 8, 4, 0, -1, -1, -1, -1, -1, -1,
-                                     -1, -1, -1, -1, -1, -1, 12, 8, 4, 0);
+simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char32_t *> ret =
+      haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(
+          buf, len, utf32_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      utf32_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  for (size_t i = 0; i < rounded_len; i += 16) {
-    __m256i in1 = _mm256_loadu_si256((__m256i *)buf);
-    __m256i in2 = _mm256_loadu_si256((__m256i *)(buf + 8));
+simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char32_t *> ret =
+      haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::BIG>(
+          buf, len, utf32_output);
+  if (ret.first.error) {
+    return ret.first;
+  } // Can return directly since scalar fallback already found correct
+    // ret.first.count
+  if (ret.first.count != len) { // All good so far, but not finished
+    result scalar_res =
+        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      utf32_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    __m256i check_combined = _mm256_or_si256(in1, in2);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
+    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
+  return convert_utf32_to_utf8(buf, len, utf8_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-    if (!_mm256_testz_si256(check_combined, high_bytes_mask)) {
-      return std::make_pair(nullptr, latin1_output);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      avx2_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf32_to_utf16::convert<endianness::LITTLE>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
     }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    // Turn UTF32 bytes into latin 1 bytes
-    __m256i shuffled1 = _mm256_shuffle_epi8(in1, shufmask);
-    __m256i shuffled2 = _mm256_shuffle_epi8(in2, shufmask);
-
-    // move Latin1 bytes to their correct spot
-    __m256i idx1 = _mm256_set_epi32(-1, -1, -1, -1, -1, -1, 4, 0);
-    __m256i idx2 = _mm256_set_epi32(-1, -1, -1, -1, 4, 0, -1, -1);
-    __m256i reshuffled1 = _mm256_permutevar8x32_epi32(shuffled1, idx1);
-    __m256i reshuffled2 = _mm256_permutevar8x32_epi32(shuffled2, idx2);
+simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  std::pair<const char32_t *, char16_t *> ret =
+      avx2_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  if (ret.first == nullptr) {
+    return 0;
+  }
+  size_t saved_bytes = ret.second - utf16_output;
+  if (ret.first != buf + len) {
+    const size_t scalar_saved_bytes =
+        scalar::utf32_to_utf16::convert<endianness::BIG>(
+            ret.first, len - (ret.first - buf), ret.second);
+    if (scalar_saved_bytes == 0) {
+      return 0;
+    }
+    saved_bytes += scalar_saved_bytes;
+  }
+  return saved_bytes;
+}
 
-    __m256i result = _mm256_or_si256(reshuffled1, reshuffled2);
-    _mm_storeu_si128((__m128i *)latin1_output, _mm256_castsi256_si128(result));
+simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(
+          buf, len, utf16_output);
+  if (ret.first.count != len) {
+    result scalar_res =
+        scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
+  }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-    latin1_output += 16;
-    buf += 16;
+simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  // ret.first.count is always the position in the buffer, not the number of
+  // code units written even if finished
+  std::pair<result, char16_t *> ret =
+      haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::BIG>(
+          buf, len, utf16_output);
+  if (ret.first.count != len) {
+    result scalar_res =
+        scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
+            buf + ret.first.count, len - ret.first.count, ret.second);
+    if (scalar_res.error) {
+      scalar_res.count += ret.first.count;
+      return scalar_res;
+    } else {
+      ret.second += scalar_res.count;
+    }
   }
+  ret.first.count =
+      ret.second -
+      utf16_output; // Set count to the number of 8-bit code units written
+  return ret.first;
+}
 
-  return std::make_pair(buf, latin1_output);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16le(buf, len, utf16_output);
 }
-std::pair<result, char *>
-avx2_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
-                                         char *latin1_output) {
-  const size_t rounded_len =
-      len & ~0x1F; // Round down to nearest multiple of 32
 
-  __m256i high_bytes_mask = _mm256_set1_epi32(0xFFFFFF00);
-  __m256i shufmask = _mm256_set_epi8(-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-                                     -1, 12, 8, 4, 0, -1, -1, -1, -1, -1, -1,
-                                     -1, -1, -1, -1, -1, -1, 12, 8, 4, 0);
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
+    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return convert_utf32_to_utf16be(buf, len, utf16_output);
+}
 
-  const char32_t *start = buf;
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return convert_utf16le_to_utf32(buf, len, utf32_output);
+}
 
-  for (size_t i = 0; i < rounded_len; i += 16) {
-    __m256i in1 = _mm256_loadu_si256((__m256i *)buf);
-    __m256i in2 = _mm256_loadu_si256((__m256i *)(buf + 8));
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
+    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
+  return convert_utf16be_to_utf32(buf, len, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    __m256i check_combined = _mm256_or_si256(in1, in2);
+#if SIMDUTF_FEATURE_UTF16
+void implementation::change_endianness_utf16(const char16_t *input,
+                                             size_t length,
+                                             char16_t *output) const noexcept {
+  utf16::change_endianness_utf16(input, length, output);
+}
 
-    if (!_mm256_testz_si256(check_combined, high_bytes_mask)) {
-      // Fallback to scalar code for handling errors
-      for (int k = 0; k < 8; k++) {
-        char32_t codepoint = buf[k];
-        if (codepoint <= 0xFF) {
-          *latin1_output++ = static_cast<char>(codepoint);
-        } else {
-          return std::make_pair(result(error_code::TOO_LARGE, buf - start + k),
-                                latin1_output);
-        }
-      }
-      buf += 8;
-    } else {
-      __m256i shuffled1 = _mm256_shuffle_epi8(in1, shufmask);
-      __m256i shuffled2 = _mm256_shuffle_epi8(in2, shufmask);
+simdutf_warn_unused size_t implementation::count_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::count_code_points<endianness::LITTLE>(input, length);
+}
 
-      __m256i idx1 = _mm256_set_epi32(-1, -1, -1, -1, -1, -1, 4, 0);
-      __m256i idx2 = _mm256_set_epi32(-1, -1, -1, -1, 4, 0, -1, -1);
-      __m256i reshuffled1 = _mm256_permutevar8x32_epi32(shuffled1, idx1);
-      __m256i reshuffled2 = _mm256_permutevar8x32_epi32(shuffled2, idx2);
+simdutf_warn_unused size_t implementation::count_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::count_code_points<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
-      __m256i result = _mm256_or_si256(reshuffled1, reshuffled2);
-      _mm_storeu_si128((__m128i *)latin1_output,
-                       _mm256_castsi256_si128(result));
+#if SIMDUTF_FEATURE_UTF8
+simdutf_warn_unused size_t
+implementation::count_utf8(const char *in, size_t size) const noexcept {
+  return utf8::count_code_points_bytemask(in, size);
+}
+#endif // SIMDUTF_FEATURE_UTF8
 
-      latin1_output += 16;
-      buf += 16;
-    }
-  }
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
+    const char *buf, size_t len) const noexcept {
+  return count_utf8(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start),
-                        latin1_output);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf8_length_from_utf16_bytemask<endianness::LITTLE>(input,
+                                                                    length);
 }
-/* end file src/haswell/avx2_convert_utf32_to_latin1.cpp */
-/* begin file src/haswell/avx2_convert_utf32_to_utf8.cpp */
-std::pair<const char32_t *, char *>
-avx2_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) {
-  const char32_t *end = buf + len;
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
-  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
-  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
-  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
-  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
-  __m256i running_max = _mm256_setzero_si256();
-  __m256i forbidden_bytemask = _mm256_setzero_si256();
 
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
+simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf8_length_from_utf16_bytemask<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
-    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
+}
 
-    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
-    // saturation
-    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
-                                        _mm256_and_si256(nextin, v_7fffffff));
-    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
+simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
+    const char16_t *input, size_t length) const noexcept {
+  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-    // Try to apply UTF-16 => UTF-8 routine on 256 bits
-    // (haswell/avx2_convert_utf16_to_utf8.cpp)
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return utf8::utf16_length_from_utf8(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
-      // 1. pack the bytes
-      const __m128i utf8_packed = _mm_packus_epi16(
-          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
-      // 2. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
-      // 3. adjust pointers
-      buf += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
+    const char *input, size_t len) const noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
+  size_t answer = len / sizeof(__m256i) * sizeof(__m256i);
+  size_t i = 0;
+  if (answer >= 2048) { // long strings optimization
+    __m256i four_64bits = _mm256_setzero_si256();
+    while (i + sizeof(__m256i) <= len) {
+      __m256i runner = _mm256_setzero_si256();
+      // We can do up to 255 loops without overflow.
+      size_t iterations = (len - i) / sizeof(__m256i);
+      if (iterations > 255) {
+        iterations = 255;
+      }
+      size_t max_i = i + iterations * sizeof(__m256i) - sizeof(__m256i);
+      for (; i + 4 * sizeof(__m256i) <= max_i; i += 4 * sizeof(__m256i)) {
+        __m256i input1 = _mm256_loadu_si256((const __m256i *)(data + i));
+        __m256i input2 =
+            _mm256_loadu_si256((const __m256i *)(data + i + sizeof(__m256i)));
+        __m256i input3 = _mm256_loadu_si256(
+            (const __m256i *)(data + i + 2 * sizeof(__m256i)));
+        __m256i input4 = _mm256_loadu_si256(
+            (const __m256i *)(data + i + 3 * sizeof(__m256i)));
+        __m256i input12 =
+            _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input1),
+                            _mm256_cmpgt_epi8(_mm256_setzero_si256(), input2));
+        __m256i input23 =
+            _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input3),
+                            _mm256_cmpgt_epi8(_mm256_setzero_si256(), input4));
+        __m256i input1234 = _mm256_add_epi8(input12, input23);
+        runner = _mm256_sub_epi8(runner, input1234);
+      }
+      for (; i <= max_i; i += sizeof(__m256i)) {
+        __m256i input_256_chunk =
+            _mm256_loadu_si256((const __m256i *)(data + i));
+        runner = _mm256_sub_epi8(
+            runner, _mm256_cmpgt_epi8(_mm256_setzero_si256(), input_256_chunk));
+      }
+      four_64bits = _mm256_add_epi64(
+          four_64bits, _mm256_sad_epu8(runner, _mm256_setzero_si256()));
     }
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
-
-    // no bits set above 11th bit
-    const __m256i one_or_two_bytes_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
-    const uint32_t one_or_two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
-    if (one_or_two_bytes_bitmask == 0xffffffff) {
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
+    answer += _mm256_extract_epi64(four_64bits, 0) +
+              _mm256_extract_epi64(four_64bits, 1) +
+              _mm256_extract_epi64(four_64bits, 2) +
+              _mm256_extract_epi64(four_64bits, 3);
+  } else if (answer > 0) {
+    for (; i + sizeof(__m256i) <= len; i += sizeof(__m256i)) {
+      __m256i latin = _mm256_loadu_si256((const __m256i *)(data + i));
+      uint32_t non_ascii = _mm256_movemask_epi8(latin);
+      answer += count_ones(non_ascii);
+    }
+  }
+  return answer + scalar::latin1::utf8_length_from_latin1(
+                      reinterpret_cast<const char *>(data + i), len - i);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const __m256i t3 = _mm256_or_si256(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const __m256i t4 = _mm256_or_si256(t3, v_c080);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  return utf32::utf8_length_from_utf32(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-      // 2. merge ASCII and 2-byte codewords
-      const __m256i utf8_unpacked =
-          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
+    const char32_t *input, size_t length) const noexcept {
+  const __m256i v_00000000 = _mm256_setzero_si256();
+  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos + 8 <= length; pos += 8) {
+    __m256i in = _mm256_loadu_si256((__m256i *)(input + pos));
+    const __m256i surrogate_bytemask =
+        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
+    const uint32_t surrogate_bitmask =
+        static_cast<uint32_t>(_mm256_movemask_epi8(surrogate_bytemask));
+    size_t surrogate_count = (32 - count_ones(surrogate_bitmask)) / 4;
+    count += 8 + surrogate_count;
+  }
+  return count +
+         scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-      // 3. prepare bitmask for 8-bit lookup
-      const uint32_t M0 = one_byte_bitmask & 0x55555555;
-      const uint32_t M1 = M0 >> 7;
-      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-      // 4. pack the bytes
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
+    const char *input, size_t length) const noexcept {
+  return utf8::count_code_points(input, length);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-      const uint8_t *row_2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
-                                                                       16)][0];
+#if SIMDUTF_FEATURE_BASE64
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  }
+}
 
-      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  }
+}
 
-      const __m256i utf8_packed = _mm256_shuffle_epi8(
-          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-      // 5. store bytes
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_castsi256_si128(utf8_packed));
-      utf8_output += row[0];
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_extractf128_si256(utf8_packed, 1));
-      utf8_output += row_2[0];
+simdutf_warn_unused result implementation::base64_to_binary(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  }
+}
 
-      // 6. adjust pointers
-      buf += 16;
-      continue;
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
-    // Must check for overflow in packing
-    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
-        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
-    if (saturation_bitmask == 0xffffffff) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
-      forbidden_bytemask = _mm256_or_si256(
-          forbidden_bytemask,
-          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
+    }
+  }
+}
 
-      const __m256i dup_even = _mm256_setr_epi16(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
+size_t implementation::binary_to_base64(const char *input, size_t length,
+                                        char *output,
+                                        base64_options options) const noexcept {
+  if (options & base64_url) {
+    return encode_base64<true>(output, input, length, options);
+  } else {
+    return encode_base64<false>(output, input, length, options);
+  }
+}
+#endif // SIMDUTF_FEATURE_BASE64
+} // namespace haswell
+} // namespace simdutf
 
-      /* In this branch we handle three cases:
-        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
+/* begin file src/simdutf/haswell/end.h */
+#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
+// nothing needed.
+#else
+SIMDUTF_UNTARGET_REGION
+#endif
 
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
+#undef SIMDUTF_SIMD_HAS_BYTEMASK
 
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+#if SIMDUTF_GCC11ORMORE // workaround for
+                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
+SIMDUTF_POP_DISABLE_WARNINGS
+#endif // end of workaround
+/* end file src/simdutf/haswell/end.h */
+/* end file src/haswell/implementation.cpp */
+#endif
+#if SIMDUTF_IMPLEMENTATION_PPC64
+/* begin file src/ppc64/implementation.cpp */
+/* begin file src/simdutf/ppc64/begin.h */
+// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
+// #define SIMDUTF_IMPLEMENTATION ppc64
+/* end file src/simdutf/ppc64/begin.h */
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+/* begin file src/ppc64/ppc64_utf16_to_utf8_tables.h */
+// Code generated automatically; DO NOT EDIT
+// file generated by scripts/ppc64_convert_utf16_to_utf8.py
+#ifndef PPC64_SIMDUTF_UTF16_TO_UTF8_TABLES_H
+#define PPC64_SIMDUTF_UTF16_TO_UTF8_TABLES_H
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+namespace simdutf {
+namespace {
+namespace tables {
+namespace ppc64_utf16_to_utf8 {
 
-      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
-      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
-      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
-      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
-      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
-      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
-      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
-      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
-                                             simdutf_vec(0b0100000000000000));
-      const __m256i s4 = _mm256_xor_si256(s3, m0);
-#undef simdutf_vec
+#if SIMDUTF_IS_BIG_ENDIAN
+// 1 byte for length, 16 bytes for mask
+const uint8_t pack_1_2_3_utf8_bytes[256][17] = {
+    {12, 1, 0, 16, 3, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80},
+    {9, 3, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 0, 16, 3, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 17, 3, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 1, 0, 16, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {11, 1, 0, 16, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 0, 16, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 17, 2, 18, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 1, 0, 16, 19, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 19, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 19, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 19, 5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 0, 16, 3, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 3, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 3, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 3, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 1, 0, 16, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 1, 0, 16, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 2, 18, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 0, 16, 19, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 19, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 19, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 19, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {11, 1, 0, 16, 3, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 3, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 0, 16, 3, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 17, 3, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 1, 0, 16, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 2, 18, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 0, 16, 19, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 19, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 19, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 19, 4, 20, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 1, 0, 16, 3, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 3, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 3, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 3, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 1, 0, 16, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 1, 0, 16, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 2, 18, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 19, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 19, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 19, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 19, 21, 7, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 1, 0, 16, 3, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 3, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 3, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 3, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 1, 0, 16, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 1, 0, 16, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 2, 18, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 0, 16, 19, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 19, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 19, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 19, 5, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 1, 0, 16, 3, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 3, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 3, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 3, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 1, 0, 16, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {2, 0, 16, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 17, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {5, 1, 0, 16, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 0, 16, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 2, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 1, 0, 16, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 0, 16, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 17, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {8, 1, 0, 16, 3, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 3, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 3, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 1, 0, 16, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 0, 16, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 1, 0, 16, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 2, 18, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 1, 0, 16, 19, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 19, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 19, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 19, 4, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 1, 0, 16, 3, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 3, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 3, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 3, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 1, 0, 16, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 0, 16, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 17, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {6, 1, 0, 16, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 2, 18, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 0, 16, 19, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 19, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 0, 16, 19, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 19, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {11, 1, 0, 16, 3, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 3, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 0, 16, 3, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 17, 3, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 1, 0, 16, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 2, 18, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 0, 16, 19, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 19, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 19, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 19, 5, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 3, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 3, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 3, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 1, 0, 16, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 0, 16, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 1, 0, 16, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 2, 18, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 1, 0, 16, 19, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 19, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 19, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 19, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {10, 1, 0, 16, 3, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 3, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 3, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 3, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 1, 0, 16, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 1, 0, 16, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 2, 18, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 19, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 19, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 19, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 19, 4, 20, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 1, 0, 16, 3, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 3, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 3, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 3, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 1, 0, 16, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 1, 0, 16, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 2, 18, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 0, 16, 19, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 19, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 19, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 19, 21, 6, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 1, 0, 16, 3, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 3, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 16, 3, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 17, 3, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 1, 0, 16, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 1, 0, 16, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 2, 18, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 1, 0, 16, 19, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 19, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 19, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 19, 5, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 0, 16, 3, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 3, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 3, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 3, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 1, 0, 16, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 0, 16, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 17, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {6, 1, 0, 16, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 2, 18, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 0, 16, 19, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 19, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 0, 16, 19, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 19, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {9, 1, 0, 16, 3, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 3, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 16, 3, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 17, 3, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 1, 0, 16, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 1, 0, 16, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 2, 18, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 0, 16, 19, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 19, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 19, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 19, 4, 20, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 0, 16, 3, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 16, 3, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 17, 3, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 1, 0, 16, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 0, 16, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 17, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 1, 0, 16, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 0, 16, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 17, 2, 18, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 1, 0, 16, 19, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 19, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 16, 19, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 17, 19, 21, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+};
+#else
+// 1 byte for length, 16 bytes for mask
+const uint8_t pack_1_2_3_utf8_bytes[256][17] = {
+    {12, 0, 1, 17, 2, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80},
+    {9, 2, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {11, 1, 17, 2, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {10, 16, 2, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 0, 1, 17, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {11, 0, 1, 17, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 1, 17, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 16, 3, 19, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 0, 1, 17, 18, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 18, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 18, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 18, 4, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 1, 17, 2, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 2, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 2, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 2, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 1, 17, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 0, 1, 17, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 3, 19, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 1, 17, 18, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 18, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 18, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 18, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {11, 0, 1, 17, 2, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 2, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 1, 17, 2, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 16, 2, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 0, 1, 17, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 3, 19, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 1, 17, 18, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 18, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 18, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 18, 5, 21, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 0, 1, 17, 2, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 2, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 2, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 2, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 1, 17, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 0, 1, 17, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 3, 19, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 18, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 18, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 18, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 18, 20, 6, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 0, 1, 17, 2, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {6, 2, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 2, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 2, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 1, 17, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 0, 1, 17, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 3, 19, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 1, 17, 18, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 18, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 18, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 18, 4, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 0, 1, 17, 2, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 2, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 2, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 2, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 0, 1, 17, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {0, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {2, 1, 17, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 16, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {5, 0, 1, 17, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 1, 17, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 3, 19, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 0, 1, 17, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 1, 17, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 16, 18, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {8, 0, 1, 17, 2, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 2, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 2, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 1, 17, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 1, 17, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 0, 1, 17, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 3, 19, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 0, 1, 17, 18, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 18, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 18, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 18, 5, 21, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 0, 1, 17, 2, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 2, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 2, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 2, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 1, 17, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 1, 17, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 16, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {6, 0, 1, 17, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 3, 19, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 1, 17, 18, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 18, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 1, 17, 18, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 18, 20, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {11, 0, 1, 17, 2, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {8, 2, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {10, 1, 17, 2, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {9, 16, 2, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 0, 1, 17, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 3, 19, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 0, 1, 17, 18, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 18, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 18, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 18, 4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 2, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 2, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 2, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 1, 17, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {4, 1, 17, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 0, 1, 17, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 3, 19, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 0, 1, 17, 18, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 18, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 18, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 18, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {10, 0, 1, 17, 2, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 2, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 2, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 2, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 1, 17, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 0, 1, 17, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 3, 19, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 18, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 18, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 18, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 18, 5, 21, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 0, 1, 17, 2, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 2, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 2, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 1, 17, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 0, 1, 17, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 3, 19, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 1, 17, 18, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 18, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 18, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 18, 20, 7, 23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {10, 0, 1, 17, 2, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80},
+    {7, 2, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {9, 1, 17, 2, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 16, 2, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 0, 1, 17, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {9, 0, 1, 17, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 3, 19, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {8, 0, 1, 17, 18, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 18, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 18, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 18, 4, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 1, 17, 2, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 2, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 2, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 2, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 0, 1, 17, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {1, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {3, 1, 17, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {2, 16, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {6, 0, 1, 17, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 3, 19, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 0, 1, 17, 18, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 18, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 1, 17, 18, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 18, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {9, 0, 1, 17, 2, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 2, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 1, 17, 2, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {7, 16, 2, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 0, 1, 17, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {8, 0, 1, 17, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 3, 19, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 0, 1, 17, 18, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 18, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 18, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 18, 5, 21, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {8, 0, 1, 17, 2, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {5, 2, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {7, 1, 17, 2, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {6, 16, 2, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 0, 1, 17, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {2, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80, 0x80},
+    {4, 1, 17, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {3, 16, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {7, 0, 1, 17, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80},
+    {4, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {6, 1, 17, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {5, 16, 3, 19, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {6, 0, 1, 17, 18, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {3, 18, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+    {5, 1, 17, 18, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80},
+    {4, 16, 18, 20, 22, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
+     0x80, 0x80, 0x80},
+};
+#endif // SIMDUTF_IS_BIG_ENDIAN
+} // namespace ppc64_utf16_to_utf8
+} // namespace tables
+} // unnamed namespace
+} // namespace simdutf
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
+#endif // PPC64_SIMDUTF_UTF16_TO_UTF8_TABLES_H
+/* end file src/ppc64/ppc64_utf16_to_utf8_tables.h */
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
+namespace simdutf {
+namespace ppc64 {
+namespace {
+#ifndef SIMDUTF_PPC64_H
+  #error "ppc64.h must be included"
+#endif
+using namespace simd;
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
+simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
+  // careful: 0x80 is not ascii.
+  return input.reduce_or().saturating_sub(0b01111111u).bits_not_set_anywhere();
+}
 
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+simdutf_really_inline simd8<bool>
+must_be_2_3_continuation(const simd8<uint8_t> prev2,
+                         const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte =
+      prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be >= 0x80
+  simd8<uint8_t> is_fourth_byte =
+      prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be >= 0x80
+  // Caller requires a bool (all 1's). All values resulting from the subtraction
+  // will be <= 64, so signed comparison is fine.
+  return simd8<bool>(is_third_byte | is_fourth_byte);
+}
 
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+/// ErrorReporting describes behaviour of a vectorized procedure regarding error
+/// checking
+enum class ErrorReporting {
+  precise,    // the procedure will report *approximate* or *precise* error
+              // position
+  at_the_end, // the procedure will only inform about an error after scanning
+              // the whole input (or its significant portion)
+  none,       // no error checking is done, we assume valid inputs
+};
 
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-    } else {
-      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
-      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
-      // wasteful to use scalar code, but being efficient with SIMD may require
-      // large, non-trivial tables?
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr, utf8_output);
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else { // 4-byte
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr, utf8_output);
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/ppc64/ppc64_validate_utf16.cpp */
+template <endianness big_endian>
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> in0,
+                                       const simd16<uint16_t> in1) {
+  if (big_endian) {
+    const vec_u8_t pack_high = {
+        0,  2,  4,  6,  8,  10, 12, 14, // in0
+        16, 18, 20, 22, 24, 26, 28, 30  // in1
+    };
 
-  // check for invalid input
-  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(
-          _mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
-    return std::make_pair(nullptr, utf8_output);
-  }
+    return vec_perm(vec_u8_t(in0.value), vec_u8_t(in1.value), pack_high);
+  } else {
+    const vec_u8_t pack_high = {
+        1,  3,  5,  7,  9,  11, 13, 15, // in0
+        17, 19, 21, 23, 25, 27, 29, 31  // in1
+    };
 
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
-    return std::make_pair(nullptr, utf8_output);
+    return vec_perm(vec_u8_t(in0.value), vec_u8_t(in1.value), pack_high);
   }
-
-  return std::make_pair(buf, utf8_output);
 }
+/* end file src/ppc64/ppc64_validate_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16
 
-std::pair<result, char *>
-avx2_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
-                                       char *utf8_output) {
-  const char32_t *end = buf + len;
-  const char32_t *start = buf;
-
-  const __m256i v_0000 = _mm256_setzero_si256();
-  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
-  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
-  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
-  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
-  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
-  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
-
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-
-  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
-    __m256i nextin = _mm256_loadu_si256((__m256i *)buf + 1);
-    // Check for too large input
-    const __m256i max_input =
-        _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
-    if (static_cast<uint32_t>(_mm256_movemask_epi8(
-            _mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
-      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
-                            utf8_output);
-    }
-
-    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
-    // saturation
-    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff),
-                                        _mm256_and_si256(nextin, v_7fffffff));
-    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);
-
-    // Try to apply UTF-16 => UTF-8 routine on 256 bits
-    // (haswell/avx2_convert_utf16_to_utf8.cpp)
-
-    if (_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
-      // 1. pack the bytes
-      const __m128i utf8_packed = _mm_packus_epi16(
-          _mm256_castsi256_si128(in_16), _mm256_extractf128_si256(in_16, 1));
-      // 2. store (16 bytes)
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_packed);
-      // 3. adjust pointers
-      buf += 16;
-      utf8_output += 16;
-      continue; // we are done for this round!
-    }
-    // no bits set above 7th bit
-    const __m256i one_byte_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
-    const uint32_t one_byte_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));
-
-    // no bits set above 11th bit
-    const __m256i one_or_two_bytes_bytemask =
-        _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
-    const uint32_t one_or_two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
-    if (one_or_two_bytes_bitmask == 0xffffffff) {
-      // 1. prepare 2-byte values
-      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
-      // expected output   : [110a|aaaa|10bb|bbbb] x 8
-      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
-      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);
-
-      // t0 = [000a|aaaa|bbbb|bb00]
-      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
-      // t1 = [000a|aaaa|0000|0000]
-      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
-      // t2 = [0000|0000|00bb|bbbb]
-      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
-      // t3 = [000a|aaaa|00bb|bbbb]
-      const __m256i t3 = _mm256_or_si256(t1, t2);
-      // t4 = [110a|aaaa|10bb|bbbb]
-      const __m256i t4 = _mm256_or_si256(t3, v_c080);
-
-      // 2. merge ASCII and 2-byte codewords
-      const __m256i utf8_unpacked =
-          _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);
-
-      // 3. prepare bitmask for 8-bit lookup
-      const uint32_t M0 = one_byte_bitmask & 0x55555555;
-      const uint32_t M1 = M0 >> 7;
-      const uint32_t M2 = (M1 | M0) & 0x00ff00ff;
-      // 4. pack the bytes
-
-      const uint8_t *row =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
-      const uint8_t *row_2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2 >>
-                                                                       16)][0];
-
-      const __m128i shuffle = _mm_loadu_si128((__m128i *)(row + 1));
-      const __m128i shuffle_2 = _mm_loadu_si128((__m128i *)(row_2 + 1));
-
-      const __m256i utf8_packed = _mm256_shuffle_epi8(
-          utf8_unpacked, _mm256_setr_m128i(shuffle, shuffle_2));
-      // 5. store bytes
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_castsi256_si128(utf8_packed));
-      utf8_output += row[0];
-      _mm_storeu_si128((__m128i *)utf8_output,
-                       _mm256_extractf128_si256(utf8_packed, 1));
-      utf8_output += row_2[0];
-
-      // 6. adjust pointers
-      buf += 16;
-      continue;
-    }
-    // Must check for overflow in packing
-    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(
-        _mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
-    if (saturation_bitmask == 0xffffffff) {
-      // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
-
-      // Check for illegal surrogate code units
-      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
-      const __m256i forbidden_bytemask =
-          _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
-      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
-          0x0) {
-        return std::make_pair(result(error_code::SURROGATE, buf - start),
-                              utf8_output);
-      }
-
-      const __m256i dup_even = _mm256_setr_epi16(
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
-          0x0000, 0x0202, 0x0404, 0x0606, 0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
-
-      /* In this branch we handle three cases:
-        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
-        single UFT-8 byte
-        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
-        UTF-8 bytes
-        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
-        three UTF-8 bytes
-
-        We expand the input word (16-bit) into two code units (32-bit), thus
-        we have room for four bytes. However, we need five distinct bit
-        layouts. Note that the last byte in cases #2 and #3 is the same.
-
-        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
-        in register t2.
+#if SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF8
+/* begin file src/ppc64/ppc64_convert_latin1_to_utf8.cpp */
+/*
+ * reads a vector of uint16 values
+ * bits after 11th are ignored
+ * first 11 bits are encoded into utf8
+ * !important! utf8_output must have at least 16 writable bytes
+ */
+simdutf_really_inline void
+write_v_u16_11bits_to_utf8(const vector_u16 v_u16, char *&utf8_output,
+                           const vector_u8 one_byte_bytemask,
+                           const uint16_t one_byte_bitmask) {
 
-        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
-        either byte 1 for case #2 or byte 2 for case #3. Note that they
-        differ by exactly one bit.
+  // 0b1100_0000_1000_0000
+  const auto v_c080 = vector_u16(0xc080);
+  // 0b0011_1111_0000_0000
+  const auto v_1f00 = vector_u16(0x1f00);
+  // 0b0000_0000_0011_1111
+  const auto v_003f = vector_u16(0x003f);
 
-        Finally from these two code units we build proper UTF-8 sequence, taking
-        into account the case (i.e, the number of bytes to write).
-      */
-      /**
-       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
-       * t2 => [0ccc|cccc] [10cc|cccc]
-       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
-       */
-#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
-      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
-      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
-      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
-      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
-      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      const __m256i t2 = _mm256_or_si256(t1, simdutf_vec(0b1000000000000000));
+  // 1. prepare 2-byte values
+  // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
+  // expected output   : [110a|aaaa|10bb|bbbb] x 8
 
-      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
-      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
-      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
-      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
-      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
-      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
-      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
-      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
-      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask,
-                                             simdutf_vec(0b0100000000000000));
-      const __m256i s4 = _mm256_xor_si256(s3, m0);
-#undef simdutf_vec
+  // t0 = [0000|0000|00bb|bbbb]
+  const auto t0 = v_u16 & v_003f;
+  // t1 = [000a|aaaa|bbbb|bb00]
+  const auto t1 = v_u16.shl<2>();
+  // t2 = [000a|aaaa|00bb|bbbb]
+  const auto t2 = select(v_1f00, t1, t0);
+  // t3 = [110a|aaaa|10bb|bbbb]
+  const auto t3 = t2 | v_c080;
 
-      // 4. expand code units 16-bit => 32-bit
-      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
-      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);
+  // 2. merge ASCII and 2-byte codewords
+  const auto utf8_unpacked1 =
+      select(one_byte_bytemask, as_vector_u8(v_u16), as_vector_u8(t3));
 
-      // 5. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
-      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
-                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
-      // Due to the wider registers, the following path is less likely to be
-      // useful.
-      /*if(mask == 0) {
-        // We only have three-byte code units. Use fast path.
-        const __m256i shuffle =
-      _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1,
-      2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1); const __m256i utf8_0 =
-      _mm256_shuffle_epi8(out0, shuffle); const __m256i utf8_1 =
-      _mm256_shuffle_epi8(out1, shuffle);
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
-        utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_0,1)); utf8_output += 12;
-        _mm_storeu_si128((__m128i*)utf8_output,
-      _mm256_extractf128_si256(utf8_1,1)); utf8_output += 12; buf += 16;
-        continue;
-      }*/
-      const uint8_t mask0 = uint8_t(mask);
-      const uint8_t *row0 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
-      const __m128i shuffle0 = _mm_loadu_si128((__m128i *)(row0 + 1));
-      const __m128i utf8_0 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);
+#if SIMDUTF_IS_BIG_ENDIAN
+  const auto tmp = as_vector_u16(utf8_unpacked1).swap_bytes();
+#else
+  const auto tmp = as_vector_u16(utf8_unpacked1);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+  const auto utf8_unpacked = as_vector_u8(tmp);
 
-      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
-      const uint8_t *row1 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
-      const __m128i shuffle1 = _mm_loadu_si128((__m128i *)(row1 + 1));
-      const __m128i utf8_1 =
-          _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);
+  // 3. prepare bitmask for 8-bit lookup
+  //    one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a
+  //    - LSB)
+  const uint16_t m0 = one_byte_bitmask & 0x5555;      // m0 = 0h0g0f0e0d0c0b0a
+  const uint16_t m1 = static_cast<uint16_t>(m0 >> 7); // m1 = 00000000h0g0f0e0
+  const uint8_t m2 = static_cast<uint8_t>((m0 | m1) & 0xff); // m2 = hdgcfbea
+  // 4. pack the bytes
+  const uint8_t *row =
+      &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
+  const auto shuffle = vector_u8::load(row + 1);
+  const auto utf8_packed = shuffle.lookup_16(utf8_unpacked);
 
-      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
-      const uint8_t *row2 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
-      const __m128i shuffle2 = _mm_loadu_si128((__m128i *)(row2 + 1));
-      const __m128i utf8_2 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out0, 1), shuffle2);
+  // 5. store bytes
+  utf8_packed.store(utf8_output);
 
-      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
-      const uint8_t *row3 =
-          &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
-      const __m128i shuffle3 = _mm_loadu_si128((__m128i *)(row3 + 1));
-      const __m128i utf8_3 =
-          _mm_shuffle_epi8(_mm256_extractf128_si256(out1, 1), shuffle3);
+  // 6. adjust pointers
+  utf8_output += row[0];
+}
 
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_0);
-      utf8_output += row0[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_1);
-      utf8_output += row1[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_2);
-      utf8_output += row2[0];
-      _mm_storeu_si128((__m128i *)utf8_output, utf8_3);
-      utf8_output += row3[0];
-      buf += 16;
-    } else {
-      // case: at least one 32-bit word is larger than 0xFFFF <=> it will
-      // produce four UTF-8 bytes. Let us do a scalar fallback. It may seem
-      // wasteful to use scalar code, but being efficient with SIMD may require
-      // large, non-trivial tables?
-      size_t forward = 15;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFFFF80) == 0) { // 1-byte (ASCII)
-          *utf8_output++ = char(word);
-        } else if ((word & 0xFFFFF800) == 0) { // 2-byte
-          *utf8_output++ = char((word >> 6) | 0b11000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else if ((word & 0xFFFF0000) == 0) { // 3-byte
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k), utf8_output);
-          }
-          *utf8_output++ = char((word >> 12) | 0b11100000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        } else { // 4-byte
-          if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k), utf8_output);
-          }
-          *utf8_output++ = char((word >> 18) | 0b11110000);
-          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
-          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
-          *utf8_output++ = char((word & 0b111111) | 0b10000000);
-        }
-      }
-      buf += k;
-    }
-  } // while
+inline void write_v_u16_11bits_to_utf8(const vector_u16 v_u16,
+                                       char *&utf8_output,
+                                       const vector_u16 v_0000,
+                                       const vector_u16 v_ff80) {
+  // no bits set above 7th bit
+  const auto one_byte_bytemask = (v_u16 & v_ff80) == v_0000;
+  const uint16_t one_byte_bitmask = one_byte_bytemask.to_bitmask();
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
+  write_v_u16_11bits_to_utf8(v_u16, utf8_output,
+                             as_vector_u8(one_byte_bytemask), one_byte_bitmask);
 }
-/* end file src/haswell/avx2_convert_utf32_to_utf8.cpp */
-/* begin file src/haswell/avx2_convert_utf32_to_utf16.cpp */
-template <endianness big_endian>
-std::pair<const char32_t *, char16_t *>
-avx2_convert_utf32_to_utf16(const char32_t *buf, size_t len,
-                            char16_t *utf16_output) {
-  const char32_t *end = buf + len;
 
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-  __m256i forbidden_bytemask = _mm256_setzero_si256();
+std::pair<const char *const, char *const>
+ppc64_convert_latin1_to_utf8(const char *latin_input,
+                             const size_t latin_input_length,
+                             char *utf8_output) {
+  const char *end = latin_input + latin_input_length;
 
-  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+  const auto v_0000 = vector_u16::zero();
+  const auto v_00 = vector_u8::zero();
 
-    const __m256i v_00000000 = _mm256_setzero_si256();
-    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
+  // 0b1111_1111_1000_0000
+  const auto v_ff80 = vector_u16(0xff80);
 
-    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
-    const __m256i saturation_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+#if SIMDUTF_IS_BIG_ENDIAN
+  const auto latin_1_half_into_u16_byte_mask =
+      vector_u8(16, 0, 16, 1, 16, 2, 16, 3, 16, 4, 16, 5, 16, 6, 16, 7);
+  const auto latin_2_half_into_u16_byte_mask =
+      vector_u8(16, 8, 16, 9, 16, 10, 16, 11, 16, 12, 16, 13, 16, 14, 16, 15);
+#else
+  const auto latin_1_half_into_u16_byte_mask =
+      vector_u8(0, 16, 1, 16, 2, 16, 3, 16, 4, 16, 5, 16, 6, 16, 7, 16);
+  const auto latin_2_half_into_u16_byte_mask =
+      vector_u8(8, 16, 9, 16, 10, 16, 11, 16, 12, 16, 13, 16, 14, 16, 15, 16);
+#endif // SIMDUTF_IS_BIG_ENDIAN
 
-    if (saturation_bitmask == 0xffffffff) {
-      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
-      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
-      forbidden_bytemask = _mm256_or_si256(
-          forbidden_bytemask,
-          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800));
+  // each latin1 takes 1-2 utf8 bytes
+  // slow path writes useful 8-15 bytes twice (eagerly writes 16 bytes and then
+  // adjust the pointer) so the last write can exceed the utf8_output size by
+  // 8-1 bytes by reserving 8 extra input bytes, we expect the output to have
+  // 8-16 bytes free
+  while (end - latin_input >= 16 + 8) {
+    // Load 16 Latin1 characters (16 bytes) into a 128-bit register
+    const auto v_latin = vector_u8::load(latin_input);
 
-      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
-                                              _mm256_extractf128_si256(in, 1));
-      if (big_endian) {
-        const __m128i swap =
-            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
-      }
-      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
-      utf16_output += 8;
-      buf += 8;
-    } else {
-      size_t forward = 7;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          *utf16_output++ =
-              big_endian
-                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
-                  : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (big_endian) {
-            high_surrogate =
-                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
-            low_surrogate =
-                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
-        }
-      }
-      buf += k;
+    if (v_latin.is_ascii()) { // ASCII fast path!!!!
+      v_latin.store(utf8_output);
+      latin_input += 16;
+      utf8_output += 16;
+      continue;
     }
+
+    // assuming a/b are bytes and A/B are uint16 of the same value
+    // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA
+    const vector_u16 v_u16_latin_1_half =
+        as_vector_u16(latin_1_half_into_u16_byte_mask.lookup_32(v_latin, v_00));
+
+    // aaaa_aaaa_bbbb_bbbb -> BBBB_BBBB
+    const vector_u16 v_u16_latin_2_half =
+        as_vector_u16(latin_2_half_into_u16_byte_mask.lookup_32(v_latin, v_00));
+
+    write_v_u16_11bits_to_utf8(v_u16_latin_1_half, utf8_output, v_0000, v_ff80);
+    write_v_u16_11bits_to_utf8(v_u16_latin_2_half, utf8_output, v_0000, v_ff80);
+    latin_input += 16;
   }
 
-  // check for invalid input
-  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) {
-    return std::make_pair(nullptr, utf16_output);
+  if (end - latin_input >= 16) {
+    // Load 16 Latin1 characters (16 bytes) into a 128-bit register
+    const auto v_latin = vector_u8::load(latin_input);
+
+    if (v_latin.is_ascii()) { // ASCII fast path!!!!
+      v_latin.store(utf8_output);
+      latin_input += 16;
+      utf8_output += 16;
+    } else {
+      // assuming a/b are bytes and A/B are uint16 of the same value
+      // aaaa_aaaa_bbbb_bbbb -> AAAA_AAAA
+      const auto v_u16_latin_1_half = as_vector_u16(
+          latin_1_half_into_u16_byte_mask.lookup_32(v_latin, v_00));
+
+      write_v_u16_11bits_to_utf8(v_u16_latin_1_half, utf8_output, v_0000,
+                                 v_ff80);
+      latin_input += 8;
+    }
   }
 
-  return std::make_pair(buf, utf16_output);
+  return std::make_pair(latin_input, utf8_output);
 }
+/* end file src/ppc64/ppc64_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF16
+/* begin file src/ppc64/ppc64_convert_latin1_to_utf16.cpp */
 template <endianness big_endian>
-std::pair<result, char16_t *>
-avx2_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
-                                        char16_t *utf16_output) {
-  const char32_t *start = buf;
-  const char32_t *end = buf + len;
-
-  const size_t safety_margin =
-      12; // to avoid overruns, see issue
-          // https://github.com/simdutf/simdutf/issues/92
-
-  while (end - buf >= std::ptrdiff_t(8 + safety_margin)) {
-    __m256i in = _mm256_loadu_si256((__m256i *)buf);
+size_t ppc64_convert_latin1_to_utf16(const char *latin1_input, size_t len,
+                                     char16_t *utf16_output) {
+  const size_t rounded_len = align_down<vector_u8::ELEMENTS>(len);
 
-    const __m256i v_00000000 = _mm256_setzero_si256();
-    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);
+  for (size_t i = 0; i < rounded_len; i += vector_u8::ELEMENTS) {
+    const auto in = vector_u8::load(&latin1_input[i]);
+    in.store_bytes_as_utf16<big_endian>(&utf16_output[i]);
+  }
 
-    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
-    const __m256i saturation_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
+  return rounded_len;
+}
+/* end file src/ppc64/ppc64_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF16
 
-    if (saturation_bitmask == 0xffffffff) {
-      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
-      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
-      const __m256i forbidden_bytemask =
-          _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800);
-      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) !=
-          0x0) {
-        return std::make_pair(result(error_code::SURROGATE, buf - start),
-                              utf16_output);
-      }
+#if SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF32
+/* begin file src/ppc64/ppc64_convert_latin1_to_utf32.cpp */
+std::pair<const char *, char32_t *>
+ppc64_convert_latin1_to_utf32(const char *buf, size_t len,
+                              char32_t *utf32_output) {
+  const size_t rounded_len = align_down<vector_u8::ELEMENTS>(len);
 
-      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),
-                                              _mm256_extractf128_si256(in, 1));
-      if (big_endian) {
-        const __m128i swap =
-            _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
-      }
-      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
-      utf16_output += 8;
-      buf += 8;
-    } else {
-      size_t forward = 7;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(
-                result(error_code::SURROGATE, buf - start + k), utf16_output);
-          }
-          *utf16_output++ =
-              big_endian
-                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
-                  : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(
-                result(error_code::TOO_LARGE, buf - start + k), utf16_output);
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (big_endian) {
-            high_surrogate =
-                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
-            low_surrogate =
-                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
-        }
-      }
-      buf += k;
-    }
+  for (size_t i = 0; i < rounded_len; i += vector_u8::ELEMENTS) {
+    const auto in = vector_u8::load(&buf[i]);
+    in.store_bytes_as_utf32(&utf32_output[i]);
   }
 
-  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
+  return std::make_pair(buf + rounded_len, utf32_output + rounded_len);
 }
-/* end file src/haswell/avx2_convert_utf32_to_utf16.cpp */
+/* end file src/ppc64/ppc64_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_LATIN1 && SIMDUTF_FEATURE_UTF32
 
-/* begin file src/haswell/avx2_convert_utf8_to_latin1.cpp */
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/ppc64/ppc64_convert_utf8_to_latin1.cpp */
 // depends on "tables/utf8_to_utf16_tables.h"
 
 // Convert up to 12 bytes from utf8 to latin1 using a mask indicating the
@@ -31095,17 +40145,15 @@ size_t convert_masked_utf8_to_latin1(const char *input,
   // have less latency. This results in more instructions but, potentially, also
   // higher speeds.
   //
-  const __m128i in = _mm_loadu_si128((__m128i *)input);
-
+  const auto in = vector_u8::load(input);
   const uint16_t input_utf8_end_of_code_point_mask =
       utf8_end_of_code_point_mask &
       0xfff; // we are only processing 12 bytes in case it is not all ASCII
-
   if (utf8_end_of_code_point_mask == 0xfff) {
     // We process the data in chunks of 12 bytes.
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(latin1_output), in);
+    in.store(latin1_output);
     latin1_output += 12; // We wrote 12 characters.
-    return 12;           // We consumed 1 bytes.
+    return 12;           // We consumed 12 bytes.
   }
   /// We do not have a fast path available, so we fallback.
   const uint8_t idx =
@@ -31122,2014 +40170,1876 @@ size_t convert_masked_utf8_to_latin1(const char *input,
   // code code units spanning between 1 and 2 bytes each is 12 bytes. On
   // processors where pdep/pext is fast, we might be able to use a small lookup
   // table.
-  const __m128i sh =
-      _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
-  const __m128i perm = _mm_shuffle_epi8(in, sh);
-  const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
-  const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
-  __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
-  const __m128i latin1_packed = _mm_packus_epi16(composed, composed);
+
+  const auto reshuffle = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+  const auto perm8 = reshuffle.lookup_32(in, vector_u8::zero());
+#if SIMDUTF_IS_BIG_ENDIAN
+  const auto perm16 = as_vector_u16(perm8).swap_bytes();
+#else
+  const auto perm16 = as_vector_u16(perm8);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+  const auto ascii = perm16 & uint16_t(0x7f);
+  const auto highbyte = perm16 & uint16_t(0x1f00);
+  const auto composed = ascii | highbyte.shr<2>();
+
+  const auto latin1_packed = vector_u16::pack(composed, composed);
+#if defined(__clang__)
+  __attribute__((aligned(16))) char buf[16];
+  latin1_packed.store(buf);
+  memcpy(latin1_output, buf, 6);
+#else
   // writing 8 bytes even though we only care about the first 6 bytes.
-  // performance note: it would be faster to use _mm_storeu_si128, we should
-  // investigate.
-  _mm_storel_epi64((__m128i *)latin1_output, latin1_packed);
+  const auto tmp = vec_u64_t(latin1_packed.value);
+  memcpy(latin1_output, &tmp[0], 8);
+#endif
   latin1_output += 6; // We wrote 6 bytes.
   return consumed;
 }
-/* end file src/haswell/avx2_convert_utf8_to_latin1.cpp */
-
-/* begin file src/haswell/avx2_base64.cpp */
-/**
- * References and further reading:
- *
- * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
- * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
- * https://arxiv.org/abs/1910.05109
- *
- * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
- * Instructions, ACM Transactions on the Web 12 (3), 2018.
- * https://arxiv.org/abs/1704.00605
- *
- * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
- * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
- * Request for Comments: 4648.
- *
- * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
- * http://www.alfredklomp.com/programming/sse-base64/. (2014).
- *
- * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
- * acceleration. https://github.com/aklomp/base64. (2014).
- *
- * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
- * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
- *
- * Nick Kopp. 2013. Base64 Encoding on a GPU.
- * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
- */
-
-template <bool base64_url>
-simdutf_really_inline __m256i lookup_pshufb_improved(const __m256i input) {
-  // credit: Wojciech Muła
-  __m256i result = _mm256_subs_epu8(input, _mm256_set1_epi8(51));
-  const __m256i less = _mm256_cmpgt_epi8(_mm256_set1_epi8(26), input);
-  result =
-      _mm256_or_si256(result, _mm256_and_si256(less, _mm256_set1_epi8(13)));
-  __m256i shift_LUT;
-  if (base64_url) {
-    shift_LUT = _mm256_setr_epi8(
-        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
-        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0,
-
-        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
-        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0);
-  } else {
-    shift_LUT = _mm256_setr_epi8(
-        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
-        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0,
-
-        'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
-        '0' - 52, '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0);
-  }
-
-  result = _mm256_shuffle_epi8(shift_LUT, result);
-  return _mm256_add_epi8(result, input);
-}
-
-template <bool isbase64url>
-size_t encode_base64(char *dst, const char *src, size_t srclen,
-                     base64_options options) {
-  // credit: Wojciech Muła
-  const uint8_t *input = (const uint8_t *)src;
-
-  uint8_t *out = (uint8_t *)dst;
-  const __m256i shuf =
-      _mm256_set_epi8(10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1,
-
-                      10, 11, 9, 10, 7, 8, 6, 7, 4, 5, 3, 4, 1, 2, 0, 1);
-  size_t i = 0;
-  for (; i + 100 <= srclen; i += 96) {
-    const __m128i lo0 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 0));
-    const __m128i hi0 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 1));
-    const __m128i lo1 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 2));
-    const __m128i hi1 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 3));
-    const __m128i lo2 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 4));
-    const __m128i hi2 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 5));
-    const __m128i lo3 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 6));
-    const __m128i hi3 = _mm_loadu_si128(
-        reinterpret_cast<const __m128i *>(input + i + 4 * 3 * 7));
-
-    __m256i in0 = _mm256_shuffle_epi8(_mm256_set_m128i(hi0, lo0), shuf);
-    __m256i in1 = _mm256_shuffle_epi8(_mm256_set_m128i(hi1, lo1), shuf);
-    __m256i in2 = _mm256_shuffle_epi8(_mm256_set_m128i(hi2, lo2), shuf);
-    __m256i in3 = _mm256_shuffle_epi8(_mm256_set_m128i(hi3, lo3), shuf);
-
-    const __m256i t0_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x0fc0fc00));
-    const __m256i t0_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x0fc0fc00));
-    const __m256i t0_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x0fc0fc00));
-    const __m256i t0_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x0fc0fc00));
-
-    const __m256i t1_0 =
-        _mm256_mulhi_epu16(t0_0, _mm256_set1_epi32(0x04000040));
-    const __m256i t1_1 =
-        _mm256_mulhi_epu16(t0_1, _mm256_set1_epi32(0x04000040));
-    const __m256i t1_2 =
-        _mm256_mulhi_epu16(t0_2, _mm256_set1_epi32(0x04000040));
-    const __m256i t1_3 =
-        _mm256_mulhi_epu16(t0_3, _mm256_set1_epi32(0x04000040));
-
-    const __m256i t2_0 = _mm256_and_si256(in0, _mm256_set1_epi32(0x003f03f0));
-    const __m256i t2_1 = _mm256_and_si256(in1, _mm256_set1_epi32(0x003f03f0));
-    const __m256i t2_2 = _mm256_and_si256(in2, _mm256_set1_epi32(0x003f03f0));
-    const __m256i t2_3 = _mm256_and_si256(in3, _mm256_set1_epi32(0x003f03f0));
-
-    const __m256i t3_0 =
-        _mm256_mullo_epi16(t2_0, _mm256_set1_epi32(0x01000010));
-    const __m256i t3_1 =
-        _mm256_mullo_epi16(t2_1, _mm256_set1_epi32(0x01000010));
-    const __m256i t3_2 =
-        _mm256_mullo_epi16(t2_2, _mm256_set1_epi32(0x01000010));
-    const __m256i t3_3 =
-        _mm256_mullo_epi16(t2_3, _mm256_set1_epi32(0x01000010));
-
-    const __m256i input0 = _mm256_or_si256(t1_0, t3_0);
-    const __m256i input1 = _mm256_or_si256(t1_1, t3_1);
-    const __m256i input2 = _mm256_or_si256(t1_2, t3_2);
-    const __m256i input3 = _mm256_or_si256(t1_3, t3_3);
-
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
-                        lookup_pshufb_improved<isbase64url>(input0));
-    out += 32;
-
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
-                        lookup_pshufb_improved<isbase64url>(input1));
-    out += 32;
-
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
-                        lookup_pshufb_improved<isbase64url>(input2));
-    out += 32;
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
-                        lookup_pshufb_improved<isbase64url>(input3));
-    out += 32;
-  }
-  for (; i + 28 <= srclen; i += 24) {
-    // lo = [xxxx|DDDC|CCBB|BAAA]
-    // hi = [xxxx|HHHG|GGFF|FEEE]
-    const __m128i lo =
-        _mm_loadu_si128(reinterpret_cast<const __m128i *>(input + i));
-    const __m128i hi =
-        _mm_loadu_si128(reinterpret_cast<const __m128i *>(input + i + 4 * 3));
-
-    // bytes from groups A, B and C are needed in separate 32-bit lanes
-    // in = [0HHH|0GGG|0FFF|0EEE[0DDD|0CCC|0BBB|0AAA]
-    __m256i in = _mm256_shuffle_epi8(_mm256_set_m128i(hi, lo), shuf);
-
-    // this part is well commented in encode.sse.cpp
-
-    const __m256i t0 = _mm256_and_si256(in, _mm256_set1_epi32(0x0fc0fc00));
-    const __m256i t1 = _mm256_mulhi_epu16(t0, _mm256_set1_epi32(0x04000040));
-    const __m256i t2 = _mm256_and_si256(in, _mm256_set1_epi32(0x003f03f0));
-    const __m256i t3 = _mm256_mullo_epi16(t2, _mm256_set1_epi32(0x01000010));
-    const __m256i indices = _mm256_or_si256(t1, t3);
-
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(out),
-                        lookup_pshufb_improved<isbase64url>(indices));
-    out += 32;
-  }
-  return i / 3 * 4 + scalar::base64::tail_encode_base64((char *)out, src + i,
-                                                        srclen - i, options);
-}
-
-static inline void compress(__m128i data, uint16_t mask, char *output) {
-  if (mask == 0) {
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data);
-    return;
-  }
-  // this particular implementation was inspired by work done by @animetosho
-  // we do it in two steps, first 8 bytes and then second 8 bytes
-  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
-  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
-  // next line just loads the 64-bit values thintable_epi8[mask1] and
-  // thintable_epi8[mask2] into a 128-bit register, using only
-  // two instructions on most compilers.
-
-  __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2],
-                                    tables::base64::thintable_epi8[mask1]);
-  // we increment by 0x08 the second half of the mask
-  shufmask =
-      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
-  // this is the version "nearly pruned"
-  __m128i pruned = _mm_shuffle_epi8(data, shufmask);
-  // we still need to put the two halves together.
-  // we compute the popcount of the first half:
-  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
-  // then load the corresponding mask, what it does is to write
-  // only the first pop1 bytes from the first 8 bytes, and then
-  // it fills in with the bytes from the second 8 bytes + some filling
-  // at the end.
-  __m128i compactmask = _mm_loadu_si128(reinterpret_cast<const __m128i *>(
-      tables::base64::pshufb_combine_table + pop1 * 8));
-  __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+/* end file src/ppc64/ppc64_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/ppc64/ppc64_convert_utf8_to_utf16.cpp */
+// depends on "tables/utf8_to_utf16_tables.h"
 
-static inline void compress(__m256i data, uint32_t mask, char *output) {
-  if (mask == 0) {
-    _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), data);
-    return;
+// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
+template <endianness big_endian>
+size_t convert_masked_utf8_to_utf16(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char16_t *&utf16_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  // We first try a few fast paths.
+  const auto in = vector_u8::load(input);
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process the data in chunks of 12 bytes.
+    // Note: using 16 bytes is unsafe, see issue_ossfuzz_71218
+    in.store_bytes_as_utf16<big_endian>(utf16_output);
+    utf16_output += 12; // We wrote 12 16-bit characters.
+    return 12;          // We consumed 12 bytes.
   }
-  compress(_mm256_castsi256_si128(data), uint16_t(mask), output);
-  compress(_mm256_extracti128_si256(data, 1), uint16_t(mask >> 16),
-           output + _mm_popcnt_u32(~mask & 0xFFFF));
-}
+  if (((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte
+    // UTF-16 code units.
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto in16 = as_vector_u16(in);
+#else
+    const auto in16 = as_vector_u16(in).swap_bytes();
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto lo = in16 & uint16_t(0x007f);
+    const auto hi = in16.shr<2>();
 
-struct block64 {
-  __m256i chunks[2];
-};
+    auto composed = select(uint16_t(0x1f00 >> 2), hi, lo);
+    if (!match_system(big_endian)) {
+      composed = composed.swap_bytes();
+    }
 
-template <bool base64_url, bool ignore_garbage>
-static inline uint32_t to_base64_mask(__m256i *src, uint32_t *error) {
-  const __m256i ascii_space_tbl =
-      _mm256_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa,
-                       0x0, 0xc, 0xd, 0x0, 0x0, 0x20, 0x0, 0x0, 0x0, 0x0, 0x0,
-                       0x0, 0x0, 0x0, 0x9, 0xa, 0x0, 0xc, 0xd, 0x0, 0x0);
-  // credit: aqrit
-  __m256i delta_asso;
-  if (base64_url) {
-    delta_asso =
-        _mm256_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0, 0x0,
-                         0x0, 0x0, 0xF, 0x0, 0xF, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1,
-                         0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF);
-  } else {
-    delta_asso = _mm256_setr_epi8(
-        0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00, 0x00, 0x00, 0x00,
-        0x00, 0x0F, 0x00, 0x0F, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F);
+    composed.store(utf16_output);
+    utf16_output += 8; // We wrote 16 bytes, 8 code points.
+    return 16;
   }
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 2-byte
+    // UTF-16 code units. There is probably a more efficient sequence, but the
+    // following might do.
 
-  __m256i delta_values;
-  if (base64_url) {
-    delta_values = _mm256_setr_epi8(
-        0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF), uint8_t(0xBF), uint8_t(0xB9),
-        uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF), uint8_t(0xE0),
-        uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF),
-        uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9), 0x0, 0x11, uint8_t(0xC3),
-        uint8_t(0xBF), uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9));
-  } else {
-    delta_values = _mm256_setr_epi8(
-        int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13), int8_t(0x04),
-        int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9), int8_t(0x00),
-        int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9),
-        int8_t(0xB9), int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13),
-        int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9),
-        int8_t(0x00), int8_t(0x10), int8_t(0xC3), int8_t(0xBF), int8_t(0xBF),
-        int8_t(0xB9), int8_t(0xB9));
-  }
-  __m256i check_asso;
+    // AltiVec: it might be done better, for now SSE translation
 
-  if (base64_url) {
-    check_asso =
-        _mm256_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x3,
-                         0x7, 0xB, 0xE, 0xB, 0x6, 0xD, 0x1, 0x1, 0x1, 0x1, 0x1,
-                         0x1, 0x1, 0x1, 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6);
-  } else {
+    const auto sh =
+        vector_u8(2, 1, 0, 16, 5, 4, 3, 16, 8, 7, 6, 16, 11, 10, 9, 16);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u32(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto b0 = perm & uint32_t(0x0000007f);
+    const auto b1 = select(uint32_t(0x00003f00 >> 2), perm.shr<2>(), b0);
+    const auto b2 = select(uint32_t(0x000f0000 >> 4), perm.shr<4>(), b1);
+    const auto composed = b2;
+    auto packed = vector_u32::pack(composed, composed);
 
-    check_asso = _mm256_setr_epi8(
-        0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x03, 0x07,
-        0x0B, 0x0B, 0x0B, 0x0F, 0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-        0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F);
-  }
-  __m256i check_values;
-  if (base64_url) {
-    check_values = _mm256_setr_epi8(
-        uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
-        uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6), uint8_t(0xA6),
-        uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0, uint8_t(0x80),
-        0x0, uint8_t(0x80), uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
-        uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF), uint8_t(0xB6),
-        uint8_t(0xA6), uint8_t(0xB5), uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0,
-        uint8_t(0x80), 0x0, uint8_t(0x80));
-  } else {
-    check_values = _mm256_setr_epi8(
-        int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0xCF),
-        int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5), int8_t(0x86),
-        int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80), int8_t(0x91),
-        int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80),
-        int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6), int8_t(0xB5),
-        int8_t(0x86), int8_t(0xD1), int8_t(0x80), int8_t(0xB1), int8_t(0x80),
-        int8_t(0x91), int8_t(0x80));
-  }
-  const __m256i shifted = _mm256_srli_epi32(*src, 3);
-  const __m256i delta_hash =
-      _mm256_avg_epu8(_mm256_shuffle_epi8(delta_asso, *src), shifted);
-  const __m256i check_hash =
-      _mm256_avg_epu8(_mm256_shuffle_epi8(check_asso, *src), shifted);
-  const __m256i out =
-      _mm256_adds_epi8(_mm256_shuffle_epi8(delta_values, delta_hash), *src);
-  const __m256i chk =
-      _mm256_adds_epi8(_mm256_shuffle_epi8(check_values, check_hash), *src);
-  const int mask = _mm256_movemask_epi8(chk);
-  if (!ignore_garbage && mask) {
-    __m256i ascii_space =
-        _mm256_cmpeq_epi8(_mm256_shuffle_epi8(ascii_space_tbl, *src), *src);
-    *error = (mask ^ _mm256_movemask_epi8(ascii_space));
-  }
-  *src = out;
-  return (uint32_t)mask;
-}
+    if (!match_system(big_endian)) {
+      packed = packed.swap_bytes();
+    }
 
-template <bool base64_url, bool ignore_garbage>
-static inline uint64_t to_base64_mask(block64 *b, uint64_t *error) {
-  uint32_t err0 = 0;
-  uint32_t err1 = 0;
-  uint64_t m0 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[0], &err0);
-  uint64_t m1 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[1], &err1);
-  if (!ignore_garbage) {
-    *error = err0 | ((uint64_t)err1 << 32);
+    packed.store(utf16_output);
+    utf16_output += 4;
+    return 12;
   }
-  return m0 | (m1 << 32);
-}
+  /// We do not have a fast path available, so we fallback.
 
-static inline void copy_block(block64 *b, char *output) {
-  _mm256_storeu_si256(reinterpret_cast<__m256i *>(output), b->chunks[0]);
-  _mm256_storeu_si256(reinterpret_cast<__m256i *>(output + 32), b->chunks[1]);
-}
+  const uint8_t idx =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
 
-static inline uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
-  uint64_t nmask = ~mask;
-  compress(b->chunks[0], uint32_t(mask), output);
-  compress(b->chunks[1], uint32_t(mask >> 32),
-           output + _mm_popcnt_u64(nmask & 0xFFFFFFFF));
-  return _mm_popcnt_u64(nmask);
-}
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
+    // processors where pdep/pext is fast, we might be able to use a small
+    // lookup table.
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u16(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u16(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto b0 = perm & uint16_t(0x007f);
+    const auto b1 = perm & uint16_t(0x1f00);
 
-// The caller of this function is responsible to ensure that there are 64 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char *src) {
-  b->chunks[0] = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src));
-  b->chunks[1] =
-      _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32));
-}
+    auto composed = b0 | b1.shr<2>();
 
-// The caller of this function is responsible to ensure that there are 128 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char16_t *src) {
-  __m256i m1 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src));
-  __m256i m2 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 16));
-  __m256i m3 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32));
-  __m256i m4 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 48));
-  __m256i m1p = _mm256_permute2x128_si256(m1, m2, 0x20);
-  __m256i m2p = _mm256_permute2x128_si256(m1, m2, 0x31);
-  __m256i m3p = _mm256_permute2x128_si256(m3, m4, 0x20);
-  __m256i m4p = _mm256_permute2x128_si256(m3, m4, 0x31);
-  b->chunks[0] = _mm256_packus_epi16(m1p, m2p);
-  b->chunks[1] = _mm256_packus_epi16(m3p, m4p);
-}
+    if (!match_system(big_endian)) {
+      composed = composed.swap_bytes();
+    }
 
-static inline void base64_decode(char *out, __m256i str) {
-  // credit: aqrit
-  const __m256i pack_shuffle =
-      _mm256_setr_epi8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1,
-                       2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, -1, -1, -1, -1);
-  const __m256i t0 = _mm256_maddubs_epi16(str, _mm256_set1_epi32(0x01400140));
-  const __m256i t1 = _mm256_madd_epi16(t0, _mm256_set1_epi32(0x00011000));
-  const __m256i t2 = _mm256_shuffle_epi8(t1, pack_shuffle);
+    composed.store(utf16_output);
+    utf16_output += 6; // We wrote 12 bytes, 6 code points.
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u32(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto b0 = perm & uint32_t(0x0000007f);
+    const auto b1 = perm & uint32_t(0x00003f00);
+    const auto b2 = perm & uint32_t(0x000f0000);
 
-  // Store the output:
-  _mm_storeu_si128((__m128i *)out, _mm256_castsi256_si128(t2));
-  _mm_storeu_si128((__m128i *)(out + 12), _mm256_extracti128_si256(t2, 1));
-}
-// decode 64 bytes and output 48 bytes
-static inline void base64_decode_block(char *out, const char *src) {
-  base64_decode(out,
-                _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src)));
-  base64_decode(out + 24, _mm256_loadu_si256(
-                              reinterpret_cast<const __m256i *>(src + 32)));
-}
-static inline void base64_decode_block_safe(char *out, const char *src) {
-  base64_decode(out,
-                _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src)));
-  char buffer[32]; // We enforce safety with a buffer.
-  base64_decode(
-      buffer, _mm256_loadu_si256(reinterpret_cast<const __m256i *>(src + 32)));
-  std::memcpy(out + 24, buffer, 24);
-}
-static inline void base64_decode_block(char *out, block64 *b) {
-  base64_decode(out, b->chunks[0]);
-  base64_decode(out + 24, b->chunks[1]);
-}
-static inline void base64_decode_block_safe(char *out, block64 *b) {
-  base64_decode(out, b->chunks[0]);
-  char buffer[32]; // We enforce safety with a buffer.
-  base64_decode(buffer, b->chunks[1]);
-  std::memcpy(out + 24, buffer, 24);
-}
+    const auto composed = b0 | b1.shr<2>() | b2.shr<4>();
 
-template <bool base64_url, bool ignore_garbage, typename chartype>
-full_result
-compress_decode_base64(char *dst, const chartype *src, size_t srclen,
-                       base64_options options,
-                       last_chunk_handling_options last_chunk_options) {
-  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
-                                        : tables::base64::to_base64_value;
-  size_t equallocation =
-      srclen; // location of the first padding character if any
-  // skip trailing spaces
-  while (!ignore_garbage && srclen > 0 &&
-         scalar::base64::is_eight_byte(src[srclen - 1]) &&
-         to_base64[uint8_t(src[srclen - 1])] == 64) {
-    srclen--;
-  }
-  size_t equalsigns = 0;
-  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
-    equallocation = srclen - 1;
-    srclen--;
-    equalsigns = 1;
-    // skip trailing spaces
-    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
-           to_base64[uint8_t(src[srclen - 1])] == 64) {
-      srclen--;
-    }
-    if (srclen > 0 && src[srclen - 1] == '=') {
-      equallocation = srclen - 1;
-      srclen--;
-      equalsigns = 2;
-    }
-  }
-  if (srclen == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    auto packed = vector_u32::pack(composed, composed);
+
+    if (!match_system(big_endian)) {
+      packed = packed.swap_bytes();
     }
-    return {SUCCESS, 0, 0};
-  }
-  char *end_of_safe_64byte_zone =
-      (srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 : dst;
 
-  const chartype *const srcinit = src;
-  const char *const dstinit = dst;
-  const chartype *const srcend = src + srclen;
+    packed.store(utf16_output);
+    utf16_output += 4;
+  } else if (idx < 209) {
+    // TWO (2) input code-code units
+    //////////////
+    // There might be garbage inputs where a leading byte mascarades as a
+    // four-byte leading byte (by being followed by 3 continuation byte), but is
+    // not greater than 0xf0. This could trigger a buffer overflow if we only
+    // counted leading bytes of the form 0xf0 as generating surrogate pairs,
+    // without further UTF-8 validation. Thus we must be careful to ensure that
+    // only leading bytes at least as large as 0xf0 generate surrogate pairs. We
+    // do as at the cost of an extra mask.
+    /////////////
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u32(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto ascii = perm & uint32_t(0x00000007f);
+    const auto middlebyte = perm & uint32_t(0x00003f00);
+    const auto middlebyte_shifted = middlebyte.shr<2>();
 
-  constexpr size_t block_size = 6;
-  static_assert(block_size >= 2, "block_size must be at least two");
-  char buffer[block_size * 64];
-  char *bufferptr = buffer;
-  if (srclen >= 64) {
-    const chartype *const srcend64 = src + srclen - 64;
-    while (src <= srcend64) {
-      block64 b;
-      load_block(&b, src);
-      src += 64;
-      uint64_t error = 0;
-      uint64_t badcharmask =
-          to_base64_mask<base64_url, ignore_garbage>(&b, &error);
-      if (!ignore_garbage && error) {
-        src -= 64;
-        size_t error_offset = _tzcnt_u64(error);
-        return {error_code::INVALID_BASE64_CHARACTER,
-                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
-      }
-      if (badcharmask != 0) {
-        // optimization opportunity: check for simple masks like those made of
-        // continuous 1s followed by continuous 0s. And masks containing a
-        // single bad character.
-        bufferptr += compress_block(&b, badcharmask, bufferptr);
-      } else if (bufferptr != buffer) {
-        copy_block(&b, bufferptr);
-        bufferptr += 64;
-      } else {
-        if (dst >= end_of_safe_64byte_zone) {
-          base64_decode_block_safe(dst, &b);
+    auto middlehighbyte = perm & uint32_t(0x003f0000);
+    // correct for spurious high bit
+
+    const auto correct = (perm & uint32_t(0x00400000)).shr<1>();
+    middlehighbyte = correct ^ middlehighbyte;
+    const auto middlehighbyte_shifted = middlehighbyte.shr<4>();
+    // We deliberately carry the leading four bits in highbyte if they are
+    // present, we remove them later when computing hightenbits.
+    const auto highbyte = perm & uint32_t(0xff000000);
+    const auto highbyte_shifted = highbyte.shr<6>();
+    // When we need to generate a surrogate pair (leading byte > 0xF0), then
+    // the corresponding 32-bit value in 'composed'  will be greater than
+    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
+    // location of the surrogate pairs.
+    const auto composed =
+        ascii | middlebyte_shifted | highbyte_shifted | middlehighbyte_shifted;
+
+    const auto composedminus = composed - uint32_t(0x10000);
+    const auto lowtenbits = composedminus & uint32_t(0x3ff);
+    // Notice the 0x3ff mask:
+    const auto hightenbits = composedminus.shr<10>() & uint32_t(0x3ff);
+    const auto lowtenbitsadd = lowtenbits + uint32_t(0xDC00);
+    const auto hightenbitsadd = hightenbits + uint32_t(0xD800);
+    const auto lowtenbitsaddshifted = lowtenbitsadd.shl<16>();
+    auto surrogates = hightenbitsadd | lowtenbitsaddshifted;
+
+    uint32_t basic_buffer[4];
+    composed.store(basic_buffer);
+    uint32_t surrogate_buffer[4];
+    surrogates.swap_bytes().store(surrogate_buffer);
+
+    for (size_t i = 0; i < 3; i++) {
+      if (basic_buffer[i] > 0x3c00000) {
+        const auto ch0 = uint16_t(surrogate_buffer[i] & 0xffff);
+        const auto ch1 = uint16_t(surrogate_buffer[i] >> 16);
+        if (match_system(big_endian)) {
+          utf16_output[1] = scalar::u16_swap_bytes(ch0);
+          utf16_output[0] = scalar::u16_swap_bytes(ch1);
         } else {
-          base64_decode_block(dst, &b);
-        }
-        dst += 48;
-      }
-      if (bufferptr >= (block_size - 1) * 64 + buffer) {
-        for (size_t i = 0; i < (block_size - 2); i++) {
-          base64_decode_block(dst, buffer + i * 64);
-          dst += 48;
+          utf16_output[1] = ch0;
+          utf16_output[0] = ch1;
         }
-        if (dst >= end_of_safe_64byte_zone) {
-          base64_decode_block_safe(dst, buffer + (block_size - 2) * 64);
+        utf16_output += 2;
+      } else {
+        const auto chr = uint16_t(basic_buffer[i]);
+        if (match_system(big_endian)) {
+          utf16_output[0] = chr;
         } else {
-          base64_decode_block(dst, buffer + (block_size - 2) * 64);
+          utf16_output[0] = scalar::u16_swap_bytes(chr);
         }
-        dst += 48;
-        std::memcpy(buffer, buffer + (block_size - 1) * 64,
-                    64); // 64 might be too much
-        bufferptr -= (block_size - 1) * 64;
+
+        utf16_output++;
       }
     }
+  } else {
+    // here we know that there is an error but we do not handle errors
   }
+  return consumed;
+}
+/* end file src/ppc64/ppc64_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-  char *buffer_start = buffer;
-  // Optimization note: if this is almost full, then it is worth our
-  // time, otherwise, we should just decode directly.
-  int last_block = (int)((bufferptr - buffer_start) % 64);
-  if (last_block != 0 && srcend - src + last_block >= 64) {
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/ppc64/ppc64_convert_utf8_to_utf32.cpp */
+// depends on "tables/utf8_to_utf16_tables.h"
 
-    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
-      uint8_t val = to_base64[uint8_t(*src)];
-      *bufferptr = char(val);
-      if (!ignore_garbage &&
-          (!scalar::base64::is_eight_byte(*src) || val > 64)) {
-        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      }
-      bufferptr += (val <= 63);
-      src++;
-    }
+// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
+// end of the code points. Only the least significant 12 bits of the mask
+// are accessed.
+// It returns how many bytes were consumed (up to 12).
+size_t convert_masked_utf8_to_utf32(const char *input,
+                                    uint64_t utf8_end_of_code_point_mask,
+                                    char32_t *&utf32_output) {
+  // we use an approach where we try to process up to 12 input bytes.
+  // Why 12 input bytes and not 16? Because we are concerned with the size of
+  // the lookup tables. Also 12 is nicely divisible by two and three.
+  //
+  //
+  // Optimization note: our main path below is load-latency dependent. Thus it
+  // is maybe beneficial to have fast paths that depend on branch prediction but
+  // have less latency. This results in more instructions but, potentially, also
+  // higher speeds.
+  //
+  // We first try a few fast paths.
+  const auto in = vector_u8::load(input);
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process the data in chunks of 12 bytes.
+    in.store_bytes_as_utf32(utf32_output);
+    utf32_output += 12; // We wrote 12 32-bit characters.
+    return 12;          // We consumed 12 bytes.
   }
+  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte
+    // UTF-32 code units.
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm = as_vector_u16(in);
+#else
+    const auto perm = as_vector_u16(in).swap_bytes();
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    // in = [110aaaaa|10bbbbbb]
+    // t0 = [00000000|00bbbbbb]
+    const auto t0 = perm & uint16_t(0x007f);
 
-  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
-    if (dst >= end_of_safe_64byte_zone) {
-      base64_decode_block_safe(dst, buffer_start);
-    } else {
-      base64_decode_block(dst, buffer_start);
-    }
-    dst += 48;
+    // t1 = [00110aaa|aabbbbbb]
+    const auto t1 = perm.shr<2>();
+    const auto composed = select(uint16_t(0x1f00 >> 2), t1, t0);
+
+    const auto composed8 = as_vector_u8(composed);
+    composed8.store_words_as_utf32(utf32_output);
+
+    utf32_output += 8; // We wrote 32 bytes, 8 code points.
+    return 16;
   }
-  if ((bufferptr - buffer_start) % 64 != 0) {
-    while (buffer_start + 4 < bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 4);
+  if (input_utf8_end_of_code_point_mask == 0x924) {
+    // We want to take 4 3-byte UTF-8 code units and turn them into 4 4-byte
+    // UTF-32 code units.
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto sh =
+        vector_u8(-1, 0, 1, 2, -1, 3, 4, 5, -1, 6, 7, 8, -1, 9, 10, 11);
+#else
+    const auto sh =
+        vector_u8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    if (buffer_start + 4 <= bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 3);
+    // in = [1110aaaa|10bbbbbb|10cccccc]
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
-    // backtrack
-    int leftover = int(bufferptr - buffer_start);
-    while (leftover > 0) {
-      if (!ignore_garbage) {
-        while (to_base64[uint8_t(*(src - 1))] == 64) {
-          src--;
-        }
-      } else {
-        while (to_base64[uint8_t(*(src - 1))] >= 64) {
-          src--;
-        }
-      }
-      src--;
-      leftover--;
-    }
+    // t0 = [00000000|00000000|00cccccc]
+    const auto t0 = perm & uint32_t(0x0000007f);
+
+    // t2 = [00000000|0000bbbb|bbcccccc]
+    const auto t1 = perm.shr<2>();
+    const auto t2 = select(uint32_t(0x00003f00 >> 2), t1, t0);
+
+    // t4 = [00000000|aaaabbbb|bbcccccc]
+    const auto t3 = perm.shr<4>();
+    const auto t4 = select(uint32_t(0x0f0000 >> 4), t3, t2);
+
+    t4.store(utf32_output);
+    utf32_output += 4;
+    return 12;
   }
-  if (src < srcend + equalsigns) {
-    full_result r = scalar::base64::base64_tail_decode(
-        dst, src, srcend - src, equalsigns, options, last_chunk_options);
-    r.input_count += size_t(src - srcinit);
-    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
-        r.error == error_code::BASE64_EXTRA_BITS) {
-      return r;
+  /// We do not have a fast path available, so we fallback.
+
+  const uint8_t idx =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
+  const uint8_t consumed =
+      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
+  if (idx < 64) {
+    // SIX (6) input code-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units spanning between 1 and 2 bytes each is 12 bytes. On
+    // processors where pdep/pext is fast, we might be able to use a small
+    // lookup table.
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u16(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u16(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto ascii = perm & uint16_t(0x7f);
+    const auto highbyte = perm & uint16_t(0x1f00);
+    const auto composed = ascii | highbyte.shr<2>();
+
+    as_vector_u8(composed).store_words_as_utf32(utf32_output);
+    utf32_output += 6; // We wrote 12 bytes, 6 code points.
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u32(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto ascii = perm & uint32_t(0x7f);
+    const auto middlebyte = perm & uint32_t(0x3f00);
+    const auto middlebyte_shifted = middlebyte.shr<2>();
+    const auto highbyte = perm & uint32_t(0x0f0000);
+    const auto highbyte_shifted = highbyte.shr<4>();
+    const auto composed = ascii | middlebyte_shifted | highbyte_shifted;
+
+    composed.store(utf32_output);
+    utf32_output += 4;
+  } else if (idx < 209) {
+    // TWO (2) input code-code units
+    const auto sh = vector_u8::load(&tables::utf8_to_utf16::shufutf8[idx]);
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto perm =
+        as_vector_u32(sh.lookup_32(in, vector_u8::zero())).swap_bytes();
+#else
+    const auto perm = as_vector_u32(sh.lookup_32(in, vector_u8::zero()));
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto ascii = perm & uint32_t(0x0000007f);
+    const auto middlebyte = perm & uint32_t(0x3f00);
+    const auto middlebyte_shifted = middlebyte.shr<2>();
+    auto middlehighbyte = perm & uint32_t(0x003f0000);
+    // correct for spurious high bit
+    const auto correct0 = perm & uint32_t(0x00400000);
+    const auto correct = correct0.shr<1>();
+    middlehighbyte = correct ^ middlehighbyte;
+    const auto middlehighbyte_shifted = middlehighbyte.shr<4>();
+    const auto highbyte = perm & uint32_t(0x07000000);
+    const auto highbyte_shifted = highbyte.shr<6>();
+    const auto composed =
+        ascii | middlebyte_shifted | highbyte_shifted | middlehighbyte_shifted;
+    composed.store(utf32_output);
+    utf32_output += 3;
+  } else {
+    // here we know that there is an error but we do not handle errors
+  }
+  return consumed;
+}
+/* end file src/ppc64/ppc64_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/ppc64/ppc64_convert_utf16_to_latin1.cpp */
+struct utf16_to_latin1_t {
+  error_code err;
+  const char16_t *input;
+  char *output;
+};
+
+template <endianness big_endian>
+utf16_to_latin1_t ppc64_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                                                char *latin1_output) {
+  const char16_t *end = buf + len;
+  while (end - buf >= 8) {
+    // Load 8 x UTF-16 characters
+    auto in = vector_u8::load(buf);
+
+    // Move low bytes of UTF-16 chars to lower half of `in`
+    // and upper bytes to upper half of `in`.
+    if (!match_system(big_endian)) {
+      const auto perm =
+          vector_u8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
+      in = perm.lookup_16(in);
     } else {
-      r.output_count += size_t(dst - dstinit);
-    }
-    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
-        r.error == error_code::SUCCESS && equalsigns > 0) {
-      // additional checks
-      if ((r.output_count % 3 == 0) ||
-          ((r.output_count % 3) + 1 + equalsigns != 4)) {
-        r.error = error_code::INVALID_BASE64_CHARACTER;
-        r.input_count = equallocation;
+      const auto perm =
+          vector_u8(1, 3, 5, 7, 9, 11, 13, 15, 0, 2, 4, 6, 8, 10, 12, 14);
+      in = perm.lookup_16(in);
+    }
+
+    // AltiVec-specific
+#if defined(__clang__)
+    __attribute__((aligned(16))) uint64_t tmp[8];
+    in.store(tmp);
+  #if SIMDUTF_IS_BIG_ENDIAN
+    memcpy(latin1_output, &tmp[0], 8);
+    const uint64_t upper = tmp[1];
+  #else
+    memcpy(latin1_output, &tmp[1], 8);
+    const uint64_t upper = tmp[0];
+  #endif // SIMDUTF_IS_BIG_ENDIAN
+#else
+    const auto tmp = vec_u64_t(in.value);
+  #if SIMDUTF_IS_BIG_ENDIAN
+    memcpy(latin1_output, &tmp[0], 8);
+    const uint64_t upper = tmp[1];
+  #else
+    memcpy(latin1_output, &tmp[1], 8);
+    const uint64_t upper = tmp[0];
+  #endif // SIMDUTF_IS_BIG_ENDIAN
+#endif   // defined(__clang__)
+    // AltiVec
+
+    if (simdutf_unlikely(upper)) {
+      uint8_t bytes[8];
+      memcpy(bytes, &upper, 8);
+      for (size_t k = 0; k < 8; k++) {
+        if (bytes[k] != 0) {
+          return utf16_to_latin1_t{error_code::TOO_LARGE, buf + k,
+                                   latin1_output};
+        }
       }
+    } else {
+      // Adjust pointers for next iteration
+      buf += 8;
+      latin1_output += 8;
     }
-    return r;
-  }
-  if (!ignore_garbage && equalsigns > 0) {
-    if ((size_t(dst - dstinit) % 3 == 0) ||
-        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
-    }
-  }
-  return {SUCCESS, srclen, size_t(dst - dstinit)};
+  } // while
+
+  return utf16_to_latin1_t{error_code::SUCCESS, buf, latin1_output};
 }
-/* end file src/haswell/avx2_base64.cpp */
+/* end file src/ppc64/ppc64_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
+/* begin file src/ppc64/ppc64_convert_utf16_to_utf8.cpp */
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
 
-/* begin file src/generic/buf_block_reader.h */
-namespace simdutf {
-namespace haswell {
-namespace {
+    We consider three cases:
+    1. an input register contains no surrogates and each value
+       is in range 0x0000 .. 0x07ff.
+    2. an input register contains no surrogates and values are
+       is in range 0x0000 .. 0xffff.
+    3. an input register contains surrogates --- i.e. codepoints
+       can have 16 or 32 bits.
 
-// Walks through a buffer in block-sized increments, loading the last part with
-// spaces
-template <size_t STEP_SIZE> struct buf_block_reader {
-public:
-  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdutf_really_inline size_t block_index();
-  simdutf_really_inline bool has_full_block() const;
-  simdutf_really_inline const uint8_t *full_block() const;
-  /**
-   * Get the last block, padded with spaces.
-   *
-   * There will always be a last block, with at least 1 byte, unless len == 0
-   * (in which case this function fills the buffer with spaces and returns 0. In
-   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
-   * block with STEP_SIZE bytes and no spaces for padding.
-   *
-   * @return the number of effective characters in the last block.
-   */
-  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdutf_really_inline void advance();
+    Ad 1.
+
+    When values are less than 0x0800, it means that a 16-bit code unit
+    can be converted into: 1) single UTF8 byte (when it is an ASCII
+    char) or 2) two UTF8 bytes.
+
+    For this case we do only some shuffle to obtain these 2-byte
+    codes and finally compress the whole SSE register with a single
+    shuffle.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+    Ad 2.
+
+    When values fit in 16-bit code units, but are above 0x07ff, then
+    a single word may produce one, two or three UTF8 bytes.
+
+    We prepare data for all these three cases in two registers.
+    The first register contains lower two UTF8 bytes (used in all
+    cases), while the second one contains just the third byte for
+    the three-UTF8-bytes case.
+
+    Finally these two registers are interleaved forming eight-element
+    array of 32-bit values. The array spans two SSE registers.
+    The bytes from the registers are compressed using two shuffles.
+
+    We need 256-entry lookup table to get a compression pattern
+    and the number of output bytes in the compressed vector register.
+    Each entry occupies 17 bytes.
+
+
+    To summarize:
+    - We need two 256-entry tables that have 8704 bytes in total.
+*/
 
-private:
-  const uint8_t *buf;
-  const size_t len;
-  const size_t lenminusstep;
-  size_t idx;
-};
+// Auxiliary procedure used by UTF-16 and UTF-32 into UTF-8.
+// Note the pointer is passed by reference, it is updated by the procedure.
+template <typename T>
+simdutf_really_inline void ppc64_convert_utf16_to_1_2_3_bytes_of_utf8(
+    const vector_u16 in, uint16_t one_byte_bitmask,
+    const T one_or_two_bytes_bytemask, uint16_t one_or_two_bytes_bitmask,
+    char *&utf8_output) {
+  // case: code units from register produce either 1, 2 or 3 UTF-8 bytes
+#if SIMDUTF_IS_BIG_ENDIAN
+  const auto dup_lsb =
+      vector_u8(1, 1, 3, 3, 5, 5, 7, 7, 9, 9, 11, 11, 13, 13, 15, 15);
+#else
+  const auto dup_lsb =
+      vector_u8(0, 0, 2, 2, 4, 4, 6, 6, 8, 8, 10, 10, 12, 12, 14, 14);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+
+  /* In this branch we handle three cases:
+     1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           -
+    single UFT-8 byte
+     2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two
+    UTF-8 bytes
+     3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] -
+    three UTF-8 bytes
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text_64(const uint8_t *text) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
-  }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
+    We expand the input word (16-bit) into two code units (32-bit), thus
+    we have room for four bytes. However, we need five distinct bit
+    layouts. Note that the last byte in cases #2 and #3 is the same.
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  in.store(reinterpret_cast<uint8_t *>(buf));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    if (buf[i] < ' ') {
-      buf[i] = '_';
-    }
-  }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
+    We precompute byte 1 for case #1 and the common byte for cases #2 & #3
+    in register t2.
 
-simdutf_unused static char *format_mask(uint64_t mask) {
-  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
-  for (size_t i = 0; i < 64; i++) {
-    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
-  }
-  buf[64] = '\0';
-  return buf;
-}
+    We precompute byte 1 for case #3 and -- **conditionally** -- precompute
+    either byte 1 for case #2 or byte 2 for case #3. Note that they
+    differ by exactly one bit.
 
-template <size_t STEP_SIZE>
-simdutf_really_inline
-buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
-    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
-      idx{0} {}
+    Finally from these two code units we build proper UTF-8 sequence, taking
+    into account the case (i.e, the number of bytes to write).
+  */
+  /**
+   * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
+   * t2 => [0ccc|cccc] [10cc|cccc]
+   * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
+   */
+  // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
+  const auto t0 = as_vector_u16(dup_lsb.lookup_16(as_vector_u8(in)));
+
+  // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
+  const auto t1 = t0 & uint16_t(0b0011111101111111);
+  // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
+  const auto t2 = t1 | uint16_t(0b1000000000000000);
+
+  // in = [aaaa|bbbb|bbcc|cccc]
+  // a0 = [0000|0000|0000|aaaa]
+  const auto a0 = in.shr<12>();
+  // b0 = [aabb|bbbb|cccc|cc00]
+  const auto b0 = in.shl<2>();
+  // s0 = [00bb|bbbb|00cc|cccc]
+  const auto s0 = select(uint16_t(0x3f00), b0, a0);
+
+  // s3 = [11bb|bbbb|1110|aaaa]
+  const auto s3 = s0 | uint16_t(0b1100000011100000);
+
+  const auto m0 =
+      ~as_vector_u16(one_or_two_bytes_bytemask) & uint16_t(0b0100000000000000);
+  const auto s4 = s3 ^ m0;
+
+  // 4. compress 32-bit code units into 1, 2 or 3 bytes -- 2 x shuffle
+  const uint16_t mask =
+      (one_byte_bitmask & 0x5555) | (one_or_two_bytes_bitmask & 0xaaaa);
+  if (mask == 0) {
+    // We only have three-byte code units. Use fast path.
+#if SIMDUTF_IS_BIG_ENDIAN
+    // Lookups produced by scripts/ppc64_convert_utf16_to_utf8.py
+    const auto shuffle0 =
+        vector_u8(1, 0, 16, 3, 2, 18, 5, 4, 20, 7, 6, 22, 9, 8, 24, 11);
+    const auto shuffle1 = vector_u8(10, 26, 13, 12, 28, 15, 14, 30, -1, -1, -1,
+                                    -1, -1, -1, -1, -1);
+#else
+    const auto shuffle0 =
+        vector_u8(0, 1, 17, 2, 3, 19, 4, 5, 21, 6, 7, 23, 8, 9, 25, 10);
+    const auto shuffle1 = vector_u8(11, 27, 12, 13, 29, 14, 15, 31, -1, -1, -1,
+                                    -1, -1, -1, -1, -1);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto utf8_0 = shuffle0.lookup_32(as_vector_u8(s4), as_vector_u8(t2));
+    const auto utf8_1 = shuffle1.lookup_32(as_vector_u8(s4), as_vector_u8(t2));
+
+    utf8_0.store(utf8_output);
+    utf8_output += 16;
+    utf8_1.store(utf8_output);
+    utf8_output += 8;
+    return;
+  }
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
-  return idx;
-}
+  const uint8_t mask0 = uint8_t(mask);
 
-template <size_t STEP_SIZE>
-simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
-  return idx < lenminusstep;
-}
+  const uint8_t *row0 =
+      &simdutf::tables::ppc64_utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
+  const auto shuffle0 = vector_u8::load(row0 + 1);
 
-template <size_t STEP_SIZE>
-simdutf_really_inline const uint8_t *
-buf_block_reader<STEP_SIZE>::full_block() const {
-  return &buf[idx];
-}
+  const auto utf8_0 = shuffle0.lookup_32(as_vector_u8(s4), as_vector_u8(t2));
+  const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t
-buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
-  if (len == idx) {
-    return 0;
-  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
-  std::memset(dst, 0x20,
-              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
-                          // to write out 8 or 16 bytes at once.
-  std::memcpy(dst, buf + idx, len - idx);
-  return len - idx;
-}
+  const uint8_t *row1 =
+      &simdutf::tables::ppc64_utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
+  const auto shuffle1 = vector_u8::load(row1 + 1) + uint8_t(8);
+  const auto utf8_1 = shuffle1.lookup_32(as_vector_u8(s4), as_vector_u8(t2));
 
-template <size_t STEP_SIZE>
-simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
-  idx += STEP_SIZE;
+  utf8_0.store(utf8_output);
+  utf8_output += row0[0];
+  utf8_1.store(utf8_output);
+  utf8_output += row1[0];
 }
 
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/buf_block_reader.h */
-/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_validation {
+struct utf16_to_utf8_t {
+  error_code err;
+  const char16_t *input;
+  char *output;
+};
 
-using namespace simd;
+/*
+  Returns utf16_to_utf8_t value
+  A scalar routine should carry on the conversion of the tail,
+  iff there was no error.
+*/
+template <endianness big_endian>
+utf16_to_utf8_t ppc64_convert_utf16_to_utf8(const char16_t *buf, size_t len,
+                                            char *utf8_output) {
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  const char16_t *end = buf + len;
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+  const auto v_f800 = vector_u16(0xf800);
+  const auto v_d800 = vector_u16(0xd800);
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
+    auto in = vector_u16::load(buf);
+    if (not match_system(big_endian)) {
+      in = in.swap_bytes();
+    }
+    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
+    if (in.is_ascii()) {
+      auto nextin = vector_u16::load(buf + vector_u16::ELEMENTS);
+      if (not match_system(big_endian)) {
+        nextin = nextin.swap_bytes();
+      }
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+      if (nextin.is_ascii()) {
+        // 1. pack the bytes
+        const auto utf8_packed = vector_u16::pack(in, nextin);
+        // 2. store (16 bytes)
+        utf8_packed.store(utf8_output);
+        // 3. adjust pointers
+        buf += 16;
+        utf8_output += 16;
+        continue; // we are done for this round!
+      }
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      // next block is not ASCII
+      const auto utf8_packed = vector_u16::pack(in, in);
+      // 2. store (16 bytes)
+      utf8_packed.store(utf8_output);
+      // 3. adjust pointers
+      buf += 8;
+      utf8_output += 8;
+      in = nextin;
+      // fallback
+    }
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
+    // no bits set above 7th bit
+    const auto one_byte_bytemask = in < uint16_t(1 << 7);
+    const uint16_t one_byte_bitmask = one_byte_bytemask.to_bitmask();
 
-//
-// Return nonzero if there are incomplete multibyte characters at the end of the
-// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
-//
-simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
-  // If the previous input's last 3 bytes match this, they're too short (they
-  // ended at EOF):
-  // ... 1111____ 111_____ 11______
-  static const uint8_t max_array[32] = {255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        0b11110000u - 1,
-                                        0b11100000u - 1,
-                                        0b11000000u - 1};
-  const simd8<uint8_t> max_value(
-      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
-  return input.gt_bits(max_value);
-}
+    // no bits set above 11th bit
+    const auto one_or_two_bytes_bytemask = in < uint16_t(1 << 11);
+    const uint16_t one_or_two_bytes_bitmask =
+        one_or_two_bytes_bytemask.to_bitmask();
 
-struct utf8_checker {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
-  // The last input we received
-  simd8<uint8_t> prev_input_block;
-  // Whether the last input we received was incomplete (used for ASCII fast
-  // path)
-  simd8<uint8_t> prev_incomplete;
+    if (one_or_two_bytes_bitmask == 0xffff) {
+      write_v_u16_11bits_to_utf8(
+          in, utf8_output, as_vector_u8(one_byte_bytemask), one_byte_bitmask);
+      buf += 8;
+      continue;
+    }
 
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
-  }
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also to deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const auto surrogates_bytemask = (in & v_f800) == v_d800;
 
-  // The only problem that can happen at EOF is that a multibyte character is
-  // too short or a byte value too large in the last bytes: check_special_cases
-  // only checks for bytes too large in the first of two bytes.
-  simdutf_really_inline void check_eof() {
-    // If the previous block had incomplete UTF-8 characters at the end, an
-    // ASCII block can't possibly finish them.
-    this->error |= this->prev_incomplete;
-  }
+    // bitmask = 0x0000 if there are no surrogates
+    //         = 0xc000 if the last word is a surrogate
+    const uint16_t surrogates_bitmask = surrogates_bytemask.to_bitmask();
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x0000) {
+      ppc64_convert_utf16_to_1_2_3_bytes_of_utf8(
+          in, one_byte_bitmask, one_or_two_bytes_bytemask,
+          one_or_two_bytes_bitmask, utf8_output);
 
-  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
-    if (simdutf_likely(is_ascii(input))) {
-      this->error |= this->prev_incomplete;
+      buf += 8;
+      // surrogate pair(s) in a register
     } else {
-      // you might think that a for-loop would work, but under Visual Studio, it
-      // is not good enough.
-      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-                    "We support either two or four chunks per 64-byte block.");
-      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
       }
-      this->prev_incomplete =
-          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
-      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
+      for (; k < forward; k++) {
+        uint16_t word = not match_system(big_endian)
+                            ? scalar::u16_swap_bytes(buf[k])
+                            : buf[k];
+        if ((word & 0xFF80) == 0) {
+          *utf8_output++ = uint8_t(word);
+        } else if ((word & 0xF800) == 0) {
+          *utf8_output++ = uint8_t((word >> 6) | 0b11000000);
+          *utf8_output++ = uint8_t((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xF800) != 0xD800) {
+          *utf8_output++ = uint8_t((word >> 12) | 0b11100000);
+          *utf8_output++ = uint8_t(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = uint8_t((word & 0b111111) | 0b10000000);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = not match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return utf16_to_utf8_t{error_code::SURROGATE, buf + k - 1,
+                                   utf8_output};
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf8_output++ = uint8_t((value >> 18) | 0b11110000);
+          *utf8_output++ = uint8_t(((value >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = uint8_t(((value >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = uint8_t((value & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
     }
-  }
-
-  // do not forget to call check_eof!
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
-  }
-
-}; // struct utf8_checker
-} // namespace utf8_validation
+  } // while
 
-using utf8_validation::utf8_checker;
+  return utf16_to_utf8_t{error_code::SUCCESS, buf, utf8_output};
+}
+/* end file src/ppc64/ppc64_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
 
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
-/* begin file src/generic/utf8_validation/utf8_validator.h */
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_validation {
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+/* begin file src/ppc64/ppc64_convert_utf16_to_utf32.cpp */
+struct utf16_to_utf32_t {
+  error_code err;        // error code
+  const char16_t *input; // last position in input buffer
+  char32_t *output;      // last position in output buffer
+};
 
-/**
- * Validates that the string is actual UTF-8.
- */
-template <class checker>
-bool generic_validate_utf8(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    reader.advance();
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  return !c.errors();
-}
+template <endianness big_endian>
+utf16_to_utf32_t ppc64_convert_utf16_to_utf32(const char16_t *buf, size_t len,
+                                              char32_t *utf32_output) {
+  const char16_t *end = buf + len;
 
-bool generic_validate_utf8(const char *input, size_t length) {
-  return generic_validate_utf8<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+  const auto v_f800 = vector_u16::splat(0xf800);
+  const auto v_d800 = vector_u16::splat(0xd800);
+  const auto zero = vector_u8::zero();
 
-/**
- * Validates that the string is actual UTF-8 and stops on errors.
- */
-template <class checker>
-result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  size_t count{0};
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    if (c.errors()) {
-      if (count != 0) {
-        count--;
-      } // Sometimes the error is only detected in the next chunk
-      result res = scalar::utf8::rewind_and_validate_with_errors(
-          reinterpret_cast<const char *>(input),
-          reinterpret_cast<const char *>(input + count), length - count);
-      res.count += count;
-      return res;
+  while (end - buf >= vector_u16::ELEMENTS) {
+    auto in = vector_u16::load(buf);
+    if (not match_system(big_endian)) {
+      in = in.swap_bytes();
     }
-    reader.advance();
-    count += 64;
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  if (c.errors()) {
-    if (count != 0) {
-      count--;
-    } // Sometimes the error is only detected in the next chunk
-    result res = scalar::utf8::rewind_and_validate_with_errors(
-        reinterpret_cast<const char *>(input),
-        reinterpret_cast<const char *>(input) + count, length - count);
-    res.count += count;
-    return res;
-  } else {
-    return result(error_code::SUCCESS, length);
-  }
-}
 
-result generic_validate_utf8_with_errors(const char *input, size_t length) {
-  return generic_validate_utf8_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+    // 1. Check if there are any surrogate word in the input chunk.
+    //    We have also deal with situation when there is a surrogate word
+    //    at the end of a chunk.
+    const auto surrogates_bytemask = (in & v_f800) == v_d800;
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
-  uint8_t blocks[64]{};
-  simd::simd8x64<uint8_t> running_or(blocks);
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    running_or |= in;
-    reader.advance();
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  running_or |= in;
-  return running_or.is_ascii();
-}
+    // bitmask = 0x0000 if there are no surrogates
+    const uint16_t surrogates_bitmask = surrogates_bytemask.to_bitmask();
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+    // It might seem like checking for surrogates_bitmask == 0xc000 could help.
+    // However, it is likely an uncommon occurrence.
+    if (surrogates_bitmask == 0x0000) {
+      // case: no surrogate pairs, extend 16-bit code units to 32-bit code units
+#if SIMDUTF_IS_BIG_ENDIAN
+      const auto lo =
+          vector_u8(16, 16, 0, 1, 16, 16, 2, 3, 16, 16, 4, 5, 16, 16, 6, 7);
+      const auto hi = vector_u8(16, 16, 8 + 0, 8 + 1, 16, 16, 8 + 2, 8 + 3, 16,
+                                16, 8 + 4, 8 + 5, 16, 16, 8 + 6, 8 + 7);
+#else
+      const auto lo =
+          vector_u8(0, 1, 16, 16, 2, 3, 16, 16, 4, 5, 16, 16, 6, 7, 16, 16);
+      const auto hi = vector_u8(8 + 0, 8 + 1, 16, 16, 8 + 2, 8 + 3, 16, 16,
+                                8 + 4, 8 + 5, 16, 16, 8 + 6, 8 + 7, 16, 16);
+#endif // SIMDUTF_IS_BIG_ENDIAN
 
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
-  size_t count{0};
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    if (!in.is_ascii()) {
-      result res = scalar::ascii::validate_with_errors(
-          reinterpret_cast<const char *>(input + count), length - count);
-      return result(res.error, count + res.count);
+      const auto utf32_0 = lo.lookup_32(as_vector_u8(in), zero);
+      const auto utf32_1 = hi.lookup_32(as_vector_u8(in), zero);
+
+      utf32_0.store(utf32_output);
+      utf32_1.store(utf32_output + 4);
+      utf32_output += 8;
+      buf += 8;
+      // surrogate pair(s) in a register
+    } else {
+      // Let us do a scalar fallback.
+      // It may seem wasteful to use scalar code, but being efficient with SIMD
+      // in the presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        const uint16_t word = not match_system(big_endian)
+                                  ? scalar::u16_swap_bytes(buf[k])
+                                  : buf[k];
+        if ((word & 0xF800) != 0xD800) {
+          *utf32_output++ = char32_t(word);
+        } else {
+          // must be a surrogate pair
+          uint16_t diff = uint16_t(word - 0xD800);
+          uint16_t next_word = not match_system(big_endian)
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
+                                   : buf[k + 1];
+          k++;
+          uint16_t diff2 = uint16_t(next_word - 0xDC00);
+          if ((diff | diff2) > 0x3FF) {
+            return utf16_to_utf32_t{error_code::SURROGATE, buf + k - 1,
+                                    utf32_output};
+          }
+          uint32_t value = (diff << 10) + diff2 + 0x10000;
+          *utf32_output++ = char32_t(value);
+        }
+      }
+      buf += k;
     }
-    reader.advance();
+  } // while
 
-    count += 64;
-  }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  if (!in.is_ascii()) {
-    result res = scalar::ascii::validate_with_errors(
-        reinterpret_cast<const char *>(input + count), length - count);
-    return result(res.error, count + res.count);
-  } else {
-    return result(error_code::SUCCESS, length);
-  }
+  return utf16_to_utf32_t{error_code::SUCCESS, buf, utf32_output};
 }
+/* end file src/ppc64/ppc64_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/ppc64/ppc64_convert_utf32_to_latin1.cpp */
+enum class ErrorChecking { disabled, enabled };
 
-} // namespace utf8_validation
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
-// transcoding from UTF-8 to UTF-16
-/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+struct utf32_to_latin1_t {
+  error_code err;
+  const char32_t *input;
+  char *output;
+};
 
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_to_utf16 {
+template <ErrorChecking ec>
+utf32_to_latin1_t simdutf_really_inline ppc64_convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) {
+  constexpr size_t N = vector_u32::ELEMENTS;
+  const size_t rounded_len = align_down<4 * N>(len);
 
-using namespace simd;
+  const auto high_bytes_mask = vector_u32::splat(0xFFFFFF00);
 
-template <endianness endian>
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char16_t *utf16_output) noexcept {
-  // The implementation is not specific to haswell and should be moved to the
-  // generic directory.
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  const size_t safety_margin = 16; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    // this loop could be unrolled further. For example, we could process the
-    // mask far more than 64 bytes.
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf16<endian>(utf16_output);
-      utf16_output += 64;
-      pos += 64;
-    } else {
-      // Slow path. We hope that the compiler will recognize that this is a slow
-      // path. Anything that is not a continuation mask is a 'leading byte',
-      // that is, the start of a new code point.
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      // The *start* of code points is not so useful, rather, we want the *end*
-      // of code points.
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      // We process in blocks of up to 12 bytes except possibly
-      // for fast paths which may process up to 16 bytes. For the
-      // slow path to work, we should have at least 12 input bytes left.
-      size_t max_starting_point = (pos + 64) - 12;
-      // Next loop is going to run at least five times when using solely
-      // the slow/regular path, and at least four times if there are fast paths.
-      while (pos < max_starting_point) {
-        // Performance note: our ability to compute 'consumed' and
-        // then shift and recompute is critical. If there is a
-        // latency of, say, 4 cycles on getting 'consumed', then
-        // the inner loop might have a total latency of about 6 cycles.
-        // Yet we process between 6 to 12 inputs bytes, thus we get
-        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-        // for this section of the code. Hence, there is a limit
-        // to how much we can further increase this latency before
-        // it seriously harms performance.
-        //
-        // Thus we may allow convert_masked_utf8_to_utf16 to process
-        // more bytes at a time under a fast-path mode where 16 bytes
-        // are consumed at once (e.g., when encountering ASCII).
-        size_t consumed = convert_masked_utf8_to_utf16<endian>(
-            input + pos, utf8_end_of_code_point_mask, utf16_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
+  for (size_t i = 0; i < rounded_len; i += 4 * N) {
+    const auto in1 = vector_u32::load(buf + 0 * N);
+    const auto in2 = vector_u32::load(buf + 1 * N);
+    const auto in3 = vector_u32::load(buf + 2 * N);
+    const auto in4 = vector_u32::load(buf + 3 * N);
+
+    if (ec == ErrorChecking::enabled) {
+      const auto combined = in1 | in2 | in3 | in4;
+      const auto too_big = (combined & high_bytes_mask) != uint32_t(0);
+
+      if (simdutf_unlikely(too_big.any())) {
+        // Scalar code will carry on from the beginning of the current block
+        // and report the exact error position.
+        return utf32_to_latin1_t{error_code::OTHER, buf, latin1_output};
       }
-      // At this point there may remain between 0 and 12 bytes in the
-      // 64-byte block. These bytes will be processed again. So we have an
-      // 80% efficiency (in the worst case). In practice we expect an
-      // 85% to 90% efficiency.
     }
-  }
-  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
-      input + pos, size - pos, utf16_output);
-  return utf16_output - start;
-}
 
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
-/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+    // Note: element #1 contains 0, and is used to mask-out elements
+#if SIMDUTF_IS_BIG_ENDIAN
+    const auto shlo = vector_u8(0 + 3, 4 + 3, 8 + 3, 12 + 3, 16 + 3, 20 + 3,
+                                24 + 3, 28 + 3, 1, 1, 1, 1, 1, 1, 1, 1);
+    const auto shhi = vector_u8(1, 1, 1, 1, 1, 1, 1, 1, 0 + 3, 4 + 3, 8 + 3,
+                                12 + 3, 16 + 3, 20 + 3, 24 + 3, 28 + 3);
+#else
+    const auto shlo =
+        vector_u8(0, 4, 8, 12, 16, 20, 24, 28, 1, 1, 1, 1, 1, 1, 1, 1);
+    const auto shhi =
+        vector_u8(1, 1, 1, 1, 1, 1, 1, 1, 0, 4, 8, 12, 16, 20, 24, 28);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+    const auto lo = shlo.lookup_32(as_vector_u8(in1), as_vector_u8(in2));
+    const auto hi = shhi.lookup_32(as_vector_u8(in3), as_vector_u8(in4));
 
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_to_utf16 {
-using namespace simd;
+    const auto merged = lo | hi;
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+    merged.store(latin1_output);
+    latin1_output += 4 * N;
+    buf += 4 * N;
+  }
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+  return utf32_to_latin1_t{error_code::SUCCESS, buf, latin1_output};
+}
+/* end file src/ppc64/ppc64_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF16
+/* begin file src/ppc64/ppc64_convert_utf32_to_utf16.cpp */
+struct utf32_to_utf16_t {
+  error_code err;
+  const char32_t *input;
+  char16_t *output;
+};
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+template <endianness big_endian, ErrorReporting er>
+utf32_to_utf16_t ppc64_convert_utf32_to_utf16(const char32_t *buf, size_t len,
+                                              char16_t *utf16_output) {
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+  const char32_t *end = buf + len;
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
+  const auto zero = vector_u32::zero();
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
 
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+  auto forbidden_global = simd16<bool>();
 
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
-  }
+  while (end - buf >= 8) {
+    const auto in0 = vector_u32::load(buf);
+    const auto in1 = vector_u32::load(buf + vector_u32::ELEMENTS);
 
-  template <endianness endian>
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (utf8_continuation_mask & 1) {
-          return 0; // error
+    const auto any_surrogate = ((in0 | in1) & v_ffff0000) != zero;
+
+    // Check if no bits set above 15th
+    if (any_surrogate.is_zero()) {
+      // Pack UTF-32 to UTF-16
+#if SIMDUTF_IS_BIG_ENDIAN
+      const auto sh = big_endian ? vector_u8(2, 3, 6, 7, 10, 11, 14, 15, 18, 19,
+                                             22, 23, 26, 27, 30, 31)
+                                 : vector_u8(3, 2, 7, 6, 11, 10, 15, 14, 19, 18,
+                                             23, 22, 27, 26, 31, 30);
+#else
+      const auto sh = big_endian ? vector_u8(1, 0, 5, 4, 9, 8, 13, 12, 17, 16,
+                                             21, 20, 25, 24, 29, 28)
+                                 : vector_u8(0, 1, 4, 5, 8, 9, 12, 13, 16, 17,
+                                             20, 21, 24, 25, 28, 29);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+      const auto packed0 = sh.lookup_32(as_vector_u8(in0), as_vector_u8(in1));
+      const auto packed = as_vector_u16(packed0);
+
+#if SIMDUTF_IS_BIG_ENDIAN
+      const auto v_f800 =
+          big_endian ? vector_u16::splat(0xf800) : vector_u16::splat(0x00f8);
+      const auto v_d800 =
+          big_endian ? vector_u16::splat(0xd800) : vector_u16::splat(0x00d8);
+#else
+      const auto v_f800 =
+          big_endian ? vector_u16::splat(0x00f8) : vector_u16::splat(0xf800);
+      const auto v_d800 =
+          big_endian ? vector_u16::splat(0x00d8) : vector_u16::splat(0xd800);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+      const auto forbidden = (packed & v_f800) == v_d800;
+
+      switch (er) {
+      case ErrorReporting::precise:
+        if (not forbidden.is_zero()) {
+          // scalar procedure will rescan the portion of buffer we've just
+          // analysed
+          return utf32_to_utf16_t{error_code::OTHER, buf, utf16_output};
         }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
+        break;
+      case ErrorReporting::at_the_end:
+        forbidden_global |= forbidden;
+        break;
+      case ErrorReporting::none:
+        break;
+      }
+
+      packed.store(utf16_output);
+      utf16_output += 8;
+      buf += 8;
+    } else {
+      size_t forward = 7;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
+      }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFF0000) == 0) {
+          // will not generate a surrogate pair
+          if (word >= 0xD800 && word <= 0xDFFF) {
+            return utf32_to_utf16_t{error_code::SURROGATE, buf + k,
+                                    utf16_output};
+          }
+          *utf16_output++ = not match_system(big_endian)
+                                ? scalar::u16_swap_bytes(uint16_t(word))
+                                : uint16_t(word);
+        } else {
+          // will generate a surrogate pair
+          if (word > 0x10FFFF) {
+            return utf32_to_utf16_t{error_code::TOO_LARGE, buf + k,
+                                    utf16_output};
+          }
+          word -= 0x10000;
+          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
+          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
+          if (not match_system(big_endian)) {
+            high_surrogate = scalar::u16_swap_bytes(high_surrogate);
+            low_surrogate = scalar::u16_swap_bytes(low_surrogate);
+          }
+          *utf16_output++ = char16_t(high_surrogate);
+          *utf16_output++ = char16_t(low_surrogate);
         }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
       }
+      buf += k;
     }
-    if (errors()) {
-      return 0;
+  }
+
+  if (er == ErrorReporting::at_the_end) {
+    // check for invalid input
+    if (not forbidden_global.is_zero()) {
+      return utf32_to_utf16_t{error_code::SURROGATE, buf, utf16_output};
     }
-    if (pos < size) {
-      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
-          in + pos, size - pos, utf16_output);
-      if (howmany == 0) {
-        return 0;
+  }
+
+  return utf32_to_utf16_t{error_code::SUCCESS, buf, utf16_output};
+}
+/* end file src/ppc64/ppc64_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF32
+/* begin file src/ppc64/ppc64_convert_utf32_to_utf8.cpp */
+struct utf32_to_utf8_t {
+  error_code err;
+  const char32_t *input;
+  char *output;
+};
+
+template <ErrorReporting er>
+utf32_to_utf8_t ppc64_convert_utf32_to_utf8(const char32_t *buf, size_t len,
+                                            char *utf8_output) {
+  const char32_t *end = buf + len;
+
+  const auto v_f800 = vector_u16::splat(0xf800);
+  const auto v_d800 = vector_u16::splat(0xd800);
+
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto v_00000000 = vector_u32::zero();
+  auto forbidden_bytemask = simd16<bool>();
+  const size_t safety_margin =
+      12; // to avoid overruns, see issue
+          // https://github.com/simdutf/simdutf/issues/92
+
+  while (end - buf >=
+         std::ptrdiff_t(
+             16 + safety_margin)) { // buf is a char32_t pointer, each char32_t
+                                    // has 4 bytes or 32 bits, thus buf + 16 *
+                                    // char_32t = 512 bits = 64 bytes
+    // We load two 16 bytes registers for a total of 32 bytes or 16 characters.
+    // These two values can hold only 8 UTF32 chars
+    auto in0 = vector_u32::load(buf);
+    auto in1 = vector_u32::load(buf + vector_u32::ELEMENTS);
+
+    // Pack 32-bit UTF-32 code units to 16-bit UTF-16 code units with unsigned
+    // saturation
+    auto in = vector_u32::pack(in0, in1);
+
+    // Try to apply UTF-16 => UTF-8 from ./ppc64_convert_utf16_to_utf8.cpp
+
+    // Check for ASCII fast path
+
+    // ASCII fast path!!!!
+    // We eagerly load another 32 bytes, hoping that they will be ASCII too.
+    // The intuition is that we try to collect 16 ASCII characters which
+    // requires a total of 64 bytes of input. If we fail, we just pass thirdin
+    // and fourthin as our new inputs.
+    if (in.is_ascii()) { // if the first two blocks are ASCII
+      const auto in2 = vector_u32::load(buf + 2 * vector_u32::ELEMENTS);
+      const auto in3 = vector_u32::load(buf + 3 * vector_u32::ELEMENTS);
+
+      const auto next = vector_u32::pack(in2, in3);
+      if (next.is_ascii()) {
+        // 1. pack the bytes
+        const auto utf8_packed = vector_u16::pack(in, next);
+        // 2. store (16 bytes)
+        utf8_packed.store(utf8_output);
+        // 3. adjust pointers
+        buf += 16;
+        utf8_output += 16;
+        continue; // we are done for this round!
       }
-      utf16_output += howmany;
+
+      // `next` is not ASCII, write `in` and carry on with next
+
+      // 1. pack the bytes
+      const auto utf8_packed = vector_u16::pack(in, in);
+      utf8_packed.store(utf8_output);
+      // 3. adjust pointers
+      buf += 8;
+      utf8_output += 8;
+
+      // Proceed with next input
+      in = next;
+      in0 = in2;
+      in1 = in3;
     }
-    return utf16_output - start;
-  }
 
-  template <endianness endian>
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
+    // no bits set above 7th bit
+    const auto one_byte_bytemask = in < uint16_t(1 << 7);
+    const uint16_t one_byte_bitmask = one_byte_bytemask.to_bitmask();
+
+    // no bits set above 11th bit
+    const auto one_or_two_bytes_bytemask = in < uint16_t(1 << 11);
+    const uint16_t one_or_two_bytes_bitmask =
+        one_or_two_bytes_bytemask.to_bitmask();
+
+    if (one_or_two_bytes_bitmask == 0xffff) {
+      write_v_u16_11bits_to_utf8(
+          in, utf8_output, as_vector_u8(one_byte_bytemask), one_byte_bitmask);
+      buf += 8;
+      continue;
     }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          // rewind_and_convert_with_errors will seek a potential error from
-          // in+pos onward, with the ability to go back up to pos bytes, and
-          // read size-pos bytes forward.
-          result res =
-              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-                  pos, in + pos, size - pos, utf16_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
+
+    // Check for overflow in packing
+    const auto saturation_bytemask = ((in0 | in1) & v_ffff0000) == v_00000000;
+    const uint16_t saturation_bitmask = saturation_bytemask.to_bitmask();
+    if (saturation_bitmask == 0xffff) {
+      switch (er) {
+      case ErrorReporting::precise: {
+        const auto forbidden = (in & v_f800) == v_d800;
+        if (forbidden.any()) {
+          // We return no error code, instead we force the scalar procedure
+          // to rescan the portion of input where we've just found an error.
+          return utf32_to_utf8_t{error_code::SUCCESS, buf, utf8_output};
         }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
+      } break;
+      case ErrorReporting::at_the_end:
+        forbidden_bytemask |= (in & v_f800) == v_d800;
+        break;
+      case ErrorReporting::none:
+        break;
       }
-    }
-    if (errors()) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      res.count += pos;
-      return res;
-    }
-    if (pos < size) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        utf16_output += res.count;
+
+      ppc64_convert_utf16_to_1_2_3_bytes_of_utf8(
+          in, one_byte_bitmask, one_or_two_bytes_bytemask,
+          one_or_two_bytes_bitmask, utf8_output);
+      buf += 8;
+    } else {
+      // case: at least one 32-bit word produce a surrogate pair in UTF-16 <=>
+      // will produce four UTF-8 bytes Let us do a scalar fallback. It may seem
+      // wasteful to use scalar code, but being efficient with SIMD in the
+      // presence of surrogate pairs may require non-trivial tables.
+      size_t forward = 15;
+      size_t k = 0;
+      if (size_t(end - buf) < forward + 1) {
+        forward = size_t(end - buf - 1);
       }
+      for (; k < forward; k++) {
+        uint32_t word = buf[k];
+        if ((word & 0xFFFFFF80) == 0) {
+          *utf8_output++ = char(word);
+        } else if ((word & 0xFFFFF800) == 0) {
+          *utf8_output++ = char((word >> 6) | 0b11000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else if ((word & 0xFFFF0000) == 0) {
+          if (er != ErrorReporting::none and
+              (word >= 0xD800 && word <= 0xDFFF)) {
+            return utf32_to_utf8_t{error_code::SURROGATE, buf + k, utf8_output};
+          }
+          *utf8_output++ = char((word >> 12) | 0b11100000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        } else {
+          if (er != ErrorReporting::none and (word > 0x10FFFF)) {
+            return utf32_to_utf8_t{error_code::TOO_LARGE, buf + k, utf8_output};
+          }
+          *utf8_output++ = char((word >> 18) | 0b11110000);
+          *utf8_output++ = char(((word >> 12) & 0b111111) | 0b10000000);
+          *utf8_output++ = char(((word >> 6) & 0b111111) | 0b10000000);
+          *utf8_output++ = char((word & 0b111111) | 0b10000000);
+        }
+      }
+      buf += k;
     }
-    return result(error_code::SUCCESS, utf16_output - start);
-  }
+  } // while
 
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+  if (er == ErrorReporting::at_the_end) {
+    if (forbidden_bytemask.any()) {
+      return utf32_to_utf8_t{error_code::SURROGATE, buf, utf8_output};
+    }
   }
 
-}; // struct utf8_checker
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
-/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+  return utf32_to_utf8_t{
+      error_code::SUCCESS,
+      buf,
+      utf8_output,
+  };
+}
+/* end file src/ppc64/ppc64_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_UTF32
 
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_to_utf32 {
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/ppc64/ppc64_utf8_length_from_latin1.cpp */
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
 
-using namespace simd;
+std::pair<const char *, size_t> ppc64_utf8_length_from_latin1(const char *input,
+                                                              size_t length) {
+  constexpr size_t N = vector_u8::ELEMENTS;
+  length = (length / N);
 
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char32_t *utf32_output) noexcept {
-  size_t pos = 0;
-  char32_t *start{utf32_output};
-  const size_t safety_margin = 16; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf32(utf32_output);
-      utf32_output += 64;
-      pos += 64;
-    } else {
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      size_t max_starting_point = (pos + 64) - 12;
-      while (pos < max_starting_point) {
-        size_t consumed = convert_masked_utf8_to_utf32(
-            input + pos, utf8_end_of_code_point_mask, utf32_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
+  size_t count = length * N;
+  while (length != 0) {
+    vector_u32 partial = vector_u32::zero();
+
+    // partial accumulator has 32 bits => this yields (2^31 / 16)
+    size_t chunk = min(length, size_t(0xffffffff / N));
+    length -= chunk;
+    while (chunk != 0) {
+      auto local = vector_u8::zero();
+      // local accumulator has 8 bits => this yields 255 max (we increment by 1
+      // in each iteration)
+      const size_t n = min(chunk, size_t(255));
+      chunk -= n;
+      for (size_t i = 0; i < n; i++) {
+        const auto in = vector_i8::load(input);
+        input += N;
+
+        local -= as_vector_u8(in < vector_i8::splat(0));
       }
+
+      partial = sum4bytes(local, partial);
+    }
+
+    for (int i = 0; i < vector_u32::ELEMENTS; i++) {
+      count += size_t(partial.value[i]);
     }
   }
-  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
-                                                       utf32_output);
-  return utf32_output - start;
+
+  return std::make_pair(input, count);
 }
+/* end file src/ppc64/ppc64_utf8_length_from_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-} // namespace utf8_to_utf32
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/ppc64/ppc64_base64.cpp */
+/*
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ *
+ * AMD XOP specific: http://0x80.pl/notesen/2016-01-12-sse-base64-encoding.html
+ * Altivec has capabilites of AMD XOP (or vice versa): shuffle using 2 vectors
+ * and variable shifts, thus this implementation shares some code solution
+ * (modulo intrisic function names).
+ */
 
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_to_utf32 {
-using namespace simd;
+constexpr bool with_base64_std = false;
+constexpr bool with_base64_url = true;
+constexpr bool with_ignore_errors = true;
+constexpr bool with_ignore_garbage = true;
+constexpr bool with_strict_checking = false;
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+// --- encoding -----------------------------------------------
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+/*
+    Procedure translates vector of bytes having 6-bit values
+    into ASCII counterparts.
+*/
+template <bool base64_url>
+vector_u8 encoding_translate_6bit_values(const vector_u8 input) {
+  // credit: Wojciech Muła
+  // reduce  0..51 -> 0
+  //        52..61 -> 1 .. 10
+  //            62 -> 11
+  //            63 -> 12
+  auto result = input.saturating_sub(vector_u8::splat(51));
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+  // distinguish between ranges 0..25 and 26..51:
+  //         0 .. 25 -> remains 13
+  //        26 .. 51 -> becomes 0
+  const auto lt = input < vector_u8::splat(26);
+  result = select(as_vector_u8(lt), vector_u8::splat(13), result);
+
+  const auto shift_LUT =
+      base64_url ? vector_u8('a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+                             '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+                             '0' - 52, '-' - 62, '_' - 63, 'A', 0, 0)
+                 : vector_u8('a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+                             '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
+                             '0' - 52, '+' - 62, '/' - 63, 'A', 0, 0);
+  // read shift
+  result = result.lookup_16(shift_LUT);
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+  return input + result;
+}
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+/*
+   Procedure expands 12 bytes (4*3 bytes) into 16 bytes,
+   each byte stores 6 bits of data
+*/
+template <typename = void>
+simdutf_really_inline vector_u8 encoding_expand_6bit_fields(vector_u8 input) {
+#if SIMDUTF_IS_BIG_ENDIAN
+  #define indices4(dx) (dx + 0), (dx + 1), (dx + 1), (dx + 2)
+  const auto expand_3_to_4 = vector_u8(indices4(0 * 3), indices4(1 * 3),
+                                       indices4(2 * 3), indices4(3 * 3));
+  #undef indices4
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
+  // input = [........|ccdddddd|bbbbcccc|aaaaaabb] as uint8_t
+  //              3        2        1        0
+  //
+  // in'   = [aaaaaabb|bbbbcccc|bbbbcccc|ccdddddd] as uint32_t
+  //              0        1        1        2
+  const auto in = as_vector_u32(expand_3_to_4.lookup_16(input));
+
+  // t0    = [00000000|00000000|00000000|00dddddd]
+  const auto t0 = in & uint32_t(0x0000003f);
+
+  // t1    = [00000000|00000000|00cccccc|00dddddd]
+  const auto t1 = select(uint32_t(0x00003f00), in.shl<2>(), t0);
+
+  // t2    = [00000000|00bbbbbb|00cccccc|00dddddd]
+  const auto t2 = select(uint32_t(0x003f0000), in.shr<4>(), t1);
+
+  // t3    = [00aaaaaa|00bbbbbb|00cccccc|00dddddd]
+  const auto t3 = select(uint32_t(0x3f000000), in.shr<2>(), t2);
+
+  return as_vector_u8(t3);
+#else
+  #define indices4(dx) (dx + 1), (dx + 0), (dx + 2), (dx + 1)
+  const auto expand_3_to_4 = vector_u8(indices4(0 * 3), indices4(1 * 3),
+                                       indices4(2 * 3), indices4(3 * 3));
+  #undef indices4
+
+  // input = [........|ccdddddd|bbbbcccc|aaaaaabb] as uint8_t
+  //              3        2        1        0
+  //
+  // in'   = [bbbbcccc|ccdddddd|aaaaaabb|bbbbcccc] as uint32_t
+  //              1        2        0        1
+  const auto in = as_vector_u32(expand_3_to_4.lookup_16(input));
+
+  // t0    = [00dddddd|00000000|00000000|00000000]
+  const auto t0 = in.shl<8>() & uint32_t(0x3f000000);
+
+  // t1    = [00dddddd|00cccccc|00000000|00000000]
+  const auto t1 = select(uint32_t(0x003f0000), in.shr<6>(), t0);
+
+  // t2    = [00dddddd|00cccccc|00bbbbbb|00000000]
+  const auto t2 = select(uint32_t(0x00003f00), in.shl<4>(), t1);
+
+  // t3    = [00dddddd|00cccccc|00bbbbbb|00aaaaaa]
+  const auto t3 = select(uint32_t(0x0000003f), in.shr<10>(), t2);
+
+  return as_vector_u8(t3);
+#endif // SIMDUTF_IS_BIG_ENDIAN
 }
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
+
+template <bool isbase64url>
+size_t encode_base64(char *dst, const char *src, size_t srclen,
+                     base64_options options) {
+
+  const uint8_t *input = (const uint8_t *)src;
+
+  uint8_t *out = (uint8_t *)dst;
+
+  size_t i = 0;
+  for (; i + 52 <= srclen; i += 48) {
+    const auto in0 = vector_u8::load(input + i + 12 * 0);
+    const auto in1 = vector_u8::load(input + i + 12 * 1);
+    const auto in2 = vector_u8::load(input + i + 12 * 2);
+    const auto in3 = vector_u8::load(input + i + 12 * 3);
+
+    const auto expanded0 = encoding_expand_6bit_fields(in0);
+    const auto expanded1 = encoding_expand_6bit_fields(in1);
+    const auto expanded2 = encoding_expand_6bit_fields(in2);
+    const auto expanded3 = encoding_expand_6bit_fields(in3);
+
+    const auto base64_0 =
+        encoding_translate_6bit_values<isbase64url>(expanded0);
+    const auto base64_1 =
+        encoding_translate_6bit_values<isbase64url>(expanded1);
+    const auto base64_2 =
+        encoding_translate_6bit_values<isbase64url>(expanded2);
+    const auto base64_3 =
+        encoding_translate_6bit_values<isbase64url>(expanded3);
+
+    base64_0.store(out);
+    out += 16;
+
+    base64_1.store(out);
+    out += 16;
+
+    base64_2.store(out);
+    out += 16;
+
+    base64_3.store(out);
+    out += 16;
+  }
+  for (; i + 16 <= srclen; i += 12) {
+    const auto in = vector_u8::load(input + i);
+    const auto expanded = encoding_expand_6bit_fields(in);
+    const auto base64 = encoding_translate_6bit_values<isbase64url>(expanded);
+
+    base64.store(out);
+    out += 16;
+  }
+
+  return i / 3 * 4 + scalar::base64::tail_encode_base64((char *)out, src + i,
+                                                        srclen - i, options);
 }
 
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
+// --- decoding -----------------------------------------------
 
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
+static simdutf_really_inline void compress(const vector_u8 data, uint16_t mask,
+                                           char *output) {
+  if (mask == 0) {
+    data.store(output);
+    return;
   }
 
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char32_t *utf32_output) {
-    size_t pos = 0;
-    char32_t *start{utf32_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 16 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (utf8_continuation_mask & 1) {
-          return 0; // we have an error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      return 0;
-    }
-    if (pos < size) {
-      size_t howmany =
-          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      utf32_output += howmany;
-    }
-    return utf32_output - start;
+  // this particular implementation was inspired by work done by @animetosho
+  // we do it in two steps, first 8 bytes and then second 8 bytes
+  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+  // next line just loads the 64-bit values thintable_epi8[mask1] and
+  // thintable_epi8[mask2] into a 128-bit register, using only
+  // two instructions on most compilers.
+
+#if SIMDUTF_IS_BIG_ENDIAN
+  vec_u64_t tmp = {
+      tables::base64::thintable_epi8[mask2],
+      tables::base64::thintable_epi8[mask1],
+  };
+
+  auto shufmask = vector_u8(vec_reve(vec_u8_t(tmp)));
+
+  // we increment by 0x08 the second half of the mask
+  shufmask =
+      shufmask + vector_u8(0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8);
+#else
+  vec_u64_t tmp = {
+      tables::base64::thintable_epi8[mask1],
+      tables::base64::thintable_epi8[mask2],
+  };
+
+  auto shufmask = vector_u8(vec_u8_t(tmp));
+
+  // we increment by 0x08 the second half of the mask
+  shufmask =
+      shufmask + vector_u8(0, 0, 0, 0, 0, 0, 0, 0, 8, 8, 8, 8, 8, 8, 8, 8);
+#endif // SIMDUTF_IS_BIG_ENDIAN
+
+  // this is the version "nearly pruned"
+  const auto pruned = shufmask.lookup_16(data);
+  // we still need to put the two halves together.
+  // we compute the popcount of the first half:
+  const int pop1 = tables::base64::BitsSetTable256mul2[mask1];
+  // then load the corresponding mask, what it does is to write
+  // only the first pop1 bytes from the first 8 bytes, and then
+  // it fills in with the bytes from the second 8 bytes + some filling
+  // at the end.
+  const auto compactmask =
+      vector_u8::load(tables::base64::pshufb_combine_table + pop1 * 8);
+
+  const auto answer = compactmask.lookup_16(pruned);
+
+  answer.store(output);
+}
+
+static simdutf_really_inline vector_u8 decoding_pack(vector_u8 input) {
+#if SIMDUTF_IS_BIG_ENDIAN
+  // in   = [00aaaaaa|00bbbbbb|00cccccc|00dddddd]
+  // want = [00000000|aaaaaabb|bbbbcccc|ccdddddd]
+
+  auto in = as_vector_u16(input);
+  // t0   = [00??aaaa|aabbbbbb|00??cccc|ccdddddd]
+  const auto t0 = in.shr<2>();
+  const auto t1 = select(uint16_t(0x0fc0), t0, in);
+
+  // t0   = [00??????|aaaaaabb|bbbbcccc|ccdddddd]
+  const auto t2 = as_vector_u32(t1);
+  const auto t3 = t2.shr<4>();
+  const auto t4 = select(uint32_t(0x00fff000), t3, t2);
+
+  const auto tmp = as_vector_u8(t4);
+
+  const auto shuffle =
+      vector_u8(1, 2, 3, 5, 6, 7, 9, 10, 11, 13, 14, 15, 0, 0, 0, 0);
+
+  const auto t = shuffle.lookup_16(tmp);
+
+  return t;
+#else
+  // in   = [00dddddd|00cccccc|00bbbbbb|00aaaaaa]
+  // want = [00000000|aaaaaabb|bbbbcccc|ccdddddd]
+
+  auto u = as_vector_u32(input).swap_bytes();
+
+  auto in = vector_u16((vec_u16_t)u.value);
+  // t0   = [00??aaaa|aabbbbbb|00??cccc|ccdddddd]
+  const auto t0 = in.shr<2>();
+  const auto t1 = select(uint16_t(0x0fc0), t0, in);
+
+  // t0   = [00??????|aaaaaabb|bbbbcccc|ccdddddd]
+  const auto t2 = as_vector_u32(t1);
+  const auto t3 = t2.shr<4>();
+  const auto t4 = select(uint32_t(0x00fff000), t3, t2);
+
+  const auto tmp = as_vector_u8(t4);
+
+  const auto shuffle =
+      vector_u8(2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, 0, 0, 0, 0);
+
+  const auto t = shuffle.lookup_16(tmp);
+
+  return t;
+#endif // SIMDUTF_IS_BIG_ENDIAN
+}
+static simdutf_really_inline void base64_decode(char *out, vector_u8 input) {
+  const auto expanded = decoding_pack(input);
+  expanded.store(out);
+}
+
+static simdutf_really_inline void base64_decode_block(char *out,
+                                                      const char *src) {
+  base64_decode(out + 12 * 0, vector_u8::load(src + 0 * 16));
+  base64_decode(out + 12 * 1, vector_u8::load(src + 1 * 16));
+  base64_decode(out + 12 * 2, vector_u8::load(src + 2 * 16));
+  base64_decode(out + 12 * 3, vector_u8::load(src + 3 * 16));
+}
+
+static simdutf_really_inline void base64_decode_block_safe(char *out,
+                                                           const char *src) {
+  base64_decode(out + 12 * 0, vector_u8::load(src + 0 * 16));
+  base64_decode(out + 12 * 1, vector_u8::load(src + 1 * 16));
+  base64_decode(out + 12 * 2, vector_u8::load(src + 2 * 16));
+
+  char buffer[16];
+  base64_decode(buffer, vector_u8::load(src + 3 * 16));
+  std::memcpy(out + 36, buffer, 12);
+}
+
+// ---base64 decoding::block64 class --------------------------
+
+class block64 {
+  simd8x64<uint8_t> b;
+
+public:
+  simdutf_really_inline block64(const char *src) : b(load_block(src)) {}
+  simdutf_really_inline block64(const char16_t *src) : b(load_block(src)) {}
+
+private:
+  // The caller of this function is responsible to ensure that there are 64
+  // bytes available from reading at src. The data is read into a block64
+  // structure.
+  static simdutf_really_inline simd8x64<uint8_t> load_block(const char *src) {
+    const auto v0 = vector_u8::load(src + 16 * 0);
+    const auto v1 = vector_u8::load(src + 16 * 1);
+    const auto v2 = vector_u8::load(src + 16 * 2);
+    const auto v3 = vector_u8::load(src + 16 * 3);
+
+    return simd8x64<uint8_t>(v0, v1, v2, v3);
+  }
+
+  // The caller of this function is responsible to ensure that there are 128
+  // bytes available from reading at src. The data is read into a block64
+  // structure.
+  static simdutf_really_inline simd8x64<uint8_t>
+  load_block(const char16_t *src) {
+    const auto m1 = vector_u16::load(src + 8 * 0);
+    const auto m2 = vector_u16::load(src + 8 * 1);
+    const auto m3 = vector_u16::load(src + 8 * 2);
+    const auto m4 = vector_u16::load(src + 8 * 3);
+    const auto m5 = vector_u16::load(src + 8 * 4);
+    const auto m6 = vector_u16::load(src + 8 * 5);
+    const auto m7 = vector_u16::load(src + 8 * 6);
+    const auto m8 = vector_u16::load(src + 8 * 7);
+
+    return simd8x64<uint8_t>(vector_u16::pack(m1, m2), vector_u16::pack(m3, m4),
+                             vector_u16::pack(m5, m6),
+                             vector_u16::pack(m7, m8));
   }
 
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char32_t *utf32_output) {
-    size_t pos = 0;
-    char32_t *start{utf32_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-              pos, in + pos, size - pos, utf32_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
-      res.count += pos;
-      return res;
+public:
+  template <bool base64_url, bool ignore_garbage>
+  static inline uint16_t to_base64_mask(vector_u8 &src, uint16_t &error) {
+    const auto ascii_space_tbl =
+        vector_u8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 0x0,
+                  0xc, 0xd, 0x0, 0x0);
+
+    // credit: aqrit
+    const auto delta_asso =
+        base64_url ? vector_u8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0,
+                               0x0, 0x0, 0x0, 0xF, 0x0, 0xF)
+                   : vector_u8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                               0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F);
+
+    const auto delta_values =
+        base64_url ? vector_u8(0x0, 0x0, 0x0, 0x13, 0x4, 0xBF, 0xBF, 0xB9, 0xB9,
+                               0x0, 0x11, 0xC3, 0xBF, 0xE0, 0xB9, 0xB9)
+                   : vector_u8(0x00, 0x00, 0x00, 0x13, 0x04, 0xBF, 0xBF, 0xB9,
+                               0xB9, 0x00, 0x10, 0xC3, 0xBF, 0xBF, 0xB9, 0xB9);
+
+    const auto check_asso =
+        base64_url ? vector_u8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1,
+                               0x3, 0x7, 0xB, 0xE, 0xB, 0x6)
+                   : vector_u8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                               0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F);
+
+    const auto check_values =
+        base64_url ? vector_u8(0x80, 0x80, 0x80, 0x80, 0xCF, 0xBF, 0xB6, 0xA6,
+                               0xB5, 0xA1, 0x0, 0x80, 0x0, 0x80, 0x0, 0x80)
+                   : vector_u8(0x80, 0x80, 0x80, 0x80, 0xCF, 0xBF, 0xD5, 0xA6,
+                               0xB5, 0x86, 0xD1, 0x80, 0xB1, 0x80, 0x91, 0x80);
+
+    const auto shifted = src.shr<3>();
+
+    const auto delta_hash = avg(src.lookup_16(delta_asso), shifted);
+    const auto check_hash = avg(src.lookup_16(check_asso), shifted);
+
+    const auto out = as_vector_i8(delta_hash.lookup_16(delta_values))
+                         .saturating_add(as_vector_i8(src));
+    const auto chk = as_vector_i8(check_hash.lookup_16(check_values))
+                         .saturating_add(as_vector_i8(src));
+
+    const uint16_t mask = chk.to_bitmask();
+    if (!ignore_garbage && mask) {
+      const auto ascii = src.lookup_16(ascii_space_tbl);
+      const auto ascii_space = (ascii == src);
+      error = (mask ^ ascii_space.to_bitmask());
     }
-    if (pos < size) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        utf32_output += res.count;
-      }
+    src = out;
+
+    return mask;
+  }
+
+  template <bool base64_url, bool ignore_garbage>
+  simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) {
+    uint16_t err0 = 0;
+    uint16_t err1 = 0;
+    uint16_t err2 = 0;
+    uint16_t err3 = 0;
+    uint64_t m0 = to_base64_mask<base64_url, ignore_garbage>(b.chunks[0], err0);
+    uint64_t m1 = to_base64_mask<base64_url, ignore_garbage>(b.chunks[1], err1);
+    uint64_t m2 = to_base64_mask<base64_url, ignore_garbage>(b.chunks[2], err2);
+    uint64_t m3 = to_base64_mask<base64_url, ignore_garbage>(b.chunks[3], err3);
+
+    if (!ignore_garbage) {
+      *error = (err0) | ((uint64_t)err1 << 16) | ((uint64_t)err2 << 32) |
+               ((uint64_t)err3 << 48);
     }
-    return result(error_code::SUCCESS, utf32_output - start);
+    return m0 | (m1 << 16) | (m2 << 32) | (m3 << 48);
   }
 
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+  simdutf_really_inline void copy_block(char *output) {
+    b.store(reinterpret_cast<uint8_t *>(output));
   }
 
-}; // struct utf8_checker
-} // namespace utf8_to_utf32
+  simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) {
+    uint64_t nmask = ~mask;
+    compress(b.chunks[0], uint16_t(mask), output);
+    compress(b.chunks[1], uint16_t(mask >> 16),
+             output + count_ones(nmask & 0xFFFF));
+    compress(b.chunks[2], uint16_t(mask >> 32),
+             output + count_ones(nmask & 0xFFFFFFFF));
+    compress(b.chunks[3], uint16_t(mask >> 48),
+             output + count_ones(nmask & 0xFFFFFFFFFFFFULL));
+    return count_ones(nmask);
+  }
+
+  simdutf_really_inline void base64_decode_block(char *out) {
+    base64_decode(out + 12 * 0, b.chunks[0]);
+    base64_decode(out + 12 * 1, b.chunks[1]);
+    base64_decode(out + 12 * 2, b.chunks[2]);
+    base64_decode(out + 12 * 3, b.chunks[3]);
+  }
+
+  simdutf_really_inline void base64_decode_block_safe(char *out) {
+    base64_decode(out + 12 * 0, b.chunks[0]);
+    base64_decode(out + 12 * 1, b.chunks[1]);
+    base64_decode(out + 12 * 2, b.chunks[2]);
+    char buffer[16];
+    base64_decode(buffer, b.chunks[3]);
+    std::memcpy(out + 12 * 3, buffer, 12);
+  }
+};
+/* end file src/ppc64/ppc64_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
+
 } // unnamed namespace
-} // namespace haswell
+} // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-// other functions
-/* begin file src/generic/utf8.h */
 
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/generic/buf_block_reader.h */
 namespace simdutf {
-namespace haswell {
+namespace ppc64 {
 namespace {
-namespace utf8 {
 
-using namespace simd;
+// Walks through a buffer in block-sized increments, loading the last part with
+// spaces
+template <size_t STEP_SIZE> struct buf_block_reader {
+public:
+  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdutf_really_inline size_t block_index();
+  simdutf_really_inline bool has_full_block() const;
+  simdutf_really_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0
+   * (in which case this function fills the buffer with spaces and returns 0. In
+   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
+   * block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
+  simdutf_really_inline void advance();
 
-simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.gt(-65);
-    count += count_ones(utf8_continuation_mask);
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text_64(const uint8_t *text) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
   }
-  return count + scalar::utf8::count_code_points(in + pos, size - pos);
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
 }
 
-simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-    // We count one word for anything that is not a continuation (so
-    // leading bytes).
-    count += 64 - count_ones(utf8_continuation_mask);
-    int64_t utf8_4byte = input.gteq_unsigned(240);
-    count += count_ones(utf8_4byte);
+// Routines to print masks and text for debugging bitmask operations
+simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
+  static char *buf =
+      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
+  in.store(reinterpret_cast<uint8_t *>(buf));
+  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') {
+      buf[i] = '_';
+    }
   }
-  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
 }
-} // namespace utf8
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8.h */
-/* begin file src/generic/utf16.h */
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf16 {
 
-template <endianness big_endian>
-simdutf_really_inline size_t count_code_points(const char16_t *in,
-                                               size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
-    }
-    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
-    count += count_ones(not_pair) / 2;
+simdutf_unused static char *format_mask(uint64_t mask) {
+  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
+  for (size_t i = 0; i < 64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
   }
-  return count +
-         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
+  buf[64] = '\0';
+  return buf;
 }
 
-template <endianness big_endian>
-simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
-    }
-    uint64_t ascii_mask = input.lteq(0x7F);
-    uint64_t twobyte_mask = input.lteq(0x7FF);
-    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
+template <size_t STEP_SIZE>
+simdutf_really_inline
+buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
+    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
+      idx{0} {}
 
-    size_t ascii_count = count_ones(ascii_mask) / 2;
-    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
-    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
-    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
-    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
-             ascii_count;
-  }
-  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
-                                                                   size - pos);
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
+  return idx;
 }
 
-template <endianness big_endian>
-simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
-                                                     size_t size) {
-  return count_code_points<big_endian>(in, size);
+template <size_t STEP_SIZE>
+simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
 }
 
-simdutf_really_inline void
-change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
-  size_t pos = 0;
+template <size_t STEP_SIZE>
+simdutf_really_inline const uint8_t *
+buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
 
-  while (pos < size / 32 * 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    input.swap_bytes();
-    input.store(reinterpret_cast<uint16_t *>(output));
-    pos += 32;
-    output += 32;
-  }
+template <size_t STEP_SIZE>
+simdutf_really_inline size_t
+buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if (len == idx) {
+    return 0;
+  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20,
+              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
+                          // to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
 
-  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
+template <size_t STEP_SIZE>
+simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
 }
 
-} // namespace utf16
 } // unnamed namespace
-} // namespace haswell
+} // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf16.h */
-
-// transcoding from UTF-8 to Latin 1
-/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
-
+/* end file src/generic/buf_block_reader.h */
+/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
 namespace simdutf {
-namespace haswell {
+namespace ppc64 {
 namespace {
-namespace utf8_to_latin1 {
+namespace utf8_validation {
+
 using namespace simd;
 
 simdutf_really_inline simd8<uint8_t>
 check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // For UTF-8 to Latin 1, we can allow any ASCII character, and any
-  // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or
-  // 0b11000010 and nothing else.
-  //
   // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
   // Bit 1 = Too Long (ASCII followed by continuation)
   // Bit 2 = Overlong 3-byte
@@ -33156,7 +42066,6 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
   // 1111011_ 1000____
   // 11111___ 1000____
   constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
-  constexpr const uint8_t FORBIDDEN = 0xff;
 
   const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
       // 0_______ ________ <ASCII in byte 1>
@@ -33167,11 +42076,11 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
       // 1100____ ________ <two byte lead in byte 1>
       TOO_SHORT | OVERLONG_2,
       // 1101____ ________ <two byte lead in byte 1>
-      FORBIDDEN,
+      TOO_SHORT,
       // 1110____ ________ <three byte lead in byte 1>
-      FORBIDDEN,
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
       // 1111____ ________ <four+ byte lead in byte 1>
-      FORBIDDEN);
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
   constexpr const uint8_t CARRY =
       TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
   const simd8<uint8_t> byte_1_low =
@@ -33185,16 +42094,23 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
               CARRY, CARRY,
 
               // ____0100 ________
-              FORBIDDEN,
+              CARRY | TOO_LARGE,
               // ____0101 ________
-              FORBIDDEN,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
               // ____011_ ________
-              FORBIDDEN, FORBIDDEN,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
 
               // ____1___ ________
-              FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
               // ____1101 ________
-              FORBIDDEN, FORBIDDEN, FORBIDDEN);
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
   const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
       // ________ 0_______ <ASCII in byte 2>
       TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
@@ -33213,1982 +42129,1639 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
       TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
   return (byte_1_high & byte_1_low & byte_2_high);
 }
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
+}
 
-struct validating_transcoder {
+//
+// Return nonzero if there are incomplete multibyte characters at the end of the
+// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
+//
+simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  // If the previous input's last 3 bytes match this, they're too short (they
+  // ended at EOF):
+  // ... 1111____ 111_____ 11______
+  static const uint8_t max_array[32] = {255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        255,
+                                        0b11110000u - 1,
+                                        0b11100000u - 1,
+                                        0b11000000u - 1};
+  const simd8<uint8_t> max_value(
+      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
+  return input.gt_bits(max_value);
+}
+
+struct utf8_checker {
   // If this is nonzero, there has been a UTF-8 error.
   simd8<uint8_t> error;
+  // The last input we received
+  simd8<uint8_t> prev_input_block;
+  // Whether the last input we received was incomplete (used for ASCII fast
+  // path)
+  simd8<uint8_t> prev_incomplete;
 
-  validating_transcoder() : error(uint8_t(0)) {}
   //
   // Check whether the current bytes are valid UTF-8.
   //
   simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    this->error |= check_special_cases(input, prev1);
-  }
-
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char *latin1_output) {
-    size_t pos = 0;
-    char *start{latin1_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 16 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 16; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) >
-                       -65); // twos complement of -65 is 1011 1111 ...
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store((int8_t *)latin1_output);
-        latin1_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask =
-            input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
-                               // this case, we also have ASCII to account for.
-        if (utf8_continuation_mask & 1) {
-          return 0; // error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_latin1(
-              in + pos, utf8_end_of_code_point_mask, latin1_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      return 0;
-    }
-    if (pos < size) {
-      size_t howmany =
-          scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      latin1_output += howmany;
-    }
-    return latin1_output - start;
-  }
-
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char *latin1_output) {
-    size_t pos = 0;
-    char *start{latin1_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store((int8_t *)latin1_output);
-        latin1_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        if (errors()) {
-          // rewind_and_convert_with_errors will seek a potential error from
-          // in+pos onward, with the ability to go back up to pos bytes, and
-          // read size-pos bytes forward.
-          result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-              pos, in + pos, size - pos, latin1_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_latin1(
-              in + pos, utf8_end_of_code_point_mask, latin1_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, latin1_output);
-      res.count += pos;
-      return res;
-    }
-    if (pos < size) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, latin1_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        latin1_output += res.count;
-      }
-    }
-    return result(error_code::SUCCESS, latin1_output - start);
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
 
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+  // The only problem that can happen at EOF is that a multibyte character is
+  // too short or a byte value too large in the last bytes: check_special_cases
+  // only checks for bytes too large in the first of two bytes.
+  simdutf_really_inline void check_eof() {
+    // If the previous block had incomplete UTF-8 characters at the end, an
+    // ASCII block can't possibly finish them.
+    this->error |= this->prev_incomplete;
   }
 
-}; // struct utf8_checker
-} // namespace utf8_to_latin1
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdutf
-/* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
-/* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-
-namespace simdutf {
-namespace haswell {
-namespace {
-namespace utf8_to_latin1 {
-using namespace simd;
-
-simdutf_really_inline size_t convert_valid(const char *in, size_t size,
-                                           char *latin1_output) {
-  size_t pos = 0;
-  char *start{latin1_output};
-  // In the worst case, we have the haswell kernel which can cause an overflow
-  // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the last
-  // 16 bytes, and if the data is valid, then it is entirely safe because 16
-  // UTF-8 bytes generate much more than 8 bytes. However, you cannot generally
-  // assume that you have valid UTF-8 input, so we are going to go back from the
-  // end counting 8 leading bytes, to give us a good margin.
-  size_t leading_byte = 0;
-  size_t margin = size;
-  for (; margin > 0 && leading_byte < 8; margin--) {
-    leading_byte += (int8_t(in[margin - 1]) >
-                     -65); // twos complement of -65 is 1011 1111 ...
-  }
-  // If the input is long enough, then we have that margin-1 is the eight last
-  // leading byte.
-  const size_t safety_margin = size - margin + 1; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    if (input.is_ascii()) {
-      input.store((int8_t *)latin1_output);
-      latin1_output += 64;
-      pos += 64;
+  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
+    if (simdutf_likely(is_ascii(input))) {
+      this->error |= this->prev_incomplete;
     } else {
       // you might think that a for-loop would work, but under Visual Studio, it
       // is not good enough.
-      uint64_t utf8_continuation_mask =
-          input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
-                             // this case, we also have ASCII to account for.
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      // We process in blocks of up to 12 bytes except possibly
-      // for fast paths which may process up to 16 bytes. For the
-      // slow path to work, we should have at least 12 input bytes left.
-      size_t max_starting_point = (pos + 64) - 12;
-      // Next loop is going to run at least five times.
-      while (pos < max_starting_point) {
-        // Performance note: our ability to compute 'consumed' and
-        // then shift and recompute is critical. If there is a
-        // latency of, say, 4 cycles on getting 'consumed', then
-        // the inner loop might have a total latency of about 6 cycles.
-        // Yet we process between 6 to 12 inputs bytes, thus we get
-        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-        // for this section of the code. Hence, there is a limit
-        // to how much we can further increase this latency before
-        // it seriously harms performance.
-        size_t consumed = convert_masked_utf8_to_latin1(
-            in + pos, utf8_end_of_code_point_mask, latin1_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
+      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                    "We support either two or four chunks per 64-byte block.");
+      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
+        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
       }
-      // At this point there may remain between 0 and 12 bytes in the
-      // 64-byte block. These bytes will be processed again. So we have an
-      // 80% efficiency (in the worst case). In practice we expect an
-      // 85% to 90% efficiency.
+      this->prev_incomplete =
+          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
+      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
     }
   }
-  if (pos < size) {
-    size_t howmany = scalar::utf8_to_latin1::convert_valid(in + pos, size - pos,
-                                                           latin1_output);
-    latin1_output += howmany;
-  }
-  return latin1_output - start;
-}
 
-} // namespace utf8_to_latin1
-} // namespace
-} // namespace haswell
-} // namespace simdutf
-  // namespace simdutf
-/* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+  // do not forget to call check_eof!
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
+  }
 
-namespace simdutf {
-namespace haswell {
+}; // struct utf8_checker
+} // namespace utf8_validation
 
-simdutf_warn_unused int
-implementation::detect_encodings(const char *input,
-                                 size_t length) const noexcept {
-  // If there is a BOM, then we trust it.
-  auto bom_encoding = simdutf::BOM::check_bom(input, length);
-  if (bom_encoding != encoding_type::unspecified) {
-    return bom_encoding;
-  }
+using utf8_validation::utf8_checker;
 
-  int out = 0;
-  uint32_t utf16_err = (length % 2);
-  uint32_t utf32_err = (length % 4);
-  uint32_t ends_with_high = 0;
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-  const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
-  const __m256i offset = _mm256_set1_epi32(0xffff2000);
-  const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
-  __m256i currentmax = _mm256_setzero_si256();
-  __m256i currentoffsetmax = _mm256_setzero_si256();
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
+/* begin file src/generic/utf8_validation/utf8_validator.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8_validation {
 
-  utf8_checker c{};
-  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template <class checker>
+bool generic_validate_utf8(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
   while (reader.has_full_block()) {
     simd::simd8x64<uint8_t> in(reader.full_block());
-    // utf8 checks
     c.check_next_input(in);
-
-    // utf16le checks
-    auto in0 = simd16<uint16_t>(in.chunks[0]);
-    auto in1 = simd16<uint16_t>(in.chunks[1]);
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
-    const auto in2 = simd16<uint16_t>::pack(t0, t1);
-    const auto surrogates_wordmask = (in2 & v_f8) == v_d8;
-    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-    const auto vL = (in2 & v_fc) == v_dc;
-    const uint32_t L = vL.to_bitmask();
-    const uint32_t H = L ^ surrogates_bitmask;
-    utf16_err |= (((H << 1) | ends_with_high) != L);
-    ends_with_high = (H & 0x80000000) != 0;
-
-    // utf32le checks
-    currentmax = _mm256_max_epu32(in.chunks[0], currentmax);
-    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[0], offset),
-                                        currentoffsetmax);
-    currentmax = _mm256_max_epu32(in.chunks[1], currentmax);
-    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[1], offset),
-                                        currentoffsetmax);
-
     reader.advance();
   }
-
   uint8_t block[64]{};
-  size_t idx = reader.block_index();
-  std::memcpy(block, &input[idx], length - idx);
+  reader.get_remainder(block);
   simd::simd8x64<uint8_t> in(block);
   c.check_next_input(in);
-
-  // utf16le last block check
-  auto in0 = simd16<uint16_t>(in.chunks[0]);
-  auto in1 = simd16<uint16_t>(in.chunks[1]);
-  const auto t0 = in0.shr<8>();
-  const auto t1 = in1.shr<8>();
-  const auto in2 = simd16<uint16_t>::pack(t0, t1);
-  const auto surrogates_wordmask = (in2 & v_f8) == v_d8;
-  const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-  const auto vL = (in2 & v_fc) == v_dc;
-  const uint32_t L = vL.to_bitmask();
-  const uint32_t H = L ^ surrogates_bitmask;
-  utf16_err |= (((H << 1) | ends_with_high) != L);
-  // this is required to check for last byte ending in high and end of input
-  // is reached
-  ends_with_high = (H & 0x80000000) != 0;
-  utf16_err |= ends_with_high;
-
-  // utf32le last block check
-  currentmax = _mm256_max_epu32(in.chunks[0], currentmax);
-  currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[0], offset),
-                                      currentoffsetmax);
-  currentmax = _mm256_max_epu32(in.chunks[1], currentmax);
-  currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in.chunks[1], offset),
-                                      currentoffsetmax);
-
   reader.advance();
-
   c.check_eof();
-  bool is_valid_utf8 = !c.errors();
-  __m256i is_zero =
-      _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
-  utf32_err |= (_mm256_testz_si256(is_zero, is_zero) == 0);
-
-  is_zero = _mm256_xor_si256(
-      _mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
-  utf32_err |= (_mm256_testz_si256(is_zero, is_zero) == 0);
-  if (is_valid_utf8) {
-    out |= encoding_type::UTF8;
-  }
-  if (utf16_err == 0) {
-    out |= encoding_type::UTF16_LE;
-  }
-  if (utf32_err == 0) {
-    out |= encoding_type::UTF32_LE;
-  }
-  return out;
-}
-
-simdutf_warn_unused bool
-implementation::validate_utf8(const char *buf, size_t len) const noexcept {
-  return haswell::utf8_validation::generic_validate_utf8(buf, len);
-}
-
-simdutf_warn_unused result implementation::validate_utf8_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return haswell::utf8_validation::generic_validate_utf8_with_errors(buf, len);
-}
-
-simdutf_warn_unused bool
-implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return haswell::utf8_validation::generic_validate_ascii(buf, len);
-}
-
-simdutf_warn_unused result implementation::validate_ascii_with_errors(
-    const char *buf, size_t len) const noexcept {
-  return haswell::utf8_validation::generic_validate_ascii_with_errors(buf, len);
-}
-
-simdutf_warn_unused bool
-implementation::validate_utf16le(const char16_t *buf,
-                                 size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid UTF-16. protect the implementation from
-    // handling nullptr
-    return true;
-  }
-  const char16_t *tail = avx2_validate_utf16<endianness::LITTLE>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::LITTLE>(tail,
-                                                       len - (tail - buf));
-  } else {
-    return false;
-  }
+  return !c.errors();
 }
 
-simdutf_warn_unused bool
-implementation::validate_utf16be(const char16_t *buf,
-                                 size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid UTF-16. protect the implementation from
-    // handling nullptr
-    return true;
-  }
-  const char16_t *tail = avx2_validate_utf16<endianness::BIG>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
-  } else {
-    return false;
-  }
+bool generic_validate_utf8(const char *input, size_t length) {
+  return generic_validate_utf8<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
 }
 
-simdutf_warn_unused result implementation::validate_utf16le_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  result res = avx2_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
-  if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
-        buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
-    return res;
+/**
+ * Validates that the string is actual UTF-8 and stops on errors.
+ */
+template <class checker>
+result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
+  checker c{};
+  buf_block_reader<64> reader(input, length);
+  size_t count{0};
+  while (reader.has_full_block()) {
+    simd::simd8x64<uint8_t> in(reader.full_block());
+    c.check_next_input(in);
+    if (c.errors()) {
+      if (count != 0) {
+        count--;
+      } // Sometimes the error is only detected in the next chunk
+      result res = scalar::utf8::rewind_and_validate_with_errors(
+          reinterpret_cast<const char *>(input),
+          reinterpret_cast<const char *>(input + count), length - count);
+      res.count += count;
+      return res;
+    }
+    reader.advance();
+    count += 64;
   }
-}
-
-simdutf_warn_unused result implementation::validate_utf16be_with_errors(
-    const char16_t *buf, size_t len) const noexcept {
-  result res = avx2_validate_utf16_with_errors<endianness::BIG>(buf, len);
-  if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
-        buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
+  uint8_t block[64]{};
+  reader.get_remainder(block);
+  simd::simd8x64<uint8_t> in(block);
+  c.check_next_input(in);
+  reader.advance();
+  c.check_eof();
+  if (c.errors()) {
+    if (count != 0) {
+      count--;
+    } // Sometimes the error is only detected in the next chunk
+    result res = scalar::utf8::rewind_and_validate_with_errors(
+        reinterpret_cast<const char *>(input),
+        reinterpret_cast<const char *>(input) + count, length - count);
+    res.count += count;
     return res;
-  }
-}
-
-simdutf_warn_unused bool
-implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid UTF-32. protect the implementation from
-    // handling nullptr
-    return true;
-  }
-  const char32_t *tail = avx2_validate_utf32le(buf, len);
-  if (tail) {
-    return scalar::utf32::validate(tail, len - (tail - buf));
-  } else {
-    return false;
-  }
-}
-
-simdutf_warn_unused result implementation::validate_utf32_with_errors(
-    const char32_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid UTF-32. protect the implementation from
-    // handling nullptr
-    return result(error_code::SUCCESS, 0);
-  }
-  result res = avx2_validate_utf32le_with_errors(buf, len);
-  if (res.count != len) {
-    result scalar_res =
-        scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
   } else {
-    return res;
-  }
-}
-
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
-    const char *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char *, char *> ret =
-      avx2_convert_latin1_to_utf8(buf, len, utf8_output);
-  size_t converted_chars = ret.second - utf8_output;
-
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    converted_chars += scalar_converted_chars;
-  }
-
-  return converted_chars;
-}
-
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char *, char16_t *> ret =
-      avx2_convert_latin1_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t converted_chars = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars =
-        scalar::latin1_to_utf16::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_converted_chars == 0) {
-      return 0;
-    }
-    converted_chars += scalar_converted_chars;
-  }
-  return converted_chars;
-}
-
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char *, char16_t *> ret =
-      avx2_convert_latin1_to_utf16<endianness::BIG>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t converted_chars = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars =
-        scalar::latin1_to_utf16::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_converted_chars == 0) {
-      return 0;
-    }
-    converted_chars += scalar_converted_chars;
-  }
-  return converted_chars;
-}
-
-simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char *, char32_t *> ret =
-      avx2_convert_latin1_to_utf32(buf, len, utf32_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t converted_chars = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_converted_chars = scalar::latin1_to_utf32::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_converted_chars == 0) {
-      return 0;
-    }
-    converted_chars += scalar_converted_chars;
+    return result(error_code::SUCCESS, length);
   }
-  return converted_chars;
-}
-
-simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  utf8_to_latin1::validating_transcoder converter;
-  return converter.convert(buf, len, latin1_output);
-}
-
-simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
-    const char *buf, size_t len, char *latin1_output) const noexcept {
-  utf8_to_latin1::validating_transcoder converter;
-  return converter.convert_with_errors(buf, len, latin1_output);
 }
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
-    const char *input, size_t size, char *latin1_output) const noexcept {
-  return utf8_to_latin1::convert_valid(input, size, latin1_output);
+result generic_validate_utf8_with_errors(const char *input, size_t length) {
+  return generic_validate_utf8_with_errors<utf8_checker>(
+      reinterpret_cast<const uint8_t *>(input), length);
 }
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
-}
+} // namespace utf8_validation
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert<endianness::BIG>(buf, len, utf16_output);
-}
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8_to_utf16 {
+using namespace simd;
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert_with_errors<endianness::LITTLE>(buf, len,
-                                                           utf16_output);
-}
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
-    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
-  utf8_to_utf16::validating_transcoder converter;
-  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
-}
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
-    const char *input, size_t size, char16_t *utf16_output) const noexcept {
-  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,
-                                                          utf16_output);
-}
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
 
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
-    const char *input, size_t size, char16_t *utf16_output) const noexcept {
-  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
-                                                       utf16_output);
-}
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
 
-simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  utf8_to_utf32::validating_transcoder converter;
-  return converter.convert(buf, len, utf32_output);
-}
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
 
-simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
-    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
-  utf8_to_utf32::validating_transcoder converter;
-  return converter.convert_with_errors(buf, len, utf32_output);
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
 }
-
-simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
-    const char *input, size_t size, char32_t *utf32_output) const noexcept {
-  return utf8_to_utf32::convert_valid(input, size, utf32_output);
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      haswell::avx2_convert_utf16_to_latin1<endianness::LITTLE>(buf, len,
-                                                                latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_latin1::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      haswell::avx2_convert_utf16_to_latin1<endianness::BIG>(buf, len,
-                                                             latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_latin1::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused result
-implementation::convert_utf16le_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<result, char *> ret =
-      avx2_convert_utf16_to_latin1_with_errors<endianness::LITTLE>(
-          buf, len, latin1_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+  template <endianness endian>
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
     }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused result
-implementation::convert_utf16be_to_latin1_with_errors(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<result, char *> ret =
-      avx2_convert_utf16_to_latin1_with_errors<endianness::BIG>(buf, len,
-                                                                latin1_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
     }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement a custom function
-  return convert_utf16be_to_latin1(buf, len, latin1_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
-    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
-  // optimization opportunity: implement a custom function
-  return convert_utf16le_to_latin1(buf, len, latin1_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      haswell::avx2_convert_utf16_to_utf8<endianness::LITTLE>(buf, len,
-                                                              utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf8::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
+    if (errors()) {
       return 0;
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
-
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char16_t *, char *> ret =
-      haswell::avx2_convert_utf16_to_utf8<endianness::BIG>(buf, len,
-                                                           utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf8::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+    if (pos < size) {
+      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
+          in + pos, size - pos, utf16_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      utf16_output += howmany;
     }
-    saved_bytes += scalar_saved_bytes;
+    return utf16_output - start;
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(
-          buf, len, utf8_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+  template <endianness endian>
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char16_t *utf16_output) {
+    size_t pos = 0;
+    char16_t *start{utf16_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
     }
-  }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::BIG>(
-          buf, len, utf8_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf16<endian>(utf16_output);
+        utf16_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res =
+              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+                  pos, in + pos, size - pos, utf16_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf16<endian>(
+              in + pos, utf8_end_of_code_point_mask, utf16_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
     }
-  }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16le_to_utf8(buf, len, utf8_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
-    const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf16be_to_utf8(buf, len, utf8_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  std::pair<const char32_t *, char *> ret =
-      avx2_convert_utf32_to_utf8(buf, len, utf8_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf8_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+    if (errors()) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      res.count += pos;
+      return res;
+    }
+    if (pos < size) {
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res =
+          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
+              pos, in + pos, size - pos, utf16_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf16_output += res.count;
+      }
     }
-    saved_bytes += scalar_saved_bytes;
+    return result(error_code::SUCCESS, utf16_output - start);
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  std::pair<const char32_t *, char *> ret =
-      avx2_convert_utf32_to_latin1(buf, len, latin1_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - latin1_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes = scalar::utf32_to_latin1::convert(
-        ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      avx2_convert_utf32_to_latin1_with_errors(buf, len, latin1_output);
-  if (ret.first.count != len) {
-    result scalar_res = scalar::utf32_to_latin1::convert_with_errors(
-        buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      latin1_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+}; // struct utf8_checker
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8_to_utf16 {
 
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
-    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
-  return convert_utf32_to_latin1(buf, len, latin1_output);
-}
+using namespace simd;
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char *> ret =
-      haswell::avx2_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
-  if (ret.first.count != len) {
-    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
-        buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
+template <endianness endian>
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char16_t *utf16_output) noexcept {
+  // The implementation is not specific to haswell and should be moved to the
+  // generic directory.
+  size_t pos = 0;
+  char16_t *start{utf16_output};
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    // this loop could be unrolled further. For example, we could process the
+    // mask far more than 64 bytes.
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf16<endian>(utf16_output);
+      utf16_output += 64;
+      pos += 64;
     } else {
-      ret.second += scalar_res.count;
+      // Slow path. We hope that the compiler will recognize that this is a slow
+      // path. Anything that is not a continuation mask is a 'leading byte',
+      // that is, the start of a new code point.
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      // The *start* of code points is not so useful, rather, we want the *end*
+      // of code points.
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
+      size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times when using solely
+      // the slow/regular path, and at least four times if there are fast paths.
+      while (pos < max_starting_point) {
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        //
+        // Thus we may allow convert_masked_utf8_to_utf16 to process
+        // more bytes at a time under a fast-path mode where 16 bytes
+        // are consumed at once (e.g., when encountering ASCII).
+        size_t consumed = convert_masked_utf8_to_utf16<endian>(
+            input + pos, utf8_end_of_code_point_mask, utf16_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
     }
   }
-  ret.first.count =
-      ret.second -
-      utf8_output; // Set count to the number of 8-bit code units written
-  return ret.first;
+  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
+      input + pos, size - pos, utf16_output);
+  return utf16_output - start;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char16_t *, char32_t *> ret =
-      haswell::avx2_convert_utf16_to_utf32<endianness::LITTLE>(buf, len,
-                                                               utf32_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
+} // namespace utf8_to_utf16
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8_to_utf32 {
+using namespace simd;
+
+simdutf_really_inline simd8<uint8_t>
+check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+  // Bit 1 = Too Long (ASCII followed by continuation)
+  // Bit 2 = Overlong 3-byte
+  // Bit 4 = Surrogate
+  // Bit 5 = Overlong 2-byte
+  // Bit 7 = Two Continuations
+  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
+                                               // 11______ 11______
+  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
+  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
+  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
+  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
+  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
+  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
+                                               // 11110100 101_____
+                                               // 11110101 1001____
+                                               // 11110101 101_____
+                                               // 1111011_ 1001____
+                                               // 1111011_ 101_____
+                                               // 11111___ 1001____
+                                               // 11111___ 101_____
+  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
+  // 11110101 1000____
+  // 1111011_ 1000____
+  // 11111___ 1000____
+  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+
+  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+  constexpr const uint8_t CARRY =
+      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+  const simd8<uint8_t> byte_1_low =
+      (prev1 & 0x0F)
+          .lookup_16<uint8_t>(
+              // ____0000 ________
+              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+              // ____0001 ________
+              CARRY | OVERLONG_2,
+              // ____001_ ________
+              CARRY, CARRY,
+
+              // ____0100 ________
+              CARRY | TOO_LARGE,
+              // ____0101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____011_ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+              // ____1___ ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              // ____1101 ________
+              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              CARRY | TOO_LARGE | TOO_LARGE_1000);
+  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
+          OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
+  return (byte_1_high & byte_1_low & byte_2_high);
+}
+simdutf_really_inline simd8<uint8_t>
+check_multibyte_lengths(const simd8<uint8_t> input,
+                        const simd8<uint8_t> prev_input,
+                        const simd8<uint8_t> sc) {
+  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+  simd8<uint8_t> must23 =
+      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
+  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+  return must23_80 ^ sc;
 }
 
-simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  std::pair<const char16_t *, char32_t *> ret =
-      haswell::avx2_convert_utf16_to_utf32<endianness::BIG>(buf, len,
-                                                            utf32_output);
-  if (ret.first == nullptr) {
-    return 0;
+struct validating_transcoder {
+  // If this is nonzero, there has been a UTF-8 error.
+  simd8<uint8_t> error;
+
+  validating_transcoder() : error(uint8_t(0)) {}
+  //
+  // Check whether the current bytes are valid UTF-8.
+  //
+  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
+                                              const simd8<uint8_t> prev_input) {
+    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
+    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
+    // small negative numbers)
+    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+    simd8<uint8_t> sc = check_special_cases(input, prev1);
+    this->error |= check_multibyte_lengths(input, prev_input, sc);
   }
-  size_t saved_bytes = ret.second - utf32_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf16_to_utf32::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
+
+  simdutf_really_inline size_t convert(const char *in, size_t size,
+                                       char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 16 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (utf8_continuation_mask & 1) {
+          return 0; // we have an error
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
       return 0;
     }
-    saved_bytes += scalar_saved_bytes;
-  }
-  return saved_bytes;
-}
-
-simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char32_t *> ret =
-      haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(
-          buf, len, utf32_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+    if (pos < size) {
+      size_t howmany =
+          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
+      if (howmany == 0) {
+        return 0;
+      }
+      utf32_output += howmany;
     }
+    return utf32_output - start;
   }
-  ret.first.count =
-      ret.second -
-      utf32_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
 
-simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char32_t *> ret =
-      haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::BIG>(
-          buf, len, utf32_output);
-  if (ret.first.error) {
-    return ret.first;
-  } // Can return directly since scalar fallback already found correct
-    // ret.first.count
-  if (ret.first.count != len) { // All good so far, but not finished
-    result scalar_res =
-        scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
+  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
+                                                   char32_t *utf32_output) {
+    size_t pos = 0;
+    char32_t *start{utf32_output};
+    // In the worst case, we have the haswell kernel which can cause an overflow
+    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
+    // last 16 bytes, and if the data is valid, then it is entirely safe because
+    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
+    // generally assume that you have valid UTF-8 input, so we are going to go
+    // back from the end counting 8 leading bytes, to give us a good margin.
+    size_t leading_byte = 0;
+    size_t margin = size;
+    for (; margin > 0 && leading_byte < 8; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) > -65);
+    }
+    // If the input is long enough, then we have that margin-1 is the fourth
+    // last leading byte.
+    const size_t safety_margin = size - margin + 1; // to avoid overruns!
+    while (pos + 64 + safety_margin <= size) {
+      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+      if (input.is_ascii()) {
+        input.store_ascii_as_utf32(utf32_output);
+        utf32_output += 64;
+        pos += 64;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio,
+        // it is not good enough.
+        static_assert(
+            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
+                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+            "We support either two or four chunks per 64-byte block.");
+        auto zero = simd8<uint8_t>{uint8_t(0)};
+        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.chunks[0], zero);
+          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
+          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
+          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+        }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        if (errors() || (utf8_continuation_mask & 1)) {
+          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, utf32_output);
+          res.count += pos;
+          return res;
+        }
+        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+        // We process in blocks of up to 12 bytes except possibly
+        // for fast paths which may process up to 16 bytes. For the
+        // slow path to work, we should have at least 12 input bytes left.
+        size_t max_starting_point = (pos + 64) - 12;
+        // Next loop is going to run at least five times.
+        while (pos < max_starting_point) {
+          // Performance note: our ability to compute 'consumed' and
+          // then shift and recompute is critical. If there is a
+          // latency of, say, 4 cycles on getting 'consumed', then
+          // the inner loop might have a total latency of about 6 cycles.
+          // Yet we process between 6 to 12 inputs bytes, thus we get
+          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+          // for this section of the code. Hence, there is a limit
+          // to how much we can further increase this latency before
+          // it seriously harms performance.
+          size_t consumed = convert_masked_utf8_to_utf32(
+              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          pos += consumed;
+          utf8_end_of_code_point_mask >>= consumed;
+        }
+        // At this point there may remain between 0 and 12 bytes in the
+        // 64-byte block. These bytes will be processed again. So we have an
+        // 80% efficiency (in the worst case). In practice we expect an
+        // 85% to 90% efficiency.
+      }
+    }
+    if (errors()) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      res.count += pos;
+      return res;
     }
-  }
-  ret.first.count =
-      ret.second -
-      utf32_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
-    const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return convert_utf32_to_utf8(buf, len, utf8_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      avx2_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::LITTLE>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
+    if (pos < size) {
+      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, utf32_output);
+      if (res.error) { // In case of error, we want the error position
+        res.count += pos;
+        return res;
+      } else { // In case of success, we want the number of word written
+        utf32_output += res.count;
+      }
     }
-    saved_bytes += scalar_saved_bytes;
+    return result(error_code::SUCCESS, utf32_output - start);
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  std::pair<const char32_t *, char16_t *> ret =
-      avx2_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
-  if (ret.first == nullptr) {
-    return 0;
-  }
-  size_t saved_bytes = ret.second - utf16_output;
-  if (ret.first != buf + len) {
-    const size_t scalar_saved_bytes =
-        scalar::utf32_to_utf16::convert<endianness::BIG>(
-            ret.first, len - (ret.first - buf), ret.second);
-    if (scalar_saved_bytes == 0) {
-      return 0;
-    }
-    saved_bytes += scalar_saved_bytes;
+  simdutf_really_inline bool errors() const {
+    return this->error.any_bits_set_anywhere();
   }
-  return saved_bytes;
-}
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(
-          buf, len, utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
-    } else {
-      ret.second += scalar_res.count;
-    }
-  }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
+}; // struct utf8_checker
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8_to_utf32 {
 
-simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  // ret.first.count is always the position in the buffer, not the number of
-  // code units written even if finished
-  std::pair<result, char16_t *> ret =
-      haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::BIG>(
-          buf, len, utf16_output);
-  if (ret.first.count != len) {
-    result scalar_res =
-        scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
-            buf + ret.first.count, len - ret.first.count, ret.second);
-    if (scalar_res.error) {
-      scalar_res.count += ret.first.count;
-      return scalar_res;
+using namespace simd;
+
+simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
+                                         char32_t *utf32_output) noexcept {
+  size_t pos = 0;
+  char32_t *start{utf32_output};
+  const size_t safety_margin = 16; // to avoid overruns!
+  while (pos + 64 + safety_margin <= size) {
+    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
+    if (in.is_ascii()) {
+      in.store_ascii_as_utf32(utf32_output);
+      utf32_output += 64;
+      pos += 64;
     } else {
-      ret.second += scalar_res.count;
+      // -65 is 0b10111111 in two-complement's, so largest possible continuation
+      // byte
+      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
+      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      size_t max_starting_point = (pos + 64) - 12;
+      while (pos < max_starting_point) {
+        size_t consumed = convert_masked_utf8_to_utf32(
+            input + pos, utf8_end_of_code_point_mask, utf32_output);
+        pos += consumed;
+        utf8_end_of_code_point_mask >>= consumed;
+      }
     }
   }
-  ret.first.count =
-      ret.second -
-      utf16_output; // Set count to the number of 8-bit code units written
-  return ret.first;
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16le(buf, len, utf16_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
-    const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return convert_utf32_to_utf16be(buf, len, utf16_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return convert_utf16le_to_utf32(buf, len, utf32_output);
-}
-
-simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
-    const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return convert_utf16be_to_utf32(buf, len, utf32_output);
-}
-
-void implementation::change_endianness_utf16(const char16_t *input,
-                                             size_t length,
-                                             char16_t *output) const noexcept {
-  utf16::change_endianness_utf16(input, length, output);
-}
-
-simdutf_warn_unused size_t implementation::count_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::count_code_points<endianness::LITTLE>(input, length);
-}
-
-simdutf_warn_unused size_t implementation::count_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::count_code_points<endianness::BIG>(input, length);
-}
-
-simdutf_warn_unused size_t
-implementation::count_utf8(const char *input, size_t length) const noexcept {
-  return utf8::count_code_points(input, length);
-}
-
-simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
-    const char *buf, size_t len) const noexcept {
-  return count_utf8(buf, len);
+  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
+                                                       utf32_output);
+  return utf32_output - start;
 }
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return scalar::utf16::latin1_length_from_utf16(length);
-}
+} // namespace utf8_to_utf32
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return scalar::utf32::latin1_length_from_utf32(length);
-}
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/generic/utf8.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf8 {
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
-}
+using namespace simd;
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
+simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.gt(-65);
+    count += count_ones(utf8_continuation_mask);
+  }
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
-}
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
 
-simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
-    const char16_t *input, size_t length) const noexcept {
-  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
-}
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf16_length_from_latin1(length);
-}
+  size_t pos = 0;
+  size_t count = 0;
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return utf8::utf16_length_from_utf8(input, length);
-}
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
 
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf32_length_from_latin1(length);
-}
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
 
-simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
-    const char *input, size_t len) const noexcept {
-  const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
-  size_t answer = len / sizeof(__m256i) * sizeof(__m256i);
-  size_t i = 0;
-  if (answer >= 2048) { // long strings optimization
-    __m256i four_64bits = _mm256_setzero_si256();
-    while (i + sizeof(__m256i) <= len) {
-      __m256i runner = _mm256_setzero_si256();
-      // We can do up to 255 loops without overflow.
-      size_t iterations = (len - i) / sizeof(__m256i);
-      if (iterations > 255) {
-        iterations = 255;
-      }
-      size_t max_i = i + iterations * sizeof(__m256i) - sizeof(__m256i);
-      for (; i + 4 * sizeof(__m256i) <= max_i; i += 4 * sizeof(__m256i)) {
-        __m256i input1 = _mm256_loadu_si256((const __m256i *)(data + i));
-        __m256i input2 =
-            _mm256_loadu_si256((const __m256i *)(data + i + sizeof(__m256i)));
-        __m256i input3 = _mm256_loadu_si256(
-            (const __m256i *)(data + i + 2 * sizeof(__m256i)));
-        __m256i input4 = _mm256_loadu_si256(
-            (const __m256i *)(data + i + 3 * sizeof(__m256i)));
-        __m256i input12 =
-            _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input1),
-                            _mm256_cmpgt_epi8(_mm256_setzero_si256(), input2));
-        __m256i input23 =
-            _mm256_add_epi8(_mm256_cmpgt_epi8(_mm256_setzero_si256(), input3),
-                            _mm256_cmpgt_epi8(_mm256_setzero_si256(), input4));
-        __m256i input1234 = _mm256_add_epi8(input12, input23);
-        runner = _mm256_sub_epi8(runner, input1234);
-      }
-      for (; i <= max_i; i += sizeof(__m256i)) {
-        __m256i input_256_chunk =
-            _mm256_loadu_si256((const __m256i *)(data + i));
-        runner = _mm256_sub_epi8(
-            runner, _mm256_cmpgt_epi8(_mm256_setzero_si256(), input_256_chunk));
-      }
-      four_64bits = _mm256_add_epi64(
-          four_64bits, _mm256_sad_epu8(runner, _mm256_setzero_si256()));
-    }
-    answer += _mm256_extract_epi64(four_64bits, 0) +
-              _mm256_extract_epi64(four_64bits, 1) +
-              _mm256_extract_epi64(four_64bits, 2) +
-              _mm256_extract_epi64(four_64bits, 3);
-  } else if (answer > 0) {
-    for (; i + sizeof(__m256i) <= len; i += sizeof(__m256i)) {
-      __m256i latin = _mm256_loadu_si256((const __m256i *)(data + i));
-      uint32_t non_ascii = _mm256_movemask_epi8(latin);
-      answer += count_ones(non_ascii);
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
     }
   }
-  return answer + scalar::latin1::utf8_length_from_latin1(
-                      reinterpret_cast<const char *>(data + i), len - i);
-}
 
-simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const __m256i v_00000000 = _mm256_setzero_si256();
-  const __m256i v_ffffff80 = _mm256_set1_epi32((uint32_t)0xffffff80);
-  const __m256i v_fffff800 = _mm256_set1_epi32((uint32_t)0xfffff800);
-  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 8 <= length; pos += 8) {
-    __m256i in = _mm256_loadu_si256((__m256i *)(input + pos));
-    const __m256i ascii_bytes_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffffff80), v_00000000);
-    const __m256i one_two_bytes_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_fffff800), v_00000000);
-    const __m256i two_bytes_bytemask =
-        _mm256_xor_si256(one_two_bytes_bytemask, ascii_bytes_bytemask);
-    const __m256i one_two_three_bytes_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const __m256i three_bytes_bytemask =
-        _mm256_xor_si256(one_two_three_bytes_bytemask, one_two_bytes_bytemask);
-    const uint32_t ascii_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(ascii_bytes_bytemask));
-    const uint32_t two_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(two_bytes_bytemask));
-    const uint32_t three_bytes_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(three_bytes_bytemask));
-
-    size_t ascii_count = count_ones(ascii_bytes_bitmask) / 4;
-    size_t two_bytes_count = count_ones(two_bytes_bitmask) / 4;
-    size_t three_bytes_count = count_ones(three_bytes_bitmask) / 4;
-    count += 32 - 3 * ascii_count - 2 * two_bytes_count - three_bytes_count;
+  if (iterations > 0) {
+    count += local.sum_bytes();
   }
-  return count +
-         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+
+  count += counters.sum();
+
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
+#endif
 
-simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
-    const char32_t *input, size_t length) const noexcept {
-  const __m256i v_00000000 = _mm256_setzero_si256();
-  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
+simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
+                                                    size_t size) {
   size_t pos = 0;
   size_t count = 0;
-  for (; pos + 8 <= length; pos += 8) {
-    __m256i in = _mm256_loadu_si256((__m256i *)(input + pos));
-    const __m256i surrogate_bytemask =
-        _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
-    const uint32_t surrogate_bitmask =
-        static_cast<uint32_t>(_mm256_movemask_epi8(surrogate_bytemask));
-    size_t surrogate_count = (32 - count_ones(surrogate_bitmask)) / 4;
-    count += 8 + surrogate_count;
+  // This algorithm could no doubt be improved!
+  for (; pos + 64 <= size; pos += 64) {
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+    // We count one word for anything that is not a continuation (so
+    // leading bytes).
+    count += 64 - count_ones(utf8_continuation_mask);
+    int64_t utf8_4byte = input.gteq_unsigned(240);
+    count += count_ones(utf8_4byte);
   }
-  return count +
-         scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
-}
-
-simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
-    const char *input, size_t length) const noexcept {
-  return utf8::count_code_points(input, length);
-}
-
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
+  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
 }
+} // namespace utf8
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
 
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
+#if SIMDUTF_FEATURE_UTF16
+/* begin file src/generic/utf16.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace utf16 {
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+template <endianness big_endian>
+simdutf_really_inline size_t count_code_points(const char16_t *in,
+                                               size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
     }
+    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
+    count += count_ones(not_pair) / 2;
   }
+  return count +
+         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
 }
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
+                                                    size_t size) {
+  size_t pos = 0;
+  size_t count = 0;
+  // This algorithm could no doubt be improved!
+  for (; pos < size / 32 * 32; pos += 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input.swap_bytes();
     }
-  }
-}
+    uint64_t ascii_mask = input.lteq(0x7F);
+    uint64_t twobyte_mask = input.lteq(0x7FF);
+    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
 
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
+    size_t ascii_count = count_ones(ascii_mask) / 2;
+    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
+    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
+    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
+    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
+             ascii_count;
   }
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
-}
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
 
-size_t implementation::binary_to_base64(const char *input, size_t length,
-                                        char *output,
-                                        base64_options options) const noexcept {
-  if (options & base64_url) {
-    return encode_base64<true>(output, input, length, options);
-  } else {
-    return encode_base64<false>(output, input, length, options);
-  }
-}
-} // namespace haswell
-} // namespace simdutf
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
 
-/* begin file src/simdutf/haswell/end.h */
-#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
-// nothing needed.
-#else
-SIMDUTF_UNTARGET_REGION
-#endif
+  const auto one = vector_u16::splat(1);
 
+  auto v_count = vector_u16::zero();
 
-#if SIMDUTF_GCC11ORMORE // workaround for
-                        // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
-SIMDUTF_POP_DISABLE_WARNINGS
-#endif // end of workaround
-/* end file src/simdutf/haswell/end.h */
-/* end file src/haswell/implementation.cpp */
-#endif
-#if SIMDUTF_IMPLEMENTATION_PPC64
-/* begin file src/ppc64/implementation.cpp */
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
 
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
 
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
 
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
 
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
 
-/* begin file src/simdutf/ppc64/begin.h */
-// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
-// #define SIMDUTF_IMPLEMENTATION ppc64
-/* end file src/simdutf/ppc64/begin.h */
-namespace simdutf {
-namespace ppc64 {
-namespace {
-#ifndef SIMDUTF_PPC64_H
-  #error "ppc64.h must be included"
-#endif
-using namespace simd;
+    /*
+        Explanation how the counting works.
 
-simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
-  // careful: 0x80 is not ascii.
-  return input.reduce_or().saturating_sub(0b01111111u).bits_not_set_anywhere();
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
+
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
+    }
+  }
+
+  if (iteration > 0) {
+    count += v_count.sum();
+  }
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
 }
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte =
-      prev1.saturating_sub(0b11000000u - 1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte =
-      prev2.saturating_sub(0b11100000u - 1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte =
-      prev3.saturating_sub(0b11110000u - 1); // Only 1111____ will be > 0
-  // Caller requires a bool (all 1's). All values resulting from the subtraction
-  // will be <= 64, so signed comparison is fine.
-  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) >
-         int8_t(0);
+template <endianness big_endian>
+simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
+                                                     size_t size) {
+  return count_code_points<big_endian>(in, size);
 }
 
-simdutf_really_inline simd8<bool>
-must_be_2_3_continuation(const simd8<uint8_t> prev2,
-                         const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte =
-      prev2.saturating_sub(0xe0u - 0x80); // Only 111_____ will be >= 0x80
-  simd8<uint8_t> is_fourth_byte =
-      prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be >= 0x80
-  // Caller requires a bool (all 1's). All values resulting from the subtraction
-  // will be <= 64, so signed comparison is fine.
-  return simd8<bool>(is_third_byte | is_fourth_byte);
+simdutf_really_inline void
+change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
+  size_t pos = 0;
+
+  while (pos < size / 32 * 32) {
+    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
+    input.swap_bytes();
+    input.store(reinterpret_cast<uint16_t *>(output));
+    pos += 32;
+    output += 32;
+  }
+
+  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
 }
 
+} // namespace utf16
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-
-/* begin file src/generic/buf_block_reader.h */
+/* end file src/generic/utf16.h */
+/* begin file src/generic/validate_utf16.h */
 namespace simdutf {
 namespace ppc64 {
 namespace {
+namespace utf16 {
+/*
+    UTF-16 validation
+    --------------------------------------------------
 
-// Walks through a buffer in block-sized increments, loading the last part with
-// spaces
-template <size_t STEP_SIZE> struct buf_block_reader {
-public:
-  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdutf_really_inline size_t block_index();
-  simdutf_really_inline bool has_full_block() const;
-  simdutf_really_inline const uint8_t *full_block() const;
-  /**
-   * Get the last block, padded with spaces.
-   *
-   * There will always be a last block, with at least 1 byte, unless len == 0
-   * (in which case this function fills the buffer with spaces and returns 0. In
-   * particular, if len == STEP_SIZE there will be 0 full_blocks and 1 remainder
-   * block with STEP_SIZE bytes and no spaces for padding.
-   *
-   * @return the number of effective characters in the last block.
-   */
-  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdutf_really_inline void advance();
+    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
 
-private:
-  const uint8_t *buf;
-  const size_t len;
-  const size_t lenminusstep;
-  size_t idx;
-};
+    In a vectorized algorithm we want to examine the most significant
+    nibble in order to select a fast path. If none of highest nibbles
+    are 0xD (13), than we are sure that UTF-16 chunk in a vector
+    register is valid.
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text_64(const uint8_t *text) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
-  }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
+    Let us analyze what we need to check if the nibble is 0xD. The
+    value of the preceding nibble determines what we have:
 
-// Routines to print masks and text for debugging bitmask operations
-simdutf_unused static char *format_input_text(const simd8x64<uint8_t> &in) {
-  static char *buf =
-      reinterpret_cast<char *>(malloc(sizeof(simd8x64<uint8_t>) + 1));
-  in.store(reinterpret_cast<uint8_t *>(buf));
-  for (size_t i = 0; i < sizeof(simd8x64<uint8_t>); i++) {
-    if (buf[i] < ' ') {
-      buf[i] = '_';
-    }
-  }
-  buf[sizeof(simd8x64<uint8_t>)] = '\0';
-  return buf;
-}
+    0xd000 .. 0xd7ff - a valid word
+    0xd800 .. 0xdbff - low surrogate
+    0xdc00 .. 0xdfff - high surrogate
 
-simdutf_unused static char *format_mask(uint64_t mask) {
-  static char *buf = reinterpret_cast<char *>(malloc(64 + 1));
-  for (size_t i = 0; i < 64; i++) {
-    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+    Other constraints we have to consider:
+    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
+    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
+    - there must not be sole low surrogate nor high surrogate
+
+    We are going to build three bitmasks based on the 3rd nibble:
+    - V = valid word,
+    - L = low surrogate (0xd800 .. 0xdbff)
+    - H = high surrogate (0xdc00 .. 0xdfff)
+
+      0   1   2   3   4   5   6   7    <--- word index
+    [ V | L | H | L | H | V | V | L ]
+      1   0   0   0   0   1   1   0     - V = valid masks
+      0   1   0   1   0   0   0   1     - L = low surrogate
+      0   0   1   0   1   0   0   0     - H high surrogate
+
+
+      1   0   0   0   0   1   1   0   V = valid masks
+      0   1   0   1   0   0   0   0   a = L & (H >> 1)
+      0   0   1   0   1   0   0   0   b = a << 1
+      1   1   1   1   1   1   1   0   c = V | a | b
+                                  ^
+                                  the last bit can be zero, we just consume 7
+   code units and recheck this word in the next iteration
+*/
+template <endianness big_endian>
+const result validate_utf16_with_errors(const char16_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    return result(error_code::SUCCESS, 0);
   }
-  buf[64] = '\0';
-  return buf;
-}
 
-template <size_t STEP_SIZE>
-simdutf_really_inline
-buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len)
-    : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE},
-      idx{0} {}
+  const char16_t *start = input;
+  const char16_t *end = input + size;
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() {
-  return idx;
-}
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
 
-template <size_t STEP_SIZE>
-simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
-  return idx < lenminusstep;
-}
+  while (input + simd16<uint16_t>::SIZE * 2 < end) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
 
-template <size_t STEP_SIZE>
-simdutf_really_inline const uint8_t *
-buf_block_reader<STEP_SIZE>::full_block() const {
-  return &buf[idx];
-}
+    // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16
+    // and yields a single vector having only higher bytes of characters.
+    const auto in = utf16_gather_high_bytes<big_endian>(in0, in1);
 
-template <size_t STEP_SIZE>
-simdutf_really_inline size_t
-buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
-  if (len == idx) {
-    return 0;
-  } // memcpy(dst, null, 0) will trigger an error with some sanitizers
-  std::memset(dst, 0x20,
-              STEP_SIZE); // std::memset STEP_SIZE because it is more efficient
-                          // to write out 8 or 16 bytes at once.
-  std::memcpy(dst, buf + idx, len - idx);
-  return len - idx;
-}
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const auto surrogates_wordmask = (in & v_f8) == v_d8;
+    const uint16_t surrogates_bitmask =
+        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
+    if (surrogates_bitmask == 0x0000) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher byte)
 
-template <size_t STEP_SIZE>
-simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
-  idx += STEP_SIZE;
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = (in & v_fc) == v_dc;
+      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
+
+      const uint16_t a = static_cast<uint16_t>(
+          L & (H >> 1)); // A low surrogate must be followed by high one.
+                         // (A low surrogate placed in the 7th register's word
+                         // is an exception we handle.)
+      const uint16_t b = static_cast<uint16_t>(
+          a << 1); // Just mark that the opinput - startite fact is hold,
+                   // thanks to that we have only two masks for valid case.
+      const uint16_t c = static_cast<uint16_t>(
+          V | a | b); // Combine all the masks into the final one.
+
+      if (c == 0xffff) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0x7fff) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return result(error_code::SURROGATE, input - start);
+      }
+    }
+  }
+
+  return result(error_code::SUCCESS, input - start);
 }
 
+} // namespace utf16
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/buf_block_reader.h */
-/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
+/* end file src/generic/validate_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf32.h */
+#include <limits>
+
 namespace simdutf {
 namespace ppc64 {
 namespace {
-namespace utf8_validation {
+namespace utf32 {
 
-using namespace simd;
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
 
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
 
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
+  const char32_t *start = input;
 
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
 
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
+  const size_t N = vector_u32::ELEMENTS;
 
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
+  size_t counter = 0;
 
-//
-// Return nonzero if there are incomplete multibyte characters at the end of the
-// block: e.g. if there is a 4-byte character, but it is 3 bytes from the end.
-//
-simdutf_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
-  // If the previous input's last 3 bytes match this, they're too short (they
-  // ended at EOF):
-  // ... 1111____ 111_____ 11______
-  static const uint8_t max_array[32] = {255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        255,
-                                        0b11110000u - 1,
-                                        0b11100000u - 1,
-                                        0b11000000u - 1};
-  const simd8<uint8_t> max_value(
-      &max_array[sizeof(max_array) - sizeof(simd8<uint8_t>)]);
-  return input.gt_bits(max_value);
-}
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
 
-struct utf8_checker {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
-  // The last input we received
-  simd8<uint8_t> prev_input_block;
-  // Whether the last input we received was incomplete (used for ASCII fast
-  // path)
-  simd8<uint8_t> prev_incomplete;
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
 
-  //
-  // Check whether the current bytes are valid UTF-8.
-  //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
-  }
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 
-  // The only problem that can happen at EOF is that a multibyte character is
-  // too short or a byte value too large in the last bytes: check_special_cases
-  // only checks for bytes too large in the first of two bytes.
-  simdutf_really_inline void check_eof() {
-    // If the previous block had incomplete UTF-8 characters at the end, an
-    // ASCII block can't possibly finish them.
-    this->error |= this->prev_incomplete;
+        input += 4 * N;
+      }
+
+      counter += acc.sum();
+    }
   }
 
-  simdutf_really_inline void check_next_input(const simd8x64<uint8_t> &input) {
-    if (simdutf_likely(is_ascii(input))) {
-      this->error |= this->prev_incomplete;
-    } else {
-      // you might think that a for-loop would work, but under Visual Studio, it
-      // is not good enough.
-      static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                        (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-                    "We support either two or four chunks per 64-byte block.");
-      if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-      } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-        this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-        this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += N;
       }
-      this->prev_incomplete =
-          is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1]);
-      this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS - 1];
+
+      counter += acc.sum();
     }
   }
 
-  // do not forget to call check_eof!
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
   }
 
-}; // struct utf8_checker
-} // namespace utf8_validation
-
-using utf8_validation::utf8_checker;
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
+}
 
+} // namespace utf32
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
-/* begin file src/generic/utf8_validation/utf8_validator.h */
+/* end file src/generic/utf32.h */
+/* begin file src/generic/validate_utf32.h */
 namespace simdutf {
 namespace ppc64 {
 namespace {
-namespace utf8_validation {
+namespace utf32 {
 
-/**
- * Validates that the string is actual UTF-8.
- */
-template <class checker>
-bool generic_validate_utf8(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    reader.advance();
+simdutf_really_inline bool validate(const char32_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return true;
   }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  return !c.errors();
-}
-
-bool generic_validate_utf8(const char *input, size_t length) {
-  return generic_validate_utf8<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-/**
- * Validates that the string is actual UTF-8 and stops on errors.
- */
-template <class checker>
-result generic_validate_utf8_with_errors(const uint8_t *input, size_t length) {
-  checker c{};
-  buf_block_reader<64> reader(input, length);
-  size_t count{0};
-  while (reader.has_full_block()) {
-    simd::simd8x64<uint8_t> in(reader.full_block());
-    c.check_next_input(in);
-    if (c.errors()) {
-      if (count != 0) {
-        count--;
-      } // Sometimes the error is only detected in the next chunk
-      result res = scalar::utf8::rewind_and_validate_with_errors(
-          reinterpret_cast<const char *>(input),
-          reinterpret_cast<const char *>(input + count), length - count);
-      res.count += count;
-      return res;
+
+  const char32_t *end = input + size;
+
+  using vector_u32 = simd32<uint32_t>;
+
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+  auto currentmax = vector_u32::zero();
+  auto currentoffsetmax = vector_u32::zero();
+
+  constexpr size_t N = vector_u32::ELEMENTS;
+
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
     }
-    reader.advance();
-    count += 64;
+
+    currentmax = max(currentmax, in);
+    currentoffsetmax = max(currentoffsetmax, in + offset);
+    input += N;
   }
-  uint8_t block[64]{};
-  reader.get_remainder(block);
-  simd::simd8x64<uint8_t> in(block);
-  c.check_next_input(in);
-  reader.advance();
-  c.check_eof();
-  if (c.errors()) {
-    if (count != 0) {
-      count--;
-    } // Sometimes the error is only detected in the next chunk
-    result res = scalar::utf8::rewind_and_validate_with_errors(
-        reinterpret_cast<const char *>(input),
-        reinterpret_cast<const char *>(input) + count, length - count);
-    res.count += count;
-    return res;
-  } else {
-    return result(error_code::SUCCESS, length);
+
+  const auto too_large = currentmax > standardmax;
+  if (too_large.any()) {
+    return false;
   }
+
+  const auto surrogate = currentoffsetmax > standardoffsetmax;
+  if (surrogate.any()) {
+    return false;
+  }
+
+  return scalar::utf32::validate(input, end - input);
 }
 
-result generic_validate_utf8_with_errors(const char *input, size_t length) {
-  return generic_validate_utf8_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
+simdutf_really_inline result validate_with_errors(const char32_t *input,
+                                                  size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return result(error_code::SUCCESS, 0);
+  }
+
+  const char32_t *start = input;
+  const char32_t *end = input + size;
+
+  using vector_u32 = simd32<uint32_t>;
+
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+
+  constexpr size_t N = vector_u32::ELEMENTS;
+
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
+    }
+
+    const auto too_large = in > standardmax;
+    const auto surrogate = (in + offset) > standardoffsetmax;
+
+    const auto combined = too_large | surrogate;
+    if (simdutf_unlikely(combined.any())) {
+      const size_t consumed = input - start;
+      auto sr = scalar::utf32::validate_with_errors(input, end - input);
+      sr.count += consumed;
+
+      return sr;
+    }
+
+    input += N;
+  }
+
+  const size_t consumed = input - start;
+  auto sr = scalar::utf32::validate_with_errors(input, end - input);
+  sr.count += consumed;
+
+  return sr;
 }
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+} // namespace utf32
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdutf
+/* end file src/generic/validate_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
+namespace simdutf {
+namespace ppc64 {
+namespace {
+namespace ascii_validation {
+
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   uint8_t blocks[64]{};
   simd::simd8x64<uint8_t> running_or(blocks);
   while (reader.has_full_block()) {
@@ -35203,14 +43776,8 @@ bool generic_validate_ascii(const uint8_t *input, size_t length) {
   return running_or.is_ascii();
 }
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   size_t count{0};
   while (reader.has_full_block()) {
     simd::simd8x64<uint8_t> in(reader.full_block());
@@ -35235,440 +43802,27 @@ result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
   }
 }
 
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-} // namespace utf8_validation
+} // namespace ascii_validation
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
-// transcoding from UTF-8 to UTF-16
-/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 namespace simdutf {
 namespace ppc64 {
 namespace {
-namespace utf8_to_utf16 {
+namespace utf8_to_latin1 {
 using namespace simd;
 
 simdutf_really_inline simd8<uint8_t>
 check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
-  // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
-  // Bit 1 = Too Long (ASCII followed by continuation)
-  // Bit 2 = Overlong 3-byte
-  // Bit 4 = Surrogate
-  // Bit 5 = Overlong 2-byte
-  // Bit 7 = Two Continuations
-  constexpr const uint8_t TOO_SHORT = 1 << 0;  // 11______ 0_______
-                                               // 11______ 11______
-  constexpr const uint8_t TOO_LONG = 1 << 1;   // 0_______ 10______
-  constexpr const uint8_t OVERLONG_3 = 1 << 2; // 11100000 100_____
-  constexpr const uint8_t SURROGATE = 1 << 4;  // 11101101 101_____
-  constexpr const uint8_t OVERLONG_2 = 1 << 5; // 1100000_ 10______
-  constexpr const uint8_t TWO_CONTS = 1 << 7;  // 10______ 10______
-  constexpr const uint8_t TOO_LARGE = 1 << 3;  // 11110100 1001____
-                                               // 11110100 101_____
-                                               // 11110101 1001____
-                                               // 11110101 101_____
-                                               // 1111011_ 1001____
-                                               // 1111011_ 101_____
-                                               // 11111___ 1001____
-                                               // 11111___ 101_____
-  constexpr const uint8_t TOO_LARGE_1000 = 1 << 6;
-  // 11110101 1000____
-  // 1111011_ 1000____
-  // 11111___ 1000____
-  constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
-
-  const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
-      // 0_______ ________ <ASCII in byte 1>
-      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
-      TOO_LONG,
-      // 10______ ________ <continuation in byte 1>
-      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
-      // 1100____ ________ <two byte lead in byte 1>
-      TOO_SHORT | OVERLONG_2,
-      // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
-      // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
-      // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
-  constexpr const uint8_t CARRY =
-      TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
-  const simd8<uint8_t> byte_1_low =
-      (prev1 & 0x0F)
-          .lookup_16<uint8_t>(
-              // ____0000 ________
-              CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
-              // ____0001 ________
-              CARRY | OVERLONG_2,
-              // ____001_ ________
-              CARRY, CARRY,
-
-              // ____0100 ________
-              CARRY | TOO_LARGE,
-              // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-
-              // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
-  const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
-      // ________ 0_______ <ASCII in byte 2>
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
-      TOO_SHORT, TOO_SHORT,
-
-      // ________ 1000____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 |
-          OVERLONG_4,
-      // ________ 1001____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
-      // ________ 101_____
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE | TOO_LARGE,
-
-      // ________ 11______
-      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
-  return (byte_1_high & byte_1_low & byte_2_high);
-}
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
-
-struct validating_transcoder {
-  // If this is nonzero, there has been a UTF-8 error.
-  simd8<uint8_t> error;
-
-  validating_transcoder() : error(uint8_t(0)) {}
-  //
-  // Check whether the current bytes are valid UTF-8.
+  // For UTF-8 to Latin 1, we can allow any ASCII character, and any
+  // continuation byte, but the non-ASCII leading bytes must be 0b11000011 or
+  // 0b11000010 and nothing else.
   //
-  simdutf_really_inline void check_utf8_bytes(const simd8<uint8_t> input,
-                                              const simd8<uint8_t> prev_input) {
-    // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+
-    // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
-    // small negative numbers)
-    simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
-  }
-
-  template <endianness endian>
-  simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (utf8_continuation_mask & 1) {
-          return 0; // error
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      return 0;
-    }
-    if (pos < size) {
-      size_t howmany = scalar::utf8_to_utf16::convert<endian>(
-          in + pos, size - pos, utf16_output);
-      if (howmany == 0) {
-        return 0;
-      }
-      utf16_output += howmany;
-    }
-    return utf16_output - start;
-  }
-
-  template <endianness endian>
-  simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char16_t *utf16_output) {
-    size_t pos = 0;
-    char16_t *start{utf16_output};
-    // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the
-    // last 16 bytes, and if the data is valid, then it is entirely safe because
-    // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
-    // generally assume that you have valid UTF-8 input, so we are going to go
-    // back from the end counting 8 leading bytes, to give us a good margin.
-    size_t leading_byte = 0;
-    size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
-    }
-    // If the input is long enough, then we have that margin-1 is the eight last
-    // leading byte.
-    const size_t safety_margin = size - margin + 1; // to avoid overruns!
-    while (pos + 64 + safety_margin <= size) {
-      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-      if (input.is_ascii()) {
-        input.store_ascii_as_utf16<endian>(utf16_output);
-        utf16_output += 64;
-        pos += 64;
-      } else {
-        // you might think that a for-loop would work, but under Visual Studio,
-        // it is not good enough.
-        static_assert(
-            (simd8x64<uint8_t>::NUM_CHUNKS == 2) ||
-                (simd8x64<uint8_t>::NUM_CHUNKS == 4),
-            "We support either two or four chunks per 64-byte block.");
-        auto zero = simd8<uint8_t>{uint8_t(0)};
-        if (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
-          this->check_utf8_bytes(input.chunks[0], zero);
-          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
-          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
-        }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          // rewind_and_convert_with_errors will seek a potential error from
-          // in+pos onward, with the ability to go back up to pos bytes, and
-          // read size-pos bytes forward.
-          result res =
-              scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-                  pos, in + pos, size - pos, utf16_output);
-          res.count += pos;
-          return res;
-        }
-        uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-        uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-        // We process in blocks of up to 12 bytes except possibly
-        // for fast paths which may process up to 16 bytes. For the
-        // slow path to work, we should have at least 12 input bytes left.
-        size_t max_starting_point = (pos + 64) - 12;
-        // Next loop is going to run at least five times.
-        while (pos < max_starting_point) {
-          // Performance note: our ability to compute 'consumed' and
-          // then shift and recompute is critical. If there is a
-          // latency of, say, 4 cycles on getting 'consumed', then
-          // the inner loop might have a total latency of about 6 cycles.
-          // Yet we process between 6 to 12 inputs bytes, thus we get
-          // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-          // for this section of the code. Hence, there is a limit
-          // to how much we can further increase this latency before
-          // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf16<endian>(
-              in + pos, utf8_end_of_code_point_mask, utf16_output);
-          pos += consumed;
-          utf8_end_of_code_point_mask >>= consumed;
-        }
-        // At this point there may remain between 0 and 12 bytes in the
-        // 64-byte block. These bytes will be processed again. So we have an
-        // 80% efficiency (in the worst case). In practice we expect an
-        // 85% to 90% efficiency.
-      }
-    }
-    if (errors()) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      res.count += pos;
-      return res;
-    }
-    if (pos < size) {
-      // rewind_and_convert_with_errors will seek a potential error from in+pos
-      // onward, with the ability to go back up to pos bytes, and read size-pos
-      // bytes forward.
-      result res =
-          scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(
-              pos, in + pos, size - pos, utf16_output);
-      if (res.error) { // In case of error, we want the error position
-        res.count += pos;
-        return res;
-      } else { // In case of success, we want the number of word written
-        utf16_output += res.count;
-      }
-    }
-    return result(error_code::SUCCESS, utf16_output - start);
-  }
-
-  simdutf_really_inline bool errors() const {
-    return this->error.any_bits_set_anywhere();
-  }
-
-}; // struct utf8_checker
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
-
-namespace simdutf {
-namespace ppc64 {
-namespace {
-namespace utf8_to_utf16 {
-
-using namespace simd;
-
-template <endianness endian>
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char16_t *utf16_output) noexcept {
-  // The implementation is not specific to haswell and should be moved to the
-  // generic directory.
-  size_t pos = 0;
-  char16_t *start{utf16_output};
-  const size_t safety_margin = 16; // to avoid overruns!
-  while (pos + 64 + safety_margin <= size) {
-    // this loop could be unrolled further. For example, we could process the
-    // mask far more than 64 bytes.
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf16<endian>(utf16_output);
-      utf16_output += 64;
-      pos += 64;
-    } else {
-      // Slow path. We hope that the compiler will recognize that this is a slow
-      // path. Anything that is not a continuation mask is a 'leading byte',
-      // that is, the start of a new code point.
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
-      // The *start* of code points is not so useful, rather, we want the *end*
-      // of code points.
-      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
-      // We process in blocks of up to 12 bytes except possibly
-      // for fast paths which may process up to 16 bytes. For the
-      // slow path to work, we should have at least 12 input bytes left.
-      size_t max_starting_point = (pos + 64) - 12;
-      // Next loop is going to run at least five times when using solely
-      // the slow/regular path, and at least four times if there are fast paths.
-      while (pos < max_starting_point) {
-        // Performance note: our ability to compute 'consumed' and
-        // then shift and recompute is critical. If there is a
-        // latency of, say, 4 cycles on getting 'consumed', then
-        // the inner loop might have a total latency of about 6 cycles.
-        // Yet we process between 6 to 12 inputs bytes, thus we get
-        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
-        // for this section of the code. Hence, there is a limit
-        // to how much we can further increase this latency before
-        // it seriously harms performance.
-        //
-        // Thus we may allow convert_masked_utf8_to_utf16 to process
-        // more bytes at a time under a fast-path mode where 16 bytes
-        // are consumed at once (e.g., when encountering ASCII).
-        size_t consumed = convert_masked_utf8_to_utf16<endian>(
-            input + pos, utf8_end_of_code_point_mask, utf16_output);
-        pos += consumed;
-        utf8_end_of_code_point_mask >>= consumed;
-      }
-      // At this point there may remain between 0 and 12 bytes in the
-      // 64-byte block. These bytes will be processed again. So we have an
-      // 80% efficiency (in the worst case). In practice we expect an
-      // 85% to 90% efficiency.
-    }
-  }
-  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(
-      input + pos, size - pos, utf16_output);
-  return utf16_output - start;
-}
-
-} // namespace utf8_to_utf16
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdutf
-/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
-/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-
-namespace simdutf {
-namespace ppc64 {
-namespace {
-namespace utf8_to_utf32 {
-using namespace simd;
-
-simdutf_really_inline simd8<uint8_t>
-check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
   // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
   // Bit 1 = Too Long (ASCII followed by continuation)
   // Bit 2 = Overlong 3-byte
@@ -35695,6 +43849,7 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
   // 1111011_ 1000____
   // 11111___ 1000____
   constexpr const uint8_t OVERLONG_4 = 1 << 6; // 11110000 1000____
+  constexpr const uint8_t FORBIDDEN = 0xff;
 
   const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
       // 0_______ ________ <ASCII in byte 1>
@@ -35705,11 +43860,11 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
       // 1100____ ________ <two byte lead in byte 1>
       TOO_SHORT | OVERLONG_2,
       // 1101____ ________ <two byte lead in byte 1>
-      TOO_SHORT,
+      FORBIDDEN,
       // 1110____ ________ <three byte lead in byte 1>
-      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      FORBIDDEN,
       // 1111____ ________ <four+ byte lead in byte 1>
-      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4);
+      FORBIDDEN);
   constexpr const uint8_t CARRY =
       TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
   const simd8<uint8_t> byte_1_low =
@@ -35723,23 +43878,16 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
               CARRY, CARRY,
 
               // ____0100 ________
-              CARRY | TOO_LARGE,
+              FORBIDDEN,
               // ____0101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              FORBIDDEN,
               // ____011_ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              FORBIDDEN, FORBIDDEN,
 
               // ____1___ ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
+              FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN, FORBIDDEN,
               // ____1101 ________
-              CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
-              CARRY | TOO_LARGE | TOO_LARGE_1000,
-              CARRY | TOO_LARGE | TOO_LARGE_1000);
+              FORBIDDEN, FORBIDDEN, FORBIDDEN);
   const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
       // ________ 0_______ <ASCII in byte 2>
       TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
@@ -35758,17 +43906,6 @@ check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
       TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT);
   return (byte_1_high & byte_1_low & byte_2_high);
 }
-simdutf_really_inline simd8<uint8_t>
-check_multibyte_lengths(const simd8<uint8_t> input,
-                        const simd8<uint8_t> prev_input,
-                        const simd8<uint8_t> sc) {
-  simd8<uint8_t> prev2 = input.prev<2>(prev_input);
-  simd8<uint8_t> prev3 = input.prev<3>(prev_input);
-  simd8<uint8_t> must23 =
-      simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
-  simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
-  return must23_80 ^ sc;
-}
 
 struct validating_transcoder {
   // If this is nonzero, there has been a UTF-8 error.
@@ -35784,33 +43921,33 @@ struct validating_transcoder {
     // lead bytes (2, 3, 4-byte leads become large positive numbers instead of
     // small negative numbers)
     simd8<uint8_t> prev1 = input.prev<1>(prev_input);
-    simd8<uint8_t> sc = check_special_cases(input, prev1);
-    this->error |= check_multibyte_lengths(input, prev_input, sc);
+    this->error |= check_special_cases(input, prev1);
   }
 
   simdutf_really_inline size_t convert(const char *in, size_t size,
-                                       char32_t *utf32_output) {
+                                       char *latin1_output) {
     size_t pos = 0;
-    char32_t *start{utf32_output};
+    char *start{latin1_output};
     // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 words when calling convert_masked_utf8_to_utf32. If you skip the
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
     // last 16 bytes, and if the data is valid, then it is entirely safe because
     // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
     // generally assume that you have valid UTF-8 input, so we are going to go
     // back from the end counting 16 leading bytes, to give us a good margin.
     size_t leading_byte = 0;
     size_t margin = size;
-    for (; margin > 0 && leading_byte < 8; margin--) {
-      leading_byte += (int8_t(in[margin - 1]) > -65);
+    for (; margin > 0 && leading_byte < 16; margin--) {
+      leading_byte += (int8_t(in[margin - 1]) >
+                       -65); // twos complement of -65 is 1011 1111 ...
     }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
     const size_t safety_margin = size - margin + 1; // to avoid overruns!
     while (pos + 64 + safety_margin <= size) {
       simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
       if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
         pos += 64;
       } else {
         // you might think that a for-loop would work, but under Visual Studio,
@@ -35829,9 +43966,11 @@ struct validating_transcoder {
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
           this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
         }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
+        uint64_t utf8_continuation_mask =
+            input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                               // this case, we also have ASCII to account for.
         if (utf8_continuation_mask & 1) {
-          return 0; // we have an error
+          return 0; // error
         }
         uint64_t utf8_leading_mask = ~utf8_continuation_mask;
         uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
@@ -35850,8 +43989,8 @@ struct validating_transcoder {
           // for this section of the code. Hence, there is a limit
           // to how much we can further increase this latency before
           // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
           pos += consumed;
           utf8_end_of_code_point_mask >>= consumed;
         }
@@ -35866,21 +44005,21 @@ struct validating_transcoder {
     }
     if (pos < size) {
       size_t howmany =
-          scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
+          scalar::utf8_to_latin1::convert(in + pos, size - pos, latin1_output);
       if (howmany == 0) {
         return 0;
       }
-      utf32_output += howmany;
+      latin1_output += howmany;
     }
-    return utf32_output - start;
+    return latin1_output - start;
   }
 
   simdutf_really_inline result convert_with_errors(const char *in, size_t size,
-                                                   char32_t *utf32_output) {
+                                                   char *latin1_output) {
     size_t pos = 0;
-    char32_t *start{utf32_output};
+    char *start{latin1_output};
     // In the worst case, we have the haswell kernel which can cause an overflow
-    // of 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the
+    // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the
     // last 16 bytes, and if the data is valid, then it is entirely safe because
     // 16 UTF-8 bytes generate much more than 8 bytes. However, you cannot
     // generally assume that you have valid UTF-8 input, so we are going to go
@@ -35890,14 +44029,14 @@ struct validating_transcoder {
     for (; margin > 0 && leading_byte < 8; margin--) {
       leading_byte += (int8_t(in[margin - 1]) > -65);
     }
-    // If the input is long enough, then we have that margin-1 is the fourth
-    // last leading byte.
+    // If the input is long enough, then we have that margin-1 is the eight last
+    // leading byte.
     const size_t safety_margin = size - margin + 1; // to avoid overruns!
     while (pos + 64 + safety_margin <= size) {
       simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
       if (input.is_ascii()) {
-        input.store_ascii_as_utf32(utf32_output);
-        utf32_output += 64;
+        input.store((int8_t *)latin1_output);
+        latin1_output += 64;
         pos += 64;
       } else {
         // you might think that a for-loop would work, but under Visual Studio,
@@ -35916,13 +44055,16 @@ struct validating_transcoder {
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
           this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
         }
-        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-        if (errors() || (utf8_continuation_mask & 1)) {
-          result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-              pos, in + pos, size - pos, utf32_output);
+        if (errors()) {
+          // rewind_and_convert_with_errors will seek a potential error from
+          // in+pos onward, with the ability to go back up to pos bytes, and
+          // read size-pos bytes forward.
+          result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+              pos, in + pos, size - pos, latin1_output);
           res.count += pos;
           return res;
         }
+        uint64_t utf8_continuation_mask = input.lt(-65 + 1);
         uint64_t utf8_leading_mask = ~utf8_continuation_mask;
         uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
         // We process in blocks of up to 12 bytes except possibly
@@ -35940,8 +44082,8 @@ struct validating_transcoder {
           // for this section of the code. Hence, there is a limit
           // to how much we can further increase this latency before
           // it seriously harms performance.
-          size_t consumed = convert_masked_utf8_to_utf32(
-              in + pos, utf8_end_of_code_point_mask, utf32_output);
+          size_t consumed = convert_masked_utf8_to_latin1(
+              in + pos, utf8_end_of_code_point_mask, latin1_output);
           pos += consumed;
           utf8_end_of_code_point_mask >>= consumed;
         }
@@ -35952,22 +44094,28 @@ struct validating_transcoder {
       }
     }
     if (errors()) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
       res.count += pos;
       return res;
     }
     if (pos < size) {
-      result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(
-          pos, in + pos, size - pos, utf32_output);
+      // rewind_and_convert_with_errors will seek a potential error from in+pos
+      // onward, with the ability to go back up to pos bytes, and read size-pos
+      // bytes forward.
+      result res = scalar::utf8_to_latin1::rewind_and_convert_with_errors(
+          pos, in + pos, size - pos, latin1_output);
       if (res.error) { // In case of error, we want the error position
         res.count += pos;
         return res;
       } else { // In case of success, we want the number of word written
-        utf32_output += res.count;
+        latin1_output += res.count;
       }
     }
-    return result(error_code::SUCCESS, utf32_output - start);
+    return result(error_code::SUCCESS, latin1_output - start);
   }
 
   simdutf_really_inline bool errors() const {
@@ -35975,180 +44123,449 @@ struct validating_transcoder {
   }
 
 }; // struct utf8_checker
-} // namespace utf8_to_utf32
+} // namespace utf8_to_latin1
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-
+/* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+/* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
 namespace simdutf {
 namespace ppc64 {
 namespace {
-namespace utf8_to_utf32 {
-
+namespace utf8_to_latin1 {
 using namespace simd;
 
-simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
-                                         char32_t *utf32_output) noexcept {
+simdutf_really_inline size_t convert_valid(const char *in, size_t size,
+                                           char *latin1_output) {
   size_t pos = 0;
-  char32_t *start{utf32_output};
-  const size_t safety_margin = 16; // to avoid overruns!
+  char *start{latin1_output};
+  // In the worst case, we have the haswell kernel which can cause an overflow
+  // of 8 bytes when calling convert_masked_utf8_to_latin1. If you skip the last
+  // 16 bytes, and if the data is valid, then it is entirely safe because 16
+  // UTF-8 bytes generate much more than 8 bytes. However, you cannot generally
+  // assume that you have valid UTF-8 input, so we are going to go back from the
+  // end counting 8 leading bytes, to give us a good margin.
+  size_t leading_byte = 0;
+  size_t margin = size;
+  for (; margin > 0 && leading_byte < 8; margin--) {
+    leading_byte += (int8_t(in[margin - 1]) >
+                     -65); // twos complement of -65 is 1011 1111 ...
+  }
+  // If the input is long enough, then we have that margin-1 is the eight last
+  // leading byte.
+  const size_t safety_margin = size - margin + 1; // to avoid overruns!
   while (pos + 64 + safety_margin <= size) {
-    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
-    if (in.is_ascii()) {
-      in.store_ascii_as_utf32(utf32_output);
-      utf32_output += 64;
+    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
+    if (input.is_ascii()) {
+      input.store((int8_t *)latin1_output);
+      latin1_output += 64;
       pos += 64;
     } else {
-      // -65 is 0b10111111 in two-complement's, so largest possible continuation
-      // byte
-      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
+      // you might think that a for-loop would work, but under Visual Studio, it
+      // is not good enough.
+      uint64_t utf8_continuation_mask =
+          input.lt(-65 + 1); // -64 is 1100 0000 in twos complement. Note: in
+                             // this case, we also have ASCII to account for.
       uint64_t utf8_leading_mask = ~utf8_continuation_mask;
       uint64_t utf8_end_of_code_point_mask = utf8_leading_mask >> 1;
+      // We process in blocks of up to 12 bytes except possibly
+      // for fast paths which may process up to 16 bytes. For the
+      // slow path to work, we should have at least 12 input bytes left.
       size_t max_starting_point = (pos + 64) - 12;
+      // Next loop is going to run at least five times.
       while (pos < max_starting_point) {
-        size_t consumed = convert_masked_utf8_to_utf32(
-            input + pos, utf8_end_of_code_point_mask, utf32_output);
+        // Performance note: our ability to compute 'consumed' and
+        // then shift and recompute is critical. If there is a
+        // latency of, say, 4 cycles on getting 'consumed', then
+        // the inner loop might have a total latency of about 6 cycles.
+        // Yet we process between 6 to 12 inputs bytes, thus we get
+        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
+        // for this section of the code. Hence, there is a limit
+        // to how much we can further increase this latency before
+        // it seriously harms performance.
+        size_t consumed = convert_masked_utf8_to_latin1(
+            in + pos, utf8_end_of_code_point_mask, latin1_output);
         pos += consumed;
         utf8_end_of_code_point_mask >>= consumed;
       }
+      // At this point there may remain between 0 and 12 bytes in the
+      // 64-byte block. These bytes will be processed again. So we have an
+      // 80% efficiency (in the worst case). In practice we expect an
+      // 85% to 90% efficiency.
     }
   }
-  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos,
-                                                       utf32_output);
-  return utf32_output - start;
+  if (pos < size) {
+    size_t howmany = scalar::utf8_to_latin1::convert_valid(in + pos, size - pos,
+                                                           latin1_output);
+    latin1_output += howmany;
+  }
+  return latin1_output - start;
 }
 
-} // namespace utf8_to_utf32
-} // unnamed namespace
+} // namespace utf8_to_latin1
+} // namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-// other functions
-/* begin file src/generic/utf16.h */
+  // namespace simdutf
+/* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/generic/base64.h */
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
 namespace simdutf {
 namespace ppc64 {
 namespace {
-namespace utf16 {
+namespace base64 {
 
-template <endianness big_endian>
-simdutf_really_inline size_t count_code_points(const char16_t *in,
-                                               size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
+/*
+    The following template function implements API for Base64 decoding.
+
+    An implementation is responsible for providing the `block64` type and
+    associated methods that perform actual conversion. Please refer
+    to any vectorized implementation to learn the API of these procedures.
+*/
+template <bool base64_url, bool ignore_garbage, typename chartype>
+full_result
+compress_decode_base64(char *dst, const chartype *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
+                                        : tables::base64::to_base64_value;
+  size_t equallocation =
+      srclen; // location of the first padding character if any
+  // skip trailing spaces
+  while (!ignore_garbage && srclen > 0 &&
+         scalar::base64::is_eight_byte(src[srclen - 1]) &&
+         to_base64[uint8_t(src[srclen - 1])] == 64) {
+    srclen--;
+  }
+  size_t equalsigns = 0;
+  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
+    equallocation = srclen - 1;
+    srclen--;
+    equalsigns = 1;
+    // skip trailing spaces
+    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+           to_base64[uint8_t(src[srclen - 1])] == 64) {
+      srclen--;
+    }
+    if (srclen > 0 && src[srclen - 1] == '=') {
+      equallocation = srclen - 1;
+      srclen--;
+      equalsigns = 2;
     }
-    uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
-    count += count_ones(not_pair) / 2;
   }
-  return count +
-         scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
-}
-
-template <endianness big_endian>
-simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos < size / 32 * 32; pos += 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    if (!match_system(big_endian)) {
-      input.swap_bytes();
+  if (srclen == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
     }
-    uint64_t ascii_mask = input.lteq(0x7F);
-    uint64_t twobyte_mask = input.lteq(0x7FF);
-    uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);
+    return {SUCCESS, 0, 0};
+  }
+  char *end_of_safe_64byte_zone =
+      (srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 : dst;
 
-    size_t ascii_count = count_ones(ascii_mask) / 2;
-    size_t twobyte_count = count_ones(twobyte_mask & ~ascii_mask) / 2;
-    size_t threebyte_count = count_ones(not_pair_mask & ~twobyte_mask) / 2;
-    size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
-    count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count +
-             ascii_count;
+  const chartype *const srcinit = src;
+  const char *const dstinit = dst;
+  const chartype *const srcend = src + srclen;
+
+  constexpr size_t block_size = 6;
+  static_assert(block_size >= 2, "block_size must be at least two");
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const chartype *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b(src);
+      src += 64;
+      uint64_t error = 0;
+      const uint64_t badcharmask =
+          b.to_base64_mask<base64_url, ignore_garbage>(&error);
+      if (!ignore_garbage && error) {
+        src -= 64;
+        const size_t error_offset = trailing_zeroes(error);
+        return {error_code::INVALID_BASE64_CHARACTER,
+                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+      }
+      if (badcharmask != 0) {
+        bufferptr += b.compress_block(badcharmask, bufferptr);
+      } else if (bufferptr != buffer) {
+        b.copy_block(bufferptr);
+        bufferptr += 64;
+      } else {
+        if (dst >= end_of_safe_64byte_zone) {
+          b.base64_decode_block_safe(dst);
+        } else {
+          b.base64_decode_block(dst);
+        }
+        dst += 48;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 2); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        if (dst >= end_of_safe_64byte_zone) {
+          base64_decode_block_safe(dst, buffer + (block_size - 2) * 64);
+        } else {
+          base64_decode_block(dst, buffer + (block_size - 2) * 64);
+        }
+        dst += 48;
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
   }
-  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
-                                                                   size - pos);
-}
 
-template <endianness big_endian>
-simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
-                                                     size_t size) {
-  return count_code_points<big_endian>(in, size);
-}
+  char *buffer_start = buffer;
+  // Optimization note: if this is almost full, then it is worth our
+  // time, otherwise, we should just decode directly.
+  int last_block = (int)((bufferptr - buffer_start) % 64);
+  if (last_block != 0 && srcend - src + last_block >= 64) {
 
-simdutf_really_inline void
-change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
-  size_t pos = 0;
+    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
+      uint8_t val = to_base64[uint8_t(*src)];
+      *bufferptr = char(val);
+      if (!ignore_garbage &&
+          (!scalar::base64::is_eight_byte(*src) || val > 64)) {
+        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      }
+      bufferptr += (val <= 63);
+      src++;
+    }
+  }
 
-  while (pos < size / 32 * 32) {
-    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
-    input.swap_bytes();
-    input.store(reinterpret_cast<uint16_t *>(output));
-    pos += 32;
-    output += 32;
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    if (dst >= end_of_safe_64byte_zone) {
+      base64_decode_block_safe(dst, buffer_start);
+    } else {
+      base64_decode_block(dst, buffer_start);
+    }
+    dst += 48;
   }
+  if ((bufferptr - buffer_start) % 64 != 0) {
+    while (buffer_start + 4 < bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
 
-  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
+      dst += 3;
+      buffer_start += 4;
+    }
+    if (buffer_start + 4 <= bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
+
+      dst += 3;
+      buffer_start += 4;
+    }
+    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
+    // backtrack
+    int leftover = int(bufferptr - buffer_start);
+    while (leftover > 0) {
+      if (!ignore_garbage) {
+        while (to_base64[uint8_t(*(src - 1))] == 64) {
+          src--;
+        }
+      } else {
+        while (to_base64[uint8_t(*(src - 1))] >= 64) {
+          src--;
+        }
+      }
+      src--;
+      leftover--;
+    }
+  }
+  if (src < srcend + equalsigns) {
+    full_result r = scalar::base64::base64_tail_decode(
+        dst, src, srcend - src, equalsigns, options, last_chunk_options);
+    r.input_count += size_t(src - srcinit);
+    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
+        r.error == error_code::BASE64_EXTRA_BITS) {
+      return r;
+    } else {
+      r.output_count += size_t(dst - dstinit);
+    }
+    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
+        r.error == error_code::SUCCESS && equalsigns > 0) {
+      // additional checks
+      if ((r.output_count % 3 == 0) ||
+          ((r.output_count % 3) + 1 + equalsigns != 4)) {
+        r.error = error_code::INVALID_BASE64_CHARACTER;
+        r.input_count = equallocation;
+      }
+    }
+    return r;
+  }
+  if (!ignore_garbage && equalsigns > 0) {
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
 }
 
-} // namespace utf16
+} // namespace base64
 } // unnamed namespace
 } // namespace ppc64
 } // namespace simdutf
-/* end file src/generic/utf16.h */
-/* begin file src/generic/utf8.h */
+/* end file src/generic/base64.h */
+#endif // SIMDUTF_FEATURE_BASE64
 
-namespace simdutf {
-namespace ppc64 {
-namespace {
-namespace utf8 {
+/* begin file src/ppc64/templates.cpp */
+/*
+    Template `convert_impl` implements generic conversion routine between
+    different encodings. Procedure returns the number of written elements,
+    or zero in the case of error.
+
+    Parameters:
+    * VectorizedConvert - vectorized procedure that returns structure having
+      three fields: error_code (err), const Source* (input), Destination*
+   (output)
+    * ScalarConvert - scalar procedure that carries on conversion of tail
+    * Source - type of input char (like char16_t, char)
+    * Destination - type of input char
+*/
+template <typename VectorizedConvert, typename ScalarConvert, typename Source,
+          typename Destination>
+size_t convert_impl(VectorizedConvert vectorized_convert,
+                    ScalarConvert scalar_convert, const Source *buf, size_t len,
+                    Destination *output) {
+  const auto vr = vectorized_convert(buf, len, output);
+  const size_t consumed = vr.input - buf;
+  const size_t written = vr.output - output;
+  if (vr.err != simdutf::error_code::SUCCESS) {
+    if (vr.err == simdutf::error_code::OTHER) {
+      // Vectorized procedure detected an error, but does not know
+      // exact position. The scalar procedure rescan the portion of
+      // input and figure out where the error is located.
+      return scalar_convert(vr.input, len - consumed, vr.output);
+    }
+    return 0;
+  }
 
-using namespace simd;
+  if (consumed == len) {
+    return written;
+  }
 
-simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.gt(-65);
-    count += count_ones(utf8_continuation_mask);
+  const auto ret = scalar_convert(vr.input, len - consumed, vr.output);
+  if (ret == 0) {
+    return 0;
   }
-  return count + scalar::utf8::count_code_points(in + pos, size - pos);
+
+  return written + ret;
 }
 
-simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
-                                                    size_t size) {
-  size_t pos = 0;
-  size_t count = 0;
-  // This algorithm could no doubt be improved!
-  for (; pos + 64 <= size; pos += 64) {
-    simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
-    uint64_t utf8_continuation_mask = input.lt(-65 + 1);
-    // We count one word for anything that is not a continuation (so
-    // leading bytes).
-    count += 64 - count_ones(utf8_continuation_mask);
-    int64_t utf8_4byte = input.gteq_unsigned(240);
-    count += count_ones(utf8_4byte);
+/*
+    Template `convert_with_errors_impl` implements generic conversion routine
+    between different encodings. Procedure returns a `result` instance ---
+    please refer to its documentation for details.
+
+    Parameters:
+    * VectorizedConvert - vectorized procedure that returns structure having
+      three fields: error_code (err), const Source* (input), Destination*
+   (output)
+    * ScalarConvert - scalar procedure that carries on conversion of tail
+    * Source - type of input char (like char16_t, char)
+    * Destination - type of input char
+*/
+template <typename VectorizedConvert, typename ScalarConvert, typename Source,
+          typename Destination>
+simdutf::result convert_with_errors_impl(VectorizedConvert vectorized_convert,
+                                         ScalarConvert scalar_convert,
+                                         const Source *buf, size_t len,
+                                         Destination *output) {
+
+  const auto vr = vectorized_convert(buf, len, output);
+  const size_t consumed = vr.input - buf;
+  const size_t written = vr.output - output;
+  if (vr.err != simdutf::error_code::SUCCESS) {
+    if (vr.err == simdutf::error_code::OTHER) {
+      // Vectorized procedure detected an error, but does not know
+      // exact position. The scalar procedure rescan the portion of
+      // input and figure out where the error is located.
+      auto sr = scalar_convert(vr.input, len - consumed, vr.output);
+      sr.count += consumed;
+      return sr;
+    }
+    return simdutf::result(vr.err, consumed);
+  }
+
+  if (consumed == len) {
+    return simdutf::result(simdutf::error_code::SUCCESS, written);
+  }
+
+  simdutf::result sr = scalar_convert(vr.input, len - consumed, vr.output);
+  if (sr.is_ok()) {
+    sr.count += written;
+  } else {
+    sr.count += consumed;
   }
-  return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
+
+  return sr;
 }
-} // namespace utf8
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdutf
-/* end file src/generic/utf8.h */
+/* end file src/ppc64/templates.cpp */
 
+#ifdef SIMDUTF_INTERNAL_TESTS
+  #if SIMDUTF_FEATURE_BASE64
+    #include "ppc64_base64_internal_tests.cpp"
+  #endif // SIMDUTF_FEATURE_BASE64
+#endif   // SIMDUTF_INTERNAL_TESTS
 //
 // Implementation-specific overrides
 //
 namespace simdutf {
 namespace ppc64 {
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused int
 implementation::detect_encodings(const char *input,
                                  size_t length) const noexcept {
@@ -36157,13 +44574,14 @@ implementation::detect_encodings(const char *input,
   if (bom_encoding != encoding_type::unspecified) {
     return bom_encoding;
   }
-  // todo: reimplement as a one-pass algorithm.
   int out = 0;
+  // todo: reimplement as a one-pass algorithm.
   if (validate_utf8(input, length)) {
     out |= encoding_type::UTF8;
   }
   if ((length % 2) == 0) {
-    if (validate_utf16(reinterpret_cast<const char16_t *>(input), length / 2)) {
+    if (validate_utf16le(reinterpret_cast<const char16_t *>(input),
+                         length / 2)) {
       out |= encoding_type::UTF16_LE;
     }
   }
@@ -36172,301 +44590,571 @@ implementation::detect_encodings(const char *input,
       out |= encoding_type::UTF32_LE;
     }
   }
-
   return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf8(const char *buf, size_t len) const noexcept {
   return ppc64::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused result implementation::validate_utf8_with_errors(
     const char *buf, size_t len) const noexcept {
   return ppc64::utf8_validation::generic_validate_utf8_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused bool
 implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return ppc64::utf8_validation::generic_validate_ascii(buf, len);
+  return ppc64::ascii_validation::generic_validate_ascii(buf, len);
 }
 
 simdutf_warn_unused result implementation::validate_ascii_with_errors(
     const char *buf, size_t len) const noexcept {
-  return ppc64::utf8_validation::generic_validate_ascii_with_errors(buf, len);
+  return ppc64::ascii_validation::generic_validate_ascii_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf16le(const char16_t *buf,
                                  size_t len) const noexcept {
-  return scalar::utf16::validate<endianness::LITTLE>(buf, len);
+  const auto res =
+      ppc64::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.is_err()) {
+    return false;
+  }
+
+  if (res.count != len) {
+    return scalar::utf16::validate<endianness::LITTLE>(buf + res.count,
+                                                       len - res.count);
+  }
+
+  return true;
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused bool
 implementation::validate_utf16be(const char16_t *buf,
                                  size_t len) const noexcept {
-  return scalar::utf16::validate<endianness::BIG>(buf, len);
+  return validate_utf16be_with_errors(buf, len).is_ok();
 }
 
 simdutf_warn_unused result implementation::validate_utf16le_with_errors(
     const char16_t *buf, size_t len) const noexcept {
-  return scalar::utf16::validate_with_errors<endianness::LITTLE>(buf, len);
+  const auto res =
+      ppc64::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.count != len) {
+    auto scalar = scalar::utf16::validate_with_errors<endianness::LITTLE>(
+        buf + res.count, len - res.count);
+    scalar.count += res.count;
+    return scalar;
+  }
+
+  return res;
 }
 
 simdutf_warn_unused result implementation::validate_utf16be_with_errors(
     const char16_t *buf, size_t len) const noexcept {
-  return scalar::utf16::validate_with_errors<endianness::BIG>(buf, len);
+  const auto res =
+      ppc64::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.count != len) {
+    auto scalar = scalar::utf16::validate_with_errors<endianness::BIG>(
+        buf + res.count, len - res.count);
+    scalar.count += res.count;
+    return scalar;
+  }
+
+  return res;
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
+  return utf32::validate(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::validate_utf32_with_errors(
     const char32_t *buf, size_t len) const noexcept {
-  return scalar::utf32::validate_with_errors(buf, len);
+  return utf32::validate_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused bool
-implementation::validate_utf32(const char16_t *buf, size_t len) const noexcept {
-  return scalar::utf32::validate(buf, len);
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
+    const char *buf, size_t len, char *utf8_output) const noexcept {
+  const auto ret = ppc64_convert_latin1_to_utf8(buf, len, utf8_output);
+  size_t converted_chars = ret.second - utf8_output;
+
+  if (ret.first != buf + len) {
+    const size_t scalar_converted_chars = scalar::latin1_to_utf8::convert(
+        ret.first, len - (ret.first - buf), ret.second);
+    converted_chars += scalar_converted_chars;
+  }
+
+  return converted_chars;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  size_t n =
+      ppc64_convert_latin1_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
+  if (n < len) {
+    n += scalar::latin1_to_utf16::convert<endianness::LITTLE>(buf + n, len - n,
+                                                              utf16_output + n);
+  }
+
+  return n;
+}
+
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  size_t n =
+      ppc64_convert_latin1_to_utf16<endianness::BIG>(buf, len, utf16_output);
+  if (n < len) {
+    n += scalar::latin1_to_utf16::convert<endianness::BIG>(buf + n, len - n,
+                                                           utf16_output + n);
+  }
+
+  return n;
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  const auto ret = ppc64_convert_latin1_to_utf32(buf, len, utf32_output);
+  if (ret.first != buf + len) {
+    const size_t processed = ret.first - buf;
+    scalar::latin1_to_utf32::convert(ret.first, len - processed, ret.second);
+  }
+
+  return len;
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert(buf, len, latin1_output);
+}
+
+simdutf_warn_unused result implementation::convert_utf8_to_latin1_with_errors(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  utf8_to_latin1::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, latin1_output);
 }
 
+simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
+    const char *buf, size_t len, char *latin1_output) const noexcept {
+  return ppc64::utf8_to_latin1::convert_valid(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert<endianness::BIG>(buf, len, utf16_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return result(error_code::OTHER, 0); // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::LITTLE>(buf, len,
+                                                           utf16_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return result(error_code::OTHER, 0); // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  utf8_to_utf16::validating_transcoder converter;
+  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::LITTLE>(buf, len,
+                                                          utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
-    const char * /*buf*/, size_t /*len*/,
-    char16_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *buf, size_t len, char16_t *utf16_output) const noexcept {
+  return utf8_to_utf16::convert_valid<endianness::BIG>(buf, len, utf16_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
-    const char * /*buf*/, size_t /*len*/,
-    char32_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert(buf, len, utf32_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(
-    const char * /*buf*/, size_t /*len*/,
-    char32_t * /*utf16_output*/) const noexcept {
-  return result(error_code::OTHER, 0); // stub
+    const char *buf, size_t len, char32_t *utf32_output) const noexcept {
+  utf8_to_utf32::validating_transcoder converter;
+  return converter.convert_with_errors(buf, len, utf32_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
-    const char * /*buf*/, size_t /*len*/,
-    char32_t * /*utf16_output*/) const noexcept {
-  return 0; // stub
+    const char *input, size_t size, char32_t *utf32_output) const noexcept {
+  return utf8_to_utf32::convert_valid(input, size, utf32_output);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+
+  return convert_impl(ppc64_convert_utf16_to_latin1<endianness::LITTLE>,
+                      scalar::utf16_to_latin1::convert<endianness::LITTLE>, buf,
+                      len, latin1_output);
 }
 
+simdutf_warn_unused size_t implementation::convert_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+
+  return convert_impl(ppc64_convert_utf16_to_latin1<endianness::BIG>,
+                      scalar::utf16_to_latin1::convert<endianness::BIG>, buf,
+                      len, latin1_output);
+}
+
+simdutf_warn_unused result
+implementation::convert_utf16le_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_latin1<endianness::LITTLE>,
+      scalar::utf16_to_latin1::convert_with_errors<endianness::LITTLE>, buf,
+      len, latin1_output);
+}
+
+simdutf_warn_unused result
+implementation::convert_utf16be_to_latin1_with_errors(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_latin1<endianness::BIG>,
+      scalar::utf16_to_latin1::convert_with_errors<endianness::BIG>, buf, len,
+      latin1_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: we could provide an optimized function.
+  return convert_utf16be_to_latin1(buf, len, latin1_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
+    const char16_t *buf, size_t len, char *latin1_output) const noexcept {
+  // optimization opportunity: we could provide an optimized function.
+  return convert_utf16le_to_latin1(buf, len, latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert<endianness::LITTLE>(buf, len,
-                                                            utf8_output);
+
+  return convert_impl(ppc64_convert_utf16_to_utf8<endianness::LITTLE>,
+                      scalar::utf16_to_utf8::convert<endianness::LITTLE>, buf,
+                      len, utf8_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert<endianness::BIG>(buf, len, utf8_output);
+
+  return convert_impl(ppc64_convert_utf16_to_utf8<endianness::BIG>,
+                      scalar::utf16_to_utf8::convert<endianness::BIG>, buf, len,
+                      utf8_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf8_output);
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_utf8<endianness::LITTLE>,
+      scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>, buf, len,
+      utf8_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
-      buf, len, utf8_output);
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_utf8<endianness::BIG>,
+      scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>, buf, len,
+      utf8_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_valid<endianness::LITTLE>(buf, len,
-                                                                  utf8_output);
+  return convert_utf16le_to_utf8(buf, len, utf8_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf16_to_utf8::convert_valid<endianness::BIG>(buf, len,
-                                                               utf8_output);
+  return convert_utf16be_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return convert_impl(ppc64_convert_utf32_to_latin1<ErrorChecking::enabled>,
+                      scalar::utf32_to_latin1::convert, buf, len,
+                      latin1_output);
+}
+
+simdutf_warn_unused result implementation::convert_utf32_to_latin1_with_errors(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return convert_with_errors_impl(
+      ppc64_convert_utf32_to_latin1<ErrorChecking::enabled>,
+      scalar::utf32_to_latin1::convert_with_errors, buf, len, latin1_output);
+}
+
+simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
+    const char32_t *buf, size_t len, char *latin1_output) const noexcept {
+  return convert_impl(ppc64_convert_utf32_to_latin1<ErrorChecking::disabled>,
+                      scalar::utf32_to_latin1::convert, buf, len,
+                      latin1_output);
+}
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert(buf, len, utf8_output);
+  return convert_impl(ppc64_convert_utf32_to_utf8<ErrorReporting::at_the_end>,
+                      scalar::utf32_to_utf8::convert, buf, len, utf8_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert_with_errors(buf, len, utf8_output);
+  return convert_with_errors_impl(
+      ppc64_convert_utf32_to_utf8<ErrorReporting::precise>,
+      scalar::utf32_to_utf8::convert_with_errors, buf, len, utf8_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
-  return scalar::utf32_to_utf8::convert_valid(buf, len, utf8_output);
+  return convert_impl(ppc64_convert_utf32_to_utf8<ErrorReporting::none>,
+                      scalar::utf32_to_utf8::convert, buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert<endianness::LITTLE>(buf, len,
-                                                             utf16_output);
+
+  return convert_impl(ppc64_convert_utf32_to_utf16<endianness::LITTLE,
+                                                   ErrorReporting::at_the_end>,
+                      scalar::utf32_to_utf16::convert<endianness::LITTLE>, buf,
+                      len, utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert<endianness::BIG>(buf, len,
-                                                          utf16_output);
+
+  return convert_impl(
+      ppc64_convert_utf32_to_utf16<endianness::BIG, ErrorReporting::at_the_end>,
+      scalar::utf32_to_utf16::convert<endianness::BIG>, buf, len, utf16_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf16_output);
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf32_to_utf16<endianness::LITTLE, ErrorReporting::precise>,
+      scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>, buf, len,
+      utf16_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
-      buf, len, utf16_output);
+
+  return convert_with_errors_impl(
+      ppc64_convert_utf32_to_utf16<endianness::BIG, ErrorReporting::precise>,
+      scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>, buf, len,
+      utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_valid<endianness::LITTLE>(
-      buf, len, utf16_output);
+
+  return convert_impl(
+      ppc64_convert_utf32_to_utf16<endianness::LITTLE, ErrorReporting::none>,
+      scalar::utf32_to_utf16::convert<endianness::LITTLE>, buf, len,
+      utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
-  return scalar::utf32_to_utf16::convert_valid<endianness::BIG>(buf, len,
-                                                                utf16_output);
+
+  return convert_impl(
+      ppc64_convert_utf32_to_utf16<endianness::BIG, ErrorReporting::none>,
+      scalar::utf32_to_utf16::convert<endianness::BIG>, buf, len, utf16_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert<endianness::LITTLE>(buf, len,
-                                                             utf32_output);
+  return convert_impl(ppc64_convert_utf16_to_utf32<endianness::LITTLE>,
+                      scalar::utf16_to_utf32::convert<endianness::LITTLE>, buf,
+                      len, utf32_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert<endianness::BIG>(buf, len,
-                                                          utf32_output);
+  return convert_impl(ppc64_convert_utf16_to_utf32<endianness::BIG>,
+                      scalar::utf16_to_utf32::convert<endianness::BIG>, buf,
+                      len, utf32_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
-      buf, len, utf32_output);
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_utf32<endianness::LITTLE>,
+      scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>, buf, len,
+      utf32_output);
 }
 
 simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
-      buf, len, utf32_output);
+  return convert_with_errors_impl(
+      ppc64_convert_utf16_to_utf32<endianness::BIG>,
+      scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>, buf, len,
+      utf32_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_valid<endianness::LITTLE>(
-      buf, len, utf32_output);
+  return convert_utf16le_to_utf32(buf, len, utf32_output);
 }
 
 simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
-  return scalar::utf16_to_utf32::convert_valid<endianness::BIG>(buf, len,
-                                                                utf32_output);
+  return convert_utf16be_to_utf32(buf, len, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 void implementation::change_endianness_utf16(const char16_t *input,
                                              size_t length,
                                              char16_t *output) const noexcept {
-  scalar::utf16::change_endianness_utf16(input, length, output);
+  utf16::change_endianness_utf16(input, length, output);
 }
 
 simdutf_warn_unused size_t implementation::count_utf16le(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::count_code_points<endianness::LITTLE>(input, length);
+  return utf16::count_code_points<endianness::LITTLE>(input, length);
 }
 
 simdutf_warn_unused size_t implementation::count_utf16be(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::count_code_points<endianness::BIG>(input, length);
+  return utf16::count_code_points<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused size_t
 implementation::count_utf8(const char *input, size_t length) const noexcept {
   return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
+    const char *buf, size_t len) const noexcept {
+  return count_utf8(buf, len);
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
+    const char *input, size_t length) const noexcept {
+  const auto ret = ppc64_utf8_length_from_latin1(input, length);
+  const size_t consumed = ret.first - input;
+
+  if (consumed == length) {
+    return ret.second;
+  }
+
+  const auto scalar =
+      scalar::latin1::utf8_length_from_latin1(ret.first, length - consumed);
+  return scalar + ret.second;
+}
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(input,
-                                                                   length);
+  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
 }
 
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
+  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf32_length_from_utf16<endianness::LITTLE>(input,
-                                                                    length);
+  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
 }
 
 simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
-  return scalar::utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
+  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
     const char *input, size_t length) const noexcept {
-  return scalar::utf8::utf16_length_from_utf8(input, length);
+  return utf8::utf16_length_from_utf8(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  return scalar::utf32::utf8_length_from_utf32(input, length);
+  return utf32::utf8_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
   return scalar::utf32::utf16_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
     const char *input, size_t length) const noexcept {
-  return scalar::utf8::count_code_points(input, length);
+  return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_BASE64
 simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
     const char *input, size_t length) const noexcept {
   return scalar::base64::maximal_binary_length_from_base64(input, length);
@@ -36475,122 +45163,120 @@ simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
 simdutf_warn_unused result implementation::base64_to_binary(
     const char *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  // skip trailing spaces
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation};
+}
+
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
-    return {SUCCESS, 0};
-  }
-  result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation};
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
-  return r;
-}
-
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
 }
 
 simdutf_warn_unused result implementation::base64_to_binary(
     const char16_t *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
-  const bool ignore_garbage =
-      (options == base64_options::base64_url_accept_garbage) ||
-      (options == base64_options::base64_default_accept_garbage);
-  // skip trailing spaces
-  while (length > 0 &&
-         scalar::base64::is_ascii_white_space(input[length - 1])) {
-    length--;
-  }
-  size_t equallocation =
-      length; // location of the first padding character if any
-  size_t equalsigns = 0;
-  if (length > 0 && input[length - 1] == '=') {
-    equallocation = length - 1;
-    length -= 1;
-    equalsigns++;
-    while (length > 0 &&
-           scalar::base64::is_ascii_white_space(input[length - 1])) {
-      length--;
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
-    if (length > 0 && input[length - 1] == '=') {
-      equallocation = length - 1;
-      equalsigns++;
-      length -= 1;
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
-  if (length == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation};
+}
+
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
-    return {SUCCESS, 0};
-  }
-  result r = scalar::base64::base64_tail_decode(
-      output, input, length, equalsigns, options, last_chunk_options);
-  if (last_chunk_options != stop_before_partial &&
-      r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-    // additional checks
-    if ((r.count % 3 == 0) || ((r.count % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation};
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
+    } else {
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
-  return r;
-}
-
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
 }
 
 size_t implementation::binary_to_base64(const char *input, size_t length,
                                         char *output,
                                         base64_options options) const noexcept {
-  return scalar::base64::binary_to_base64(input, length, output, options);
+  if (options & base64_url) {
+    return encode_base64<true>(output, input, length, options);
+  } else {
+    return encode_base64<false>(output, input, length, options);
+  }
+}
+#endif // SIMDUTF_FEATURE_BASE64
+
+#ifdef SIMDUTF_INTERNAL_TESTS
+std::vector<implementation::TestProcedure>
+implementation::internal_tests() const {
+  #define entry(proc)                                                          \
+    TestProcedure { #proc, proc }
+  return {entry(base64_encoding_translate_6bit_values),
+          entry(base64_encoding_expand_6bit_fields),
+          entry(base64_decoding_valid),
+          entry(base64_decoding_invalid_ignore_errors),
+          entry(base64url_decoding_invalid_ignore_errors),
+          entry(base64_decoding_invalid_strict_errors),
+          entry(base64url_decoding_invalid_strict_errors),
+          entry(base64_decoding_pack),
+          entry(base64_compress)};
+  #undef entry
 }
+#endif
+
 } // namespace ppc64
 } // namespace simdutf
 
@@ -36600,11 +45286,6 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
 #endif
 #if SIMDUTF_IMPLEMENTATION_RVV
 /* begin file src/rvv/implementation.cpp */
-
-
-
-
-
 /* begin file src/simdutf/rvv/begin.h */
 // redefining SIMDUTF_IMPLEMENTATION to "rvv"
 // #define SIMDUTF_IMPLEMENTATION rvv
@@ -36658,7 +45339,7 @@ rvv_utf32_store_utf16_m4(uint16_t *dst, vuint32m4_t utf32, size_t vl,
 /* end file src/rvv/rvv_helpers.inl.cpp */
 
 /* begin file src/rvv/rvv_length_from.inl.cpp */
-
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t
 implementation::count_utf16le(const char16_t *src, size_t len) const noexcept {
   return utf32_length_from_utf16le(src, len);
@@ -36668,37 +45349,23 @@ simdutf_warn_unused size_t
 implementation::count_utf16be(const char16_t *src, size_t len) const noexcept {
   return utf32_length_from_utf16be(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused size_t
 implementation::count_utf8(const char *src, size_t len) const noexcept {
   return utf32_length_from_utf8(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
     const char *src, size_t len) const noexcept {
   return utf32_length_from_utf8(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t len) const noexcept {
-  return len;
-}
-
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t len) const noexcept {
-  return len;
-}
-
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t len) const noexcept {
-  return len;
-}
-
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t len) const noexcept {
-  return len;
-}
-
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
     const char *src, size_t len) const noexcept {
   size_t count = 0;
@@ -36710,7 +45377,9 @@ simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
   }
   return count;
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static size_t
 rvv_utf32_length_from_utf16(const char16_t *src, size_t len) {
@@ -36739,7 +45408,9 @@ simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
   else
     return rvv_utf32_length_from_utf16<simdutf_ByteFlip::V>(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
     const char *src, size_t len) const noexcept {
   size_t count = len;
@@ -36750,7 +45421,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
   }
   return count;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static size_t
 rvv_utf8_length_from_utf16(const char16_t *src, size_t len) {
@@ -36782,7 +45455,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
   else
     return rvv_utf8_length_from_utf16<simdutf_ByteFlip::V>(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
     const char32_t *src, size_t len) const noexcept {
   size_t count = 0;
@@ -36797,7 +45472,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
   }
   return count;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
     const char *src, size_t len) const noexcept {
   size_t count = 0;
@@ -36811,7 +45488,9 @@ simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
   }
   return count;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
     const char32_t *src, size_t len) const noexcept {
   size_t count = 0;
@@ -36823,10 +45502,10 @@ simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
   }
   return count;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* end file src/rvv/rvv_length_from.inl.cpp */
 /* begin file src/rvv/rvv_validate.inl.cpp */
-
-
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused bool
 implementation::validate_ascii(const char *src, size_t len) const noexcept {
   size_t vlmax = __riscv_vsetvlmax_e8m8();
@@ -36852,7 +45531,9 @@ simdutf_warn_unused result implementation::validate_ascii_with_errors(
   }
   return result(error_code::SUCCESS, src - beg);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* Returns a close estimation of the number of valid UTF-8 bytes up to the
  * first invalid one, but never overestimating. */
 simdutf_really_inline static size_t rvv_count_valid_utf8(const char *src,
@@ -36864,7 +45545,7 @@ simdutf_really_inline static size_t rvv_count_valid_utf8(const char *src,
   /* validate first three bytes */
   {
     size_t idx = 3;
-    while (idx < len && (src[idx] >> 6) == 0b10)
+    while (idx < len && (uint8_t(src[idx]) >> 6) == 0b10)
       ++idx;
     if (idx > 3 + 3 || !scalar::utf8::validate(src, idx))
       return 0;
@@ -36931,7 +45612,7 @@ simdutf_really_inline static size_t rvv_count_valid_utf8(const char *src,
   }
 
   /* we need to validate the last character */
-  while (tail < len && (src[0] >> 6) == 0b10)
+  while (tail < len && (uint8_t(src[0]) >> 6) == 0b10)
     --src, ++tail;
   return src - beg;
 }
@@ -36941,26 +45622,18 @@ implementation::validate_utf8(const char *src, size_t len) const noexcept {
   size_t count = rvv_count_valid_utf8(src, len);
   return scalar::utf8::validate(src + count, len - count);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused result implementation::validate_utf8_with_errors(
     const char *src, size_t len) const noexcept {
   size_t count = rvv_count_valid_utf8(src, len);
   result res = scalar::utf8::validate_with_errors(src + count, len - count);
   return result(res.error, count + res.count);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
-simdutf_warn_unused bool
-implementation::validate_utf16le(const char16_t *src,
-                                 size_t len) const noexcept {
-  return validate_utf16le_with_errors(src, len).error == error_code::SUCCESS;
-}
-
-simdutf_warn_unused bool
-implementation::validate_utf16be(const char16_t *src,
-                                 size_t len) const noexcept {
-  return validate_utf16be_with_errors(src, len).error == error_code::SUCCESS;
-}
-
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static result
 rvv_validate_utf16_with_errors(const char16_t *src, size_t len) {
@@ -36993,7 +45666,26 @@ rvv_validate_utf16_with_errors(const char16_t *src, size_t len) {
     return result(error_code::SUCCESS, src - beg);
   }
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+simdutf_warn_unused bool
+implementation::validate_utf16le(const char16_t *src,
+                                 size_t len) const noexcept {
+  return rvv_validate_utf16_with_errors<simdutf_ByteFlip::NONE>(src, len)
+             .error == error_code::SUCCESS;
+}
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF16
+simdutf_warn_unused bool
+implementation::validate_utf16be(const char16_t *src,
+                                 size_t len) const noexcept {
+  return validate_utf16be_with_errors(src, len).error == error_code::SUCCESS;
+}
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused result implementation::validate_utf16le_with_errors(
     const char16_t *src, size_t len) const noexcept {
   return rvv_validate_utf16_with_errors<simdutf_ByteFlip::NONE>(src, len);
@@ -37006,7 +45698,9 @@ simdutf_warn_unused result implementation::validate_utf16be_with_errors(
   else
     return rvv_validate_utf16_with_errors<simdutf_ByteFlip::V>(src, len);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf32(const char32_t *src, size_t len) const noexcept {
   size_t vlmax = __riscv_vsetvlmax_e32m8();
@@ -37025,7 +45719,9 @@ implementation::validate_utf32(const char32_t *src, size_t len) const noexcept {
                  __riscv_vmsne_vx_u32m8_b4(maxOff, 0xFFFFF7FF, vlmax), vlmax),
              vlmax) < 0;
 }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::validate_utf32_with_errors(
     const char32_t *src, size_t len) const noexcept {
   const char32_t *beg = src;
@@ -37053,10 +45749,11 @@ simdutf_warn_unused result implementation::validate_utf32_with_errors(
   }
   return result(error_code::SUCCESS, src - beg);
 }
+#endif // SIMDUTF_FEATURE_UTF32
 /* end file src/rvv/rvv_validate.inl.cpp */
 
 /* begin file src/rvv/rvv_latin1_to.inl.cpp */
-
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
     const char *src, size_t len, char *dst) const noexcept {
   char *beg = dst;
@@ -37087,7 +45784,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
   }
   return dst - beg;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
     const char *src, size_t len, char16_t *dst) const noexcept {
   char16_t *beg = dst;
@@ -37111,7 +45810,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
   }
   return dst - beg;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
     const char *src, size_t len, char32_t *dst) const noexcept {
   char32_t *beg = dst;
@@ -37122,10 +45823,10 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
   }
   return dst - beg;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* end file src/rvv/rvv_latin1_to.inl.cpp */
 /* begin file src/rvv/rvv_utf16_to.inl.cpp */
-#include <cstdio>
-
+#if SIMDUTF_FEATURE_UTF16
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static result
 rvv_utf16_to_latin1_with_errors(const char16_t *src, size_t len, char *dst) {
@@ -37141,7 +45842,9 @@ rvv_utf16_to_latin1_with_errors(const char16_t *src, size_t len, char *dst) {
   }
   return result(error_code::SUCCESS, src - beg);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
     const char16_t *src, size_t len, char *dst) const noexcept {
   result res = convert_utf16le_to_latin1_with_errors(src, len, dst);
@@ -37191,7 +45894,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_latin1(
   }
   return src - beg;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static result
 rvv_utf16_to_utf8_with_errors(const char16_t *src, size_t len, char *dst) {
@@ -37369,7 +46074,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
     const char16_t *src, size_t len, char *dst) const noexcept {
   return convert_utf16be_to_utf8(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static result
 rvv_utf16_to_utf32_with_errors(const char16_t *src, size_t len, char32_t *dst) {
@@ -37517,9 +46224,11 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
     const char16_t *src, size_t len, char32_t *dst) const noexcept {
   return convert_utf16be_to_utf32(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* end file src/rvv/rvv_utf16_to.inl.cpp */
-/* begin file src/rvv/rvv_utf32_to.inl.cpp */
 
+/* begin file src/rvv/rvv_utf32_to.inl.cpp */
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
     const char32_t *src, size_t len, char *dst) const noexcept {
   result res = convert_utf32_to_latin1_with_errors(src, len, dst);
@@ -37549,7 +46258,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
     const char32_t *src, size_t len, char *dst) const noexcept {
   return convert_utf32_to_latin1(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
     const char32_t *src, size_t len, char *dst) const noexcept {
   size_t n = len;
@@ -37701,7 +46412,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
     const char32_t *src, size_t len, char *dst) const noexcept {
   return convert_utf32_to_utf8(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static result
 rvv_convert_utf32_to_utf16_with_errors(const char32_t *src, size_t len,
@@ -37808,8 +46521,10 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(
   else
     return rvv_convert_valid_utf32_to_utf16<simdutf_ByteFlip::V>(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* end file src/rvv/rvv_utf32_to.inl.cpp */
 /* begin file src/rvv/rvv_utf8_to.inl.cpp */
+#if SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32)
 template <typename Tdst, simdutf_ByteFlip bflip, bool validate = true>
 simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
                                                        size_t len, Tdst *dst) {
@@ -37831,7 +46546,7 @@ simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
   /* validate first three bytes */
   if (validate) {
     size_t idx = 3;
-    while (idx < len && (src[idx] >> 6) == 0b10)
+    while (idx < len && (uint8_t(src[idx]) >> 6) == 0b10)
       ++idx;
     if (idx > 3 + 3 || !scalar::utf8::validate(src, idx))
       return 0;
@@ -37852,6 +46567,7 @@ simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
   const vuint8m1_t err3tbl =
       __riscv_vreinterpret_v_u64m1_u8m1(__riscv_vle64_v_u64m1(err3m, 2));
 
+  size_t vl8m1 = __riscv_vsetvlmax_e8m1();
   size_t vl8m2 = __riscv_vsetvlmax_e8m2();
   vbool4_t m4even = __riscv_vmseq_vx_u8m2_b4(
       __riscv_vand_vx_u8m2(__riscv_vid_v_u8m2(vl8m2), 1, vl8m2), 0, vl8m2);
@@ -38000,51 +46716,51 @@ simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
      * vssubu.vx v, 10, (max(x-10, 0)) almost gives us what we want, we
      * just need to manually detect and handle the one special case:
      */
-#define SIMDUTF_RVV_UTF8_TO_COMMON_M1(idx)                                     \
-  vuint8m1_t c1 = __riscv_vget_v_u8m2_u8m1(b1, idx);                           \
-  vuint8m1_t c2 = __riscv_vget_v_u8m2_u8m1(b2, idx);                           \
-  vuint8m1_t c3 = __riscv_vget_v_u8m2_u8m1(b3, idx);                           \
-  vuint8m1_t c4 = __riscv_vget_v_u8m2_u8m1(b4, idx);                           \
-  /* remove prefix from trailing bytes */                                      \
-  c2 = __riscv_vand_vx_u8m1(c2, 0b00111111, vlOut);                            \
-  c3 = __riscv_vand_vx_u8m1(c3, 0b00111111, vlOut);                            \
-  c4 = __riscv_vand_vx_u8m1(c4, 0b00111111, vlOut);                            \
-  vuint8m1_t shift = __riscv_vsrl_vx_u8m1(c1, 4, vlOut);                       \
-  shift = __riscv_vmerge_vxm_u8m1(__riscv_vssubu_vx_u8m1(shift, 10, vlOut), 3, \
-                                  __riscv_vmseq_vx_u8m1_b8(shift, 12, vlOut),  \
-                                  vlOut);                                      \
-  c1 = __riscv_vsll_vv_u8m1(c1, shift, vlOut);                                 \
-  c1 = __riscv_vsrl_vv_u8m1(c1, shift, vlOut);                                 \
-  /* unconditionally widen and combine to c1234 */                             \
-  vuint16m2_t c34 = __riscv_vwaddu_wv_u16m2(                                   \
-      __riscv_vwmulu_vx_u16m2(c3, 1 << 6, vlOut), c4, vlOut);                  \
-  vuint16m2_t c12 = __riscv_vwaddu_wv_u16m2(                                   \
-      __riscv_vwmulu_vx_u16m2(c1, 1 << 6, vlOut), c2, vlOut);                  \
-  vuint32m4_t c1234 = __riscv_vwaddu_wv_u32m4(                                 \
-      __riscv_vwmulu_vx_u32m4(c12, 1 << 12, vlOut), c34, vlOut);               \
-  /* derive required right-shift amount from `shift` to reduce                 \
-   * c1234 to the required number of bytes */                                  \
-  c1234 = __riscv_vsrl_vv_u32m4(                                               \
-      c1234,                                                                   \
-      __riscv_vzext_vf4_u32m4(                                                 \
-          __riscv_vmul_vx_u8m1(                                                \
-              __riscv_vrsub_vx_u8m1(__riscv_vssubu_vx_u8m1(shift, 2, vlOut),   \
-                                    3, vlOut),                                 \
-              6, vlOut),                                                       \
-          vlOut),                                                              \
-      vlOut);                                                                  \
-  /* store result in desired format */                                         \
-  if (is16)                                                                    \
-    vlDst = rvv_utf32_store_utf16_m4<bflip>((uint16_t *)dst, c1234, vlOut,     \
-                                            m4even);                           \
-  else                                                                         \
-    vlDst = vlOut, __riscv_vse32_v_u32m4((uint32_t *)dst, c1234, vlOut);
+  #define SIMDUTF_RVV_UTF8_TO_COMMON_M1(idx)                                   \
+    vuint8m1_t c1 = __riscv_vget_v_u8m2_u8m1(b1, idx);                         \
+    vuint8m1_t c2 = __riscv_vget_v_u8m2_u8m1(b2, idx);                         \
+    vuint8m1_t c3 = __riscv_vget_v_u8m2_u8m1(b3, idx);                         \
+    vuint8m1_t c4 = __riscv_vget_v_u8m2_u8m1(b4, idx);                         \
+    /* remove prefix from trailing bytes */                                    \
+    c2 = __riscv_vand_vx_u8m1(c2, 0b00111111, vlOut);                          \
+    c3 = __riscv_vand_vx_u8m1(c3, 0b00111111, vlOut);                          \
+    c4 = __riscv_vand_vx_u8m1(c4, 0b00111111, vlOut);                          \
+    vuint8m1_t shift = __riscv_vsrl_vx_u8m1(c1, 4, vlOut);                     \
+    shift = __riscv_vmerge_vxm_u8m1(                                           \
+        __riscv_vssubu_vx_u8m1(shift, 10, vlOut), 3,                           \
+        __riscv_vmseq_vx_u8m1_b8(shift, 12, vlOut), vlOut);                    \
+    c1 = __riscv_vsll_vv_u8m1(c1, shift, vlOut);                               \
+    c1 = __riscv_vsrl_vv_u8m1(c1, shift, vlOut);                               \
+    /* unconditionally widen and combine to c1234 */                           \
+    vuint16m2_t c34 = __riscv_vwaddu_wv_u16m2(                                 \
+        __riscv_vwmulu_vx_u16m2(c3, 1 << 6, vlOut), c4, vlOut);                \
+    vuint16m2_t c12 = __riscv_vwaddu_wv_u16m2(                                 \
+        __riscv_vwmulu_vx_u16m2(c1, 1 << 6, vlOut), c2, vlOut);                \
+    vuint32m4_t c1234 = __riscv_vwaddu_wv_u32m4(                               \
+        __riscv_vwmulu_vx_u32m4(c12, 1 << 12, vlOut), c34, vlOut);             \
+    /* derive required right-shift amount from `shift` to reduce               \
+     * c1234 to the required number of bytes */                                \
+    c1234 = __riscv_vsrl_vv_u32m4(                                             \
+        c1234,                                                                 \
+        __riscv_vzext_vf4_u32m4(                                               \
+            __riscv_vmul_vx_u8m1(                                              \
+                __riscv_vrsub_vx_u8m1(__riscv_vssubu_vx_u8m1(shift, 2, vlOut), \
+                                      3, vlOut),                               \
+                6, vlOut),                                                     \
+            vlOut),                                                            \
+        vlOut);                                                                \
+    /* store result in desired format */                                       \
+    if (is16)                                                                  \
+      vlDst = rvv_utf32_store_utf16_m4<bflip>((uint16_t *)dst, c1234, vlOut,   \
+                                              m4even);                         \
+    else                                                                       \
+      vlDst = vlOut, __riscv_vse32_v_u32m4((uint32_t *)dst, c1234, vlOut);
 
     /* Unrolling this manually reduces register pressure and allows
      * us to terminate early. */
     {
       size_t vlOutm2 = vlOut, vlDst;
-      vlOut = __riscv_vsetvl_e8m1(vlOut);
+      vlOut = __riscv_vsetvl_e8m1(vlOut < vl8m1 ? vlOut : vl8m1);
       SIMDUTF_RVV_UTF8_TO_COMMON_M1(0)
       if (vlOutm2 == vlOut) {
         vlOut = vlDst;
@@ -38060,14 +46776,14 @@ simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
       vlOut = vlDst;
     }
 
-#undef SIMDUTF_RVV_UTF8_TO_COMMON_M1
+  #undef SIMDUTF_RVV_UTF8_TO_COMMON_M1
   }
 
   /* validate the last character and reparse it + tail */
   if (len > tail) {
-    if ((src[0] >> 6) == 0b10)
+    if ((uint8_t(src[0]) >> 6) == 0b10)
       --dst;
-    while ((src[0] >> 6) == 0b10 && tail < len)
+    while ((uint8_t(src[0]) >> 6) == 0b10 && tail < len)
       --src, ++tail;
     if (is16) {
       /* go back one more, when on high surrogate */
@@ -38081,7 +46797,10 @@ simdutf_really_inline static size_t rvv_utf8_to_common(char const *src,
     return 0;
   return (size_t)(dst - beg) + ret;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32)
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
     const char *src, size_t len, char *dst) const noexcept {
   const char *beg = dst;
@@ -38170,7 +46889,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
   }
   return dst - beg;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
     const char *src, size_t len, char16_t *dst) const noexcept {
   return rvv_utf8_to_common<uint16_t, simdutf_ByteFlip::NONE>(src, len,
@@ -38220,7 +46941,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
     return rvv_utf8_to_common<uint16_t, simdutf_ByteFlip::V, false>(
         src, len, (uint16_t *)dst);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
     const char *src, size_t len, char32_t *dst) const noexcept {
   return rvv_utf8_to_common<uint32_t, simdutf_ByteFlip::NONE>(src, len,
@@ -38240,8 +46963,10 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
   return rvv_utf8_to_common<uint32_t, simdutf_ByteFlip::NONE, false>(
       src, len, (uint32_t *)dst);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* end file src/rvv/rvv_utf8_to.inl.cpp */
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused int
 implementation::detect_encodings(const char *input,
                                  size_t length) const noexcept {
@@ -38254,7 +46979,7 @@ implementation::detect_encodings(const char *input,
   if (validate_utf8(input, length))
     out |= encoding_type::UTF8;
   if (length % 2 == 0) {
-    if (validate_utf16(reinterpret_cast<const char16_t *>(input), length / 2))
+    if (validate_utf16le(reinterpret_cast<const char16_t *>(input), length / 2))
       out |= encoding_type::UTF16_LE;
   }
   if (length % 4 == 0) {
@@ -38264,7 +46989,9 @@ implementation::detect_encodings(const char *input,
 
   return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 template <simdutf_ByteFlip bflip>
 simdutf_really_inline static void
 rvv_change_endianness_utf16(const char16_t *src, size_t len, char16_t *dst) {
@@ -38282,12 +47009,9 @@ void implementation::change_endianness_utf16(const char16_t *src, size_t len,
   else
     return rvv_change_endianness_utf16<simdutf_ByteFlip::V>(src, len, dst);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
+#if SIMDUTF_FEATURE_BASE64
 simdutf_warn_unused result implementation::base64_to_binary(
     const char *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
@@ -38391,11 +47115,6 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
   return r;
 }
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
 simdutf_warn_unused result implementation::base64_to_binary(
     const char16_t *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
@@ -38499,16 +47218,13 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
   return r;
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
-}
-
 size_t implementation::binary_to_base64(const char *input, size_t length,
                                         char *output,
                                         base64_options options) const noexcept {
   return scalar::base64::tail_encode_base64(output, input, length, options);
 }
+#endif // SIMDUTF_FEATURE_BASE64
+
 } // namespace rvv
 } // namespace simdutf
 
@@ -38527,6 +47243,7 @@ SIMDUTF_UNTARGET_REGION
 /* begin file src/simdutf/westmere/begin.h */
 // redefining SIMDUTF_IMPLEMENTATION to "westmere"
 // #define SIMDUTF_IMPLEMENTATION westmere
+#define SIMDUTF_SIMD_HAS_BYTEMASK 1
 
 #if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
 // nothing needed.
@@ -38534,6 +47251,7 @@ SIMDUTF_UNTARGET_REGION
 SIMDUTF_TARGET_WESTMERE
 #endif
 /* end file src/simdutf/westmere/begin.h */
+
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -38542,25 +47260,15 @@ namespace {
 #endif
 using namespace simd;
 
+#if SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    SIMDUTF_FEATURE_UTF8
 simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
   return input.reduce_or().is_ascii();
 }
+#endif // SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // SIMDUTF_FEATURE_UTF8
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte =
-      prev1.saturating_sub(0b11000000u - 1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte =
-      prev2.saturating_sub(0b11100000u - 1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte =
-      prev3.saturating_sub(0b11110000u - 1); // Only 1111____ will be > 0
-  // Caller requires a bool (all 1's). All values resulting from the subtraction
-  // will be <= 64, so signed comparison is fine.
-  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) >
-         int8_t(0);
-}
-
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_really_inline simd8<bool>
 must_be_2_3_continuation(const simd8<uint8_t> prev2,
                          const simd8<uint8_t> prev3) {
@@ -38570,7 +47278,9 @@ must_be_2_3_continuation(const simd8<uint8_t> prev2,
       prev3.saturating_sub(0xf0u - 0x80); // Only 1111____ will be >= 0x80
   return simd8<bool>(is_third_byte | is_fourth_byte);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 /* begin file src/westmere/internal/loader.cpp */
 namespace internal {
 namespace westmere {
@@ -38647,292 +47357,31 @@ inline void write_v_u16_11bits_to_utf8(const __m128i v_u16, char *&utf8_output,
 } // namespace westmere
 } // namespace internal
 /* end file src/westmere/internal/loader.cpp */
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/westmere/sse_validate_utf16.cpp */
-/*
-    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
-
-    In a vectorized algorithm we want to examine the most significant
-    nibble in order to select a fast path. If none of highest nibbles
-    are 0xD (13), than we are sure that UTF-16 chunk in a vector
-    register is valid.
-
-    Let us analyze what we need to check if the nibble is 0xD. The
-    value of the preceding nibble determines what we have:
-
-    0xd000 .. 0xd7ff - a valid word
-    0xd800 .. 0xdbff - low surrogate
-    0xdc00 .. 0xdfff - high surrogate
-
-    Other constraints we have to consider:
-    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
-    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
-    - there must not be sole low surrogate nor high surrogate
-
-    We are going to build three bitmasks based on the 3rd nibble:
-    - V = valid word,
-    - L = low surrogate (0xd800 .. 0xdbff)
-    - H = high surrogate (0xdc00 .. 0xdfff)
-
-      0   1   2   3   4   5   6   7    <--- word index
-    [ V | L | H | L | H | V | V | L ]
-      1   0   0   0   0   1   1   0     - V = valid masks
-      0   1   0   1   0   0   0   1     - L = low surrogate
-      0   0   1   0   1   0   0   0     - H high surrogate
-
-
-      1   0   0   0   0   1   1   0   V = valid masks
-      0   1   0   1   0   0   0   0   a = L & (H >> 1)
-      0   0   1   0   1   0   0   0   b = a << 1
-      1   1   1   1   1   1   1   0   c = V | a | b
-                                  ^
-                                  the last bit can be zero, we just consume 7
-   code units and recheck this word in the next iteration
-*/
-
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
 template <endianness big_endian>
-const char16_t *sse_validate_utf16(const char16_t *input, size_t size) {
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::SIZE * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-
-    const auto t0 = in0.shr<8>();
-    const auto t1 = in1.shr<8>();
-
-    const auto in = simd16<uint16_t>::pack(t0, t1);
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint16_t surrogates_bitmask =
-        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
-    if (surrogates_bitmask == 0x0000) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
-
-      const uint16_t a = static_cast<uint16_t>(
-          L & (H >> 1)); // A low surrogate must be followed by high one.
-                         // (A low surrogate placed in the 7th register's word
-                         // is an exception we handle.)
-      const uint16_t b = static_cast<uint16_t>(
-          a << 1); // Just mark that the opinput - startite fact is hold,
-                   // thanks to that we have only two masks for valid case.
-      const uint16_t c = static_cast<uint16_t>(
-          V | a | b); // Combine all the masks into the final one.
-
-      if (c == 0xffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0x7fff) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return nullptr;
-      }
-    }
-  }
-
-  return input;
-}
-
-template <endianness big_endian>
-const result sse_validate_utf16_with_errors(const char16_t *input,
-                                            size_t size) {
-  if (simdutf_unlikely(size == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  const char16_t *start = input;
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::SIZE * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> in0,
+                                       const simd16<uint16_t> in1) {
+  if (big_endian) {
+    // we want lower bytes
+    const auto mask = simd16<uint16_t>(0x00ff);
+    const auto t0 = in0 & mask;
+    const auto t1 = in1 & mask;
 
+    return simd16<uint16_t>::pack(t0, t1);
+  } else {
     const auto t0 = in0.shr<8>();
     const auto t1 = in1.shr<8>();
 
-    const auto in = simd16<uint16_t>::pack(t0, t1);
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint16_t surrogates_bitmask =
-        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
-    if (surrogates_bitmask == 0x0000) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
-
-      const uint16_t a = static_cast<uint16_t>(
-          L & (H >> 1)); // A low surrogate must be followed by high one.
-                         // (A low surrogate placed in the 7th register's word
-                         // is an exception we handle.)
-      const uint16_t b = static_cast<uint16_t>(
-          a << 1); // Just mark that the opinput - startite fact is hold,
-                   // thanks to that we have only two masks for valid case.
-      const uint16_t c = static_cast<uint16_t>(
-          V | a | b); // Combine all the masks into the final one.
-
-      if (c == 0xffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0x7fff) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return result(error_code::SURROGATE, input - start);
-      }
-    }
+    return simd16<uint16_t>::pack(t0, t1);
   }
-
-  return result(error_code::SUCCESS, input - start);
 }
 /* end file src/westmere/sse_validate_utf16.cpp */
-/* begin file src/westmere/sse_validate_utf32le.cpp */
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
-const char32_t *sse_validate_utf32le(const char32_t *input, size_t size) {
-  const char32_t *end = input + size;
-
-  const __m128i standardmax = _mm_set1_epi32(0x10ffff);
-  const __m128i offset = _mm_set1_epi32(0xffff2000);
-  const __m128i standardoffsetmax = _mm_set1_epi32(0xfffff7ff);
-  __m128i currentmax = _mm_setzero_si128();
-  __m128i currentoffsetmax = _mm_setzero_si128();
-
-  while (input + 4 < end) {
-    const __m128i in = _mm_loadu_si128((__m128i *)input);
-    currentmax = _mm_max_epu32(in, currentmax);
-    currentoffsetmax =
-        _mm_max_epu32(_mm_add_epi32(in, offset), currentoffsetmax);
-    input += 4;
-  }
-  __m128i is_zero =
-      _mm_xor_si128(_mm_max_epu32(currentmax, standardmax), standardmax);
-  if (_mm_test_all_zeros(is_zero, is_zero) == 0) {
-    return nullptr;
-  }
-
-  is_zero = _mm_xor_si128(_mm_max_epu32(currentoffsetmax, standardoffsetmax),
-                          standardoffsetmax);
-  if (_mm_test_all_zeros(is_zero, is_zero) == 0) {
-    return nullptr;
-  }
-
-  return input;
-}
-
-const result sse_validate_utf32le_with_errors(const char32_t *input,
-                                              size_t size) {
-  const char32_t *start = input;
-  const char32_t *end = input + size;
-
-  const __m128i standardmax = _mm_set1_epi32(0x10ffff);
-  const __m128i offset = _mm_set1_epi32(0xffff2000);
-  const __m128i standardoffsetmax = _mm_set1_epi32(0xfffff7ff);
-  __m128i currentmax = _mm_setzero_si128();
-  __m128i currentoffsetmax = _mm_setzero_si128();
-
-  while (input + 4 < end) {
-    const __m128i in = _mm_loadu_si128((__m128i *)input);
-    currentmax = _mm_max_epu32(in, currentmax);
-    currentoffsetmax =
-        _mm_max_epu32(_mm_add_epi32(in, offset), currentoffsetmax);
-
-    __m128i is_zero =
-        _mm_xor_si128(_mm_max_epu32(currentmax, standardmax), standardmax);
-    if (_mm_test_all_zeros(is_zero, is_zero) == 0) {
-      return result(error_code::TOO_LARGE, input - start);
-    }
-
-    is_zero = _mm_xor_si128(_mm_max_epu32(currentoffsetmax, standardoffsetmax),
-                            standardoffsetmax);
-    if (_mm_test_all_zeros(is_zero, is_zero) == 0) {
-      return result(error_code::SURROGATE, input - start);
-    }
-    input += 4;
-  }
-
-  return result(error_code::SUCCESS, input - start);
-}
-/* end file src/westmere/sse_validate_utf32le.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_latin1_to_utf8.cpp */
 std::pair<const char *const, char *const>
 sse_convert_latin1_to_utf8(const char *latin_input,
@@ -39006,6 +47455,9 @@ sse_convert_latin1_to_utf8(const char *latin_input,
   return std::make_pair(latin_input, utf8_output);
 }
 /* end file src/westmere/sse_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_latin1_to_utf16.cpp */
 template <endianness big_endian>
 std::pair<const char *, char16_t *>
@@ -39029,6 +47481,9 @@ sse_convert_latin1_to_utf16(const char *latin1_input, size_t len,
   return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len);
 }
 /* end file src/westmere/sse_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_latin1_to_utf32.cpp */
 std::pair<const char *, char32_t *>
 sse_convert_latin1_to_utf32(const char *buf, size_t len,
@@ -39062,7 +47517,9 @@ sse_convert_latin1_to_utf32(const char *buf, size_t len,
   return std::make_pair(buf, utf32_output);
 }
 /* end file src/westmere/sse_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /* begin file src/westmere/sse_convert_utf8_to_utf16.cpp */
 // depends on "tables/utf8_to_utf16_tables.h"
 
@@ -39262,6 +47719,9 @@ size_t convert_masked_utf8_to_utf16(const char *input,
   return consumed;
 }
 /* end file src/westmere/sse_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/westmere/sse_convert_utf8_to_utf32.cpp */
 // depends on "tables/utf8_to_utf16_tables.h"
 
@@ -39405,6 +47865,9 @@ size_t convert_masked_utf8_to_utf32(const char *input,
   return consumed;
 }
 /* end file src/westmere/sse_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_utf8_to_latin1.cpp */
 // depends on "tables/utf8_to_utf16_tables.h"
 
@@ -39465,7 +47928,9 @@ size_t convert_masked_utf8_to_latin1(const char *input,
   return consumed;
 }
 /* end file src/westmere/sse_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_utf16_to_latin1.cpp */
 template <endianness big_endian>
 std::pair<const char16_t *, char *>
@@ -39523,9 +47988,8 @@ sse_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
     } else {
       // Fallback to scalar code for handling errors
       for (int k = 0; k < 8; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if (word <= 0xff) {
           *latin1_output++ = char(word);
         } else {
@@ -39540,6 +48004,9 @@ sse_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/westmere/sse_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /* begin file src/westmere/sse_convert_utf16_to_utf8.cpp */
 /*
     The vectorized algorithm works on single SSE register i.e., it
@@ -39781,7 +48248,7 @@ sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) {
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -39795,7 +48262,7 @@ sse_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_output) {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
           if ((diff | diff2) > 0x3FF) {
@@ -40010,7 +48477,7 @@ sse_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -40024,7 +48491,7 @@ sse_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
           if ((diff | diff2) > 0x3FF) {
@@ -40046,6 +48513,9 @@ sse_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
   return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
 }
 /* end file src/westmere/sse_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/westmere/sse_convert_utf16_to_utf32.cpp */
 /*
     The vectorized algorithm works on single SSE register i.e., it
@@ -40098,7 +48568,7 @@ sse_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
 
 /*
   Returns a pair: the first unprocessed byte from buf and utf8_output
-  A scalar routing should carry on the conversion of the tail.
+  A scalar routine should carry on the conversion of the tail.
 */
 template <endianness big_endian>
 std::pair<const char16_t *, char32_t *>
@@ -40149,14 +48619,14 @@ sse_convert_utf16_to_utf32(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
           if ((diff | diff2) > 0x3FF) {
@@ -40229,14 +48699,14 @@ sse_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
+        uint16_t word = big_endian ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word =
-              big_endian ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
+              big_endian ? scalar::u16_swap_bytes(buf[k + 1]) : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
           if ((diff | diff2) > 0x3FF) {
@@ -40254,7 +48724,9 @@ sse_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
   return std::make_pair(result(error_code::SUCCESS, buf - start), utf32_output);
 }
 /* end file src/westmere/sse_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/westmere/sse_convert_utf32_to_latin1.cpp */
 std::pair<const char32_t *, char *>
 sse_convert_utf32_to_latin1(const char32_t *buf, size_t len,
@@ -40339,6 +48811,9 @@ sse_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/westmere/sse_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/westmere/sse_convert_utf32_to_utf8.cpp */
 std::pair<const char32_t *, char *>
 sse_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_output) {
@@ -40931,7 +49406,64 @@ sse_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
   return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
 }
 /* end file src/westmere/sse_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/westmere/sse_convert_utf32_to_utf16.cpp */
+struct expansion_result_t {
+  size_t u16count;
+  __m128i compressed;
+};
+
+// Function sse_expand_surrogate takes four **valid** UTF-32 characters
+// having at least one code-point producing a surrogate pair.
+template <endianness byte_order>
+expansion_result_t sse_expand_surrogate(const __m128i x) {
+  using vector_u32 = simd32<uint32_t>;
+  using vector_u8 = simd8<uint8_t>;
+
+  const auto in = vector_u32(x);
+
+  const auto non_surrogate_mask = (in & uint32_t(0xffff0000)) == uint32_t(0);
+  const auto mask = (~non_surrogate_mask.to_4bit_bitmask()) & 0xf;
+
+  const auto t0 = in - uint32_t(0x00010000);
+  const auto hi = t0.shr<10>() & uint32_t(0x000003ff);
+  const auto lo = t0.shl<16>() & uint32_t(0x03ff0000);
+  const auto surrogates = (lo | hi) | uint32_t(0xdc00d800);
+
+  const auto merged = as_vector_u8(select(non_surrogate_mask, in, surrogates));
+
+  const auto shuffle = vector_u8::load(
+      (byte_order == endianness::LITTLE)
+          ? tables::utf32_to_utf16::pack_utf32_to_utf16le[mask]
+          : tables::utf32_to_utf16::pack_utf32_to_utf16be[mask]);
+
+  const size_t u16count = (4 + count_ones(mask));
+  const auto compressed = shuffle.lookup_16(merged);
+
+  return {u16count, compressed};
+}
+
+// Function `validate_utf32` checks 2 x 4 UTF-32 characters for their validity.
+simdutf_really_inline bool validate_utf32(const __m128i a, const __m128i b) {
+  using vector_u32 = simd32<uint32_t>;
+
+  const auto in0 = vector_u32(a);
+  const auto in1 = vector_u32(b);
+
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+
+  const auto too_large = max(in0, in1) > standardmax;
+  const auto surrogate0 = (in0 + offset) > standardoffsetmax;
+  const auto surrogate1 = (in1 + offset) > standardoffsetmax;
+
+  const auto combined = too_large | surrogate0 | surrogate1;
+  return !combined.any();
+}
+
 template <endianness big_endian>
 std::pair<const char32_t *, char16_t *>
 sse_convert_utf32_to_utf16(const char32_t *buf, size_t len,
@@ -40939,74 +49471,64 @@ sse_convert_utf32_to_utf16(const char32_t *buf, size_t len,
 
   const char32_t *end = buf + len;
 
-  const __m128i v_0000 = _mm_setzero_si128();
   const __m128i v_ffff0000 = _mm_set1_epi32((int32_t)0xffff0000);
   __m128i forbidden_bytemask = _mm_setzero_si128();
 
-  while (end - buf >= 8) {
-    __m128i in = _mm_loadu_si128((__m128i *)buf);
-    __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1);
-    const __m128i saturation_bytemask = _mm_cmpeq_epi32(
-        _mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));
+  while (end - buf >= 16 + 8) {
+    const __m128i *ptr = reinterpret_cast<const __m128i *>(buf);
+    const __m128i in0 = _mm_loadu_si128(ptr + 0);
+    const __m128i in1 = _mm_loadu_si128(ptr + 1);
+    const __m128i in2 = _mm_loadu_si128(ptr + 2);
+    const __m128i in3 = _mm_loadu_si128(ptr + 3);
 
-    // Check if no bits set above 16th
-    if (saturation_bitmask == 0xffff) {
-      // Pack UTF-32 to UTF-16
-      __m128i utf16_packed = _mm_packus_epi32(in, nextin);
+    const __m128i combined =
+        _mm_or_si128(_mm_or_si128(in2, in3), _mm_or_si128(in0, in1));
+    if (simdutf_likely(_mm_testz_si128(combined, v_ffff0000))) {
+      // No bits set above 16th, directly pack UTF-32 to UTF-16
+      __m128i utf16_packed0 = _mm_packus_epi32(in0, in1);
+      __m128i utf16_packed1 = _mm_packus_epi32(in2, in3);
 
       const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
       const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800);
       forbidden_bytemask = _mm_or_si128(
           forbidden_bytemask,
-          _mm_cmpeq_epi16(_mm_and_si128(utf16_packed, v_f800), v_d800));
+          _mm_or_si128(
+              _mm_cmpeq_epi16(_mm_and_si128(utf16_packed0, v_f800), v_d800),
+              _mm_cmpeq_epi16(_mm_and_si128(utf16_packed1, v_f800), v_d800)));
 
       if (big_endian) {
         const __m128i swap =
             _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
-        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
+        utf16_packed0 = _mm_shuffle_epi8(utf16_packed0, swap);
+        utf16_packed1 = _mm_shuffle_epi8(utf16_packed1, swap);
       }
 
-      _mm_storeu_si128((__m128i *)utf16_output, utf16_packed);
-      utf16_output += 8;
-      buf += 8;
+      _mm_storeu_si128((__m128i *)utf16_output + 0, utf16_packed0);
+      _mm_storeu_si128((__m128i *)utf16_output + 1, utf16_packed1);
+      utf16_output += 16;
+      buf += 16;
     } else {
-      size_t forward = 7;
-      size_t k = 0;
-      if (size_t(end - buf) < forward + 1) {
-        forward = size_t(end - buf - 1);
-      }
-      for (; k < forward; k++) {
-        uint32_t word = buf[k];
-        if ((word & 0xFFFF0000) == 0) {
-          // will not generate a surrogate pair
-          if (word >= 0xD800 && word <= 0xDFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          *utf16_output++ =
-              big_endian
-                  ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8))
-                  : char16_t(word);
-        } else {
-          // will generate a surrogate pair
-          if (word > 0x10FFFF) {
-            return std::make_pair(nullptr, utf16_output);
-          }
-          word -= 0x10000;
-          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
-          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
-          if (big_endian) {
-            high_surrogate =
-                uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
-            low_surrogate =
-                uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
-          }
-          *utf16_output++ = char16_t(high_surrogate);
-          *utf16_output++ = char16_t(low_surrogate);
-        }
+      if (!validate_utf32(in0, in1) || !validate_utf32(in2, in3)) {
+        return std::make_pair(nullptr, utf16_output);
       }
-      buf += k;
+
+      const auto ret0 = sse_expand_surrogate<big_endian>(in0);
+      _mm_storeu_si128((__m128i *)utf16_output, ret0.compressed);
+      utf16_output += ret0.u16count;
+
+      const auto ret1 = sse_expand_surrogate<big_endian>(in1);
+      _mm_storeu_si128((__m128i *)utf16_output, ret1.compressed);
+      utf16_output += ret1.u16count;
+
+      const auto ret2 = sse_expand_surrogate<big_endian>(in2);
+      _mm_storeu_si128((__m128i *)utf16_output, ret2.compressed);
+      utf16_output += ret2.u16count;
+
+      const auto ret3 = sse_expand_surrogate<big_endian>(in3);
+      _mm_storeu_si128((__m128i *)utf16_output, ret3.compressed);
+      utf16_output += ret3.u16count;
+
+      buf += 16;
     }
   }
 
@@ -41025,20 +49547,15 @@ sse_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
   const char32_t *start = buf;
   const char32_t *end = buf + len;
 
-  const __m128i v_0000 = _mm_setzero_si128();
   const __m128i v_ffff0000 = _mm_set1_epi32((int32_t)0xffff0000);
 
   while (end - buf >= 8) {
-    __m128i in = _mm_loadu_si128((__m128i *)buf);
-    __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1);
-    const __m128i saturation_bytemask = _mm_cmpeq_epi32(
-        _mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
-    const uint32_t saturation_bitmask =
-        static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));
+    const __m128i in = _mm_loadu_si128((__m128i *)buf);
+    const __m128i nextin = _mm_loadu_si128((__m128i *)buf + 1);
 
-    // Check if no bits set above 16th
-    if (saturation_bitmask == 0xffff) {
-      // Pack UTF-32 to UTF-16
+    const __m128i combined = _mm_or_si128(in, nextin);
+    if (simdutf_likely(_mm_testz_si128(combined, v_ffff0000))) {
+      // No bits set above 16th, directly pack UTF-32 to UTF-16
       __m128i utf16_packed = _mm_packus_epi32(in, nextin);
 
       const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
@@ -41103,6 +49620,9 @@ sse_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
   return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
 }
 /* end file src/westmere/sse_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_BASE64
 /* begin file src/westmere/sse_base64.cpp */
 /**
  * References and further reading:
@@ -41131,6 +49651,9 @@ sse_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
  * Nick Kopp. 2013. Base64 Encoding on a GPU.
  * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
  */
+
+// --- encoding ----------------------------------------------------
+
 template <bool base64_url> __m128i lookup_pshufb_improved(const __m128i input) {
   // credit: Wojciech Muła
   // reduce  0..51 -> 0
@@ -41272,202 +49795,47 @@ size_t encode_base64(char *dst, const char *src, size_t srclen,
   }
 
   return i / 3 * 4 + scalar::base64::tail_encode_base64((char *)out, src + i,
-                                                        srclen - i, options);
-}
-static inline void compress(__m128i data, uint16_t mask, char *output) {
-  if (mask == 0) {
-    _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data);
-    return;
-  }
-
-  // this particular implementation was inspired by work done by @animetosho
-  // we do it in two steps, first 8 bytes and then second 8 bytes
-  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
-  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
-  // next line just loads the 64-bit values thintable_epi8[mask1] and
-  // thintable_epi8[mask2] into a 128-bit register, using only
-  // two instructions on most compilers.
-
-  __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2],
-                                    tables::base64::thintable_epi8[mask1]);
-  // we increment by 0x08 the second half of the mask
-  shufmask =
-      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
-  // this is the version "nearly pruned"
-  __m128i pruned = _mm_shuffle_epi8(data, shufmask);
-  // we still need to put the two halves together.
-  // we compute the popcount of the first half:
-  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
-  // then load the corresponding mask, what it does is to write
-  // only the first pop1 bytes from the first 8 bytes, and then
-  // it fills in with the bytes from the second 8 bytes + some filling
-  // at the end.
-  __m128i compactmask = _mm_loadu_si128(reinterpret_cast<const __m128i *>(
-      tables::base64::pshufb_combine_table + pop1 * 8));
-  __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
-}
-
-struct block64 {
-  __m128i chunks[4];
-};
-
-template <bool base64_url, bool ignore_garbage>
-static inline uint16_t to_base64_mask(__m128i *src, uint32_t *error) {
-  const __m128i ascii_space_tbl =
-      _mm_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa, 0x0,
-                    0xc, 0xd, 0x0, 0x0);
-  // credit: aqrit
-  __m128i delta_asso;
-  if (base64_url) {
-    delta_asso = _mm_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0, 0x0,
-                               0x0, 0x0, 0x0, 0xF, 0x0, 0xF);
-  } else {
-
-    delta_asso = _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                               0x00, 0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F);
-  }
-  __m128i delta_values;
-  if (base64_url) {
-    delta_values = _mm_setr_epi8(0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF),
-                                 uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9),
-                                 0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF),
-                                 uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9));
-  } else {
-
-    delta_values =
-        _mm_setr_epi8(int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13),
-                      int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9),
-                      int8_t(0xB9), int8_t(0x00), int8_t(0x10), int8_t(0xC3),
-                      int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9));
-  }
-  __m128i check_asso;
-  if (base64_url) {
-    check_asso = _mm_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1,
-                               0x3, 0x7, 0xB, 0xE, 0xB, 0x6);
-  } else {
-
-    check_asso = _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                               0x01, 0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F);
-  }
-  __m128i check_values;
-  if (base64_url) {
-    check_values = _mm_setr_epi8(uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
-                                 uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF),
-                                 uint8_t(0xB6), uint8_t(0xA6), uint8_t(0xB5),
-                                 uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0,
-                                 uint8_t(0x80), 0x0, uint8_t(0x80));
-  } else {
-
-    check_values =
-        _mm_setr_epi8(int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80),
-                      int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6),
-                      int8_t(0xB5), int8_t(0x86), int8_t(0xD1), int8_t(0x80),
-                      int8_t(0xB1), int8_t(0x80), int8_t(0x91), int8_t(0x80));
-  }
-  const __m128i shifted = _mm_srli_epi32(*src, 3);
-
-  const __m128i delta_hash =
-      _mm_avg_epu8(_mm_shuffle_epi8(delta_asso, *src), shifted);
-  const __m128i check_hash =
-      _mm_avg_epu8(_mm_shuffle_epi8(check_asso, *src), shifted);
-
-  const __m128i out =
-      _mm_adds_epi8(_mm_shuffle_epi8(delta_values, delta_hash), *src);
-  const __m128i chk =
-      _mm_adds_epi8(_mm_shuffle_epi8(check_values, check_hash), *src);
-  const int mask = _mm_movemask_epi8(chk);
-  if (!ignore_garbage && mask) {
-    __m128i ascii_space =
-        _mm_cmpeq_epi8(_mm_shuffle_epi8(ascii_space_tbl, *src), *src);
-    *error = (mask ^ _mm_movemask_epi8(ascii_space));
-  }
-  *src = out;
-  return (uint16_t)mask;
-}
-
-template <bool base64_url, bool ignore_garbage>
-static inline uint64_t to_base64_mask(block64 *b, uint64_t *error) {
-  uint32_t err0 = 0;
-  uint32_t err1 = 0;
-  uint32_t err2 = 0;
-  uint32_t err3 = 0;
-  uint64_t m0 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[0], &err0);
-  uint64_t m1 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[1], &err1);
-  uint64_t m2 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[2], &err2);
-  uint64_t m3 =
-      to_base64_mask<base64_url, ignore_garbage>(&b->chunks[3], &err3);
-  if (!ignore_garbage) {
-    *error = (err0) | ((uint64_t)err1 << 16) | ((uint64_t)err2 << 32) |
-             ((uint64_t)err3 << 48);
-  }
-  return m0 | (m1 << 16) | (m2 << 32) | (m3 << 48);
-}
-
-#if defined(_MSC_VER) && !defined(__clang__)
-static inline size_t simdutf_tzcnt_u64(uint64_t num) {
-  unsigned long ret;
-  if (num == 0) {
-    return 64;
-  }
-  _BitScanForward64(&ret, num);
-  return ret;
-}
-#else // GCC or Clang
-static inline size_t simdutf_tzcnt_u64(uint64_t num) {
-  return num ? __builtin_ctzll(num) : 64;
-}
-#endif
-
-static inline void copy_block(block64 *b, char *output) {
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output), b->chunks[0]);
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 16), b->chunks[1]);
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 32), b->chunks[2]);
-  _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 48), b->chunks[3]);
+                                                        srclen - i, options);
 }
 
-static inline uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
-  uint64_t nmask = ~mask;
-  compress(b->chunks[0], uint16_t(mask), output);
-  compress(b->chunks[1], uint16_t(mask >> 16),
-           output + _mm_popcnt_u64(nmask & 0xFFFF));
-  compress(b->chunks[2], uint16_t(mask >> 32),
-           output + _mm_popcnt_u64(nmask & 0xFFFFFFFF));
-  compress(b->chunks[3], uint16_t(mask >> 48),
-           output + _mm_popcnt_u64(nmask & 0xFFFFFFFFFFFFULL));
-  return _mm_popcnt_u64(nmask);
-}
+// --- decoding -----------------------------------------------
 
-// The caller of this function is responsible to ensure that there are 64 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char *src) {
-  b->chunks[0] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src));
-  b->chunks[1] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 16));
-  b->chunks[2] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 32));
-  b->chunks[3] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48));
-}
+static simdutf_really_inline void compress(__m128i data, uint16_t mask,
+                                           char *output) {
+  if (mask == 0) {
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output), data);
+    return;
+  }
 
-// The caller of this function is responsible to ensure that there are 128 bytes
-// available from reading at src. The data is read into a block64 structure.
-static inline void load_block(block64 *b, const char16_t *src) {
-  __m128i m1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src));
-  __m128i m2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 8));
-  __m128i m3 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 16));
-  __m128i m4 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 24));
-  __m128i m5 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 32));
-  __m128i m6 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 40));
-  __m128i m7 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48));
-  __m128i m8 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 56));
-  b->chunks[0] = _mm_packus_epi16(m1, m2);
-  b->chunks[1] = _mm_packus_epi16(m3, m4);
-  b->chunks[2] = _mm_packus_epi16(m5, m6);
-  b->chunks[3] = _mm_packus_epi16(m7, m8);
+  // this particular implementation was inspired by work done by @animetosho
+  // we do it in two steps, first 8 bytes and then second 8 bytes
+  uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+  uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+  // next line just loads the 64-bit values thintable_epi8[mask1] and
+  // thintable_epi8[mask2] into a 128-bit register, using only
+  // two instructions on most compilers.
+
+  __m128i shufmask = _mm_set_epi64x(tables::base64::thintable_epi8[mask2],
+                                    tables::base64::thintable_epi8[mask1]);
+  // we increment by 0x08 the second half of the mask
+  shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+  // this is the version "nearly pruned"
+  __m128i pruned = _mm_shuffle_epi8(data, shufmask);
+  // we still need to put the two halves together.
+  // we compute the popcount of the first half:
+  int pop1 = tables::base64::BitsSetTable256mul2[mask1];
+  // then load the corresponding mask, what it does is to write
+  // only the first pop1 bytes from the first 8 bytes, and then
+  // it fills in with the bytes from the second 8 bytes + some filling
+  // at the end.
+  __m128i compactmask = _mm_loadu_si128(reinterpret_cast<const __m128i *>(
+      tables::base64::pshufb_combine_table + pop1 * 8));
+  __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+  _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
 }
 
-static inline void base64_decode(char *out, __m128i str) {
+static simdutf_really_inline void base64_decode(char *out, __m128i str) {
   // credit: aqrit
 
   const __m128i pack_shuffle =
@@ -41480,6 +49848,7 @@ static inline void base64_decode(char *out, __m128i str) {
   // this writes 16 bytes, but we only need 12.
   _mm_storeu_si128((__m128i *)out, t2);
 }
+
 // decode 64 bytes and output 48 bytes
 static inline void base64_decode_block(char *out, const char *src) {
   base64_decode(out, _mm_loadu_si128(reinterpret_cast<const __m128i *>(src)));
@@ -41490,6 +49859,7 @@ static inline void base64_decode_block(char *out, const char *src) {
   base64_decode(out + 36,
                 _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48)));
 }
+
 static inline void base64_decode_block_safe(char *out, const char *src) {
   base64_decode(out, _mm_loadu_si128(reinterpret_cast<const __m128i *>(src)));
   base64_decode(out + 12,
@@ -41501,222 +49871,250 @@ static inline void base64_decode_block_safe(char *out, const char *src) {
                 _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48)));
   std::memcpy(out + 36, buffer, 12);
 }
-static inline void base64_decode_block(char *out, block64 *b) {
-  base64_decode(out, b->chunks[0]);
-  base64_decode(out + 12, b->chunks[1]);
-  base64_decode(out + 24, b->chunks[2]);
-  base64_decode(out + 36, b->chunks[3]);
-}
-static inline void base64_decode_block_safe(char *out, block64 *b) {
-  base64_decode(out, b->chunks[0]);
-  base64_decode(out + 12, b->chunks[1]);
-  base64_decode(out + 24, b->chunks[2]);
-  char buffer[16];
-  base64_decode(buffer, b->chunks[3]);
-  std::memcpy(out + 36, buffer, 12);
-}
 
-template <bool base64_url, bool ignore_garbage, typename chartype>
-full_result
-compress_decode_base64(char *dst, const chartype *src, size_t srclen,
-                       base64_options options,
-                       last_chunk_handling_options last_chunk_options) {
-  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
-                                        : tables::base64::to_base64_value;
-  size_t equallocation =
-      srclen; // location of the first padding character if any
-  // skip trailing spaces
-  while (!ignore_garbage && srclen > 0 &&
-         scalar::base64::is_eight_byte(src[srclen - 1]) &&
-         to_base64[uint8_t(src[srclen - 1])] == 64) {
-    srclen--;
-  }
-  size_t equalsigns = 0;
-  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
-    equallocation = srclen - 1;
-    srclen--;
-    equalsigns = 1;
-    // skip trailing spaces
-    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
-           to_base64[uint8_t(src[srclen - 1])] == 64) {
-      srclen--;
-    }
-    if (srclen > 0 && src[srclen - 1] == '=') {
-      equallocation = srclen - 1;
-      srclen--;
-      equalsigns = 2;
-    }
+// --- decoding - base64 class --------------------------------
+
+class block64 {
+  __m128i chunks[4];
+
+public:
+  // The caller of this function is responsible to ensure that there are 64
+  // bytes available from reading at src.
+  simdutf_really_inline block64(const char *src) {
+    chunks[0] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src));
+    chunks[1] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 16));
+    chunks[2] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 32));
+    chunks[3] = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48));
   }
-  if (srclen == 0) {
-    if (!ignore_garbage && equalsigns > 0) {
-      if (last_chunk_options == last_chunk_handling_options::strict) {
-        return {BASE64_INPUT_REMAINDER, 0, 0};
-      } else if (last_chunk_options ==
-                 last_chunk_handling_options::stop_before_partial) {
-        return {SUCCESS, 0, 0};
-      }
-      return {INVALID_BASE64_CHARACTER, equallocation, 0};
-    }
-    return {SUCCESS, 0, 0};
+
+public:
+  // The caller of this function is responsible to ensure that there are 128
+  // bytes available from reading at src. The data is read into a block64
+  // structure.
+  simdutf_really_inline block64(const char16_t *src) {
+    const auto m1 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src));
+    const auto m2 = _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 8));
+    const auto m3 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 16));
+    const auto m4 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 24));
+    const auto m5 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 32));
+    const auto m6 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 40));
+    const auto m7 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 48));
+    const auto m8 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(src + 56));
+    chunks[0] = _mm_packus_epi16(m1, m2);
+    chunks[1] = _mm_packus_epi16(m3, m4);
+    chunks[2] = _mm_packus_epi16(m5, m6);
+    chunks[3] = _mm_packus_epi16(m7, m8);
   }
-  char *end_of_safe_64byte_zone =
-      (srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 : dst;
 
-  const chartype *const srcinit = src;
-  const char *const dstinit = dst;
-  const chartype *const srcend = src + srclen;
+public:
+  simdutf_really_inline void copy_block(char *output) {
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output), chunks[0]);
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 16), chunks[1]);
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 32), chunks[2]);
+    _mm_storeu_si128(reinterpret_cast<__m128i *>(output + 48), chunks[3]);
+  }
 
-  constexpr size_t block_size = 6;
-  static_assert(block_size >= 2, "block should of size 2 or more");
-  char buffer[block_size * 64];
-  char *bufferptr = buffer;
-  if (srclen >= 64) {
-    const chartype *const srcend64 = src + srclen - 64;
-    while (src <= srcend64) {
-      block64 b;
-      load_block(&b, src);
-      src += 64;
-      uint64_t error = 0;
-      uint64_t badcharmask =
-          to_base64_mask<base64_url, ignore_garbage>(&b, &error);
-      if (error && !ignore_garbage) {
-        src -= 64;
-        size_t error_offset = simdutf_tzcnt_u64(error);
-        return {error_code::INVALID_BASE64_CHARACTER,
-                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
-      }
-      if (badcharmask != 0) {
-        // optimization opportunity: check for simple masks like those made of
-        // continuous 1s followed by continuous 0s. And masks containing a
-        // single bad character.
-        bufferptr += compress_block(&b, badcharmask, bufferptr);
-      } else if (bufferptr != buffer) {
-        copy_block(&b, bufferptr);
-        bufferptr += 64;
-      } else {
-        if (dst >= end_of_safe_64byte_zone) {
-          base64_decode_block_safe(dst, &b);
-        } else {
-          base64_decode_block(dst, &b);
-        }
-        dst += 48;
-      }
-      if (bufferptr >= (block_size - 1) * 64 + buffer) {
-        for (size_t i = 0; i < (block_size - 2); i++) {
-          base64_decode_block(dst, buffer + i * 64);
-          dst += 48;
-        }
-        if (dst >= end_of_safe_64byte_zone) {
-          base64_decode_block_safe(dst, buffer + (block_size - 2) * 64);
-        } else {
-          base64_decode_block(dst, buffer + (block_size - 2) * 64);
-        }
-        dst += 48;
-        std::memcpy(buffer, buffer + (block_size - 1) * 64,
-                    64); // 64 might be too much
-        bufferptr -= (block_size - 1) * 64;
-      }
+public:
+  simdutf_really_inline uint64_t compress_block(uint64_t mask, char *output) {
+    if (is_power_of_two(mask)) {
+      return compress_block_single(mask, output);
     }
+
+    uint64_t nmask = ~mask;
+    compress(chunks[0], uint16_t(mask), output);
+    compress(chunks[1], uint16_t(mask >> 16),
+             output + count_ones(nmask & 0xFFFF));
+    compress(chunks[2], uint16_t(mask >> 32),
+             output + count_ones(nmask & 0xFFFFFFFF));
+    compress(chunks[3], uint16_t(mask >> 48),
+             output + count_ones(nmask & 0xFFFFFFFFFFFFULL));
+    return count_ones(nmask);
   }
 
-  char *buffer_start = buffer;
-  // Optimization note: if this is almost full, then it is worth our
-  // time, otherwise, we should just decode directly.
-  int last_block = (int)((bufferptr - buffer_start) % 64);
-  if (last_block != 0 && srcend - src + last_block >= 64) {
-    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
-      uint8_t val = to_base64[uint8_t(*src)];
-      *bufferptr = char(val);
-      if ((!scalar::base64::is_eight_byte(*src) || val > 64) &&
-          !ignore_garbage) {
-        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
-                size_t(dst - dstinit)};
-      }
-      bufferptr += (val <= 63);
-      src++;
-    }
+private:
+  simdutf_really_inline size_t compress_block_single(uint64_t mask,
+                                                     char *output) {
+    const size_t pos64 = trailing_zeroes(mask);
+    const int8_t pos = pos64 & 0xf;
+    switch (pos64 >> 4) {
+    case 0b00: {
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(chunks[0], sh);
+
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), compressed);
+      _mm_storeu_si128((__m128i *)(output + 1 * 16 - 1), chunks[1]);
+      _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]);
+      _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]);
+    } break;
+    case 0b01: {
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(chunks[1], sh);
+
+      _mm_storeu_si128((__m128i *)(output + 1 * 16), compressed);
+      _mm_storeu_si128((__m128i *)(output + 2 * 16 - 1), chunks[2]);
+      _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]);
+    } break;
+    case 0b10: {
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]);
+      _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(chunks[2], sh);
+
+      _mm_storeu_si128((__m128i *)(output + 2 * 16), compressed);
+      _mm_storeu_si128((__m128i *)(output + 3 * 16 - 1), chunks[3]);
+    } break;
+    case 0b11: {
+      _mm_storeu_si128((__m128i *)(output + 0 * 16), chunks[0]);
+      _mm_storeu_si128((__m128i *)(output + 1 * 16), chunks[1]);
+      _mm_storeu_si128((__m128i *)(output + 2 * 16), chunks[2]);
+
+      const __m128i v0 = _mm_set1_epi8(char(pos - 1));
+      const __m128i v1 =
+          _mm_setr_epi8(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+      const __m128i v2 = _mm_cmpgt_epi8(v1, v0);
+      const __m128i sh = _mm_sub_epi8(v1, v2);
+      const __m128i compressed = _mm_shuffle_epi8(chunks[3], sh);
+
+      _mm_storeu_si128((__m128i *)(output + 3 * 16), compressed);
+    } break;
+    }
+
+    return 63;
   }
 
-  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
-    if (dst >= end_of_safe_64byte_zone) {
-      base64_decode_block_safe(dst, buffer_start);
-    } else {
-      base64_decode_block(dst, buffer_start);
-    }
-    dst += 48;
+public:
+  template <bool base64_url, bool ignore_garbage>
+  simdutf_really_inline uint64_t to_base64_mask(uint64_t *error) {
+    uint32_t err0 = 0;
+    uint32_t err1 = 0;
+    uint32_t err2 = 0;
+    uint32_t err3 = 0;
+    uint64_t m0 = to_base64_mask<base64_url, ignore_garbage>(&chunks[0], &err0);
+    uint64_t m1 = to_base64_mask<base64_url, ignore_garbage>(&chunks[1], &err1);
+    uint64_t m2 = to_base64_mask<base64_url, ignore_garbage>(&chunks[2], &err2);
+    uint64_t m3 = to_base64_mask<base64_url, ignore_garbage>(&chunks[3], &err3);
+    if (!ignore_garbage) {
+      *error = (err0) | ((uint64_t)err1 << 16) | ((uint64_t)err2 << 32) |
+               ((uint64_t)err3 << 48);
+    }
+    return m0 | (m1 << 16) | (m2 << 32) | (m3 << 48);
   }
-  if ((bufferptr - buffer_start) % 64 != 0) {
-    while (buffer_start + 4 < bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 4);
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    if (buffer_start + 4 <= bufferptr) {
-      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
-                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
-                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
-                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
-                        << 8;
-      triple = scalar::utf32::swap_bytes(triple);
-      std::memcpy(dst, &triple, 3);
+private:
+  template <bool base64_url, bool ignore_garbage>
+  simdutf_really_inline uint16_t to_base64_mask(__m128i *src, uint32_t *error) {
+    const __m128i ascii_space_tbl =
+        _mm_setr_epi8(0x20, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x9, 0xa,
+                      0x0, 0xc, 0xd, 0x0, 0x0);
+    // credit: aqrit
+    __m128i delta_asso;
+    if (base64_url) {
+      delta_asso = _mm_setr_epi8(0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x0,
+                                 0x0, 0x0, 0x0, 0x0, 0xF, 0x0, 0xF);
+    } else {
 
-      dst += 3;
-      buffer_start += 4;
-    }
-    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
-    // backtrack
-    int leftover = int(bufferptr - buffer_start);
-    while (leftover > 0) {
-      if (!ignore_garbage) {
-        while (to_base64[uint8_t(*(src - 1))] == 64) {
-          src--;
-        }
-      } else {
-        while (to_base64[uint8_t(*(src - 1))] >= 64) {
-          src--;
-        }
-      }
-      src--;
-      leftover--;
+      delta_asso =
+          _mm_setr_epi8(0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x00,
+                        0x00, 0x00, 0x00, 0x00, 0x0F, 0x00, 0x0F);
     }
-  }
-  if (src < srcend + equalsigns) {
-    full_result r = scalar::base64::base64_tail_decode(
-        dst, src, srcend - src, equalsigns, options, last_chunk_options);
-    r.input_count += size_t(src - srcinit);
-    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
-        r.error == error_code::BASE64_EXTRA_BITS) {
-      return r;
+    __m128i delta_values;
+    if (base64_url) {
+      delta_values = _mm_setr_epi8(0x0, 0x0, 0x0, 0x13, 0x4, uint8_t(0xBF),
+                                   uint8_t(0xBF), uint8_t(0xB9), uint8_t(0xB9),
+                                   0x0, 0x11, uint8_t(0xC3), uint8_t(0xBF),
+                                   uint8_t(0xE0), uint8_t(0xB9), uint8_t(0xB9));
     } else {
-      r.output_count += size_t(dst - dstinit);
+      delta_values =
+          _mm_setr_epi8(int8_t(0x00), int8_t(0x00), int8_t(0x00), int8_t(0x13),
+                        int8_t(0x04), int8_t(0xBF), int8_t(0xBF), int8_t(0xB9),
+                        int8_t(0xB9), int8_t(0x00), int8_t(0x10), int8_t(0xC3),
+                        int8_t(0xBF), int8_t(0xBF), int8_t(0xB9), int8_t(0xB9));
+    }
+    __m128i check_asso;
+    if (base64_url) {
+      check_asso = _mm_setr_epi8(0xD, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1, 0x1,
+                                 0x1, 0x3, 0x7, 0xB, 0xE, 0xB, 0x6);
+    } else {
+      check_asso =
+          _mm_setr_epi8(0x0D, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                        0x01, 0x03, 0x07, 0x0B, 0x0B, 0x0B, 0x0F);
+    }
+    __m128i check_values;
+    if (base64_url) {
+      check_values = _mm_setr_epi8(uint8_t(0x80), uint8_t(0x80), uint8_t(0x80),
+                                   uint8_t(0x80), uint8_t(0xCF), uint8_t(0xBF),
+                                   uint8_t(0xB6), uint8_t(0xA6), uint8_t(0xB5),
+                                   uint8_t(0xA1), 0x0, uint8_t(0x80), 0x0,
+                                   uint8_t(0x80), 0x0, uint8_t(0x80));
+    } else {
+      check_values =
+          _mm_setr_epi8(int8_t(0x80), int8_t(0x80), int8_t(0x80), int8_t(0x80),
+                        int8_t(0xCF), int8_t(0xBF), int8_t(0xD5), int8_t(0xA6),
+                        int8_t(0xB5), int8_t(0x86), int8_t(0xD1), int8_t(0x80),
+                        int8_t(0xB1), int8_t(0x80), int8_t(0x91), int8_t(0x80));
     }
-    if (last_chunk_options != stop_before_partial &&
-        r.error == error_code::SUCCESS && equalsigns > 0 && !ignore_garbage) {
-      // additional checks
-      if ((r.output_count % 3 == 0) ||
-          ((r.output_count % 3) + 1 + equalsigns != 4)) {
-        r.error = error_code::INVALID_BASE64_CHARACTER;
-        r.input_count = equallocation;
-      }
+    const __m128i shifted = _mm_srli_epi32(*src, 3);
+
+    const __m128i delta_hash =
+        _mm_avg_epu8(_mm_shuffle_epi8(delta_asso, *src), shifted);
+    const __m128i check_hash =
+        _mm_avg_epu8(_mm_shuffle_epi8(check_asso, *src), shifted);
+
+    const __m128i out =
+        _mm_adds_epi8(_mm_shuffle_epi8(delta_values, delta_hash), *src);
+    const __m128i chk =
+        _mm_adds_epi8(_mm_shuffle_epi8(check_values, check_hash), *src);
+    const int mask = _mm_movemask_epi8(chk);
+    if (!ignore_garbage && mask) {
+      __m128i ascii_space =
+          _mm_cmpeq_epi8(_mm_shuffle_epi8(ascii_space_tbl, *src), *src);
+      *error = (mask ^ _mm_movemask_epi8(ascii_space));
     }
-    return r;
+    *src = out;
+    return (uint16_t)mask;
   }
-  if (equalsigns > 0 && !ignore_garbage) {
-    if ((size_t(dst - dstinit) % 3 == 0) ||
-        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
-      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
-    }
+
+public:
+  simdutf_really_inline void base64_decode_block(char *out) {
+    base64_decode(out, chunks[0]);
+    base64_decode(out + 12, chunks[1]);
+    base64_decode(out + 24, chunks[2]);
+    base64_decode(out + 36, chunks[3]);
   }
-  return {SUCCESS, srclen, size_t(dst - dstinit)};
-}
+
+public:
+  simdutf_really_inline void base64_decode_block_safe(char *out) {
+    base64_decode(out, chunks[0]);
+    base64_decode(out + 12, chunks[1]);
+    base64_decode(out + 24, chunks[2]);
+    char buffer[16];
+    base64_decode(buffer, chunks[3]);
+    std::memcpy(out + 36, buffer, 12);
+  }
+};
 /* end file src/westmere/sse_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
 
 } // unnamed namespace
 } // namespace westmere
@@ -41833,6 +50231,7 @@ simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
 } // namespace westmere
 } // namespace simdutf
 /* end file src/generic/buf_block_reader.h */
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
 namespace simdutf {
 namespace westmere {
@@ -42139,9 +50538,21 @@ result generic_validate_utf8_with_errors(const char *input, size_t length) {
       reinterpret_cast<const uint8_t *>(input), length);
 }
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+} // namespace utf8_validation
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace ascii_validation {
+
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   uint8_t blocks[64]{};
   simd::simd8x64<uint8_t> running_or(blocks);
   while (reader.has_full_block()) {
@@ -42156,14 +50567,8 @@ bool generic_validate_ascii(const uint8_t *input, size_t length) {
   return running_or.is_ascii();
 }
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   size_t count{0};
   while (reader.has_full_block()) {
     simd::simd8x64<uint8_t> in(reader.full_block());
@@ -42188,19 +50593,16 @@ result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
   }
 }
 
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-} // namespace utf8_validation
+} // namespace ascii_validation
 } // unnamed namespace
 } // namespace westmere
 } // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
-// transcoding from UTF-8 to UTF-16
-/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  // transcoding from UTF-8 to UTF-16
+/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -42277,7 +50679,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 /* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -42611,9 +51012,10 @@ struct validating_transcoder {
 } // namespace westmere
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
-/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -42658,7 +51060,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
 /* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -42978,9 +51379,148 @@ struct validating_transcoder {
 } // namespace westmere
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-// other functions
-/* begin file src/generic/utf8.h */
+/* begin file src/generic/utf32.h */
+#include <limits>
+
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace utf32 {
+
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
+
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
+
+  const char32_t *start = input;
+
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
+
+  const size_t N = vector_u32::ELEMENTS;
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+  size_t counter = 0;
+
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
+
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
+
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += 4 * N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
+  }
+
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
+}
 
+} // namespace utf32
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+/* end file src/generic/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+
+#if SIMDUTF_FEATURE_UTF8
+/* begin file src/generic/utf8.h */
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -42999,6 +51539,59 @@ simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
   return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
+
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
+
+  size_t pos = 0;
+  size_t count = 0;
+
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
+
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
+
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
+    }
+  }
+
+  if (iterations > 0) {
+    count += local.sum_bytes();
+  }
+
+  count += counters.sum();
+
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
+}
+#endif
+
 simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
                                                     size_t size) {
   size_t pos = 0;
@@ -43020,6 +51613,8 @@ simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
 } // namespace westmere
 } // namespace simdutf
 /* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF16
 /* begin file src/generic/utf16.h */
 namespace simdutf {
 namespace westmere {
@@ -43069,6 +51664,87 @@ simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
                                                                    size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
+
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
+
+  const auto one = vector_u16::splat(1);
+
+  auto v_count = vector_u16::zero();
+
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
+
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
+
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
+
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
+
+    /*
+        Explanation how the counting works.
+
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
+
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
+    }
+  }
+
+  if (iteration > 0) {
+    count += v_count.sum();
+  }
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
+}
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
+
 template <endianness big_endian>
 simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
                                                      size_t size) {
@@ -43095,9 +51771,144 @@ change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
 } // namespace westmere
 } // namespace simdutf
 /* end file src/generic/utf16.h */
-// transcoding from UTF-8 to Latin 1
-/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf16.h */
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace utf16 {
+/*
+    UTF-16 validation
+    --------------------------------------------------
+
+    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
+
+    In a vectorized algorithm we want to examine the most significant
+    nibble in order to select a fast path. If none of highest nibbles
+    are 0xD (13), than we are sure that UTF-16 chunk in a vector
+    register is valid.
+
+    Let us analyze what we need to check if the nibble is 0xD. The
+    value of the preceding nibble determines what we have:
+
+    0xd000 .. 0xd7ff - a valid word
+    0xd800 .. 0xdbff - low surrogate
+    0xdc00 .. 0xdfff - high surrogate
+
+    Other constraints we have to consider:
+    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
+    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
+    - there must not be sole low surrogate nor high surrogate
+
+    We are going to build three bitmasks based on the 3rd nibble:
+    - V = valid word,
+    - L = low surrogate (0xd800 .. 0xdbff)
+    - H = high surrogate (0xdc00 .. 0xdfff)
+
+      0   1   2   3   4   5   6   7    <--- word index
+    [ V | L | H | L | H | V | V | L ]
+      1   0   0   0   0   1   1   0     - V = valid masks
+      0   1   0   1   0   0   0   1     - L = low surrogate
+      0   0   1   0   1   0   0   0     - H high surrogate
+
+
+      1   0   0   0   0   1   1   0   V = valid masks
+      0   1   0   1   0   0   0   0   a = L & (H >> 1)
+      0   0   1   0   1   0   0   0   b = a << 1
+      1   1   1   1   1   1   1   0   c = V | a | b
+                                  ^
+                                  the last bit can be zero, we just consume 7
+   code units and recheck this word in the next iteration
+*/
+template <endianness big_endian>
+const result validate_utf16_with_errors(const char16_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+
+  const char16_t *start = input;
+  const char16_t *end = input + size;
+
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+
+  while (input + simd16<uint16_t>::SIZE * 2 < end) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
+
+    // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16
+    // and yields a single vector having only higher bytes of characters.
+    const auto in = utf16_gather_high_bytes<big_endian>(in0, in1);
+
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const auto surrogates_wordmask = (in & v_f8) == v_d8;
+    const uint16_t surrogates_bitmask =
+        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
+    if (surrogates_bitmask == 0x0000) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher byte)
+
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = (in & v_fc) == v_dc;
+      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
+
+      const uint16_t a = static_cast<uint16_t>(
+          L & (H >> 1)); // A low surrogate must be followed by high one.
+                         // (A low surrogate placed in the 7th register's word
+                         // is an exception we handle.)
+      const uint16_t b = static_cast<uint16_t>(
+          a << 1); // Just mark that the opinput - startite fact is hold,
+                   // thanks to that we have only two masks for valid case.
+      const uint16_t c = static_cast<uint16_t>(
+          V | a | b); // Combine all the masks into the final one.
+
+      if (c == 0xffff) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0x7fff) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return result(error_code::SURROGATE, input - start);
+      }
+    }
+  }
+
+  return result(error_code::SUCCESS, input - start);
+}
 
+} // namespace utf16
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+/* end file src/generic/validate_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+/* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -43416,7 +52227,6 @@ struct validating_transcoder {
 } // namespace simdutf
 /* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 /* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-
 namespace simdutf {
 namespace westmere {
 namespace {
@@ -43496,6 +52306,361 @@ simdutf_really_inline size_t convert_valid(const char *in, size_t size,
 } // namespace simdutf
   // namespace simdutf
 /* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf32.h */
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace utf32 {
+
+simdutf_really_inline bool validate(const char32_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return true;
+  }
+
+  const char32_t *end = input + size;
+
+  using vector_u32 = simd32<uint32_t>;
+
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+  auto currentmax = vector_u32::zero();
+  auto currentoffsetmax = vector_u32::zero();
+
+  constexpr size_t N = vector_u32::ELEMENTS;
+
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
+    }
+
+    currentmax = max(currentmax, in);
+    currentoffsetmax = max(currentoffsetmax, in + offset);
+    input += N;
+  }
+
+  const auto too_large = currentmax > standardmax;
+  if (too_large.any()) {
+    return false;
+  }
+
+  const auto surrogate = currentoffsetmax > standardoffsetmax;
+  if (surrogate.any()) {
+    return false;
+  }
+
+  return scalar::utf32::validate(input, end - input);
+}
+
+simdutf_really_inline result validate_with_errors(const char32_t *input,
+                                                  size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    // empty input is valid UTF-32. protect the implementation from
+    // handling nullptr
+    return result(error_code::SUCCESS, 0);
+  }
+
+  const char32_t *start = input;
+  const char32_t *end = input + size;
+
+  using vector_u32 = simd32<uint32_t>;
+
+  const auto standardmax = vector_u32::splat(0x10ffff);
+  const auto offset = vector_u32::splat(0xffff2000);
+  const auto standardoffsetmax = vector_u32::splat(0xfffff7ff);
+
+  constexpr size_t N = vector_u32::ELEMENTS;
+
+  while (input + N < end) {
+    auto in = vector_u32(input);
+    if (!match_system(endianness::BIG)) {
+      in.swap_bytes();
+    }
+
+    const auto too_large = in > standardmax;
+    const auto surrogate = (in + offset) > standardoffsetmax;
+
+    const auto combined = too_large | surrogate;
+    if (simdutf_unlikely(combined.any())) {
+      const size_t consumed = input - start;
+      auto sr = scalar::utf32::validate_with_errors(input, end - input);
+      sr.count += consumed;
+
+      return sr;
+    }
+
+    input += N;
+  }
+
+  const size_t consumed = input - start;
+  auto sr = scalar::utf32::validate_with_errors(input, end - input);
+  sr.count += consumed;
+
+  return sr;
+}
+
+} // namespace utf32
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+/* end file src/generic/validate_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_BASE64
+/* begin file src/generic/base64.h */
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
+namespace simdutf {
+namespace westmere {
+namespace {
+namespace base64 {
+
+/*
+    The following template function implements API for Base64 decoding.
+
+    An implementation is responsible for providing the `block64` type and
+    associated methods that perform actual conversion. Please refer
+    to any vectorized implementation to learn the API of these procedures.
+*/
+template <bool base64_url, bool ignore_garbage, typename chartype>
+full_result
+compress_decode_base64(char *dst, const chartype *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  const uint8_t *to_base64 = base64_url ? tables::base64::to_base64_url_value
+                                        : tables::base64::to_base64_value;
+  size_t equallocation =
+      srclen; // location of the first padding character if any
+  // skip trailing spaces
+  while (!ignore_garbage && srclen > 0 &&
+         scalar::base64::is_eight_byte(src[srclen - 1]) &&
+         to_base64[uint8_t(src[srclen - 1])] == 64) {
+    srclen--;
+  }
+  size_t equalsigns = 0;
+  if (!ignore_garbage && srclen > 0 && src[srclen - 1] == '=') {
+    equallocation = srclen - 1;
+    srclen--;
+    equalsigns = 1;
+    // skip trailing spaces
+    while (srclen > 0 && scalar::base64::is_eight_byte(src[srclen - 1]) &&
+           to_base64[uint8_t(src[srclen - 1])] == 64) {
+      srclen--;
+    }
+    if (srclen > 0 && src[srclen - 1] == '=') {
+      equallocation = srclen - 1;
+      srclen--;
+      equalsigns = 2;
+    }
+  }
+  if (srclen == 0) {
+    if (!ignore_garbage && equalsigns > 0) {
+      if (last_chunk_options == last_chunk_handling_options::strict) {
+        return {BASE64_INPUT_REMAINDER, 0, 0};
+      } else if (last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial) {
+        return {SUCCESS, 0, 0};
+      }
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, 0, 0};
+  }
+  char *end_of_safe_64byte_zone =
+      (srclen + 3) / 4 * 3 >= 63 ? dst + (srclen + 3) / 4 * 3 - 63 : dst;
+
+  const chartype *const srcinit = src;
+  const char *const dstinit = dst;
+  const chartype *const srcend = src + srclen;
+
+  constexpr size_t block_size = 6;
+  static_assert(block_size >= 2, "block_size must be at least two");
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const chartype *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b(src);
+      src += 64;
+      uint64_t error = 0;
+      const uint64_t badcharmask =
+          b.to_base64_mask<base64_url, ignore_garbage>(&error);
+      if (!ignore_garbage && error) {
+        src -= 64;
+        const size_t error_offset = trailing_zeroes(error);
+        return {error_code::INVALID_BASE64_CHARACTER,
+                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+      }
+      if (badcharmask != 0) {
+        bufferptr += b.compress_block(badcharmask, bufferptr);
+      } else if (bufferptr != buffer) {
+        b.copy_block(bufferptr);
+        bufferptr += 64;
+      } else {
+        if (dst >= end_of_safe_64byte_zone) {
+          b.base64_decode_block_safe(dst);
+        } else {
+          b.base64_decode_block(dst);
+        }
+        dst += 48;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 2); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        if (dst >= end_of_safe_64byte_zone) {
+          base64_decode_block_safe(dst, buffer + (block_size - 2) * 64);
+        } else {
+          base64_decode_block(dst, buffer + (block_size - 2) * 64);
+        }
+        dst += 48;
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
+  }
+
+  char *buffer_start = buffer;
+  // Optimization note: if this is almost full, then it is worth our
+  // time, otherwise, we should just decode directly.
+  int last_block = (int)((bufferptr - buffer_start) % 64);
+  if (last_block != 0 && srcend - src + last_block >= 64) {
+
+    while ((bufferptr - buffer_start) % 64 != 0 && src < srcend) {
+      uint8_t val = to_base64[uint8_t(*src)];
+      *bufferptr = char(val);
+      if (!ignore_garbage &&
+          (!scalar::base64::is_eight_byte(*src) || val > 64)) {
+        return {error_code::INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+                size_t(dst - dstinit)};
+      }
+      bufferptr += (val <= 63);
+      src++;
+    }
+  }
+
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    if (dst >= end_of_safe_64byte_zone) {
+      base64_decode_block_safe(dst, buffer_start);
+    } else {
+      base64_decode_block(dst, buffer_start);
+    }
+    dst += 48;
+  }
+  if ((bufferptr - buffer_start) % 64 != 0) {
+    while (buffer_start + 4 < bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
+
+      dst += 3;
+      buffer_start += 4;
+    }
+    if (buffer_start + 4 <= bufferptr) {
+      uint32_t triple = ((uint32_t(uint8_t(buffer_start[0])) << 3 * 6) +
+                         (uint32_t(uint8_t(buffer_start[1])) << 2 * 6) +
+                         (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
+                         (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
+                        << 8;
+#if !SIMDUTF_IS_BIG_ENDIAN
+      triple = scalar::u32_swap_bytes(triple);
+#endif
+      std::memcpy(dst, &triple, 3);
+
+      dst += 3;
+      buffer_start += 4;
+    }
+    // we may have 1, 2 or 3 bytes left and we need to decode them so let us
+    // backtrack
+    int leftover = int(bufferptr - buffer_start);
+    while (leftover > 0) {
+      if (!ignore_garbage) {
+        while (to_base64[uint8_t(*(src - 1))] == 64) {
+          src--;
+        }
+      } else {
+        while (to_base64[uint8_t(*(src - 1))] >= 64) {
+          src--;
+        }
+      }
+      src--;
+      leftover--;
+    }
+  }
+  if (src < srcend + equalsigns) {
+    full_result r = scalar::base64::base64_tail_decode(
+        dst, src, srcend - src, equalsigns, options, last_chunk_options);
+    r.input_count += size_t(src - srcinit);
+    if (r.error == error_code::INVALID_BASE64_CHARACTER ||
+        r.error == error_code::BASE64_EXTRA_BITS) {
+      return r;
+    } else {
+      r.output_count += size_t(dst - dstinit);
+    }
+    if (!ignore_garbage && last_chunk_options != stop_before_partial &&
+        r.error == error_code::SUCCESS && equalsigns > 0) {
+      // additional checks
+      if ((r.output_count % 3 == 0) ||
+          ((r.output_count % 3) + 1 + equalsigns != 4)) {
+        r.error = error_code::INVALID_BASE64_CHARACTER;
+        r.input_count = equallocation;
+      }
+    }
+    return r;
+  }
+  if (!ignore_garbage && equalsigns > 0) {
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + equalsigns != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
+}
+
+} // namespace base64
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdutf
+/* end file src/generic/base64.h */
+#endif // SIMDUTF_FEATURE_BASE64
 
 //
 // Implementation-specific overrides
@@ -43504,6 +52669,7 @@ simdutf_really_inline size_t convert_valid(const char *in, size_t size,
 namespace simdutf {
 namespace westmere {
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused int
 implementation::detect_encodings(const char *input,
                                  size_t length) const noexcept {
@@ -43644,28 +52810,38 @@ implementation::detect_encodings(const char *input,
   }
   return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf8(const char *buf, size_t len) const noexcept {
   return westmere::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused result implementation::validate_utf8_with_errors(
     const char *buf, size_t len) const noexcept {
   return westmere::utf8_validation::generic_validate_utf8_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused bool
 implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return westmere::utf8_validation::generic_validate_ascii(buf, len);
+  return westmere::ascii_validation::generic_validate_ascii(buf, len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused result implementation::validate_ascii_with_errors(
     const char *buf, size_t len) const noexcept {
-  return westmere::utf8_validation::generic_validate_ascii_with_errors(buf,
-                                                                       len);
+  return westmere::ascii_validation::generic_validate_ascii_with_errors(buf,
+                                                                        len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf16le(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -43674,15 +52850,21 @@ implementation::validate_utf16le(const char16_t *buf,
     // handling nullptr
     return true;
   }
-  const char16_t *tail = sse_validate_utf16<endianness::LITTLE>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::LITTLE>(tail,
-                                                       len - (tail - buf));
-  } else {
+  const auto res =
+      westmere::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count == len)
+    return true;
+
+  return scalar::utf16::validate<endianness::LITTLE>(buf + res.count,
+                                                     len - res.count);
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused bool
 implementation::validate_utf16be(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -43691,20 +52873,27 @@ implementation::validate_utf16be(const char16_t *buf,
     // handling nullptr
     return true;
   }
-  const char16_t *tail = sse_validate_utf16<endianness::BIG>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
-  } else {
+  const auto res =
+      westmere::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count == len)
+    return true;
+
+  return scalar::utf16::validate<endianness::BIG>(buf + res.count,
+                                                  len - res.count);
 }
 
 simdutf_warn_unused result implementation::validate_utf16le_with_errors(
     const char16_t *buf, size_t len) const noexcept {
-  result res = sse_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  const result res =
+      westmere::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
   if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
-        buf + res.count, len - res.count);
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::LITTLE>(
+            buf + res.count, len - res.count);
     return result(scalar_res.error, res.count + scalar_res.count);
   } else {
     return res;
@@ -43713,7 +52902,8 @@ simdutf_warn_unused result implementation::validate_utf16le_with_errors(
 
 simdutf_warn_unused result implementation::validate_utf16be_with_errors(
     const char16_t *buf, size_t len) const noexcept {
-  result res = sse_validate_utf16_with_errors<endianness::BIG>(buf, len);
+  const result res =
+      westmere::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
   if (res.count != len) {
     result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
         buf + res.count, len - res.count);
@@ -43722,39 +52912,23 @@ simdutf_warn_unused result implementation::validate_utf16be_with_errors(
     return res;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
-  if (simdutf_unlikely(len == 0)) {
-    // empty input is valid UTF-32. protect the implementation from
-    // handling nullptr
-    return true;
-  }
-  const char32_t *tail = sse_validate_utf32le(buf, len);
-  if (tail) {
-    return scalar::utf32::validate(tail, len - (tail - buf));
-  } else {
-    return false;
-  }
+  return utf32::validate(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::validate_utf32_with_errors(
     const char32_t *buf, size_t len) const noexcept {
-  if (len == 0) {
-    // empty input is valid UTF-32. protect the implementation from
-    // handling nullptr
-    return result(error_code::SUCCESS, 0);
-  }
-  result res = sse_validate_utf32le_with_errors(buf, len);
-  if (res.count != len) {
-    result scalar_res =
-        scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
-    return result(scalar_res.error, res.count + scalar_res.count);
-  } else {
-    return res;
-  }
+  return utf32::validate_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
     const char *buf, size_t len, char *utf8_output) const noexcept {
 
@@ -43770,7 +52944,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
 
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char *, char16_t *> ret =
@@ -43810,7 +52986,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char *, char32_t *> ret =
@@ -43829,7 +53007,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
     const char *buf, size_t len, char *latin1_output) const noexcept {
   utf8_to_latin1::validating_transcoder converter;
@@ -43846,7 +53026,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
     const char *buf, size_t len, char *latin1_output) const noexcept {
   return westmere::utf8_to_latin1::convert_valid(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   utf8_to_utf16::validating_transcoder converter;
@@ -43883,7 +53065,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
   return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
                                                        utf16_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   utf8_to_utf32::validating_transcoder converter;
@@ -43900,7 +53084,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
     const char *input, size_t size, char32_t *utf32_output) const noexcept {
   return utf8_to_utf32::convert_valid(input, size, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
     const char16_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -44008,7 +53194,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
   // optimization opportunity: we could provide an optimized function.
   return convert_utf16le_to_latin1(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -44114,7 +53302,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   return convert_utf16be_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
     const char32_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char32_t *, char *> ret =
@@ -44163,7 +53353,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
   // optimization opportunity: we could provide an optimized function.
   return convert_utf32_to_latin1(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char32_t *, char *> ret =
@@ -44204,7 +53396,9 @@ simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
       utf8_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char16_t *, char32_t *> ret =
@@ -44300,12 +53494,16 @@ simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
       utf32_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   return convert_utf32_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char32_t *, char16_t *> ret =
@@ -44413,7 +53611,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   return convert_utf16be_to_utf32(buf, len, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 void implementation::change_endianness_utf16(const char16_t *input,
                                              size_t length,
                                              char16_t *output) const noexcept {
@@ -44429,47 +53629,36 @@ simdutf_warn_unused size_t implementation::count_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::count_code_points<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused size_t
 implementation::count_utf8(const char *input, size_t length) const noexcept {
-  return utf8::count_code_points(input, length);
+  return utf8::count_code_points_bytemask(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
     const char *buf, size_t len) const noexcept {
   return count_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return scalar::utf16::latin1_length_from_utf16(length);
-}
-
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return scalar::utf32::latin1_length_from_utf32(length);
-}
-
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
+  return utf16::utf8_length_from_utf16_bytemask<endianness::LITTLE>(input,
+                                                                    length);
 }
 
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
-  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
-}
-
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf16_length_from_latin1(length);
-}
-
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return scalar::latin1::utf32_length_from_latin1(length);
+  return utf16::utf8_length_from_utf16_bytemask<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
     const char *input, size_t len) const noexcept {
   const uint8_t *str = reinterpret_cast<const uint8_t *>(input);
@@ -44529,7 +53718,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
   return answer + scalar::latin1::utf8_length_from_latin1(
                       reinterpret_cast<const char *>(str + i), len - i);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
@@ -44539,48 +53730,23 @@ simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::utf16_length_from_utf8(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  const __m128i v_00000000 = _mm_setzero_si128();
-  const __m128i v_ffffff80 = _mm_set1_epi32((uint32_t)0xffffff80);
-  const __m128i v_fffff800 = _mm_set1_epi32((uint32_t)0xfffff800);
-  const __m128i v_ffff0000 = _mm_set1_epi32((uint32_t)0xffff0000);
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 4 <= length; pos += 4) {
-    __m128i in = _mm_loadu_si128((__m128i *)(input + pos));
-    const __m128i ascii_bytes_bytemask =
-        _mm_cmpeq_epi32(_mm_and_si128(in, v_ffffff80), v_00000000);
-    const __m128i one_two_bytes_bytemask =
-        _mm_cmpeq_epi32(_mm_and_si128(in, v_fffff800), v_00000000);
-    const __m128i two_bytes_bytemask =
-        _mm_xor_si128(one_two_bytes_bytemask, ascii_bytes_bytemask);
-    const __m128i one_two_three_bytes_bytemask =
-        _mm_cmpeq_epi32(_mm_and_si128(in, v_ffff0000), v_00000000);
-    const __m128i three_bytes_bytemask =
-        _mm_xor_si128(one_two_three_bytes_bytemask, one_two_bytes_bytemask);
-    const uint16_t ascii_bytes_bitmask =
-        static_cast<uint16_t>(_mm_movemask_epi8(ascii_bytes_bytemask));
-    const uint16_t two_bytes_bitmask =
-        static_cast<uint16_t>(_mm_movemask_epi8(two_bytes_bytemask));
-    const uint16_t three_bytes_bitmask =
-        static_cast<uint16_t>(_mm_movemask_epi8(three_bytes_bytemask));
-
-    size_t ascii_count = count_ones(ascii_bytes_bitmask) / 4;
-    size_t two_bytes_count = count_ones(two_bytes_bitmask) / 4;
-    size_t three_bytes_count = count_ones(three_bytes_bitmask) / 4;
-    count += 16 - 3 * ascii_count - 2 * two_bytes_count - three_bytes_count;
-  }
-  return count +
-         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+  return utf32::utf8_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
   const __m128i v_00000000 = _mm_setzero_si128();
@@ -44599,35 +53765,34 @@ simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
   return count +
          scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
+#if SIMDUTF_FEATURE_BASE64
 simdutf_warn_unused result implementation::base64_to_binary(
     const char *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
   if (options & base64_url) {
     if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
   } else {
     if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
 }
@@ -44637,46 +53802,41 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
     last_chunk_handling_options last_chunk_options) const noexcept {
   if (options & base64_url) {
     if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
   } else {
     if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
 }
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
 simdutf_warn_unused result implementation::base64_to_binary(
     const char16_t *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
   if (options & base64_url) {
     if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
   } else {
     if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
 }
@@ -44686,28 +53846,23 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
     last_chunk_handling_options last_chunk_options) const noexcept {
   if (options & base64_url) {
     if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
+      return base64::compress_decode_base64<true, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<true, false>(
+          output, input, length, options, last_chunk_options);
     }
   } else {
     if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
+      return base64::compress_decode_base64<false, true>(
+          output, input, length, options, last_chunk_options);
     } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
+      return base64::compress_decode_base64<false, false>(
+          output, input, length, options, last_chunk_options);
     }
   }
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
-}
-
 size_t implementation::binary_to_base64(const char *input, size_t length,
                                         char *output,
                                         base64_options options) const noexcept {
@@ -44717,6 +53872,8 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
     return encode_base64<false>(output, input, length, options);
   }
 }
+#endif // SIMDUTF_FEATURE_BASE64
+
 } // namespace westmere
 } // namespace simdutf
 
@@ -44727,6 +53884,7 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
 SIMDUTF_UNTARGET_REGION
 #endif
 
+#undef SIMDUTF_SIMD_HAS_BYTEMASK
 /* end file src/simdutf/westmere/end.h */
 /* end file src/westmere/implementation.cpp */
 #endif
@@ -44735,6 +53893,7 @@ SIMDUTF_UNTARGET_REGION
 /* begin file src/simdutf/lsx/begin.h */
 // redefining SIMDUTF_IMPLEMENTATION to "lsx"
 // #define SIMDUTF_IMPLEMENTATION lsx
+#define SIMDUTF_SIMD_HAS_UNSIGNED_CMP 1
 /* end file src/simdutf/lsx/begin.h */
 namespace simdutf {
 namespace lsx {
@@ -44744,6 +53903,7 @@ namespace {
 #endif
 using namespace simd;
 
+#if SIMDUTF_FEATURE_UTF8
 // convert vmskltz/vmskgez/vmsknz to
 // simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes index
 const uint8_t lsx_1_2_utf8_bytes_mask[] = {
@@ -44765,33 +53925,23 @@ const uint8_t lsx_1_2_utf8_bytes_mask[] = {
     169, 172, 173, 184, 185, 188, 189, 232, 233, 236, 237, 248, 249, 252, 253,
     170, 171, 174, 175, 186, 187, 190, 191, 234, 235, 238, 239, 250, 251, 254,
     255};
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 simdutf_really_inline __m128i lsx_swap_bytes(__m128i vec) {
-  // const v16u8 shuf = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
-  // return __lsx_vshuf_b(__lsx_vldi(0), vec, shuf);
   return __lsx_vshuf4i_b(vec, 0b10110001);
-  // return __lsx_vor_v(__lsx_vslli_h(vec, 8), __lsx_vsrli_h(vec, 8));
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    SIMDUTF_FEATURE_UTF8
 simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
   return input.is_ascii();
 }
+#endif // SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // SIMDUTF_FEATURE_UTF8
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<bool> is_second_byte = prev1 >= uint8_t(0b11000000u);
-  simd8<bool> is_third_byte = prev2 >= uint8_t(0b11100000u);
-  simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
-  // Use ^ instead of | for is_*_byte, because ^ is commutative, and the caller
-  // is using ^ as well. This will work fine because we only have to report
-  // errors for cases with 0-1 lead bytes. Multiple lead bytes implies 2
-  // overlapping multibyte characters, and if that happens, there is guaranteed
-  // to be at least *one* lead byte that is part of only 1 other multibyte
-  // character. The error will be detected there.
-  return is_second_byte ^ is_third_byte ^ is_fourth_byte;
-}
-
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_really_inline simd8<bool>
 must_be_2_3_continuation(const simd8<uint8_t> prev2,
                          const simd8<uint8_t> prev3) {
@@ -44799,7 +53949,9 @@ must_be_2_3_continuation(const simd8<uint8_t> prev2,
   simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
   return is_third_byte ^ is_fourth_byte;
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32)
 // common functions for utf8 conversions
 simdutf_really_inline __m128i convert_utf8_3_byte_to_utf16(__m128i in) {
   // Low half contains  10bbbbbb|10cccccc
@@ -44849,7 +54001,7 @@ convert_utf8_1_to_2_byte_to_utf16(__m128i in, size_t shufutf8_idx) {
   __m128i ascii = __lsx_vand_v(perm, __lsx_vrepli_h(0x7f)); // 6 or 7 bits
   // 1 byte: 00000000 00000000
   // 2 byte: 00000aaa aa000000
-  const __m128i v1f00 = __lsx_vldi(-2785); // -2785(13bit) => 151f
+  const __m128i v1f00 = lsx_splat_u16(0x1f00);
   __m128i composed = __lsx_vsrli_h(__lsx_vand_v(perm, v1f00), 2); // 5 bits
   // Combine with a shift right accumulate
   // 1 byte: 00000000 0bbbbbbb
@@ -44857,220 +54009,36 @@ convert_utf8_1_to_2_byte_to_utf16(__m128i in, size_t shufutf8_idx) {
   composed = __lsx_vadd_h(ascii, composed);
   return composed;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32)
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/lsx/lsx_validate_utf16.cpp */
-/*
-    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
-
-    In a vectorized algorithm we want to examine the most significant
-    nibble in order to select a fast path. If none of highest nibbles
-    are 0xD (13), than we are sure that UTF-16 chunk in a vector
-    register is valid.
-
-    Let us analyze what we need to check if the nibble is 0xD. The
-    value of the preceding nibble determines what we have:
-
-    0xd000 .. 0xd7ff - a valid word
-    0xd800 .. 0xdbff - low surrogate
-    0xdc00 .. 0xdfff - high surrogate
-
-    Other constraints we have to consider:
-    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
-    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
-    - there must not be sole low surrogate nor high surrogate
-
-    We're going to build three bitmasks based on the 3rd nibble:
-    - V = valid word,
-    - L = low surrogate (0xd800 .. 0xdbff)
-    - H = high surrogate (0xdc00 .. 0xdfff)
-
-      0   1   2   3   4   5   6   7    <--- word index
-    [ V | L | H | L | H | V | V | L ]
-      1   0   0   0   0   1   1   0     - V = valid masks
-      0   1   0   1   0   0   0   1     - L = low surrogate
-      0   0   1   0   1   0   0   0     - H high surrogate
-
-
-      1   0   0   0   0   1   1   0   V = valid masks
-      0   1   0   1   0   0   0   0   a = L & (H >> 1)
-      0   0   1   0   1   0   0   0   b = a << 1
-      1   1   1   1   1   1   1   0   c = V | a | b
-                                  ^
-                                  the last bit can be zero, we just consume 7
-   code units and recheck this word in the next iteration
-*/
-
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
-template <endianness big_endian>
-const char16_t *lsx_validate_utf16(const char16_t *input, size_t size) {
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::SIZE * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-    const auto in = simd8<uint8_t>(__lsx_vssrlni_bu_h(in1.value, in0.value, 8));
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint16_t surrogates_bitmask =
-        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
-    if (surrogates_bitmask == 0x0000) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
-
-      const uint16_t a = static_cast<uint16_t>(
-          L & (H >> 1)); // A low surrogate must be followed by high one.
-                         // (A low surrogate placed in the 7th register's word
-                         // is an exception we handle.)
-      const uint16_t b = static_cast<uint16_t>(
-          a << 1); // Just mark that the opinput - startite fact is hold,
-                   // thanks to that we have only two masks for valid case.
-      const uint16_t c = static_cast<uint16_t>(
-          V | a | b); // Combine all the masks into the final one.
-
-      if (c == 0xffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0x7fff) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return nullptr;
-      }
-    }
-  }
-
-  return input;
-}
-
 template <endianness big_endian>
-const result lsx_validate_utf16_with_errors(const char16_t *input,
-                                            size_t size) {
-  const char16_t *start = input;
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::SIZE * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 =
-        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-
-    const auto in = simd8<uint8_t>(__lsx_vssrlni_bu_h(in1.value, in0.value, 8));
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint16_t surrogates_bitmask =
-        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
-    if (surrogates_bitmask == 0x0000) {
-      input += 16;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
-
-      const uint16_t a = static_cast<uint16_t>(
-          L & (H >> 1)); // A low surrogate must be followed by high one.
-                         // (A low surrogate placed in the 7th register's word
-                         // is an exception we handle.)
-      const uint16_t b = static_cast<uint16_t>(
-          a << 1); // Just mark that the opinput - startite fact is hold,
-                   // thanks to that we have only two masks for valid case.
-      const uint16_t c = static_cast<uint16_t>(
-          V | a | b); // Combine all the masks into the final one.
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> in0,
+                                       const simd16<uint16_t> in1) {
+  if (big_endian) {
+    const auto mask = simd16<uint16_t>(0x00ff);
+    const auto t0 = in0 & mask;
+    const auto t1 = in1 & mask;
 
-      if (c == 0xffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += 16;
-      } else if (c == 0x7fff) {
-        // The 15 lower code units of the input register contains valid UTF-16.
-        // The 15th word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += 15;
-      } else {
-        return result(error_code::SURROGATE, input - start);
-      }
-    }
+    return simd16<uint16_t>::pack(t0, t1);
+  } else {
+    return simd8<uint8_t>(__lsx_vssrlni_bu_h(in1.value, in0.value, 8));
   }
-
-  return result(error_code::SUCCESS, input - start);
 }
 /* end file src/lsx/lsx_validate_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/lsx/lsx_validate_utf32le.cpp */
-
 const char32_t *lsx_validate_utf32le(const char32_t *input, size_t size) {
   const char32_t *end = input + size;
 
-  __m128i offset = __lsx_vreplgr2vr_w(uint32_t(0xffff2000));
-  __m128i standardoffsetmax = __lsx_vreplgr2vr_w(uint32_t(0xfffff7ff));
-  __m128i standardmax = __lsx_vldi(-2288); /*0x10ffff*/
-  __m128i currentmax = __lsx_vldi(0x0);
-  __m128i currentoffsetmax = __lsx_vldi(0x0);
+  __m128i offset = lsx_splat_u32(0xffff2000);
+  __m128i standardoffsetmax = lsx_splat_u32(0xfffff7ff);
+  __m128i standardmax = lsx_splat_u32(0x10ffff);
+  __m128i currentmax = lsx_splat_u32(0);
+  __m128i currentoffsetmax = lsx_splat_u32(0);
 
   while (input + 4 < end) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint32_t *>(input), 0);
@@ -45102,11 +54070,11 @@ const result lsx_validate_utf32le_with_errors(const char32_t *input,
   const char32_t *start = input;
   const char32_t *end = input + size;
 
-  __m128i offset = __lsx_vreplgr2vr_w(uint32_t(0xffff2000));
-  __m128i standardoffsetmax = __lsx_vreplgr2vr_w(uint32_t(0xfffff7ff));
-  __m128i standardmax = __lsx_vldi(-2288); /*0x10ffff*/
-  __m128i currentmax = __lsx_vldi(0x0);
-  __m128i currentoffsetmax = __lsx_vldi(0x0);
+  __m128i offset = lsx_splat_u32(0xffff2000);
+  __m128i standardoffsetmax = lsx_splat_u32(0xfffff7ff);
+  __m128i standardmax = lsx_splat_u32(0x10ffff);
+  __m128i currentmax = lsx_splat_u32(0);
+  __m128i currentoffsetmax = lsx_splat_u32(0);
 
   while (input + 4 < end) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint32_t *>(input), 0);
@@ -45132,7 +54100,9 @@ const result lsx_validate_utf32le_with_errors(const char32_t *input,
   return result(error_code::SUCCESS, input - start);
 }
 /* end file src/lsx/lsx_validate_utf32le.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_latin1_to_utf8.cpp */
 /*
   Returns a pair: the first unprocessed byte from buf and utf8_output
@@ -45148,7 +54118,7 @@ lsx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
   __m128i zero = __lsx_vldi(0);
   // We always write 16 bytes, of which more than the first 8 bytes
   // are valid. A safety margin of 8 is more than sufficient.
-  while (latin1_input + 16 <= end) {
+  while (end - latin1_input >= 16) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(latin1_input), 0);
     uint32_t ascii = __lsx_vpickve2gr_hu(__lsx_vmskgez_b(in8), 0);
     if (ascii == 0xffff) { // ASCII fast path!!!!
@@ -45166,7 +54136,7 @@ lsx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
     // t0 = [0000|00aa|bbbb|bb00]
     __m128i t0 = __lsx_vslli_h(in16, 2);
     // t1 = [0000|00aa|0000|0000]
-    __m128i t1 = __lsx_vand_v(t0, __lsx_vldi(-2785));
+    __m128i t1 = __lsx_vand_v(t0, lsx_splat_u16(0x300));
     // t3 = [0000|00aa|00bb|bbbb]
     __m128i t2 = __lsx_vbitsel_v(t1, in16, __lsx_vrepli_h(0x3f));
     // t4 = [1100|00aa|10bb|bbbb]
@@ -45191,6 +54161,8 @@ lsx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
   return std::make_pair(latin1_input, reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lsx/lsx_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_latin1_to_utf16.cpp */
 std::pair<const char *, char16_t *>
 lsx_convert_latin1_to_utf16le(const char *buf, size_t len,
@@ -45198,7 +54170,7 @@ lsx_convert_latin1_to_utf16le(const char *buf, size_t len,
   const char *end = buf + len;
 
   __m128i zero = __lsx_vldi(0);
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m128i inlow = __lsx_vilvl_b(zero, in8);
@@ -45218,7 +54190,7 @@ lsx_convert_latin1_to_utf16be(const char *buf, size_t len,
                               char16_t *utf16_output) {
   const char *end = buf + len;
   __m128i zero = __lsx_vldi(0);
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m128i inlow = __lsx_vilvl_b(in8, zero);
@@ -45232,13 +54204,15 @@ lsx_convert_latin1_to_utf16be(const char *buf, size_t len,
   return std::make_pair(buf, utf16_output);
 }
 /* end file src/lsx/lsx_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_latin1_to_utf32.cpp */
 std::pair<const char *, char32_t *>
 lsx_convert_latin1_to_utf32(const char *buf, size_t len,
                             char32_t *utf32_output) {
   const char *end = buf + len;
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m128i zero = __lsx_vldi(0);
@@ -45261,7 +54235,9 @@ lsx_convert_latin1_to_utf32(const char *buf, size_t len,
   return std::make_pair(buf, utf32_output);
 }
 /* end file src/lsx/lsx_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /* begin file src/lsx/lsx_convert_utf8_to_utf16.cpp */
 // Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
 // end of the code points. Only the least significant 12 bits of the mask
@@ -45376,7 +54352,7 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // 1 byte: 00000000 00000000
     // 2 byte: xx0bbbbb 00000000
     // 3 byte: xxbbbbbb 00000000
-    __m128i middlebyte = __lsx_vand_v(lowperm, __lsx_vldi(-2561) /*0xFF00*/);
+    __m128i middlebyte = __lsx_vand_v(lowperm, lsx_splat_u16(0xFF00));
     // 1 byte: 00000000 0ccccccc
     // 2 byte: 0010bbbb bbcccccc
     // 3 byte: 0010bbbb bbcccccc
@@ -45403,6 +54379,15 @@ size_t convert_masked_utf8_to_utf16(const char *input,
       // of the extra memory access is less important than the early branch
       // overhead in shorter sequences.
 
+      __m128i expected_mask =
+          (__m128i)v16u8{0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0,
+                         0xf8, 0xc0, 0xc0, 0xc0, 0x0,  0x0,  0x0,  0x0};
+      __m128i expected =
+          (__m128i)v16u8{0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80,
+                         0xf0, 0x80, 0x80, 0x80, 0x0,  0x0,  0x0,  0x0};
+      __m128i check = __lsx_vseq_b(__lsx_vand_v(in, expected_mask), expected);
+      if (__lsx_bz_b(check))
+        return 12;
       // Swap byte pairs
       // 10dddddd 10cccccc|10bbbbbb 11110aaa
       // 10cccccc 10dddddd|11110aaa 10bbbbbb
@@ -45419,10 +54404,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
       //                   = +0xDC00|0xE7C0
       __m128i magic = __lsx_vreplgr2vr_w(uint32_t(0xDC00E7C0));
       // Generate unadjusted trail surrogate minus lowest 2 bits
-      // vec(0000FF00) = __lsx_vldi(-1758)
       // xxxxxxxx xxxxxxxx|11110aaa bbbbbb00
-      __m128i trail =
-          __lsx_vbitsel_v(shift, swap, __lsx_vldi(-1758 /*0000FF00*/));
+      __m128i trail = __lsx_vbitsel_v(shift, swap, lsx_splat_u32(0x0000ff00));
       // Insert low 2 bits of trail surrogate to magic number for later
       // 11011100 00000000 11100111 110000cc
       __m128i magic_with_low_2 = __lsx_vor_v(__lsx_vsrli_w(shift, 30), magic);
@@ -45435,10 +54418,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
           __lsx_vrepli_h(0x3f /* 0x003f*/));
 
       // Blend pairs
-      // __lsx_vldi(-1741) => vec(0x0000FFFF)
       // 000000cc ccdddddd|11110aaa bbbbbb00
-      __m128i blend =
-          __lsx_vbitsel_v(lead, trail, __lsx_vldi(-1741) /* (0x0000FFFF)*4 */);
+      __m128i blend = __lsx_vbitsel_v(lead, trail, lsx_splat_u32(0x0000FFFF));
 
       // Add magic number to finish the result
       // 110111CC CCDDDDDD|110110AA BBBBBBCC
@@ -45477,13 +54458,12 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // first.
     __m128i middlehigh = __lsx_vslli_w(perm, 2);
     // 00000000 00000000 00cccccc 00000000
-    __m128i middlebyte = __lsx_vand_v(perm, __lsx_vldi(-3777) /* 0x00003F00 */);
+    __m128i middlebyte = __lsx_vand_v(perm, lsx_splat_u32(0x00003F00));
     // Start assembling the sequence. Since the 4th byte is in the same position
     // as it would be in a surrogate and there is no dependency, shift left
     // instead of right. 3 byte: 00000000 10bbbbxx xxxxxxxx xxxxxxxx 4 byte:
     // 11110aaa bbbbbbxx xxxxxxxx xxxxxxxx
-    __m128i ab =
-        __lsx_vbitsel_v(middlehigh, perm, __lsx_vldi(-1656) /*0xFF000000*/);
+    __m128i ab = __lsx_vbitsel_v(middlehigh, perm, lsx_splat_u32(0xFF000000));
     // Top 16 bits contains the high ten bits of the surrogate pair before
     // correction 3 byte: 00000000 10bbbbcc|cccc0000 00000000 4 byte: 11110aaa
     // bbbbbbcc|cccc0000 00000000 - high 10 bits correct w/o correction
@@ -45497,8 +54477,7 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // After this is for surrogates
     // Blend the low and high surrogates
     // 4 byte: 11110aaa bbbbbbcc|bbbbcccc ccdddddd
-    __m128i mixed =
-        __lsx_vbitsel_v(abc, composed, __lsx_vldi(-1741) /*0x0000FFFF*/);
+    __m128i mixed = __lsx_vbitsel_v(abc, composed, lsx_splat_u32(0x0000FFFF));
     // Clear the upper 6 bits of the low surrogate. Don't clear the upper bits
     // yet as 0x10000 was not subtracted from the codepoint yet. 4 byte:
     // 11110aaa bbbbbbcc|000000cc ccdddddd
@@ -45552,6 +54531,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
   }
 }
 /* end file src/lsx/lsx_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lsx/lsx_convert_utf8_to_utf32.cpp */
 // Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
 // end of the code points. Only the least significant 12 bits of the mask
@@ -45650,14 +54631,13 @@ size_t convert_masked_utf8_to_utf32(const char *input,
     // The top bits will be corrected later in the bsl
     // 00000000 10bbbbbb 00000000
     __m128i middle =
-        __lsx_vand_v(perm, __lsx_vldi(-1758 /*0x0000FF00*/)); // 5 or 6 bits
+        __lsx_vand_v(perm, lsx_splat_u32(0x0000FF00)); // 5 or 6 bits
     // Combine low and middle with shift right accumulate
     // 00000000 00xxbbbb bbcccccc
     __m128i lowmid = __lsx_vor_v(ascii, __lsx_vsrli_w(middle, 2));
     // Insert top 4 bits from high byte with bitwise select
     // 00000000 aaaabbbb bbcccccc
-    __m128i composed =
-        __lsx_vbitsel_v(lowmid, high, __lsx_vldi(-3600 /*0x0000F000*/));
+    __m128i composed = __lsx_vbitsel_v(lowmid, high, lsx_splat_u32(0x0000F000));
     __lsx_vst(composed, utf32_output, 0);
     utf32_output += 4; // We wrote 4 32-bit characters.
     return consumed;
@@ -45686,10 +54666,10 @@ size_t convert_masked_utf8_to_utf32(const char *input,
       __m128i merge2 =
           __lsx_vbitsel_v(__lsx_vslli_w(merge1, 12), /* merge1 << 12 */
                           __lsx_vsrli_w(merge1, 16), /* merge1 >> 16 */
-                          __lsx_vldi(-2545));        /*0x00000FFF*/
+                          lsx_splat_u32(0x00000FFF));
       // Clear the garbage
       // 00000000 000aaabb bbbbcccc ccdddddd
-      __m128i composed = __lsx_vand_v(merge2, __lsx_vldi(-2273 /*0x1FFFFF*/));
+      __m128i composed = __lsx_vand_v(merge2, lsx_splat_u32(0x1FFFFF));
       // Store
       __lsx_vst(composed, utf32_output, 0);
       utf32_output += 3; // We wrote 3 32-bit characters.
@@ -45709,12 +54689,11 @@ size_t convert_masked_utf8_to_utf32(const char *input,
 
     // Ascii
     __m128i ascii = __lsx_vand_v(perm, __lsx_vrepli_w(0x7F));
-    __m128i middle = __lsx_vand_v(perm, __lsx_vldi(-3777 /*0x00003f00*/));
+    __m128i middle = __lsx_vand_v(perm, lsx_splat_u32(0x00003f00));
     // 00000000 00000000 0000cccc ccdddddd
-    __m128i cd =
-        __lsx_vbitsel_v(__lsx_vsrli_w(middle, 2), ascii, __lsx_vrepli_w(0x3f));
+    __m128i cd = __lsx_vor_v(__lsx_vsrli_w(middle, 2), ascii);
 
-    __m128i correction = __lsx_vand_v(perm, __lsx_vldi(-3520 /*0x00400000*/));
+    __m128i correction = __lsx_vand_v(perm, lsx_splat_u32(0x00400000));
     __m128i corrected = __lsx_vadd_b(perm, __lsx_vsrli_w(correction, 1));
     // Insert twice
     // 00000000 000aaabb bbbbxxxx xxxxxxxx
@@ -45724,8 +54703,7 @@ size_t convert_masked_utf8_to_utf32(const char *input,
         __lsx_vbitsel_v(corrected_srli2, corrected, __lsx_vrepli_h(0x3f));
     ab = __lsx_vsrli_w(ab, 4);
     // 00000000 000aaabb bbbbcccc ccdddddd
-    __m128i composed =
-        __lsx_vbitsel_v(ab, cd, __lsx_vldi(-2545 /*0x00000FFF*/));
+    __m128i composed = __lsx_vbitsel_v(ab, cd, lsx_splat_u32(0x00000FFF));
     // Store
     __lsx_vst(composed, utf32_output, 0);
     utf32_output += 3; // We wrote 3 32-bit characters.
@@ -45736,6 +54714,8 @@ size_t convert_masked_utf8_to_utf32(const char *input,
   }
 }
 /* end file src/lsx/lsx_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_utf8_to_latin1.cpp */
 size_t convert_masked_utf8_to_latin1(const char *input,
                                      uint64_t utf8_end_of_code_point_mask,
@@ -45813,14 +54793,16 @@ size_t convert_masked_utf8_to_latin1(const char *input,
   return consumed;
 }
 /* end file src/lsx/lsx_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_utf16_to_latin1.cpp */
 template <endianness big_endian>
 std::pair<const char16_t *, char *>
 lsx_convert_utf16_to_latin1(const char16_t *buf, size_t len,
                             char *latin1_output) {
   const char16_t *end = buf + len;
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 16);
     if (!match_system(big_endian)) {
@@ -45848,7 +54830,7 @@ lsx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
                                         char *latin1_output) {
   const char16_t *start = buf;
   const char16_t *end = buf + len;
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 16);
     if (!match_system(big_endian)) {
@@ -45866,9 +54848,8 @@ lsx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
     } else {
       // Let us do a scalar fallback.
       for (int k = 0; k < 16; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if (word <= 0xff) {
           *latin1_output++ = char(word);
         } else {
@@ -45882,6 +54863,8 @@ lsx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/lsx/lsx_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
 /* begin file src/lsx/lsx_convert_utf16_to_utf8.cpp */
 /*
     The vectorized algorithm works on single SSE register i.e., it
@@ -45946,7 +54929,7 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
           // https://github.com/simdutf/simdutf/issues/92
 
   __m128i v_07ff = __lsx_vreplgr2vr_h(uint16_t(0x7ff));
-  while (buf + 16 + safety_margin <= end) {
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lsx_swap_bytes(in);
@@ -45990,7 +54973,7 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       // t0 = [000a|aaaa|bbbb|bb00]
       __m128i t0 = __lsx_vslli_h(in, 2);
       // t1 = [000a|aaaa|0000|0000]
-      __m128i t1 = __lsx_vand_v(t0, __lsx_vldi(-2785 /*0x1f00*/));
+      __m128i t1 = __lsx_vand_v(t0, lsx_splat_u16(0x1f00));
       // t2 = [0000|0000|00bb|bbbb]
       __m128i t2 = __lsx_vand_v(in, __lsx_vrepli_h(0x3f));
       // t3 = [000a|aaaa|00bb|bbbb]
@@ -46016,9 +54999,8 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       utf8_output += row[0];
       continue;
     }
-    __m128i surrogates_bytemask =
-        __lsx_vseq_h(__lsx_vand_v(in, __lsx_vldi(-2568 /*0xF800*/)),
-                     __lsx_vldi(-2600 /*0xD800*/));
+    __m128i surrogates_bytemask = __lsx_vseq_h(
+        __lsx_vand_v(in, lsx_splat_u16(0xf800)), lsx_splat_u16(0xd800));
     // It might seem like checking for surrogates_bitmask == 0xc000 could help.
     // However, it is likely an uncommon occurrence.
     if (__lsx_bz_v(surrogates_bytemask)) {
@@ -46058,14 +55040,14 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       __m128i v_3f7f = __lsx_vreplgr2vr_h(uint16_t(0x3F7F));
       __m128i t1 = __lsx_vand_v(t0, v_3f7f);
       // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      __m128i t2 = __lsx_vor_v(t1, __lsx_vldi(-2688 /*0x8000*/));
+      __m128i t2 = __lsx_vor_v(t1, lsx_splat_u16(0x8000));
 
       // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
       __m128i s0 = __lsx_vsrli_h(in, 12);
       // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
       __m128i s1 = __lsx_vslli_h(in, 2);
       // s1: [aabb|bbbb|cccc|cc00] => [00bb|bbbb|0000|0000]
-      s1 = __lsx_vand_v(s1, __lsx_vldi(-2753 /*0x3F00*/));
+      s1 = __lsx_vand_v(s1, lsx_splat_u16(0x3f00));
 
       // [00bb|bbbb|0000|aaaa]
       __m128i s2 = __lsx_vor_v(s0, s1);
@@ -46073,8 +55055,8 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       __m128i v_c0e0 = __lsx_vreplgr2vr_h(uint16_t(0xC0E0));
       __m128i s3 = __lsx_vor_v(s2, v_c0e0);
       __m128i one_or_two_bytes_bytemask = __lsx_vsle_hu(in, v_07ff);
-      __m128i m0 = __lsx_vandn_v(one_or_two_bytes_bytemask,
-                                 __lsx_vldi(-2752 /*0x4000*/));
+      __m128i m0 =
+          __lsx_vandn_v(one_or_two_bytes_bytemask, lsx_splat_u16(0x4000));
       __m128i s4 = __lsx_vxor_v(s3, m0);
 
       // 4. expand code units 16-bit => 32-bit
@@ -46130,9 +55112,8 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -46146,7 +55127,7 @@ lsx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -46185,7 +55166,7 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
   const size_t safety_margin =
       12; // to avoid overruns, see issue
           // https://github.com/simdutf/simdutf/issues/92
-  while (buf + 16 + safety_margin <= end) {
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lsx_swap_bytes(in);
@@ -46230,7 +55211,7 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       // t0 = [000a|aaaa|bbbb|bb00]
       __m128i t0 = __lsx_vslli_h(in, 2);
       // t1 = [000a|aaaa|0000|0000]
-      __m128i t1 = __lsx_vand_v(t0, __lsx_vldi(-2785 /*0x1f00*/));
+      __m128i t1 = __lsx_vand_v(t0, lsx_splat_u16(0x1f00));
       // t2 = [0000|0000|00bb|bbbb]
       __m128i t2 = __lsx_vand_v(in, __lsx_vrepli_h(0x3f));
       // t3 = [000a|aaaa|00bb|bbbb]
@@ -46256,9 +55237,8 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       utf8_output += row[0];
       continue;
     }
-    __m128i surrogates_bytemask =
-        __lsx_vseq_h(__lsx_vand_v(in, __lsx_vldi(-2568 /*0xF800*/)),
-                     __lsx_vldi(-2600 /*0xD800*/));
+    __m128i surrogates_bytemask = __lsx_vseq_h(
+        __lsx_vand_v(in, lsx_splat_u16(0xf800)), lsx_splat_u16(0xd800));
     // It might seem like checking for surrogates_bitmask == 0xc000 could help.
     // However, it is likely an uncommon occurrence.
     if (__lsx_bz_v(surrogates_bytemask)) {
@@ -46298,14 +55278,14 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       __m128i v_3f7f = __lsx_vreplgr2vr_h(uint16_t(0x3F7F));
       __m128i t1 = __lsx_vand_v(t0, v_3f7f);
       // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      __m128i t2 = __lsx_vor_v(t1, __lsx_vldi(-2688));
+      __m128i t2 = __lsx_vor_v(t1, lsx_splat_u16(0x8000));
 
       // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
       __m128i s0 = __lsx_vsrli_h(in, 12);
       // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
       __m128i s1 = __lsx_vslli_h(in, 2);
       // s1: [aabb|bbbb|cccc|cc00] => [00bb|bbbb|0000|0000]
-      s1 = __lsx_vand_v(s1, __lsx_vldi(-2753 /*0x3F00*/));
+      s1 = __lsx_vand_v(s1, lsx_splat_u16(0x3f00));
 
       // [00bb|bbbb|0000|aaaa]
       __m128i s2 = __lsx_vor_v(s0, s1);
@@ -46313,8 +55293,8 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       __m128i v_c0e0 = __lsx_vreplgr2vr_h(uint16_t(0xC0E0));
       __m128i s3 = __lsx_vor_v(s2, v_c0e0);
       __m128i one_or_two_bytes_bytemask = __lsx_vsle_hu(in, v_07ff);
-      __m128i m0 = __lsx_vandn_v(one_or_two_bytes_bytemask,
-                                 __lsx_vldi(-2752 /*0x4000*/));
+      __m128i m0 =
+          __lsx_vandn_v(one_or_two_bytes_bytemask, lsx_splat_u16(0x4000));
       __m128i s4 = __lsx_vxor_v(s3, m0);
 
       // 4. expand code units 16-bit => 32-bit
@@ -46370,9 +55350,8 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -46386,7 +55365,7 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -46410,6 +55389,8 @@ lsx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
                         reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lsx/lsx_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF8
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lsx/lsx_convert_utf16_to_utf32.cpp */
 template <endianness big_endian>
 std::pair<const char16_t *, char32_t *>
@@ -46419,10 +55400,10 @@ lsx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
   const char16_t *end = buf + len;
 
   __m128i zero = __lsx_vldi(0);
-  __m128i v_f800 = __lsx_vldi(-2568); /*0xF800*/
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
+  __m128i v_f800 = lsx_splat_u16(0xf800);
+  __m128i v_d800 = lsx_splat_u16(0xd800);
 
-  while (buf + 8 <= end) {
+  while (end - buf >= 8) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lsx_swap_bytes(in);
@@ -46450,16 +55431,15 @@ lsx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -46493,10 +55473,10 @@ lsx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
   const char16_t *end = buf + len;
 
   __m128i zero = __lsx_vldi(0);
-  __m128i v_f800 = __lsx_vldi(-2568); /*0xF800*/
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
+  __m128i v_f800 = lsx_splat_u16(0xf800);
+  __m128i v_d800 = lsx_splat_u16(0xd800);
 
-  while (buf + 8 <= end) {
+  while (end - buf >= 8) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lsx_swap_bytes(in);
@@ -46522,16 +55502,15 @@ lsx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -46551,7 +55530,9 @@ lsx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
                         reinterpret_cast<char32_t *>(utf32_output));
 }
 /* end file src/lsx/lsx_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lsx/lsx_convert_utf32_to_latin1.cpp */
 std::pair<const char32_t *, char *>
 lsx_convert_utf32_to_latin1(const char32_t *buf, size_t len,
@@ -46560,7 +55541,7 @@ lsx_convert_utf32_to_latin1(const char32_t *buf, size_t len,
   const v16u8 shuf_mask = {0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0};
   __m128i v_ff = __lsx_vrepli_w(0xFF);
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i in2 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
 
@@ -46589,7 +55570,7 @@ lsx_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
   const v16u8 shuf_mask = {0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0};
   __m128i v_ff = __lsx_vrepli_w(0xFF);
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i in2 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
 
@@ -46620,23 +55601,25 @@ lsx_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/lsx/lsx_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lsx/lsx_convert_utf32_to_utf8.cpp */
 std::pair<const char32_t *, char *>
 lsx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
   uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
   const char32_t *end = buf + len;
 
-  __m128i v_c080 = __lsx_vreplgr2vr_h(uint16_t(0xC080));
-  __m128i v_07ff = __lsx_vreplgr2vr_h(uint16_t(0x7FF));
-  __m128i v_dfff = __lsx_vreplgr2vr_h(uint16_t(0xDFFF));
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
+  __m128i v_c080 = lsx_splat_u16(0xc080);
+  __m128i v_07ff = lsx_splat_u16(0x07ff);
+  __m128i v_dfff = lsx_splat_u16(0xdfff);
+  __m128i v_d800 = lsx_splat_u16(0xd800);
   __m128i forbidden_bytemask = __lsx_vldi(0x0);
 
   const size_t safety_margin =
       12; // to avoid overruns, see issue
           // https://github.com/simdutf/simdutf/issues/92
 
-  while (buf + 16 + safety_margin < end) {
+  while (end - buf > std::ptrdiff_t(16 + safety_margin)) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i nextin = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
 
@@ -46667,7 +55650,7 @@ lsx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
         // t0 = [000a|aaaa|bbbb|bb00]
         const __m128i t0 = __lsx_vslli_h(utf16_packed, 2);
         // t1 = [000a|aaaa|0000|0000]
-        const __m128i t1 = __lsx_vand_v(t0, __lsx_vldi(-2785 /*0x1f00*/));
+        const __m128i t1 = __lsx_vand_v(t0, lsx_splat_u16(0x1f00));
         // t2 = [0000|0000|00bb|bbbb]
         const __m128i t2 = __lsx_vand_v(utf16_packed, __lsx_vrepli_h(0x3f));
         // t3 = [000a|aaaa|00bb|bbbb]
@@ -46736,23 +55719,22 @@ lsx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
         __m128i v_3f7f = __lsx_vreplgr2vr_h(uint16_t(0x3F7F));
         __m128i t1 = __lsx_vand_v(t0, v_3f7f);
         // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        __m128i t2 = __lsx_vor_v(t1, __lsx_vldi(-2688 /*0x8000*/));
+        __m128i t2 = __lsx_vor_v(t1, lsx_splat_u16(0x8000));
 
         // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
         __m128i s0 = __lsx_vsrli_h(utf16_packed, 12);
         // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
         __m128i s1 = __lsx_vslli_h(utf16_packed, 2);
         // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        s1 = __lsx_vand_v(s1, __lsx_vldi(-2753 /*0x3F00*/));
+        s1 = __lsx_vand_v(s1, lsx_splat_u16(0x3F00));
         // [00bb|bbbb|0000|aaaa]
         __m128i s2 = __lsx_vor_v(s0, s1);
         // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
         __m128i v_c0e0 = __lsx_vreplgr2vr_h(uint16_t(0xC0E0));
         __m128i s3 = __lsx_vor_v(s2, v_c0e0);
-        // __m128i v_07ff = vmovq_n_u16((uint16_t)0x07FF);
         __m128i one_or_two_bytes_bytemask = __lsx_vsle_hu(utf16_packed, v_07ff);
-        __m128i m0 = __lsx_vandn_v(one_or_two_bytes_bytemask,
-                                   __lsx_vldi(-2752 /*0x4000*/));
+        __m128i m0 =
+            __lsx_vandn_v(one_or_two_bytes_bytemask, lsx_splat_u16(0x4000));
         __m128i s4 = __lsx_vxor_v(s3, m0);
 
         // 4. expand code units 16-bit => 32-bit
@@ -46846,6 +55828,7 @@ lsx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
   if (__lsx_bnz_v(forbidden_bytemask)) {
     return std::make_pair(nullptr, reinterpret_cast<char *>(utf8_output));
   }
+
   return std::make_pair(buf, reinterpret_cast<char *>(utf8_output));
 }
 
@@ -46856,16 +55839,16 @@ lsx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
   const char32_t *start = buf;
   const char32_t *end = buf + len;
 
-  __m128i v_c080 = __lsx_vreplgr2vr_h(uint16_t(0xC080));
-  __m128i v_07ff = __lsx_vreplgr2vr_h(uint16_t(0x7FF));
-  __m128i v_dfff = __lsx_vreplgr2vr_h(uint16_t(0xDFFF));
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
+  __m128i v_c080 = lsx_splat_u16(0xc080);
+  __m128i v_07ff = lsx_splat_u16(0x07ff);
+  __m128i v_dfff = lsx_splat_u16(0xdfff);
+  __m128i v_d800 = lsx_splat_u16(0xd800);
   __m128i forbidden_bytemask = __lsx_vldi(0x0);
   const size_t safety_margin =
       12; // to avoid overruns, see issue
           // https://github.com/simdutf/simdutf/issues/92
 
-  while (buf + 16 + safety_margin < end) {
+  while (end - buf > std::ptrdiff_t(16 + safety_margin)) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i nextin = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
 
@@ -46896,7 +55879,7 @@ lsx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         // t0 = [000a|aaaa|bbbb|bb00]
         const __m128i t0 = __lsx_vslli_h(utf16_packed, 2);
         // t1 = [000a|aaaa|0000|0000]
-        const __m128i t1 = __lsx_vand_v(t0, __lsx_vldi(-2785 /*0x1f00*/));
+        const __m128i t1 = __lsx_vand_v(t0, lsx_splat_u16(0x1f00));
         // t2 = [0000|0000|00bb|bbbb]
         const __m128i t2 = __lsx_vand_v(utf16_packed, __lsx_vrepli_h(0x3f));
         // t3 = [000a|aaaa|00bb|bbbb]
@@ -46969,14 +55952,14 @@ lsx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         __m128i v_3f7f = __lsx_vreplgr2vr_h(uint16_t(0x3F7F));
         __m128i t1 = __lsx_vand_v(t0, v_3f7f);
         // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        __m128i t2 = __lsx_vor_v(t1, __lsx_vldi(-2688 /*0x8000*/));
+        __m128i t2 = __lsx_vor_v(t1, lsx_splat_u16(0x8000));
 
         // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
         __m128i s0 = __lsx_vsrli_h(utf16_packed, 12);
         // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
         __m128i s1 = __lsx_vslli_h(utf16_packed, 2);
         // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        s1 = __lsx_vand_v(s1, __lsx_vldi(-2753 /*0x3F00*/));
+        s1 = __lsx_vand_v(s1, lsx_splat_u16(0x3F00));
         // [00bb|bbbb|0000|aaaa]
         __m128i s2 = __lsx_vor_v(s0, s1);
         // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
@@ -46984,8 +55967,8 @@ lsx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         __m128i s3 = __lsx_vor_v(s2, v_c0e0);
         // __m128i v_07ff = vmovq_n_u16((uint16_t)0x07FF);
         __m128i one_or_two_bytes_bytemask = __lsx_vsle_hu(utf16_packed, v_07ff);
-        __m128i m0 = __lsx_vandn_v(one_or_two_bytes_bytemask,
-                                   __lsx_vldi(-2752 /*0x4000*/));
+        __m128i m0 =
+            __lsx_vandn_v(one_or_two_bytes_bytemask, lsx_splat_u16(0x4000));
         __m128i s4 = __lsx_vxor_v(s3, m0);
 
         // 4. expand code units 16-bit => 32-bit
@@ -47081,6 +56064,8 @@ lsx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
                         reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lsx/lsx_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lsx/lsx_convert_utf32_to_utf16.cpp */
 template <endianness big_endian>
 std::pair<const char32_t *, char16_t *>
@@ -47090,9 +56075,9 @@ lsx_convert_utf32_to_utf16(const char32_t *buf, size_t len,
   const char32_t *end = buf + len;
 
   __m128i forbidden_bytemask = __lsx_vrepli_h(0);
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
-  __m128i v_dfff = __lsx_vreplgr2vr_h(uint16_t(0xdfff));
-  while (buf + 8 <= end) {
+  __m128i v_d800 = lsx_splat_u16(0xd800);
+  __m128i v_dfff = lsx_splat_u16(0xdfff);
+  while (end - buf >= 8) {
     __m128i in0 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
 
@@ -47166,10 +56151,10 @@ lsx_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
   const char32_t *end = buf + len;
 
   __m128i forbidden_bytemask = __lsx_vrepli_h(0);
-  __m128i v_d800 = __lsx_vldi(-2600); /*0xD800*/
-  __m128i v_dfff = __lsx_vreplgr2vr_h(uint16_t(0xdfff));
+  __m128i v_d800 = lsx_splat_u16(0xd800);
+  __m128i v_dfff = lsx_splat_u16(0xdfff);
 
-  while (buf + 8 <= end) {
+  while (end - buf >= 8) {
     __m128i in0 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint32_t *>(buf), 16);
     // Check if no bits set above 16th
@@ -47238,6 +56223,8 @@ lsx_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
                         reinterpret_cast<char16_t *>(utf16_output));
 }
 /* end file src/lsx/lsx_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_BASE64
 /* begin file src/lsx/lsx_base64.cpp */
 /**
  * References and further reading:
@@ -47608,9 +56595,8 @@ static inline void load_block(block64 *b, const char16_t *src) {
 static inline void base64_decode(char *out, __m128i str) {
   __m128i t0 = __lsx_vor_v(
       __lsx_vslli_w(str, 26),
-      __lsx_vslli_w(__lsx_vand_v(str, __lsx_vldi(-1758 /*0x0000FF00*/)), 12));
-  __m128i t1 =
-      __lsx_vsrli_w(__lsx_vand_v(str, __lsx_vldi(-3521 /*0x003F0000*/)), 2);
+      __lsx_vslli_w(__lsx_vand_v(str, lsx_splat_u32(0x0000FF00)), 12));
+  __m128i t1 = __lsx_vsrli_w(__lsx_vand_v(str, lsx_splat_u32(0x003F0000)), 2);
   __m128i t2 = __lsx_vor_v(t0, t1);
   __m128i t3 = __lsx_vor_v(t2, __lsx_vsrli_w(str, 16));
   const v16u8 pack_shuffle = {3, 2,  1,  7,  6, 5, 11, 10,
@@ -47768,7 +56754,7 @@ compress_decode_base64(char *dst, const char_type *src, size_t srclen,
                          (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
                          (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
                         << 8;
-      triple = scalar::utf32::swap_bytes(triple);
+      triple = scalar::u32_swap_bytes(triple);
       std::memcpy(dst, &triple, 4);
 
       dst += 3;
@@ -47780,7 +56766,7 @@ compress_decode_base64(char *dst, const char_type *src, size_t srclen,
                          (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
                          (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
                         << 8;
-      triple = scalar::utf32::swap_bytes(triple);
+      triple = scalar::u32_swap_bytes(triple);
       std::memcpy(dst, &triple, 3);
 
       dst += 3;
@@ -47833,6 +56819,7 @@ compress_decode_base64(char *dst, const char_type *src, size_t srclen,
   return {SUCCESS, srclen, size_t(dst - dstinit)};
 }
 /* end file src/lsx/lsx_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
 
 } // namespace
 } // namespace lsx
@@ -47949,6 +56936,7 @@ simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
 } // namespace lsx
 } // namespace simdutf
 /* end file src/generic/buf_block_reader.h */
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
 namespace simdutf {
 namespace lsx {
@@ -48255,9 +57243,21 @@ result generic_validate_utf8_with_errors(const char *input, size_t length) {
       reinterpret_cast<const uint8_t *>(input), length);
 }
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+} // namespace utf8_validation
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
+namespace simdutf {
+namespace lsx {
+namespace {
+namespace ascii_validation {
+
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   uint8_t blocks[64]{};
   simd::simd8x64<uint8_t> running_or(blocks);
   while (reader.has_full_block()) {
@@ -48272,14 +57272,8 @@ bool generic_validate_ascii(const uint8_t *input, size_t length) {
   return running_or.is_ascii();
 }
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   size_t count{0};
   while (reader.has_full_block()) {
     simd::simd8x64<uint8_t> in(reader.full_block());
@@ -48304,20 +57298,16 @@ result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
   }
 }
 
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-} // namespace utf8_validation
+} // namespace ascii_validation
 } // unnamed namespace
 } // namespace lsx
 } // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
 
-// transcoding from UTF-8 to Latin 1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  // transcoding from UTF-8 to Latin 1
 /* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
-
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -48636,7 +57626,6 @@ struct validating_transcoder {
 } // namespace simdutf
 /* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 /* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -48716,9 +57705,11 @@ simdutf_really_inline size_t convert_valid(const char *in, size_t size,
 } // namespace simdutf
   // namespace simdutf
 /* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-// transcoding from UTF-8 to UTF-16
-/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  // transcoding from UTF-8 to UTF-16
+/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -48795,7 +57786,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 /* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -49129,9 +58119,11 @@ struct validating_transcoder {
 } // namespace lsx
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
-/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  // transcoding from UTF-8 to UTF-32
+/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -49176,7 +58168,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
 /* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -49496,11 +58487,10 @@ struct validating_transcoder {
 } // namespace lsx
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-
-// other functions
+#if SIMDUTF_FEATURE_UTF8
 /* begin file src/generic/utf8.h */
-
 namespace simdutf {
 namespace lsx {
 namespace {
@@ -49519,6 +58509,59 @@ simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
   return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
+
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
+
+  size_t pos = 0;
+  size_t count = 0;
+
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
+
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
+
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
+    }
+  }
+
+  if (iterations > 0) {
+    count += local.sum_bytes();
+  }
+
+  count += counters.sum();
+
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
+}
+#endif
+
 simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
                                                     size_t size) {
   size_t pos = 0;
@@ -49540,6 +58583,9 @@ simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
 } // namespace lsx
 } // namespace simdutf
 /* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
 /* begin file src/generic/utf16.h */
 namespace simdutf {
 namespace lsx {
@@ -49589,6 +58635,87 @@ simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
                                                                    size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
+
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
+
+  const auto one = vector_u16::splat(1);
+
+  auto v_count = vector_u16::zero();
+
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
+
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
+
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
+
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
+
+    /*
+        Explanation how the counting works.
+
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
+
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
+    }
+  }
+
+  if (iteration > 0) {
+    count += v_count.sum();
+  }
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
+}
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
+
 template <endianness big_endian>
 simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
                                                      size_t size) {
@@ -49615,6 +58742,284 @@ change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
 } // namespace lsx
 } // namespace simdutf
 /* end file src/generic/utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf16.h */
+namespace simdutf {
+namespace lsx {
+namespace {
+namespace utf16 {
+/*
+    UTF-16 validation
+    --------------------------------------------------
+
+    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
+
+    In a vectorized algorithm we want to examine the most significant
+    nibble in order to select a fast path. If none of highest nibbles
+    are 0xD (13), than we are sure that UTF-16 chunk in a vector
+    register is valid.
+
+    Let us analyze what we need to check if the nibble is 0xD. The
+    value of the preceding nibble determines what we have:
+
+    0xd000 .. 0xd7ff - a valid word
+    0xd800 .. 0xdbff - low surrogate
+    0xdc00 .. 0xdfff - high surrogate
+
+    Other constraints we have to consider:
+    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
+    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
+    - there must not be sole low surrogate nor high surrogate
+
+    We are going to build three bitmasks based on the 3rd nibble:
+    - V = valid word,
+    - L = low surrogate (0xd800 .. 0xdbff)
+    - H = high surrogate (0xdc00 .. 0xdfff)
+
+      0   1   2   3   4   5   6   7    <--- word index
+    [ V | L | H | L | H | V | V | L ]
+      1   0   0   0   0   1   1   0     - V = valid masks
+      0   1   0   1   0   0   0   1     - L = low surrogate
+      0   0   1   0   1   0   0   0     - H high surrogate
+
+
+      1   0   0   0   0   1   1   0   V = valid masks
+      0   1   0   1   0   0   0   0   a = L & (H >> 1)
+      0   0   1   0   1   0   0   0   b = a << 1
+      1   1   1   1   1   1   1   0   c = V | a | b
+                                  ^
+                                  the last bit can be zero, we just consume 7
+   code units and recheck this word in the next iteration
+*/
+template <endianness big_endian>
+const result validate_utf16_with_errors(const char16_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+
+  const char16_t *start = input;
+  const char16_t *end = input + size;
+
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+
+  while (input + simd16<uint16_t>::SIZE * 2 < end) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
+
+    // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16
+    // and yields a single vector having only higher bytes of characters.
+    const auto in = utf16_gather_high_bytes<big_endian>(in0, in1);
+
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const auto surrogates_wordmask = (in & v_f8) == v_d8;
+    const uint16_t surrogates_bitmask =
+        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
+    if (surrogates_bitmask == 0x0000) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher byte)
+
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = (in & v_fc) == v_dc;
+      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
+
+      const uint16_t a = static_cast<uint16_t>(
+          L & (H >> 1)); // A low surrogate must be followed by high one.
+                         // (A low surrogate placed in the 7th register's word
+                         // is an exception we handle.)
+      const uint16_t b = static_cast<uint16_t>(
+          a << 1); // Just mark that the opinput - startite fact is hold,
+                   // thanks to that we have only two masks for valid case.
+      const uint16_t c = static_cast<uint16_t>(
+          V | a | b); // Combine all the masks into the final one.
+
+      if (c == 0xffff) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0x7fff) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return result(error_code::SURROGATE, input - start);
+      }
+    }
+  }
+
+  return result(error_code::SUCCESS, input - start);
+}
+
+} // namespace utf16
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdutf
+/* end file src/generic/validate_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf32.h */
+#include <limits>
+
+namespace simdutf {
+namespace lsx {
+namespace {
+namespace utf32 {
+
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
+
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
+
+  const char32_t *start = input;
+
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
+
+  const size_t N = vector_u32::ELEMENTS;
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+  size_t counter = 0;
+
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
+
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
+
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += 4 * N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
+  }
+
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
+}
+
+} // namespace utf32
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdutf
+/* end file src/generic/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32
 
 //
 // Implementation-specific overrides
@@ -49622,6 +59027,7 @@ change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
 namespace simdutf {
 namespace lsx {
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused int
 implementation::detect_encodings(const char *input,
                                  size_t length) const noexcept {
@@ -49648,27 +59054,35 @@ implementation::detect_encodings(const char *input,
   }
   return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf8(const char *buf, size_t len) const noexcept {
   return lsx::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused result implementation::validate_utf8_with_errors(
     const char *buf, size_t len) const noexcept {
   return lsx::utf8_validation::generic_validate_utf8_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused bool
 implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return lsx::utf8_validation::generic_validate_ascii(buf, len);
+  return lsx::ascii_validation::generic_validate_ascii(buf, len);
 }
 
 simdutf_warn_unused result implementation::validate_ascii_with_errors(
     const char *buf, size_t len) const noexcept {
-  return lsx::utf8_validation::generic_validate_ascii_with_errors(buf, len);
+  return lsx::ascii_validation::generic_validate_ascii_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf16le(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -49676,15 +59090,23 @@ implementation::validate_utf16le(const char16_t *buf,
     // empty input is valid. protected the implementation from nullptr.
     return true;
   }
-  const char16_t *tail = lsx_validate_utf16<endianness::LITTLE>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::LITTLE>(tail,
-                                                       len - (tail - buf));
-  } else {
+  const auto res =
+      lsx::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count != len) {
+    return scalar::utf16::validate<endianness::LITTLE>(buf + res.count,
+                                                       len - res.count);
+  }
+
+  return true;
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused bool
 implementation::validate_utf16be(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -49692,12 +59114,19 @@ implementation::validate_utf16be(const char16_t *buf,
     // empty input is valid. protected the implementation from nullptr.
     return true;
   }
-  const char16_t *tail = lsx_validate_utf16<endianness::BIG>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
-  } else {
+  const auto res =
+      lsx::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count != len) {
+    return scalar::utf16::validate<endianness::BIG>(buf + res.count,
+                                                    len - res.count);
+  }
+
+  return true;
 }
 
 simdutf_warn_unused result implementation::validate_utf16le_with_errors(
@@ -49705,10 +59134,12 @@ simdutf_warn_unused result implementation::validate_utf16le_with_errors(
   if (simdutf_unlikely(len == 0)) {
     return result(error_code::SUCCESS, 0);
   }
-  result res = lsx_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  const result res =
+      lsx::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
   if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
-        buf + res.count, len - res.count);
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::LITTLE>(
+            buf + res.count, len - res.count);
     return result(scalar_res.error, res.count + scalar_res.count);
   } else {
     return res;
@@ -49720,16 +59151,20 @@ simdutf_warn_unused result implementation::validate_utf16be_with_errors(
   if (simdutf_unlikely(len == 0)) {
     return result(error_code::SUCCESS, 0);
   }
-  result res = lsx_validate_utf16_with_errors<endianness::BIG>(buf, len);
+  const result res =
+      lsx::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
   if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
-        buf + res.count, len - res.count);
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count,
+                                                             len - res.count);
     return result(scalar_res.error, res.count + scalar_res.count);
   } else {
     return res;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -49743,7 +59178,9 @@ implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
     return false;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::validate_utf32_with_errors(
     const char32_t *buf, size_t len) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -49758,7 +59195,9 @@ simdutf_warn_unused result implementation::validate_utf32_with_errors(
     return res;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
     const char *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char *, char *> ret =
@@ -49772,7 +59211,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char *, char16_t *> ret =
@@ -49800,7 +59241,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char *, char32_t *> ret =
@@ -49813,7 +59256,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
     const char *buf, size_t len, char *latin1_output) const noexcept {
   utf8_to_latin1::validating_transcoder converter;
@@ -49830,7 +59275,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
     const char *buf, size_t len, char *latin1_output) const noexcept {
   return lsx::utf8_to_latin1::convert_valid(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   utf8_to_utf16::validating_transcoder converter;
@@ -49867,7 +59314,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
   return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
                                                        utf16_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   utf8_to_utf32::validating_transcoder converter;
@@ -49884,7 +59333,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
     const char *input, size_t size, char32_t *utf32_output) const noexcept {
   return utf8_to_utf32::convert_valid(input, size, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
     const char16_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -49992,7 +59443,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
   // optimization opportunity: implement a custom function.
   return convert_utf16le_to_latin1(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -50098,7 +59551,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   return convert_utf16be_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -50145,7 +59600,9 @@ simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
       utf8_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char16_t *, char32_t *> ret =
@@ -50241,7 +59698,9 @@ simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
       utf32_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
     const char32_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char32_t *, char *> ret =
@@ -50302,13 +59761,17 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
   }
   return saved_bytes;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   // optimization opportunity: implement a custom function.
   return convert_utf32_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char32_t *, char16_t *> ret =
@@ -50417,7 +59880,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   return convert_utf16be_to_utf32(buf, len, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 void implementation::change_endianness_utf16(const char16_t *input,
                                              size_t length,
                                              char16_t *output) const noexcept {
@@ -50433,33 +59898,29 @@ simdutf_warn_unused size_t implementation::count_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::count_code_points<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused size_t
 implementation::count_utf8(const char *input, size_t length) const noexcept {
   return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
     const char *buf, size_t len) const noexcept {
   return count_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return length;
-}
-
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return length;
-}
-
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
     const char *input, size_t length) const noexcept {
   const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
   const uint8_t *data_end = data + length;
   uint64_t result = 0;
-  while (data + 16 < data_end) {
+  while (data_end - data > 16) {
     uint64_t two_bytes = 0;
     __m128i input_vec = __lsx_vld(data, 0);
     two_bytes =
@@ -50470,7 +59931,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
   return result + scalar::latin1::utf8_length_from_latin1((const char *)data,
                                                           data_end - data);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
@@ -50480,17 +59943,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return length;
-}
-
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return length;
-}
-
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
@@ -50500,45 +59955,26 @@ simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::utf16_length_from_utf8(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  const __m128i v_80 = __lsx_vrepli_w(0x80); /*0x00000080*/
-  const __m128i v_800 = __lsx_vldi(-3832);   /*0x00000800*/
-  const __m128i v_10000 = __lsx_vldi(-3583); /*0x00010000*/
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 4 <= length; pos += 4) {
-    __m128i in = __lsx_vld(reinterpret_cast<const uint32_t *>(input + pos), 0);
-    const __m128i ascii_bytes_bytemask = __lsx_vslt_w(in, v_80);
-    const __m128i one_two_bytes_bytemask = __lsx_vslt_w(in, v_800);
-    const __m128i two_bytes_bytemask =
-        __lsx_vxor_v(one_two_bytes_bytemask, ascii_bytes_bytemask);
-    const __m128i three_bytes_bytemask =
-        __lsx_vxor_v(__lsx_vslt_w(in, v_10000), one_two_bytes_bytemask);
-
-    const uint32_t ascii_bytes_count = __lsx_vpickve2gr_bu(
-        __lsx_vpcnt_b(__lsx_vmskltz_w(ascii_bytes_bytemask)), 0);
-    const uint32_t two_bytes_count = __lsx_vpickve2gr_bu(
-        __lsx_vpcnt_b(__lsx_vmskltz_w(two_bytes_bytemask)), 0);
-    const uint32_t three_bytes_count = __lsx_vpickve2gr_bu(
-        __lsx_vpcnt_b(__lsx_vmskltz_w(three_bytes_bytemask)), 0);
-
-    count +=
-        16 - 3 * ascii_bytes_count - 2 * two_bytes_count - three_bytes_count;
-  }
-  return count +
-         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+  return utf32::utf8_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  const __m128i v_ffff = __lsx_vldi(-2304); /*0x0000ffff*/
+  const __m128i v_ffff = lsx_splat_u32(0x0000ffff);
   size_t pos = 0;
   size_t count = 0;
   for (; pos + 4 <= length; pos += 4) {
@@ -50551,17 +59987,16 @@ simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
   return count +
          scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
+#if SIMDUTF_FEATURE_BASE64
 simdutf_warn_unused result implementation::base64_to_binary(
     const char *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
@@ -50585,55 +60020,28 @@ simdutf_warn_unused result implementation::base64_to_binary(
 }
 
 simdutf_warn_unused full_result implementation::base64_to_binary_details(
-    const char *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
-
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
-simdutf_warn_unused result implementation::base64_to_binary(
-    const char16_t *input, size_t length, char *output, base64_options options,
-    last_chunk_handling_options last_chunk_options) const noexcept {
-  if (options & base64_url) {
-    if (options == base64_options::base64_url_accept_garbage) {
-      return compress_decode_base64<true, true>(output, input, length, options,
-                                                last_chunk_options);
-    } else {
-      return compress_decode_base64<true, false>(output, input, length, options,
-                                                 last_chunk_options);
-    }
-  } else {
-    if (options == base64_options::base64_default_accept_garbage) {
-      return compress_decode_base64<false, true>(output, input, length, options,
-                                                 last_chunk_options);
-    } else {
-      return compress_decode_base64<false, false>(output, input, length,
-                                                  options, last_chunk_options);
-    }
-  }
-}
-
-simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
+}
+
+simdutf_warn_unused result implementation::base64_to_binary(
     const char16_t *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
   if (options & base64_url) {
@@ -50655,9 +60063,26 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
   }
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
+simdutf_warn_unused full_result implementation::base64_to_binary_details(
+    const char16_t *input, size_t length, char *output, base64_options options,
+    last_chunk_handling_options last_chunk_options) const noexcept {
+  if (options & base64_url) {
+    if (options == base64_options::base64_url_accept_garbage) {
+      return compress_decode_base64<true, true>(output, input, length, options,
+                                                last_chunk_options);
+    } else {
+      return compress_decode_base64<true, false>(output, input, length, options,
+                                                 last_chunk_options);
+    }
+  } else {
+    if (options == base64_options::base64_default_accept_garbage) {
+      return compress_decode_base64<false, true>(output, input, length, options,
+                                                 last_chunk_options);
+    } else {
+      return compress_decode_base64<false, false>(output, input, length,
+                                                  options, last_chunk_options);
+    }
+  }
 }
 
 size_t implementation::binary_to_base64(const char *input, size_t length,
@@ -50669,10 +60094,12 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
     return encode_base64<false>(output, input, length, options);
   }
 }
+#endif // SIMDUTF_FEATURE_BASE64
 } // namespace lsx
 } // namespace simdutf
 
 /* begin file src/simdutf/lsx/end.h */
+#undef SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 /* end file src/simdutf/lsx/end.h */
 /* end file src/lsx/implementation.cpp */
 #endif
@@ -50681,6 +60108,7 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
 /* begin file src/simdutf/lasx/begin.h */
 // redefining SIMDUTF_IMPLEMENTATION to "lasx"
 // #define SIMDUTF_IMPLEMENTATION lasx
+#define SIMDUTF_SIMD_HAS_UNSIGNED_CMP 1
 /* end file src/simdutf/lasx/begin.h */
 namespace simdutf {
 namespace lasx {
@@ -50690,6 +60118,7 @@ namespace {
 #endif
 using namespace simd;
 
+#if SIMDUTF_FEATURE_UTF8
 // convert vmskltz/vmskgez/vmsknz to
 // simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes index
 const uint8_t lasx_1_2_utf8_bytes_mask[] = {
@@ -50711,33 +60140,26 @@ const uint8_t lasx_1_2_utf8_bytes_mask[] = {
     169, 172, 173, 184, 185, 188, 189, 232, 233, 236, 237, 248, 249, 252, 253,
     170, 171, 174, 175, 186, 187, 190, 191, 234, 235, 238, 239, 250, 251, 254,
     255};
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 simdutf_really_inline __m128i lsx_swap_bytes(__m128i vec) {
   return __lsx_vshuf4i_b(vec, 0b10110001);
 }
 simdutf_really_inline __m256i lasx_swap_bytes(__m256i vec) {
   return __lasx_xvshuf4i_b(vec, 0b10110001);
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||                \
+    SIMDUTF_FEATURE_UTF8
 simdutf_really_inline bool is_ascii(const simd8x64<uint8_t> &input) {
   return input.is_ascii();
 }
+#endif // SIMDUTF_FEATURE_ASCII || SIMDUTF_FEATURE_DETECT_ENCODING ||
+       // SIMDUTF_FEATURE_UTF8
 
-simdutf_unused simdutf_really_inline simd8<bool>
-must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2,
-                     const simd8<uint8_t> prev3) {
-  simd8<bool> is_second_byte = prev1 >= uint8_t(0b11000000u);
-  simd8<bool> is_third_byte = prev2 >= uint8_t(0b11100000u);
-  simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
-  // Use ^ instead of | for is_*_byte, because ^ is commutative, and the caller
-  // is using ^ as well. This will work fine because we only have to report
-  // errors for cases with 0-1 lead bytes. Multiple lead bytes implies 2
-  // overlapping multibyte characters, and if that happens, there is guaranteed
-  // to be at least *one* lead byte that is part of only 1 other multibyte
-  // character. The error will be detected there.
-  return is_second_byte ^ is_third_byte ^ is_fourth_byte;
-}
-
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_really_inline simd8<bool>
 must_be_2_3_continuation(const simd8<uint8_t> prev2,
                          const simd8<uint8_t> prev3) {
@@ -50745,7 +60167,9 @@ must_be_2_3_continuation(const simd8<uint8_t> prev2,
   simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
   return is_third_byte ^ is_fourth_byte;
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_UTF32)
 // common functions for utf8 conversions
 simdutf_really_inline __m128i convert_utf8_3_byte_to_utf16(__m128i in) {
   // Low half contains  10bbbbbb|10cccccc
@@ -50795,7 +60219,7 @@ convert_utf8_1_to_2_byte_to_utf16(__m128i in, size_t shufutf8_idx) {
   __m128i ascii = __lsx_vand_v(perm, __lsx_vrepli_h(0x7f)); // 6 or 7 bits
   // 1 byte: 00000000 00000000
   // 2 byte: 00000aaa aa000000
-  __m128i v1f00 = __lsx_vldi(-2785); // -2785(13bit) => 151f
+  __m128i v1f00 = lsx_splat_u16(0x1f00);
   __m128i composed = __lsx_vsrli_h(__lsx_vand_v(perm, v1f00), 2); // 5 bits
   // Combine with a shift right accumulate
   // 1 byte: 00000000 0bbbbbbb
@@ -50803,212 +60227,28 @@ convert_utf8_1_to_2_byte_to_utf16(__m128i in, size_t shufutf8_idx) {
   composed = __lsx_vadd_h(ascii, composed);
   return composed;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && (SIMDUTF_FEATURE_UTF16 ||
+       // SIMDUTF_FEATURE_UTF32)
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/lasx/lasx_validate_utf16.cpp */
-/*
-    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
-
-    In a vectorized algorithm we want to examine the most significant
-    nibble in order to select a fast path. If none of highest nibbles
-    are 0xD (13), than we are sure that UTF-16 chunk in a vector
-    register is valid.
-
-    Let us analyze what we need to check if the nibble is 0xD. The
-    value of the preceding nibble determines what we have:
-
-    0xd000 .. 0xd7ff - a valid word
-    0xd800 .. 0xdbff - low surrogate
-    0xdc00 .. 0xdfff - high surrogate
-
-    Other constraints we have to consider:
-    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
-    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
-    - there must not be sole low surrogate nor high surrogate
-
-    We're going to build three bitmasks based on the 3rd nibble:
-    - V = valid word,
-    - L = low surrogate (0xd800 .. 0xdbff)
-    - H = high surrogate (0xdc00 .. 0xdfff)
-
-      0   1   2   3   4   5   6   7    <--- word index
-    [ V | L | H | L | H | V | V | L ]
-      1   0   0   0   0   1   1   0     - V = valid masks
-      0   1   0   1   0   0   0   1     - L = low surrogate
-      0   0   1   0   1   0   0   0     - H high surrogate
-
-
-      1   0   0   0   0   1   1   0   V = valid masks
-      0   1   0   1   0   0   0   0   a = L & (H >> 1)
-      0   0   1   0   1   0   0   0   b = a << 1
-      1   1   1   1   1   1   1   0   c = V | a | b
-                                  ^
-                                  the last bit can be zero, we just consume 7
-   code units and recheck this word in the next iteration
-*/
-
-/* Returns:
-   - pointer to the last unprocessed character (a scalar fallback should check
-   the rest);
-   - nullptr if an error was detected.
-*/
 template <endianness big_endian>
-const char16_t *lasx_validate_utf16(const char16_t *input, size_t size) {
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-
-    const auto in = simd8<uint8_t>(__lasx_xvpermi_d(
-        __lasx_xvssrlni_bu_h(in1.value, in0.value, 8), 0b11011000));
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-    if (surrogates_bitmask == 0x0) {
-      input += simd16<uint16_t>::ELEMENTS * 2;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint32_t V = ~surrogates_bitmask;
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint32_t H = vH.to_bitmask();
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint32_t L = ~H & surrogates_bitmask;
-
-      const uint32_t a =
-          L & (H >> 1); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint32_t b =
-          a << 1; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint32_t c = V | a | b; // Combine all the masks into the final one.
-
-      if (c == 0xffffffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += simd16<uint16_t>::ELEMENTS * 2;
-      } else if (c == 0x7fffffff) {
-        // The 31 lower code units of the input register contains valid UTF-16.
-        // The 31 word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += simd16<uint16_t>::ELEMENTS * 2 - 1;
-      } else {
-        return nullptr;
-      }
-    }
-  }
-
-  return input;
-}
-
-template <endianness big_endian>
-const result lasx_validate_utf16_with_errors(const char16_t *input,
-                                             size_t size) {
-  if (simdutf_unlikely(size == 0)) {
-    return result(error_code::SUCCESS, 0);
-  }
-  const char16_t *start = input;
-  const char16_t *end = input + size;
-
-  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
-  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
-  const auto v_fc = simd8<uint8_t>::splat(0xfc);
-  const auto v_dc = simd8<uint8_t>::splat(0xdc);
-
-  while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
-    // 0. Load data: since the validation takes into account only higher
-    //    byte of each word, we compress the two vectors into one which
-    //    consists only the higher bytes.
-    auto in0 = simd16<uint16_t>(input);
-    auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);
-
-    if (big_endian) {
-      in0 = in0.swap_bytes();
-      in1 = in1.swap_bytes();
-    }
-    const auto in = simd8<uint8_t>(__lasx_xvpermi_d(
-        __lasx_xvssrlni_bu_h(in1.value, in0.value, 8), 0b11011000));
-
-    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
-    const auto surrogates_wordmask = (in & v_f8) == v_d8;
-    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
-    if (surrogates_bitmask == 0x0) {
-      input += simd16<uint16_t>::ELEMENTS * 2;
-    } else {
-      // 2. We have some surrogates that have to be distinguished:
-      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
-      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
-      //
-      //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)
-
-      // V - non-surrogate code units
-      //     V = not surrogates_wordmask
-      const uint32_t V = ~surrogates_bitmask;
-
-      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
-      const auto vH = (in & v_fc) == v_dc;
-      const uint32_t H = vH.to_bitmask();
-
-      // L - word mask for low surrogates
-      //     L = not H and surrogates_wordmask
-      const uint32_t L = ~H & surrogates_bitmask;
-
-      const uint32_t a =
-          L & (H >> 1); // A low surrogate must be followed by high one.
-                        // (A low surrogate placed in the 7th register's word
-                        // is an exception we handle.)
-      const uint32_t b =
-          a << 1; // Just mark that the opposite fact is hold,
-                  // thanks to that we have only two masks for valid case.
-      const uint32_t c = V | a | b; // Combine all the masks into the final one.
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> in0,
+                                       const simd16<uint16_t> in1) {
+  if (big_endian) {
+    const auto mask = simd16<uint16_t>(0x00ff);
+    const auto t0 = in0 & mask;
+    const auto t1 = in1 & mask;
 
-      if (c == 0xffffffff) {
-        // The whole input register contains valid UTF-16, i.e.,
-        // either single code units or proper surrogate pairs.
-        input += simd16<uint16_t>::ELEMENTS * 2;
-      } else if (c == 0x7fffffff) {
-        // The 31 lower code units of the input register contains valid UTF-16.
-        // The 31 word may be either a low or high surrogate. It the next
-        // iteration we 1) check if the low surrogate is followed by a high
-        // one, 2) reject sole high surrogate.
-        input += simd16<uint16_t>::ELEMENTS * 2 - 1;
-      } else {
-        return result(error_code::SURROGATE, input - start);
-      }
-    }
+    return simd16<uint16_t>::pack(t0, t1);
+  } else {
+    return simd16<uint16_t>::pack_shifted_right<8>(in0, in1);
   }
-
-  return result(error_code::SUCCESS, input - start);
 }
 /* end file src/lasx/lasx_validate_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/lasx/lasx_validate_utf32le.cpp */
-
 const char32_t *lasx_validate_utf32le(const char32_t *input, size_t size) {
   const char32_t *end = input + size;
 
@@ -51020,9 +60260,9 @@ const char32_t *lasx_validate_utf32le(const char32_t *input, size_t size) {
     }
   }
 
-  __m256i offset = __lasx_xvreplgr2vr_w(uint32_t(0xffff2000));
-  __m256i standardoffsetmax = __lasx_xvreplgr2vr_w(uint32_t(0xfffff7ff));
-  __m256i standardmax = __lasx_xvldi(-2288); /*0x10ffff*/
+  __m256i offset = lasx_splat_u32(0xffff2000);
+  __m256i standardoffsetmax = lasx_splat_u32(0xfffff7ff);
+  __m256i standardmax = lasx_splat_u32(0x10ffff);
   __m256i currentmax = __lasx_xvldi(0x0);
   __m256i currentoffsetmax = __lasx_xvldi(0x0);
 
@@ -51065,9 +60305,9 @@ const result lasx_validate_utf32le_with_errors(const char32_t *input,
     input++;
   }
 
-  __m256i offset = __lasx_xvreplgr2vr_w(uint32_t(0xffff2000));
-  __m256i standardoffsetmax = __lasx_xvreplgr2vr_w(uint32_t(0xfffff7ff));
-  __m256i standardmax = __lasx_xvldi(-2288); /*0x10ffff*/
+  __m256i offset = lasx_splat_u32(0xffff2000);
+  __m256i standardoffsetmax = lasx_splat_u32(0xfffff7ff);
+  __m256i standardmax = lasx_splat_u32(0x10ffff);
   __m256i currentmax = __lasx_xvldi(0x0);
   __m256i currentoffsetmax = __lasx_xvldi(0x0);
 
@@ -51094,7 +60334,9 @@ const result lasx_validate_utf32le_with_errors(const char32_t *input,
   return result(error_code::SUCCESS, input - start);
 }
 /* end file src/lasx/lasx_validate_utf32le.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_latin1_to_utf8.cpp */
 /*
   Returns a pair: the first unprocessed byte from buf and utf8_output
@@ -51106,11 +60348,11 @@ lasx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
                             char *utf8_out) {
   uint8_t *utf8_output = reinterpret_cast<uint8_t *>(utf8_out);
   const size_t safety_margin = 12;
-  const char *end = latin1_input + len - safety_margin;
+  const char *end = latin1_input + len;
 
   // We always write 16 bytes, of which more than the first 8 bytes
   // are valid. A safety margin of 8 is more than sufficient.
-  while (latin1_input + 16 <= end) {
+  while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(latin1_input), 0);
     uint32_t ascii_mask = __lsx_vpickve2gr_wu(__lsx_vmskgez_b(in8), 0);
     if (ascii_mask == 0xFFFF) {
@@ -51128,7 +60370,7 @@ lasx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
     // t0 = [0000|00aa|bbbb|bb00]
     __m256i t0 = __lasx_xvslli_h(in16, 2);
     // t1 = [0000|00aa|0000|0000]
-    __m256i t1 = __lasx_xvand_v(t0, __lasx_xvldi(-2785));
+    __m256i t1 = __lasx_xvand_v(t0, lasx_splat_u16(0x300));
     // t3 = [0000|00aa|00bb|bbbb]
     __m256i t2 = __lasx_xvbitsel_v(t1, in16, __lasx_xvrepli_h(0x3f));
     // t4 = [1100|00aa|10bb|bbbb]
@@ -51162,6 +60404,8 @@ lasx_convert_latin1_to_utf8(const char *latin1_input, size_t len,
   return std::make_pair(latin1_input, reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lasx/lasx_convert_latin1_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_latin1_to_utf16.cpp */
 std::pair<const char *, char16_t *>
 lasx_convert_latin1_to_utf16le(const char *buf, size_t len,
@@ -51174,7 +60418,7 @@ lasx_convert_latin1_to_utf16le(const char *buf, size_t len,
     buf++;
   }
 
-  while (buf + 32 <= end) {
+  while (end - buf >= 32) {
     __m256i in8 = __lasx_xvld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m256i inlow = __lasx_vext2xv_hu_bu(in8);
@@ -51187,7 +60431,7 @@ lasx_convert_latin1_to_utf16le(const char *buf, size_t len,
     buf += 32;
   }
 
-  if (buf + 16 <= end) {
+  if (end - buf >= 16) {
     __m128i zero = __lsx_vldi(0);
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
@@ -51212,7 +60456,7 @@ lasx_convert_latin1_to_utf16be(const char *buf, size_t len,
   }
 
   __m256i zero = __lasx_xvldi(0);
-  while (buf + 32 <= end) {
+  while (end - buf >= 32) {
     __m256i in8 = __lasx_xvld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m256i in8_shuf = __lasx_xvpermi_d(in8, 0b11011000);
@@ -51225,7 +60469,7 @@ lasx_convert_latin1_to_utf16be(const char *buf, size_t len,
     buf += 32;
   }
 
-  if (buf + 16 <= end) {
+  if (end - buf >= 16) {
     __m128i zero_128 = __lsx_vldi(0);
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
@@ -51240,6 +60484,8 @@ lasx_convert_latin1_to_utf16be(const char *buf, size_t len,
   return std::make_pair(buf, utf16_output);
 }
 /* end file src/lasx/lasx_convert_latin1_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_latin1_to_utf32.cpp */
 std::pair<const char *, char32_t *>
 lasx_convert_latin1_to_utf32(const char *buf, size_t len,
@@ -51252,7 +60498,7 @@ lasx_convert_latin1_to_utf32(const char *buf, size_t len,
     buf++;
   }
 
-  while (buf + 32 <= end) {
+  while (end - buf >= 32) {
     __m256i in8 = __lasx_xvld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m256i in32_0 = __lasx_vext2xv_wu_bu(in8);
@@ -51274,7 +60520,7 @@ lasx_convert_latin1_to_utf32(const char *buf, size_t len,
     buf += 32;
   }
 
-  if (buf + 16 <= end) {
+  if (end - buf >= 16) {
     __m128i in8 = __lsx_vld(reinterpret_cast<const uint8_t *>(buf), 0);
 
     __m128i zero = __lsx_vldi(0);
@@ -51297,7 +60543,9 @@ lasx_convert_latin1_to_utf32(const char *buf, size_t len,
   return std::make_pair(buf, utf32_output);
 }
 /* end file src/lasx/lasx_convert_latin1_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /* begin file src/lasx/lasx_convert_utf8_to_utf16.cpp */
 // Convert up to 16 bytes from utf8 to utf16 using a mask indicating the
 // end of the code points. Only the least significant 12 bits of the mask
@@ -51419,7 +60667,7 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // 1 byte: 00000000 00000000
     // 2 byte: xx0bbbbb 00000000
     // 3 byte: xxbbbbbb 00000000
-    __m128i middlebyte = __lsx_vand_v(lowperm, __lsx_vldi(-2561) /*0xFF00*/);
+    __m128i middlebyte = __lsx_vand_v(lowperm, lsx_splat_u16(0xFF00));
     // 1 byte: 00000000 0ccccccc
     // 2 byte: 0010bbbb bbcccccc
     // 3 byte: 0010bbbb bbcccccc
@@ -51439,6 +60687,15 @@ size_t convert_masked_utf8_to_utf16(const char *input,
   } else if (idx < 209) {
     // THREE (3) input code-code units
     if (input_utf8_end_of_code_point_mask == 0x888) {
+      __m128i expected_mask =
+          (__m128i)v16u8{0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0,
+                         0xf8, 0xc0, 0xc0, 0xc0, 0x0,  0x0,  0x0,  0x0};
+      __m128i expected =
+          (__m128i)v16u8{0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80,
+                         0xf0, 0x80, 0x80, 0x80, 0x0,  0x0,  0x0,  0x0};
+      __m128i check = __lsx_vseq_b(__lsx_vand_v(in, expected_mask), expected);
+      if (__lsx_bz_b(check))
+        return 12;
       // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
       // UTF-16 pairs. Generating surrogate pairs is a little tricky though, but
       // it is easier when we can assume they are all pairs. This version does
@@ -51462,10 +60719,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
       //                   = +0xDC00|0xE7C0
       __m128i magic = __lsx_vreplgr2vr_w(uint32_t(0xDC00E7C0));
       // Generate unadjusted trail surrogate minus lowest 2 bits
-      // vec(0000FF00) = __lsx_vldi(-1758)
       // xxxxxxxx xxxxxxxx|11110aaa bbbbbb00
-      __m128i trail =
-          __lsx_vbitsel_v(shift, swap, __lsx_vldi(-1758 /*0000FF00*/));
+      __m128i trail = __lsx_vbitsel_v(shift, swap, lsx_splat_u32(0x0000FF00));
       // Insert low 2 bits of trail surrogate to magic number for later
       // 11011100 00000000 11100111 110000cc
       __m128i magic_with_low_2 = __lsx_vor_v(__lsx_vsrli_w(shift, 30), magic);
@@ -51478,10 +60733,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
           __lsx_vrepli_h(0x3f /* 0x003f*/));
 
       // Blend pairs
-      // __lsx_vldi(-1741) => vec(0x0000FFFF)
       // 000000cc ccdddddd|11110aaa bbbbbb00
-      __m128i blend =
-          __lsx_vbitsel_v(lead, trail, __lsx_vldi(-1741) /* (0x0000FFFF)*4 */);
+      __m128i blend = __lsx_vbitsel_v(lead, trail, lsx_splat_u32(0x0000FFFF));
 
       // Add magic number to finish the result
       // 110111CC CCDDDDDD|110110AA BBBBBBCC
@@ -51518,13 +60771,12 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // first.
     __m128i middlehigh = __lsx_vslli_w(perm, 2);
     // 00000000 00000000 00cccccc 00000000
-    __m128i middlebyte = __lsx_vand_v(perm, __lsx_vldi(-3777) /* 0x00003F00 */);
+    __m128i middlebyte = __lsx_vand_v(perm, lsx_splat_u32(0x00003F00));
     // Start assembling the sequence. Since the 4th byte is in the same position
     // as it would be in a surrogate and there is no dependency, shift left
     // instead of right. 3 byte: 00000000 10bbbbxx xxxxxxxx xxxxxxxx 4 byte:
     // 11110aaa bbbbbbxx xxxxxxxx xxxxxxxx
-    __m128i ab =
-        __lsx_vbitsel_v(middlehigh, perm, __lsx_vldi(-1656) /*0xFF000000*/);
+    __m128i ab = __lsx_vbitsel_v(middlehigh, perm, lsx_splat_u32(0xFF000000));
     // Top 16 bits contains the high ten bits of the surrogate pair before
     // correction 3 byte: 00000000 10bbbbcc|cccc0000 00000000 4 byte: 11110aaa
     // bbbbbbcc|cccc0000 00000000 - high 10 bits correct w/o correction
@@ -51538,8 +60790,7 @@ size_t convert_masked_utf8_to_utf16(const char *input,
     // After this is for surrogates
     // Blend the low and high surrogates
     // 4 byte: 11110aaa bbbbbbcc|bbbbcccc ccdddddd
-    __m128i mixed =
-        __lsx_vbitsel_v(abc, composed, __lsx_vldi(-1741) /*0x0000FFFF*/);
+    __m128i mixed = __lsx_vbitsel_v(abc, composed, lsx_splat_u32(0x0000FFFF));
     // Clear the upper 6 bits of the low surrogate. Don't clear the upper bits
     // yet as 0x10000 was not subtracted from the codepoint yet. 4 byte:
     // 11110aaa bbbbbbcc|000000cc ccdddddd
@@ -51593,6 +60844,8 @@ size_t convert_masked_utf8_to_utf16(const char *input,
   }
 }
 /* end file src/lasx/lasx_convert_utf8_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lasx/lasx_convert_utf8_to_utf32.cpp */
 // Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
 // end of the code points. Only the least significant 12 bits of the mask
@@ -51702,14 +60955,13 @@ size_t convert_masked_utf8_to_utf32(const char *input,
     // The top bits will be corrected later in the bsl
     // 00000000 10bbbbbb 00000000
     __m128i middle =
-        __lsx_vand_v(perm, __lsx_vldi(-1758 /*0x0000FF00*/)); // 5 or 6 bits
+        __lsx_vand_v(perm, lsx_splat_u32(0x0000FF00)); // 5 or 6 bits
     // Combine low and middle with shift right accumulate
     // 00000000 00xxbbbb bbcccccc
     __m128i lowmid = __lsx_vor_v(ascii, __lsx_vsrli_w(middle, 2));
     // Insert top 4 bits from high byte with bitwise select
     // 00000000 aaaabbbb bbcccccc
-    __m128i composed =
-        __lsx_vbitsel_v(lowmid, high, __lsx_vldi(-3600 /*0x0000F000*/));
+    __m128i composed = __lsx_vbitsel_v(lowmid, high, lsx_splat_u32(0x0000F000));
     __lsx_vst(composed, utf32_output, 0);
     utf32_output += 4; // We wrote 4 32-bit characters.
     return consumed;
@@ -51738,10 +60990,10 @@ size_t convert_masked_utf8_to_utf32(const char *input,
       __m128i merge2 =
           __lsx_vbitsel_v(__lsx_vslli_w(merge1, 12), /* merge1 << 12 */
                           __lsx_vsrli_w(merge1, 16), /* merge1 >> 16 */
-                          __lsx_vldi(-2545));        /*0x00000FFF*/
+                          lsx_splat_u32(0x00000FFF));
       // Clear the garbage
       // 00000000 000aaabb bbbbcccc ccdddddd
-      __m128i composed = __lsx_vand_v(merge2, __lsx_vldi(-2273 /*0x1FFFFF*/));
+      __m128i composed = __lsx_vand_v(merge2, lsx_splat_u32(0x1FFFFF));
       // Store
       __lsx_vst(composed, utf32_output, 0);
       utf32_output += 3; // We wrote 3 32-bit characters.
@@ -51761,12 +61013,11 @@ size_t convert_masked_utf8_to_utf32(const char *input,
 
     // Ascii
     __m128i ascii = __lsx_vand_v(perm, __lsx_vrepli_w(0x7F));
-    __m128i middle = __lsx_vand_v(perm, __lsx_vldi(-3777 /*0x00003f00*/));
+    __m128i middle = __lsx_vand_v(perm, lsx_splat_u32(0x00003f00));
     // 00000000 00000000 0000cccc ccdddddd
-    __m128i cd =
-        __lsx_vbitsel_v(__lsx_vsrli_w(middle, 2), ascii, __lsx_vrepli_w(0x3f));
+    __m128i cd = __lsx_vor_v(__lsx_vsrli_w(middle, 2), ascii);
 
-    __m128i correction = __lsx_vand_v(perm, __lsx_vldi(-3520 /*0x00400000*/));
+    __m128i correction = __lsx_vand_v(perm, lsx_splat_u32(0x00400000));
     __m128i corrected = __lsx_vadd_b(perm, __lsx_vsrli_w(correction, 1));
     // Insert twice
     // 00000000 000aaabb bbbbxxxx xxxxxxxx
@@ -51776,8 +61027,7 @@ size_t convert_masked_utf8_to_utf32(const char *input,
         __lsx_vbitsel_v(corrected_srli2, corrected, __lsx_vrepli_h(0x3f));
     ab = __lsx_vsrli_w(ab, 4);
     // 00000000 000aaabb bbbbcccc ccdddddd
-    __m128i composed =
-        __lsx_vbitsel_v(ab, cd, __lsx_vldi(-2545 /*0x00000FFF*/));
+    __m128i composed = __lsx_vbitsel_v(ab, cd, lsx_splat_u32(0x00000FFF));
     // Store
     __lsx_vst(composed, utf32_output, 0);
     utf32_output += 3; // We wrote 3 32-bit characters.
@@ -51788,6 +61038,8 @@ size_t convert_masked_utf8_to_utf32(const char *input,
   }
 }
 /* end file src/lasx/lasx_convert_utf8_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_utf8_to_latin1.cpp */
 size_t convert_masked_utf8_to_latin1(const char *input,
                                      uint64_t utf8_end_of_code_point_mask,
@@ -51862,14 +61114,16 @@ size_t convert_masked_utf8_to_latin1(const char *input,
   return consumed;
 }
 /* end file src/lasx/lasx_convert_utf8_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_utf16_to_latin1.cpp */
 template <endianness big_endian>
 std::pair<const char16_t *, char *>
 lasx_convert_utf16_to_latin1(const char16_t *buf, size_t len,
                              char *latin1_output) {
   const char16_t *end = buf + len;
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 16);
     if (!match_system(big_endian)) {
@@ -51897,7 +61151,7 @@ lasx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
                                          char *latin1_output) {
   const char16_t *start = buf;
   const char16_t *end = buf + len;
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m128i in = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 0);
     __m128i in1 = __lsx_vld(reinterpret_cast<const uint16_t *>(buf), 16);
     if (!match_system(big_endian)) {
@@ -51915,9 +61169,8 @@ lasx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
     } else {
       // Let us do a scalar fallback.
       for (int k = 0; k < 16; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if (word <= 0xff) {
           *latin1_output++ = char(word);
         } else {
@@ -51931,6 +61184,8 @@ lasx_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/lasx/lasx_convert_utf16_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /* begin file src/lasx/lasx_convert_utf16_to_utf8.cpp */
 /*
     The vectorized algorithm works on single LASX register i.e., it
@@ -51998,7 +61253,7 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
   __m256i v_07ff = __lasx_xvreplgr2vr_h(uint16_t(0x7ff));
   __m256i zero = __lasx_xvldi(0);
   __m128i zero_128 = __lsx_vldi(0);
-  while (buf + 16 + safety_margin <= end) {
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lasx_swap_bytes(in);
@@ -52023,7 +61278,7 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       // t0 = [000a|aaaa|bbbb|bb00]
       __m256i t0 = __lasx_xvslli_h(in, 2);
       // t1 = [000a|aaaa|0000|0000]
-      __m256i t1 = __lasx_xvand_v(t0, __lasx_xvldi(-2785 /*0x1f00*/));
+      __m256i t1 = __lasx_xvand_v(t0, lasx_splat_u16(0x1f00));
       // t2 = [0000|0000|00bb|bbbb]
       __m256i t2 = __lasx_xvand_v(in, __lasx_xvrepli_h(0x3f));
       // t3 = [000a|aaaa|00bb|bbbb]
@@ -52061,9 +61316,8 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       buf += 16;
       continue;
     }
-    __m256i surrogates_bytemask =
-        __lasx_xvseq_h(__lasx_xvand_v(in, __lasx_xvldi(-2568 /*0xF800*/)),
-                       __lasx_xvldi(-2600 /*0xD800*/));
+    __m256i surrogates_bytemask = __lasx_xvseq_h(
+        __lasx_xvand_v(in, lasx_splat_u16(0xf800)), lasx_splat_u16(0xd800));
     // It might seem like checking for surrogates_bitmask == 0xc000 could help.
     // However, it is likely an uncommon occurrence.
     if (__lasx_xbz_v(surrogates_bytemask)) {
@@ -52103,14 +61357,14 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       __m256i v_3f7f = __lasx_xvreplgr2vr_h(uint16_t(0x3F7F));
       __m256i t1 = __lasx_xvand_v(t0, v_3f7f);
       // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      __m256i t2 = __lasx_xvor_v(t1, __lasx_xvldi(-2688));
+      __m256i t2 = __lasx_xvor_v(t1, lasx_splat_u16(0x8000));
 
       // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
       __m256i s0 = __lasx_xvsrli_h(in, 12);
       // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
       __m256i s1 = __lasx_xvslli_h(in, 2);
       // s1: [aabb|bbbb|cccc|cc00] => [00bb|bbbb|0000|0000]
-      s1 = __lasx_xvand_v(s1, __lasx_xvldi(-2753 /*0x3F00*/));
+      s1 = __lasx_xvand_v(s1, lasx_splat_u16(0x3f00));
 
       // [00bb|bbbb|0000|aaaa]
       __m256i s2 = __lasx_xvor_v(s0, s1);
@@ -52118,8 +61372,8 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
       __m256i v_c0e0 = __lasx_xvreplgr2vr_h(uint16_t(0xC0E0));
       __m256i s3 = __lasx_xvor_v(s2, v_c0e0);
       __m256i one_or_two_bytes_bytemask = __lasx_xvsle_hu(in, v_07ff);
-      __m256i m0 = __lasx_xvandn_v(one_or_two_bytes_bytemask,
-                                   __lasx_xvldi(-2752 /*0x4000*/));
+      __m256i m0 =
+          __lasx_xvandn_v(one_or_two_bytes_bytemask, lasx_splat_u16(0x4000));
       __m256i s4 = __lasx_xvxor_v(s3, m0);
 
       // 4. expand code units 16-bit => 32-bit
@@ -52195,9 +61449,8 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -52211,7 +61464,7 @@ lasx_convert_utf16_to_utf8(const char16_t *buf, size_t len, char *utf8_out) {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -52254,7 +61507,7 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
   __m256i v_07ff = __lasx_xvreplgr2vr_h(uint16_t(0x7ff));
   __m256i zero = __lasx_xvldi(0);
   __m128i zero_128 = __lsx_vldi(0);
-  while (buf + 16 + safety_margin <= end) {
+  while (end - buf >= std::ptrdiff_t(16 + safety_margin)) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lasx_swap_bytes(in);
@@ -52279,7 +61532,7 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       // t0 = [000a|aaaa|bbbb|bb00]
       __m256i t0 = __lasx_xvslli_h(in, 2);
       // t1 = [000a|aaaa|0000|0000]
-      __m256i t1 = __lasx_xvand_v(t0, __lasx_xvldi(-2785 /*0x1f00*/));
+      __m256i t1 = __lasx_xvand_v(t0, lasx_splat_u16(0x1f00));
       // t2 = [0000|0000|00bb|bbbb]
       __m256i t2 = __lasx_xvand_v(in, __lasx_xvrepli_h(0x3f));
       // t3 = [000a|aaaa|00bb|bbbb]
@@ -52317,9 +61570,8 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       buf += 16;
       continue;
     }
-    __m256i surrogates_bytemask =
-        __lasx_xvseq_h(__lasx_xvand_v(in, __lasx_xvldi(-2568 /*0xF800*/)),
-                       __lasx_xvldi(-2600 /*0xD800*/));
+    __m256i surrogates_bytemask = __lasx_xvseq_h(
+        __lasx_xvand_v(in, lasx_splat_u16(0xf800)), lasx_splat_u16(0xd800));
     // It might seem like checking for surrogates_bitmask == 0xc000 could help.
     // However, it is likely an uncommon occurrence.
     if (__lasx_xbz_v(surrogates_bytemask)) {
@@ -52359,14 +61611,14 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       __m256i v_3f7f = __lasx_xvreplgr2vr_h(uint16_t(0x3F7F));
       __m256i t1 = __lasx_xvand_v(t0, v_3f7f);
       // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-      __m256i t2 = __lasx_xvor_v(t1, __lasx_xvldi(-2688));
+      __m256i t2 = __lasx_xvor_v(t1, lasx_splat_u16(0x8000));
 
       // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
       __m256i s0 = __lasx_xvsrli_h(in, 12);
       // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
       __m256i s1 = __lasx_xvslli_h(in, 2);
       // s1: [aabb|bbbb|cccc|cc00] => [00bb|bbbb|0000|0000]
-      s1 = __lasx_xvand_v(s1, __lasx_xvldi(-2753 /*0x3F00*/));
+      s1 = __lasx_xvand_v(s1, lasx_splat_u16(0x3f00));
 
       // [00bb|bbbb|0000|aaaa]
       __m256i s2 = __lasx_xvor_v(s0, s1);
@@ -52374,8 +61626,8 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
       __m256i v_c0e0 = __lasx_xvreplgr2vr_h(uint16_t(0xC0E0));
       __m256i s3 = __lasx_xvor_v(s2, v_c0e0);
       __m256i one_or_two_bytes_bytemask = __lasx_xvsle_hu(in, v_07ff);
-      __m256i m0 = __lasx_xvandn_v(one_or_two_bytes_bytemask,
-                                   __lasx_xvldi(-2752 /*0x4000*/));
+      __m256i m0 =
+          __lasx_xvandn_v(one_or_two_bytes_bytemask, lasx_splat_u16(0x4000));
       __m256i s4 = __lasx_xvxor_v(s3, m0);
 
       // 4. expand code units 16-bit => 32-bit
@@ -52451,9 +61703,8 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xFF80) == 0) {
           *utf8_output++ = char(word);
         } else if ((word & 0xF800) == 0) {
@@ -52467,7 +61718,7 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -52491,6 +61742,8 @@ lasx_convert_utf16_to_utf8_with_errors(const char16_t *buf, size_t len,
                         reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lasx/lasx_convert_utf16_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lasx/lasx_convert_utf16_to_utf32.cpp */
 template <endianness big_endian>
 std::pair<const char16_t *, char32_t *>
@@ -52502,7 +61755,7 @@ lasx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
   // Performance degradation when memory address is not 32-byte aligned
   while (((uint64_t)utf32_output & 0x1f) && buf < end) {
     uint16_t word =
-        !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[0]) : buf[0];
+        !match_system(big_endian) ? scalar::u16_swap_bytes(buf[0]) : buf[0];
     if ((word & 0xF800) != 0xD800) {
       *utf32_output++ = char32_t(word);
       buf++;
@@ -52513,9 +61766,8 @@ lasx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
       }
       // must be a surrogate pair
       uint16_t diff = uint16_t(word - 0xD800);
-      uint16_t next_word = !match_system(big_endian)
-                               ? scalar::utf16::swap_bytes(buf[1])
-                               : buf[1];
+      uint16_t next_word =
+          !match_system(big_endian) ? scalar::u16_swap_bytes(buf[1]) : buf[1];
       uint16_t diff2 = uint16_t(next_word - 0xDC00);
       if ((diff | diff2) > 0x3FF) {
         return std::make_pair(nullptr,
@@ -52527,10 +61779,10 @@ lasx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
     }
   }
 
-  __m256i v_f800 = __lasx_xvldi(-2568); /*0xF800*/
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
+  __m256i v_f800 = lasx_splat_u16(0xf800);
+  __m256i v_d800 = lasx_splat_u16(0xd800);
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lasx_swap_bytes(in);
@@ -52559,16 +61811,15 @@ lasx_convert_utf16_to_utf32(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -52604,16 +61855,15 @@ lasx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
   // Performance degradation when memory address is not 32-byte aligned
   while (((uint64_t)utf32_output & 0x1f) && buf < end) {
     uint16_t word =
-        !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[0]) : buf[0];
+        !match_system(big_endian) ? scalar::u16_swap_bytes(buf[0]) : buf[0];
     if ((word & 0xF800) != 0xD800) {
       *utf32_output++ = char32_t(word);
       buf++;
     } else if (buf + 1 < end) {
       // must be a surrogate pair
       uint16_t diff = uint16_t(word - 0xD800);
-      uint16_t next_word = !match_system(big_endian)
-                               ? scalar::utf16::swap_bytes(buf[1])
-                               : buf[1];
+      uint16_t next_word =
+          !match_system(big_endian) ? scalar::u16_swap_bytes(buf[1]) : buf[1];
       uint16_t diff2 = uint16_t(next_word - 0xDC00);
       if ((diff | diff2) > 0x3FF) {
         return std::make_pair(result(error_code::SURROGATE, buf - start),
@@ -52628,9 +61878,9 @@ lasx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
     }
   }
 
-  __m256i v_f800 = __lasx_xvldi(-2568); /*0xF800*/
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
-  while (buf + 16 <= end) {
+  __m256i v_f800 = lasx_splat_u16(0xf800);
+  __m256i v_d800 = lasx_splat_u16(0xd800);
+  while (end - buf >= 16) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint16_t *>(buf), 0);
     if (!match_system(big_endian)) {
       in = lasx_swap_bytes(in);
@@ -52659,16 +61909,15 @@ lasx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
         forward = size_t(end - buf - 1);
       }
       for (; k < forward; k++) {
-        uint16_t word = !match_system(big_endian)
-                            ? scalar::utf16::swap_bytes(buf[k])
-                            : buf[k];
+        uint16_t word =
+            !match_system(big_endian) ? scalar::u16_swap_bytes(buf[k]) : buf[k];
         if ((word & 0xF800) != 0xD800) {
           *utf32_output++ = char32_t(word);
         } else {
           // must be a surrogate pair
           uint16_t diff = uint16_t(word - 0xD800);
           uint16_t next_word = !match_system(big_endian)
-                                   ? scalar::utf16::swap_bytes(buf[k + 1])
+                                   ? scalar::u16_swap_bytes(buf[k + 1])
                                    : buf[k + 1];
           k++;
           uint16_t diff2 = uint16_t(next_word - 0xDC00);
@@ -52688,7 +61937,9 @@ lasx_convert_utf16_to_utf32_with_errors(const char16_t *buf, size_t len,
                         reinterpret_cast<char32_t *>(utf32_output));
 }
 /* end file src/lasx/lasx_convert_utf16_to_utf32.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /* begin file src/lasx/lasx_convert_utf32_to_latin1.cpp */
 std::pair<const char32_t *, char *>
 lasx_convert_utf32_to_latin1(const char32_t *buf, size_t len,
@@ -52698,7 +61949,7 @@ lasx_convert_utf32_to_latin1(const char32_t *buf, size_t len,
       (__m128i)v16u8{0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0});
   __m256i v_ff = __lasx_xvrepli_w(0xFF);
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m256i in1 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i in2 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -52731,7 +61982,7 @@ lasx_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
       (__m128i)v16u8{0, 4, 8, 12, 16, 20, 24, 28, 0, 0, 0, 0, 0, 0, 0, 0});
   __m256i v_ff = __lasx_xvrepli_w(0xFF);
 
-  while (buf + 16 <= end) {
+  while (end - buf >= 16) {
     __m256i in1 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i in2 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -52764,6 +62015,8 @@ lasx_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
                         latin1_output);
 }
 /* end file src/lasx/lasx_convert_utf32_to_latin1.cpp */
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lasx/lasx_convert_utf32_to_utf8.cpp */
 std::pair<const char32_t *, char *>
 lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
@@ -52797,10 +62050,10 @@ lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
     buf++;
   }
 
-  __m256i v_c080 = __lasx_xvreplgr2vr_h(uint16_t(0xC080));
-  __m256i v_07ff = __lasx_xvreplgr2vr_h(uint16_t(0x7FF));
-  __m256i v_dfff = __lasx_xvreplgr2vr_h(uint16_t(0xDFFF));
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
+  __m256i v_c080 = lasx_splat_u16(0xc080);
+  __m256i v_07ff = lasx_splat_u16(0x07ff);
+  __m256i v_dfff = lasx_splat_u16(0xdfff);
+  __m256i v_d800 = lasx_splat_u16(0xd800);
   __m256i zero = __lasx_xvldi(0);
   __m128i zero_128 = __lsx_vldi(0);
   __m256i forbidden_bytemask = __lasx_xvldi(0x0);
@@ -52809,7 +62062,7 @@ lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
       12; // to avoid overruns, see issue
           // https://github.com/simdutf/simdutf/issues/92
 
-  while (buf + 16 + safety_margin < end) {
+  while (end - buf > std::ptrdiff_t(16 + safety_margin)) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i nextin = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -52842,7 +62095,7 @@ lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
         // t0 = [000a|aaaa|bbbb|bb00]
         const __m256i t0 = __lasx_xvslli_h(utf16_packed, 2);
         // t1 = [000a|aaaa|0000|0000]
-        const __m256i t1 = __lasx_xvand_v(t0, __lasx_xvldi(-2785 /*0x1f00*/));
+        const __m256i t1 = __lasx_xvand_v(t0, lasx_splat_u16(0x1f00));
         // t2 = [0000|0000|00bb|bbbb]
         const __m256i t2 = __lasx_xvand_v(utf16_packed, __lasx_xvrepli_h(0x3f));
         // t3 = [000a|aaaa|00bb|bbbb]
@@ -52922,14 +62175,14 @@ lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
         __m256i v_3f7f = __lasx_xvreplgr2vr_h(uint16_t(0x3F7F));
         __m256i t1 = __lasx_xvand_v(t0, v_3f7f);
         // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        __m256i t2 = __lasx_xvor_v(t1, __lasx_xvldi(-2688 /*0x8000*/));
+        __m256i t2 = __lasx_xvor_v(t1, lasx_splat_u16(0x8000));
 
         // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
         __m256i s0 = __lasx_xvsrli_h(utf16_packed, 12);
         // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
         __m256i s1 = __lasx_xvslli_h(utf16_packed, 2);
         // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        s1 = __lasx_xvand_v(s1, __lasx_xvldi(-2753 /*0x3F00*/));
+        s1 = __lasx_xvand_v(s1, lasx_splat_u16(0x3f00));
         // [00bb|bbbb|0000|aaaa]
         __m256i s2 = __lasx_xvor_v(s0, s1);
         // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
@@ -52938,8 +62191,8 @@ lasx_convert_utf32_to_utf8(const char32_t *buf, size_t len, char *utf8_out) {
         // __m256i v_07ff = vmovq_n_u16((uint16_t)0x07FF);
         __m256i one_or_two_bytes_bytemask =
             __lasx_xvsle_hu(utf16_packed, v_07ff);
-        __m256i m0 = __lasx_xvandn_v(one_or_two_bytes_bytemask,
-                                     __lasx_xvldi(-2752 /*0x4000*/));
+        __m256i m0 =
+            __lasx_xvandn_v(one_or_two_bytes_bytemask, lasx_splat_u16(0x4000));
         __m256i s4 = __lasx_xvxor_v(s3, m0);
 
         // 4. expand code units 16-bit => 32-bit
@@ -53093,10 +62346,10 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
     buf++;
   }
 
-  __m256i v_c080 = __lasx_xvreplgr2vr_h(uint16_t(0xC080));
-  __m256i v_07ff = __lasx_xvreplgr2vr_h(uint16_t(0x7FF));
-  __m256i v_dfff = __lasx_xvreplgr2vr_h(uint16_t(0xDFFF));
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
+  __m256i v_c080 = lasx_splat_u16(0xc080);
+  __m256i v_07ff = lasx_splat_u16(0x07ff);
+  __m256i v_dfff = lasx_splat_u16(0xdfff);
+  __m256i v_d800 = lasx_splat_u16(0xd800);
   __m256i zero = __lasx_xvldi(0);
   __m128i zero_128 = __lsx_vldi(0);
   __m256i forbidden_bytemask = __lasx_xvldi(0x0);
@@ -53104,7 +62357,7 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
       12; // to avoid overruns, see issue
           // https://github.com/simdutf/simdutf/issues/92
 
-  while (buf + 16 + safety_margin < end) {
+  while (end - buf > std::ptrdiff_t(16 + safety_margin)) {
     __m256i in = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i nextin = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -53137,7 +62390,7 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         // t0 = [000a|aaaa|bbbb|bb00]
         const __m256i t0 = __lasx_xvslli_h(utf16_packed, 2);
         // t1 = [000a|aaaa|0000|0000]
-        const __m256i t1 = __lasx_xvand_v(t0, __lasx_xvldi(-2785 /*0x1f00*/));
+        const __m256i t1 = __lasx_xvand_v(t0, lasx_splat_u16(0x1f00));
         // t2 = [0000|0000|00bb|bbbb]
         const __m256i t2 = __lasx_xvand_v(utf16_packed, __lasx_xvrepli_h(0x3f));
         // t3 = [000a|aaaa|00bb|bbbb]
@@ -53221,14 +62474,14 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         __m256i v_3f7f = __lasx_xvreplgr2vr_h(uint16_t(0x3F7F));
         __m256i t1 = __lasx_xvand_v(t0, v_3f7f);
         // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
-        __m256i t2 = __lasx_xvor_v(t1, __lasx_xvldi(-2688 /*0x8000*/));
+        __m256i t2 = __lasx_xvor_v(t1, lasx_splat_u16(0x8000));
 
         // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
         __m256i s0 = __lasx_xvsrli_h(utf16_packed, 12);
         // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
         __m256i s1 = __lasx_xvslli_h(utf16_packed, 2);
         // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
-        s1 = __lasx_xvand_v(s1, __lasx_xvldi(-2753 /*0x3F00*/));
+        s1 = __lasx_xvand_v(s1, lasx_splat_u16(0x3F00));
         // [00bb|bbbb|0000|aaaa]
         __m256i s2 = __lasx_xvor_v(s0, s1);
         // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
@@ -53237,8 +62490,8 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
         // __m256i v_07ff = vmovq_n_u16((uint16_t)0x07FF);
         __m256i one_or_two_bytes_bytemask =
             __lasx_xvsle_hu(utf16_packed, v_07ff);
-        __m256i m0 = __lasx_xvandn_v(one_or_two_bytes_bytemask,
-                                     __lasx_xvldi(-2752 /*0x4000*/));
+        __m256i m0 =
+            __lasx_xvandn_v(one_or_two_bytes_bytemask, lasx_splat_u16(0x4000));
         __m256i s4 = __lasx_xvxor_v(s3, m0);
 
         // 4. expand code units 16-bit => 32-bit
@@ -53355,6 +62608,8 @@ lasx_convert_utf32_to_utf8_with_errors(const char32_t *buf, size_t len,
                         reinterpret_cast<char *>(utf8_output));
 }
 /* end file src/lasx/lasx_convert_utf32_to_utf8.cpp */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /* begin file src/lasx/lasx_convert_utf32_to_utf16.cpp */
 template <endianness big_endian>
 std::pair<const char32_t *, char16_t *>
@@ -53396,9 +62651,9 @@ lasx_convert_utf32_to_utf16(const char32_t *buf, size_t len,
   }
 
   __m256i forbidden_bytemask = __lasx_xvrepli_h(0);
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
-  __m256i v_dfff = __lasx_xvreplgr2vr_h(uint16_t(0xdfff));
-  while (buf + 16 <= end) {
+  __m256i v_d800 = lasx_splat_u16(0xd800);
+  __m256i v_dfff = lasx_splat_u16(0xdfff);
+  while (end - buf >= 16) {
     __m256i in0 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i in1 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -53503,9 +62758,9 @@ lasx_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
   }
 
   __m256i forbidden_bytemask = __lasx_xvrepli_h(0);
-  __m256i v_d800 = __lasx_xvldi(-2600); /*0xD800*/
-  __m256i v_dfff = __lasx_xvreplgr2vr_h(uint16_t(0xdfff));
-  while (buf + 16 <= end) {
+  __m256i v_d800 = lasx_splat_u16(0xd800);
+  __m256i v_dfff = lasx_splat_u16(0xdfff);
+  while (end - buf >= 16) {
     __m256i in0 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 0);
     __m256i in1 = __lasx_xvld(reinterpret_cast<const uint32_t *>(buf), 32);
 
@@ -53575,6 +62830,8 @@ lasx_convert_utf32_to_utf16_with_errors(const char32_t *buf, size_t len,
                         reinterpret_cast<char16_t *>(utf16_output));
 }
 /* end file src/lasx/lasx_convert_utf32_to_utf16.cpp */
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
+#if SIMDUTF_FEATURE_BASE64
 /* begin file src/lasx/lasx_base64.cpp */
 /**
  * References and further reading:
@@ -53943,10 +63200,9 @@ static inline void load_block(block64 *b, const char16_t *src) {
 static inline void base64_decode(char *out, __m256i str) {
   __m256i t0 = __lasx_xvor_v(
       __lasx_xvslli_w(str, 26),
-      __lasx_xvslli_w(__lasx_xvand_v(str, __lasx_xvldi(-1758 /*0x0000FF00*/)),
-                      12));
-  __m256i t1 = __lasx_xvsrli_w(
-      __lasx_xvand_v(str, __lasx_xvldi(-3521 /*0x003F0000*/)), 2);
+      __lasx_xvslli_w(__lasx_xvand_v(str, lasx_splat_u32(0x0000ff00)), 12));
+  __m256i t1 =
+      __lasx_xvsrli_w(__lasx_xvand_v(str, lasx_splat_u32(0x003f0000)), 2);
   __m256i t2 = __lasx_xvor_v(t0, t1);
   __m256i t3 = __lasx_xvor_v(t2, __lasx_xvsrli_w(str, 16));
   __m256i pack_shuffle = ____m256i(
@@ -54121,7 +63377,7 @@ compress_decode_base64(char *dst, const chartype *src, size_t srclen,
                          (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
                          (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
                         << 8;
-      triple = scalar::utf32::swap_bytes(triple);
+      triple = scalar::u32_swap_bytes(triple);
       std::memcpy(dst, &triple, 4);
 
       dst += 3;
@@ -54133,7 +63389,7 @@ compress_decode_base64(char *dst, const chartype *src, size_t srclen,
                          (uint32_t(uint8_t(buffer_start[2])) << 1 * 6) +
                          (uint32_t(uint8_t(buffer_start[3])) << 0 * 6))
                         << 8;
-      triple = scalar::utf32::swap_bytes(triple);
+      triple = scalar::u32_swap_bytes(triple);
       std::memcpy(dst, &triple, 3);
 
       dst += 3;
@@ -54186,6 +63442,7 @@ compress_decode_base64(char *dst, const chartype *src, size_t srclen,
   return {SUCCESS, srclen, size_t(dst - dstinit)};
 }
 /* end file src/lasx/lasx_base64.cpp */
+#endif // SIMDUTF_FEATURE_BASE64
 
 } // namespace
 } // namespace lasx
@@ -54302,6 +63559,7 @@ simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
 } // namespace lasx
 } // namespace simdutf
 /* end file src/generic/buf_block_reader.h */
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 /* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
 namespace simdutf {
 namespace lasx {
@@ -54608,9 +63866,21 @@ result generic_validate_utf8_with_errors(const char *input, size_t length) {
       reinterpret_cast<const uint8_t *>(input), length);
 }
 
-template <class checker>
-bool generic_validate_ascii(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+} // namespace utf8_validation
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdutf
+/* end file src/generic/utf8_validation/utf8_validator.h */
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
+#if SIMDUTF_FEATURE_ASCII
+/* begin file src/generic/ascii_validation.h */
+namespace simdutf {
+namespace lasx {
+namespace {
+namespace ascii_validation {
+
+bool generic_validate_ascii(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   uint8_t blocks[64]{};
   simd::simd8x64<uint8_t> running_or(blocks);
   while (reader.has_full_block()) {
@@ -54625,14 +63895,8 @@ bool generic_validate_ascii(const uint8_t *input, size_t length) {
   return running_or.is_ascii();
 }
 
-bool generic_validate_ascii(const char *input, size_t length) {
-  return generic_validate_ascii<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-template <class checker>
-result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
-  buf_block_reader<64> reader(input, length);
+result generic_validate_ascii_with_errors(const char *input, size_t length) {
+  buf_block_reader<64> reader(reinterpret_cast<const uint8_t *>(input), length);
   size_t count{0};
   while (reader.has_full_block()) {
     simd::simd8x64<uint8_t> in(reader.full_block());
@@ -54657,20 +63921,16 @@ result generic_validate_ascii_with_errors(const uint8_t *input, size_t length) {
   }
 }
 
-result generic_validate_ascii_with_errors(const char *input, size_t length) {
-  return generic_validate_ascii_with_errors<utf8_checker>(
-      reinterpret_cast<const uint8_t *>(input), length);
-}
-
-} // namespace utf8_validation
+} // namespace ascii_validation
 } // unnamed namespace
 } // namespace lasx
 } // namespace simdutf
-/* end file src/generic/utf8_validation/utf8_validator.h */
+/* end file src/generic/ascii_validation.h */
+#endif // SIMDUTF_FEATURE_ASCII
 
-// transcoding from UTF-8 to Latin 1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+  // transcoding from UTF-8 to Latin 1
 /* begin file src/generic/utf8_to_latin1/utf8_to_latin1.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -54989,7 +64249,6 @@ struct validating_transcoder {
 } // namespace simdutf
 /* end file src/generic/utf8_to_latin1/utf8_to_latin1.h */
 /* begin file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55069,9 +64328,10 @@ simdutf_really_inline size_t convert_valid(const char *in, size_t size,
 } // namespace simdutf
   // namespace simdutf
 /* end file src/generic/utf8_to_latin1/valid_utf8_to_latin1.h */
-// transcoding from UTF-8 to UTF-16
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+  // transcoding from UTF-8 to UTF-16
 /* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55148,7 +64408,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
 /* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55482,9 +64741,10 @@ struct validating_transcoder {
 } // namespace lasx
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
-// transcoding from UTF-8 to UTF-32
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
+  // transcoding from UTF-8 to UTF-32
 /* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55529,7 +64789,6 @@ simdutf_warn_unused size_t convert_valid(const char *input, size_t size,
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
 /* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55849,11 +65108,10 @@ struct validating_transcoder {
 } // namespace lasx
 } // namespace simdutf
 /* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-
-// other functions
+#if SIMDUTF_FEATURE_UTF8
 /* begin file src/generic/utf8.h */
-
 namespace simdutf {
 namespace lasx {
 namespace {
@@ -55872,6 +65130,59 @@ simdutf_really_inline size_t count_code_points(const char *in, size_t size) {
   return count + scalar::utf8::count_code_points(in + pos, size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+simdutf_really_inline size_t count_code_points_bytemask(const char *in,
+                                                        size_t size) {
+  using vector_i8 = simd8<int8_t>;
+  using vector_u8 = simd8<uint8_t>;
+  using vector_u64 = simd64<uint64_t>;
+
+  constexpr size_t N = vector_i8::SIZE;
+  constexpr size_t max_iterations = 255 / 4;
+
+  size_t pos = 0;
+  size_t count = 0;
+
+  auto counters = vector_u64::zero();
+  auto local = vector_u8::zero();
+  size_t iterations = 0;
+  for (; pos + 4 * N <= size; pos += 4 * N) {
+    const auto input0 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 0 * N));
+    const auto input1 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 1 * N));
+    const auto input2 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 2 * N));
+    const auto input3 =
+        simd8<int8_t>::load(reinterpret_cast<const int8_t *>(in + pos + 3 * N));
+    const auto mask0 = input0 > int8_t(-65);
+    const auto mask1 = input1 > int8_t(-65);
+    const auto mask2 = input2 > int8_t(-65);
+    const auto mask3 = input3 > int8_t(-65);
+
+    local -= vector_u8(mask0);
+    local -= vector_u8(mask1);
+    local -= vector_u8(mask2);
+    local -= vector_u8(mask3);
+
+    iterations += 1;
+    if (iterations == max_iterations) {
+      counters += sum_8bytes(local);
+      local = vector_u8::zero();
+      iterations = 0;
+    }
+  }
+
+  if (iterations > 0) {
+    count += local.sum_bytes();
+  }
+
+  count += counters.sum();
+
+  return count + scalar::utf8::count_code_points(in + pos, size - pos);
+}
+#endif
+
 simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
                                                     size_t size) {
   size_t pos = 0;
@@ -55893,6 +65204,9 @@ simdutf_really_inline size_t utf16_length_from_utf8(const char *in,
 } // namespace lasx
 } // namespace simdutf
 /* end file src/generic/utf8.h */
+#endif // SIMDUTF_FEATURE_UTF8
+
+#if SIMDUTF_FEATURE_UTF16
 /* begin file src/generic/utf16.h */
 namespace simdutf {
 namespace lasx {
@@ -55942,6 +65256,87 @@ simdutf_really_inline size_t utf8_length_from_utf16(const char16_t *in,
                                                                    size - pos);
 }
 
+#ifdef SIMDUTF_SIMD_HAS_BYTEMASK
+template <endianness big_endian>
+simdutf_really_inline size_t utf8_length_from_utf16_bytemask(const char16_t *in,
+                                                             size_t size) {
+  size_t pos = 0;
+
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS;
+
+  const auto one = vector_u16::splat(1);
+
+  auto v_count = vector_u16::zero();
+
+  // each char16 yields at least one byte
+  size_t count = size / N * N;
+
+  // in a single iteration the increment is 0, 1 or 2, despite we have
+  // three additions
+  constexpr size_t max_iterations = 65535 / 2;
+  size_t iteration = max_iterations;
+
+  for (; pos < size / N * N; pos += N) {
+    auto input = vector_u16::load(reinterpret_cast<const uint16_t *>(in + pos));
+    if (!match_system(big_endian)) {
+      input = input.swap_bytes();
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    const auto is_surrogate = ((input & uint16_t(0xf800)) == uint16_t(0xd800));
+
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    const auto c0 = min(input & uint16_t(0xff80), one);
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    const auto c1 = min(input & uint16_t(0xf800), one);
+
+    /*
+        Explanation how the counting works.
+
+        In the case of a non-surrogate character we count:
+        * always 1 -- see how `count` is initialized above;
+        * c0 = 1 if the current char yields 2 or 3 bytes;
+        * c1 = 1 if the current char yields 3 bytes.
+
+        Thus, we always have correct count for the current char:
+        from 1, 2 or 3 bytes.
+
+        A trickier part is how we count surrogate pairs. Whether
+        we encounter a surrogate (low or high), we count it as
+        3 chars and then minus 1 (`is_surrogate` is -1 or 0).
+        Each surrogate char yields 2. A surrogate pair, that
+        is a low surrogate followed by a high one, yields
+        the expected 4 bytes.
+
+        It also correctly handles cases when low surrogate is
+        processed by the this loop, but high surrogate is counted
+        by the scalar procedure. The scalar procedure uses exactly
+        the described approach, thanks to that for valid UTF-16
+        strings it always count correctly.
+    */
+    v_count += c0;
+    v_count += c1;
+    v_count += vector_u16(is_surrogate);
+
+    iteration -= 1;
+    if (iteration == 0) {
+      count += v_count.sum();
+      v_count = vector_u16::zero();
+      iteration = max_iterations;
+    }
+  }
+
+  if (iteration > 0) {
+    count += v_count.sum();
+  }
+
+  return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos,
+                                                                   size - pos);
+}
+#endif // SIMDUTF_SIMD_HAS_BYTEMASK
+
 template <endianness big_endian>
 simdutf_really_inline size_t utf32_length_from_utf16(const char16_t *in,
                                                      size_t size) {
@@ -55968,6 +65363,284 @@ change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
 } // namespace lasx
 } // namespace simdutf
 /* end file src/generic/utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16
+
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+/* begin file src/generic/validate_utf16.h */
+namespace simdutf {
+namespace lasx {
+namespace {
+namespace utf16 {
+/*
+    UTF-16 validation
+    --------------------------------------------------
+
+    In UTF-16 code units in range 0xD800 to 0xDFFF have special meaning.
+
+    In a vectorized algorithm we want to examine the most significant
+    nibble in order to select a fast path. If none of highest nibbles
+    are 0xD (13), than we are sure that UTF-16 chunk in a vector
+    register is valid.
+
+    Let us analyze what we need to check if the nibble is 0xD. The
+    value of the preceding nibble determines what we have:
+
+    0xd000 .. 0xd7ff - a valid word
+    0xd800 .. 0xdbff - low surrogate
+    0xdc00 .. 0xdfff - high surrogate
+
+    Other constraints we have to consider:
+    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
+    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
+    - there must not be sole low surrogate nor high surrogate
+
+    We are going to build three bitmasks based on the 3rd nibble:
+    - V = valid word,
+    - L = low surrogate (0xd800 .. 0xdbff)
+    - H = high surrogate (0xdc00 .. 0xdfff)
+
+      0   1   2   3   4   5   6   7    <--- word index
+    [ V | L | H | L | H | V | V | L ]
+      1   0   0   0   0   1   1   0     - V = valid masks
+      0   1   0   1   0   0   0   1     - L = low surrogate
+      0   0   1   0   1   0   0   0     - H high surrogate
+
+
+      1   0   0   0   0   1   1   0   V = valid masks
+      0   1   0   1   0   0   0   0   a = L & (H >> 1)
+      0   0   1   0   1   0   0   0   b = a << 1
+      1   1   1   1   1   1   1   0   c = V | a | b
+                                  ^
+                                  the last bit can be zero, we just consume 7
+   code units and recheck this word in the next iteration
+*/
+template <endianness big_endian>
+const result validate_utf16_with_errors(const char16_t *input, size_t size) {
+  if (simdutf_unlikely(size == 0)) {
+    return result(error_code::SUCCESS, 0);
+  }
+
+  const char16_t *start = input;
+  const char16_t *end = input + size;
+
+  const auto v_d8 = simd8<uint8_t>::splat(0xd8);
+  const auto v_f8 = simd8<uint8_t>::splat(0xf8);
+  const auto v_fc = simd8<uint8_t>::splat(0xfc);
+  const auto v_dc = simd8<uint8_t>::splat(0xdc);
+
+  while (input + simd16<uint16_t>::SIZE * 2 < end) {
+    // 0. Load data: since the validation takes into account only higher
+    //    byte of each word, we compress the two vectors into one which
+    //    consists only the higher bytes.
+    auto in0 = simd16<uint16_t>(input);
+    auto in1 =
+        simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
+
+    // Function `utf16_gather_high_bytes` consumes two vectors of UTF-16
+    // and yields a single vector having only higher bytes of characters.
+    const auto in = utf16_gather_high_bytes<big_endian>(in0, in1);
+
+    // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
+    const auto surrogates_wordmask = (in & v_f8) == v_d8;
+    const uint16_t surrogates_bitmask =
+        static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
+    if (surrogates_bitmask == 0x0000) {
+      input += 16;
+    } else {
+      // 2. We have some surrogates that have to be distinguished:
+      //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
+      //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
+      //
+      //    Fact: high surrogate has 11th bit set (3rd bit in the higher byte)
+
+      // V - non-surrogate code units
+      //     V = not surrogates_wordmask
+      const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);
+
+      // H - word-mask for high surrogates: the six highest bits are 0b1101'11
+      const auto vH = (in & v_fc) == v_dc;
+      const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());
+
+      // L - word mask for low surrogates
+      //     L = not H and surrogates_wordmask
+      const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);
+
+      const uint16_t a = static_cast<uint16_t>(
+          L & (H >> 1)); // A low surrogate must be followed by high one.
+                         // (A low surrogate placed in the 7th register's word
+                         // is an exception we handle.)
+      const uint16_t b = static_cast<uint16_t>(
+          a << 1); // Just mark that the opinput - startite fact is hold,
+                   // thanks to that we have only two masks for valid case.
+      const uint16_t c = static_cast<uint16_t>(
+          V | a | b); // Combine all the masks into the final one.
+
+      if (c == 0xffff) {
+        // The whole input register contains valid UTF-16, i.e.,
+        // either single code units or proper surrogate pairs.
+        input += 16;
+      } else if (c == 0x7fff) {
+        // The 15 lower code units of the input register contains valid UTF-16.
+        // The 15th word may be either a low or high surrogate. It the next
+        // iteration we 1) check if the low surrogate is followed by a high
+        // one, 2) reject sole high surrogate.
+        input += 15;
+      } else {
+        return result(error_code::SURROGATE, input - start);
+      }
+    }
+  }
+
+  return result(error_code::SUCCESS, input - start);
+}
+
+} // namespace utf16
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdutf
+/* end file src/generic/validate_utf16.h */
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
+
+#if SIMDUTF_FEATURE_UTF32
+/* begin file src/generic/utf32.h */
+#include <limits>
+
+namespace simdutf {
+namespace lasx {
+namespace {
+namespace utf32 {
+
+template <typename T> T min(T a, T b) { return a <= b ? a : b; }
+
+simdutf_really_inline size_t utf8_length_from_utf32(const char32_t *input,
+                                                    size_t length) {
+  using vector_u32 = simd32<uint32_t>;
+
+  const char32_t *start = input;
+
+  // we add up to three ones in a single iteration (see the vectorized loop in
+  // section #2 below)
+  const size_t max_increment = 3;
+
+  const size_t N = vector_u32::ELEMENTS;
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+  const auto v_0000007f = vector_u32::splat(0x0000007f);
+  const auto v_000007ff = vector_u32::splat(0x000007ff);
+  const auto v_0000ffff = vector_u32::splat(0x0000ffff);
+#else
+  const auto v_ffffff80 = vector_u32::splat(0xffffff80);
+  const auto v_fffff800 = vector_u32::splat(0xfffff800);
+  const auto v_ffff0000 = vector_u32::splat(0xffff0000);
+  const auto one = vector_u32::splat(1);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+  size_t counter = 0;
+
+  // 1. vectorized loop unrolled 4 times
+  {
+    // we use vector of uint32 counters, this is why this limit is used
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / (max_increment * 4);
+    size_t blocks = length / (N * 4);
+    length -= blocks * (N * 4);
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      simd32<uint32_t> acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in0 = vector_u32(input + 0 * N);
+        const auto in1 = vector_u32(input + 1 * N);
+        const auto in2 = vector_u32(input + 2 * N);
+        const auto in3 = vector_u32(input + 3 * N);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in0 > v_0000007f);
+        acc -= as_vector_u32(in1 > v_0000007f);
+        acc -= as_vector_u32(in2 > v_0000007f);
+        acc -= as_vector_u32(in3 > v_0000007f);
+
+        acc -= as_vector_u32(in0 > v_000007ff);
+        acc -= as_vector_u32(in1 > v_000007ff);
+        acc -= as_vector_u32(in2 > v_000007ff);
+        acc -= as_vector_u32(in3 > v_000007ff);
+
+        acc -= as_vector_u32(in0 > v_0000ffff);
+        acc -= as_vector_u32(in1 > v_0000ffff);
+        acc -= as_vector_u32(in2 > v_0000ffff);
+        acc -= as_vector_u32(in3 > v_0000ffff);
+#else
+        acc += min(one, in0 & v_ffffff80);
+        acc += min(one, in1 & v_ffffff80);
+        acc += min(one, in2 & v_ffffff80);
+        acc += min(one, in3 & v_ffffff80);
+
+        acc += min(one, in0 & v_fffff800);
+        acc += min(one, in1 & v_fffff800);
+        acc += min(one, in2 & v_fffff800);
+        acc += min(one, in3 & v_fffff800);
+
+        acc += min(one, in0 & v_ffff0000);
+        acc += min(one, in1 & v_ffff0000);
+        acc += min(one, in2 & v_ffff0000);
+        acc += min(one, in3 & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += 4 * N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  // 2. vectorized loop for tail
+  {
+    const size_t max_iterations =
+        std::numeric_limits<uint32_t>::max() / max_increment;
+    size_t blocks = length / N;
+    length -= blocks * N;
+    while (blocks != 0) {
+      const size_t iterations = min(blocks, max_iterations);
+      blocks -= iterations;
+
+      auto acc = vector_u32::zero();
+      for (size_t i = 0; i < iterations; i++) {
+        const auto in = vector_u32(input);
+
+#if SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+        acc -= as_vector_u32(in > v_0000007f);
+        acc -= as_vector_u32(in > v_000007ff);
+        acc -= as_vector_u32(in > v_0000ffff);
+#else
+        acc += min(one, in & v_ffffff80);
+        acc += min(one, in & v_fffff800);
+        acc += min(one, in & v_ffff0000);
+#endif // SIMDUTF_SIMD_HAS_UNSIGNED_CMP
+
+        input += N;
+      }
+
+      counter += acc.sum();
+    }
+  }
+
+  const size_t consumed = input - start;
+  if (consumed != 0) {
+    // We don't count 0th bytes in the vectorized loops above, this
+    // is why we need to count them in the end.
+    counter += consumed;
+  }
+
+  return counter + scalar::utf32::utf8_length_from_utf32(input, length);
+}
+
+} // namespace utf32
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdutf
+/* end file src/generic/utf32.h */
+#endif // SIMDUTF_FEATURE_UTF32
 
 //
 // Implementation-specific overrides
@@ -55975,6 +65648,7 @@ change_endianness_utf16(const char16_t *in, size_t size, char16_t *output) {
 namespace simdutf {
 namespace lasx {
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused int
 implementation::detect_encodings(const char *input,
                                  size_t length) const noexcept {
@@ -56001,27 +65675,35 @@ implementation::detect_encodings(const char *input,
   }
   return out;
 }
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf8(const char *buf, size_t len) const noexcept {
   return lasx::utf8_validation::generic_validate_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused result implementation::validate_utf8_with_errors(
     const char *buf, size_t len) const noexcept {
   return lasx::utf8_validation::generic_validate_utf8_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
 simdutf_warn_unused bool
 implementation::validate_ascii(const char *buf, size_t len) const noexcept {
-  return lasx::utf8_validation::generic_validate_ascii(buf, len);
+  return lasx::ascii_validation::generic_validate_ascii(buf, len);
 }
 
 simdutf_warn_unused result implementation::validate_ascii_with_errors(
     const char *buf, size_t len) const noexcept {
-  return lasx::utf8_validation::generic_validate_ascii_with_errors(buf, len);
+  return lasx::ascii_validation::generic_validate_ascii_with_errors(buf, len);
 }
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf16le(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -56029,15 +65711,22 @@ implementation::validate_utf16le(const char16_t *buf,
     // empty input is valid. protected the implementation from nullptr.
     return true;
   }
-  const char16_t *tail = lasx_validate_utf16<endianness::LITTLE>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::LITTLE>(tail,
-                                                       len - (tail - buf));
-  } else {
+  const auto res =
+      lasx::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count != len) {
+    return scalar::utf16::validate<endianness::LITTLE>(buf + res.count,
+                                                       len - res.count);
+  }
+
+  return true;
 }
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused bool
 implementation::validate_utf16be(const char16_t *buf,
                                  size_t len) const noexcept {
@@ -56045,12 +65734,19 @@ implementation::validate_utf16be(const char16_t *buf,
     // empty input is valid. protected the implementation from nullptr.
     return true;
   }
-  const char16_t *tail = lasx_validate_utf16<endianness::BIG>(buf, len);
-  if (tail) {
-    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
-  } else {
+
+  const auto res =
+      lasx::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
+  if (res.is_err()) {
     return false;
   }
+
+  if (res.count != len) {
+    return scalar::utf16::validate<endianness::BIG>(buf + res.count,
+                                                    len - res.count);
+  }
+
+  return true;
 }
 
 simdutf_warn_unused result implementation::validate_utf16le_with_errors(
@@ -56058,10 +65754,12 @@ simdutf_warn_unused result implementation::validate_utf16le_with_errors(
   if (simdutf_unlikely(len == 0)) {
     return result(error_code::SUCCESS, 0);
   }
-  result res = lasx_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
+  const result res =
+      lasx::utf16::validate_utf16_with_errors<endianness::LITTLE>(buf, len);
   if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(
-        buf + res.count, len - res.count);
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::LITTLE>(
+            buf + res.count, len - res.count);
     return result(scalar_res.error, res.count + scalar_res.count);
   } else {
     return res;
@@ -56073,16 +65771,20 @@ simdutf_warn_unused result implementation::validate_utf16be_with_errors(
   if (simdutf_unlikely(len == 0)) {
     return result(error_code::SUCCESS, 0);
   }
-  result res = lasx_validate_utf16_with_errors<endianness::BIG>(buf, len);
+  const result res =
+      lasx::utf16::validate_utf16_with_errors<endianness::BIG>(buf, len);
   if (res.count != len) {
-    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(
-        buf + res.count, len - res.count);
+    const result scalar_res =
+        scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count,
+                                                             len - res.count);
     return result(scalar_res.error, res.count + scalar_res.count);
   } else {
     return res;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 simdutf_warn_unused bool
 implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -56096,7 +65798,9 @@ implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
     return false;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused result implementation::validate_utf32_with_errors(
     const char32_t *buf, size_t len) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -56111,7 +65815,9 @@ simdutf_warn_unused result implementation::validate_utf32_with_errors(
     return res;
   }
 }
+#endif // SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
     const char *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char *, char *> ret =
@@ -56125,7 +65831,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf8(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char *, char16_t *> ret =
@@ -56153,7 +65861,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf16be(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char *, char32_t *> ret =
@@ -56166,7 +65876,9 @@ simdutf_warn_unused size_t implementation::convert_latin1_to_utf32(
   }
   return converted_chars;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf8_to_latin1(
     const char *buf, size_t len, char *latin1_output) const noexcept {
   size_t pos = 0;
@@ -56281,7 +65993,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_latin1(
       lasx::utf8_to_latin1::convert_valid(buf + pos, len - pos, latin1_output);
   return convert_result ? convert_size + convert_result : 0;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(
     const char *buf, size_t len, char16_t *utf16_output) const noexcept {
   utf8_to_utf16::validating_transcoder converter;
@@ -56318,7 +66032,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(
   return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,
                                                        utf16_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(
     const char *buf, size_t len, char32_t *utf32_output) const noexcept {
   utf8_to_utf32::validating_transcoder converter;
@@ -56335,7 +66051,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(
     const char *input, size_t size, char32_t *utf32_output) const noexcept {
   return utf8_to_utf32::convert_valid(input, size, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf16le_to_latin1(
     const char16_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -56443,7 +66161,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_latin1(
   // optimization opportunity: implement a custom function.
   return convert_utf16le_to_latin1(buf, len, latin1_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   std::pair<const char16_t *, char *> ret =
@@ -56549,7 +66269,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(
     const char16_t *buf, size_t len, char *utf8_output) const noexcept {
   return convert_utf16be_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   if (simdutf_unlikely(len == 0)) {
@@ -56596,7 +66318,9 @@ simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(
       utf8_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   std::pair<const char16_t *, char32_t *> ret =
@@ -56692,7 +66416,9 @@ simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(
       utf32_output; // Set count to the number of 8-bit code units written
   return ret.first;
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::convert_utf32_to_latin1(
     const char32_t *buf, size_t len, char *latin1_output) const noexcept {
   std::pair<const char32_t *, char *> ret =
@@ -56753,13 +66479,17 @@ simdutf_warn_unused size_t implementation::convert_valid_utf32_to_latin1(
   }
   return saved_bytes;
 }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(
     const char32_t *buf, size_t len, char *utf8_output) const noexcept {
   // optimization opportunity: implement a custom function.
   return convert_utf32_to_utf8(buf, len, utf8_output);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(
     const char32_t *buf, size_t len, char16_t *utf16_output) const noexcept {
   std::pair<const char32_t *, char16_t *> ret =
@@ -56868,7 +66598,9 @@ simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(
     const char16_t *buf, size_t len, char32_t *utf32_output) const noexcept {
   return convert_utf16be_to_utf32(buf, len, utf32_output);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 void implementation::change_endianness_utf16(const char16_t *input,
                                              size_t length,
                                              char16_t *output) const noexcept {
@@ -56884,7 +66616,9 @@ simdutf_warn_unused size_t implementation::count_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::count_code_points<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 simdutf_warn_unused size_t
 implementation::count_utf8(const char *input, size_t length) const noexcept {
   size_t pos = 0;
@@ -56905,28 +66639,22 @@ implementation::count_utf8(const char *input, size_t length) const noexcept {
   }
   return count + scalar::utf8::count_code_points(input + pos, length - pos);
 }
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::latin1_length_from_utf8(
     const char *buf, size_t len) const noexcept {
   return count_utf8(buf, len);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf16(size_t length) const noexcept {
-  return length;
-}
-
-simdutf_warn_unused size_t
-implementation::latin1_length_from_utf32(size_t length) const noexcept {
-  return length;
-}
-
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
     const char *input, size_t length) const noexcept {
   const uint8_t *data = reinterpret_cast<const uint8_t *>(input);
   const uint8_t *data_end = data + length;
   uint64_t result = 0;
-  while (data + 16 < data_end) {
+  while (data_end - data > 16) {
     uint64_t two_bytes = 0;
     __m128i input_vec = __lsx_vld(data, 0);
     two_bytes =
@@ -56937,7 +66665,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_latin1(
   return result + scalar::latin1::utf8_length_from_latin1((const char *)data,
                                                           data_end - data);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
@@ -56947,17 +66677,9 @@ simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
-simdutf_warn_unused size_t
-implementation::utf16_length_from_latin1(size_t length) const noexcept {
-  return length;
-}
-
-simdutf_warn_unused size_t
-implementation::utf32_length_from_latin1(size_t length) const noexcept {
-  return length;
-}
-
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
@@ -56967,51 +66689,26 @@ simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(
     const char16_t *input, size_t length) const noexcept {
   return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 simdutf_warn_unused size_t implementation::utf16_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::utf16_length_from_utf8(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf8_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  __m256i v_80 = __lasx_xvrepli_w(0x80); /*0x00000080*/
-  __m256i v_800 = __lasx_xvldi(-3832);   /*0x00000800*/
-  __m256i v_10000 = __lasx_xvldi(-3583); /*0x00010000*/
-  size_t pos = 0;
-  size_t count = 0;
-  for (; pos + 8 <= length; pos += 8) {
-    __m256i in =
-        __lasx_xvld(reinterpret_cast<const uint32_t *>(input + pos), 0);
-    __m256i ascii_bytes_bytemask = __lasx_xvslt_w(in, v_80);
-    __m256i one_two_bytes_bytemask = __lasx_xvslt_w(in, v_800);
-    __m256i two_bytes_bytemask =
-        __lasx_xvxor_v(one_two_bytes_bytemask, ascii_bytes_bytemask);
-    __m256i three_bytes_bytemask =
-        __lasx_xvxor_v(__lasx_xvslt_w(in, v_10000), one_two_bytes_bytemask);
-
-    __m256i ascii_bytes =
-        __lasx_xvpcnt_w(__lasx_xvmskltz_w(ascii_bytes_bytemask));
-    const uint32_t ascii_bytes_count = __lasx_xvpickve2gr_wu(ascii_bytes, 0) +
-                                       __lasx_xvpickve2gr_wu(ascii_bytes, 4);
-    __m256i two_bytes = __lasx_xvpcnt_w(__lasx_xvmskltz_w(two_bytes_bytemask));
-    const uint32_t two_bytes_count = __lasx_xvpickve2gr_wu(two_bytes, 0) +
-                                     __lasx_xvpickve2gr_wu(two_bytes, 4);
-    __m256i three_bytes =
-        __lasx_xvpcnt_w(__lasx_xvmskltz_w(three_bytes_bytemask));
-    const uint32_t three_bytes_count = __lasx_xvpickve2gr_wu(three_bytes, 0) +
-                                       __lasx_xvpickve2gr_wu(three_bytes, 4);
-
-    count +=
-        32 - 3 * ascii_bytes_count - 2 * two_bytes_count - three_bytes_count;
-  }
-  return count +
-         scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
+  return utf32::utf8_length_from_utf32(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
     const char32_t *input, size_t length) const noexcept {
-  __m128i v_ffff = __lsx_vldi(-2304); /*0x0000ffff*/
+  __m128i v_ffff = lsx_splat_u32(0x0000ffff);
   size_t pos = 0;
   size_t count = 0;
   for (; pos + 4 <= length; pos += 4) {
@@ -57024,17 +66721,16 @@ simdutf_warn_unused size_t implementation::utf16_length_from_utf32(
   return count +
          scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
 }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 simdutf_warn_unused size_t implementation::utf32_length_from_utf8(
     const char *input, size_t length) const noexcept {
   return utf8::count_code_points(input, length);
 }
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
+#if SIMDUTF_FEATURE_BASE64
 simdutf_warn_unused result implementation::base64_to_binary(
     const char *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
@@ -57079,11 +66775,6 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
   }
 }
 
-simdutf_warn_unused size_t implementation::maximal_binary_length_from_base64(
-    const char16_t *input, size_t length) const noexcept {
-  return scalar::base64::maximal_binary_length_from_base64(input, length);
-}
-
 simdutf_warn_unused result implementation::base64_to_binary(
     const char16_t *input, size_t length, char *output, base64_options options,
     last_chunk_handling_options last_chunk_options) const noexcept {
@@ -57128,11 +66819,6 @@ simdutf_warn_unused full_result implementation::base64_to_binary_details(
   }
 }
 
-simdutf_warn_unused size_t implementation::base64_length_from_binary(
-    size_t length, base64_options options) const noexcept {
-  return scalar::base64::base64_length_from_binary(length, options);
-}
-
 size_t implementation::binary_to_base64(const char *input, size_t length,
                                         char *output,
                                         base64_options options) const noexcept {
@@ -57142,10 +66828,12 @@ size_t implementation::binary_to_base64(const char *input, size_t length,
     return encode_base64<false>(output, input, length, options);
   }
 }
+#endif // SIMDUTF_FEATURE_BASE64
 } // namespace lasx
 } // namespace simdutf
 
 /* begin file src/simdutf/lasx/end.h */
+#undef SIMDUTF_SIMD_HAS_UNSIGNED_CMP
 /* end file src/simdutf/lasx/end.h */
 /* end file src/lasx/implementation.cpp */
 #endif
diff --git a/deps/simdutf/simdutf.h b/deps/simdutf/simdutf.h
index 4bec0cf300292a..02ee47d388e129 100644
--- a/deps/simdutf/simdutf.h
+++ b/deps/simdutf/simdutf.h
@@ -1,4 +1,4 @@
-/* auto-generated on 2025-01-08 17:51:07 -0500. Do not edit! */
+/* auto-generated on 2025-04-03 18:47:20 -0400. Do not edit! */
 /* begin file include/simdutf.h */
 #ifndef SIMDUTF_H
 #define SIMDUTF_H
@@ -81,7 +81,10 @@
   #if __cpp_concepts >= 201907L && __cpp_lib_span >= 202002L &&                \
       !defined(SIMDUTF_SPAN_DISABLED)
     #define SIMDUTF_SPAN 1
-  #endif
+  #endif // __cpp_concepts >= 201907L && __cpp_lib_span >= 202002L
+  #if __cpp_lib_atomic_ref >= 201806L
+    #define SIMDUTF_ATOMIC_REF 1
+  #endif // __cpp_lib_atomic_ref
 #endif
 
 /**
@@ -160,9 +163,9 @@
 #elif defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
   #define SIMDUTF_IS_ARM64 1
 #elif defined(__PPC64__) || defined(_M_PPC64)
-// #define SIMDUTF_IS_PPC64 1
-//  The simdutf library does yet support SIMD acceleration under
-//  POWER processors. Please see https://github.com/lemire/simdutf/issues/51
+  #if defined(__VEC__) && defined(__ALTIVEC__)
+    #define SIMDUTF_IS_PPC64 1
+  #endif
 #elif defined(__s390__)
 // s390 IBM system. Big endian.
 #elif (defined(__riscv) || defined(__riscv__)) && __riscv_xlen == 64
@@ -602,6 +605,14 @@ struct result {
 
   simdutf_really_inline result(error_code err, size_t pos)
       : error{err}, count{pos} {}
+
+  simdutf_really_inline bool is_ok() const {
+    return error == error_code::SUCCESS;
+  }
+
+  simdutf_really_inline bool is_err() const {
+    return error != error_code::SUCCESS;
+  }
 };
 
 struct full_result {
@@ -641,7 +652,7 @@ SIMDUTF_DISABLE_UNDESIRED_WARNINGS
 #define SIMDUTF_SIMDUTF_VERSION_H
 
 /** The version of simdutf being used (major.minor.revision) */
-#define SIMDUTF_VERSION "6.0.3"
+#define SIMDUTF_VERSION "6.4.2"
 
 namespace simdutf {
 enum {
@@ -652,11 +663,11 @@ enum {
   /**
    * The minor version (major.MINOR.revision) of simdutf being used.
    */
-  SIMDUTF_VERSION_MINOR = 0,
+  SIMDUTF_VERSION_MINOR = 4,
   /**
    * The revision (major.minor.REVISION) of simdutf being used.
    */
-  SIMDUTF_VERSION_REVISION = 3
+  SIMDUTF_VERSION_REVISION = 2
 };
 } // namespace simdutf
 
@@ -742,6 +753,13 @@ struct simdutf_riscv_hwprobe {
   #define SIMDUTF_RISCV_HWPROBE_EXT_ZVBB (1 << 17)
 #endif // SIMDUTF_IS_RISCV64 && defined(__linux__)
 
+#if defined(__loongarch__) && defined(__linux__)
+  #include <sys/auxv.h>
+// bits/hwcap.h
+// #define HWCAP_LOONGARCH_LSX             (1 << 4)
+// #define HWCAP_LOONGARCH_LASX            (1 << 5)
+#endif
+
 namespace simdutf {
 namespace internal {
 
@@ -960,12 +978,6 @@ static inline uint32_t detect_supported_architectures() {
   return host_isa;
 }
 #elif defined(__loongarch__)
-  #if defined(__linux__)
-    #include <sys/auxv.h>
-  // bits/hwcap.h
-  // #define HWCAP_LOONGARCH_LSX             (1 << 4)
-  // #define HWCAP_LOONGARCH_LASX            (1 << 5)
-  #endif
 
 static inline uint32_t detect_supported_architectures() {
   uint32_t host_isa = instruction_set::DEFAULT;
@@ -1002,6 +1014,23 @@ static inline uint32_t detect_supported_architectures() {
   #include <span>
 #endif
 
+// The following defines are conditionally enabled/disabled during amalgamation.
+// By default all features are enabled, regular code shouldn't check them. Only
+// when user code really relies of a selected subset, it's good to verify these
+// flags, like:
+//
+//      #if !SIMDUTF_FEATURE_UTF16
+//      #   error("Please amalgamate simdutf with UTF-16 support")
+//      #endif
+//
+#define SIMDUTF_FEATURE_DETECT_ENCODING 1
+#define SIMDUTF_FEATURE_ASCII 1
+#define SIMDUTF_FEATURE_LATIN1 1
+#define SIMDUTF_FEATURE_UTF8 1
+#define SIMDUTF_FEATURE_UTF16 1
+#define SIMDUTF_FEATURE_UTF32 1
+#define SIMDUTF_FEATURE_BASE64 1
+
 namespace simdutf {
 
 #if SIMDUTF_SPAN
@@ -1051,6 +1080,7 @@ concept output_span_of_byte_like = requires(T &t) {
 } // namespace detail
 #endif
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
 /**
  * Autodetect the encoding of the input, a single encoding is recommended.
  * E.g., the function might return simdutf::encoding_type::UTF8,
@@ -1067,7 +1097,7 @@ simdutf_really_inline simdutf_warn_unused simdutf::encoding_type
 autodetect_encoding(const uint8_t *input, size_t length) noexcept {
   return autodetect_encoding(reinterpret_cast<const char *>(input), length);
 }
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 /**
  * Autodetect the encoding of the input, a single encoding is recommended.
  * E.g., the function might return simdutf::encoding_type::UTF8,
@@ -1085,7 +1115,7 @@ autodetect_encoding(
   return autodetect_encoding(reinterpret_cast<const char *>(input.data()),
                              input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Autodetect the possible encodings of the input in one pass.
@@ -1104,14 +1134,16 @@ simdutf_really_inline simdutf_warn_unused int
 detect_encodings(const uint8_t *input, size_t length) noexcept {
   return detect_encodings(reinterpret_cast<const char *>(input), length);
 }
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused int
 detect_encodings(const detail::input_span_of_byte_like auto &input) noexcept {
   return detect_encodings(reinterpret_cast<const char *>(input.data()),
                           input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 /**
  * Validate the UTF-8 string. This function may be best when you expect
  * the input to be almost always valid. Otherwise, consider using
@@ -1124,14 +1156,16 @@ detect_encodings(const detail::input_span_of_byte_like auto &input) noexcept {
  * @return true if and only if the string is valid UTF-8.
  */
 simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_utf8(const detail::input_span_of_byte_like auto &input) noexcept {
   return validate_utf8(reinterpret_cast<const char *>(input.data()),
                        input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
 /**
  * Validate the UTF-8 string and stop on error.
  *
@@ -1146,14 +1180,16 @@ validate_utf8(const detail::input_span_of_byte_like auto &input) noexcept {
  */
 simdutf_warn_unused result validate_utf8_with_errors(const char *buf,
                                                      size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result validate_utf8_with_errors(
     const detail::input_span_of_byte_like auto &input) noexcept {
   return validate_utf8_with_errors(reinterpret_cast<const char *>(input.data()),
                                    input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
 /**
  * Validate the ASCII string.
  *
@@ -1164,13 +1200,13 @@ simdutf_really_inline simdutf_warn_unused result validate_utf8_with_errors(
  * @return true if and only if the string is valid ASCII.
  */
 simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_ascii(const detail::input_span_of_byte_like auto &input) noexcept {
   return validate_ascii(reinterpret_cast<const char *>(input.data()),
                         input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Validate the ASCII string and stop on error. It might be faster than
@@ -1187,14 +1223,16 @@ validate_ascii(const detail::input_span_of_byte_like auto &input) noexcept {
  */
 simdutf_warn_unused result validate_ascii_with_errors(const char *buf,
                                                       size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result validate_ascii_with_errors(
     const detail::input_span_of_byte_like auto &input) noexcept {
   return validate_ascii_with_errors(
       reinterpret_cast<const char *>(input.data()), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness; Validate the UTF-16 string.
  * This function may be best when you expect the input to be almost always
@@ -1211,13 +1249,15 @@ simdutf_really_inline simdutf_warn_unused result validate_ascii_with_errors(
  */
 simdutf_warn_unused bool validate_utf16(const char16_t *buf,
                                         size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_utf16(std::span<const char16_t> input) noexcept {
   return validate_utf16(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 /**
  * Validate the UTF-16LE string. This function may be best when you expect
  * the input to be almost always valid. Otherwise, consider using
@@ -1234,13 +1274,15 @@ validate_utf16(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused bool validate_utf16le(const char16_t *buf,
                                           size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_utf16le(std::span<const char16_t> input) noexcept {
   return validate_utf16le(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
 /**
  * Validate the UTF-16BE string. This function may be best when you expect
  * the input to be almost always valid. Otherwise, consider using
@@ -1257,12 +1299,12 @@ validate_utf16le(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused bool validate_utf16be(const char16_t *buf,
                                           size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_utf16be(std::span<const char16_t> input) noexcept {
   return validate_utf16be(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness; Validate the UTF-16 string and stop on error.
@@ -1283,12 +1325,12 @@ validate_utf16be(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused result validate_utf16_with_errors(const char16_t *buf,
                                                       size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 validate_utf16_with_errors(std::span<const char16_t> input) noexcept {
   return validate_utf16_with_errors(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Validate the UTF-16LE string and stop on error. It might be faster than
@@ -1308,12 +1350,12 @@ validate_utf16_with_errors(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf,
                                                         size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 validate_utf16le_with_errors(std::span<const char16_t> input) noexcept {
   return validate_utf16le_with_errors(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Validate the UTF-16BE string and stop on error. It might be faster than
@@ -1333,13 +1375,15 @@ validate_utf16le_with_errors(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf,
                                                         size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 validate_utf16be_with_errors(std::span<const char16_t> input) noexcept {
   return validate_utf16be_with_errors(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 /**
  * Validate the UTF-32 string. This function may be best when you expect
  * the input to be almost always valid. Otherwise, consider using
@@ -1356,13 +1400,15 @@ validate_utf16be_with_errors(std::span<const char16_t> input) noexcept {
  */
 simdutf_warn_unused bool validate_utf32(const char32_t *buf,
                                         size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused bool
 validate_utf32(std::span<const char32_t> input) noexcept {
   return validate_utf32(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
 /**
  * Validate the UTF-32 string and stop on error. It might be faster than
  * validate_utf32 when an error is expected to occur early.
@@ -1381,13 +1427,15 @@ validate_utf32(std::span<const char32_t> input) noexcept {
  */
 simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf,
                                                       size_t len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 validate_utf32_with_errors(std::span<const char32_t> input) noexcept {
   return validate_utf32_with_errors(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert Latin1 string into UTF8 string.
  *
@@ -1401,7 +1449,7 @@ validate_utf32_with_errors(std::span<const char32_t> input) noexcept {
 simdutf_warn_unused size_t convert_latin1_to_utf8(const char *input,
                                                   size_t length,
                                                   char *utf8_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8(
     const detail::input_span_of_byte_like auto &latin1_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -1409,7 +1457,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8(
       reinterpret_cast<const char *>(latin1_input.data()), latin1_input.size(),
       utf8_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert Latin1 string into UTF8 string with output limit.
@@ -1425,7 +1473,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8(
 simdutf_warn_unused size_t
 convert_latin1_to_utf8_safe(const char *input, size_t length, char *utf8_output,
                             size_t utf8_len) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8_safe(
     const detail::input_span_of_byte_like auto &input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -1439,8 +1487,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8_safe(
       input.data(), input.size(), reinterpret_cast<char *>(utf8_output.data()),
       utf8_output.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert possibly Latin1 string into UTF-16LE string.
  *
@@ -1453,7 +1503,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf8_safe(
  */
 simdutf_warn_unused size_t convert_latin1_to_utf16le(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf16le(
     const detail::input_span_of_byte_like auto &latin1_input,
     std::span<char16_t> utf16_output) noexcept {
@@ -1461,7 +1511,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf16le(
       reinterpret_cast<const char *>(latin1_input.data()), latin1_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert Latin1 string into UTF-16BE string.
@@ -1475,15 +1525,36 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf16le(
  */
 simdutf_warn_unused size_t convert_latin1_to_utf16be(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_latin1_to_utf16be(const detail::input_span_of_byte_like auto &input,
                           std::span<char16_t> output) noexcept {
   return convert_latin1_to_utf16be(reinterpret_cast<const char *>(input.data()),
                                    input.size(), output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+/**
+ * Compute the number of bytes that this UTF-16 string would require in Latin1
+ * format.
+ *
+ * @param length        the length of the string in Latin1 code units (char)
+ * @return the length of the string in Latin1 code units (char) required to
+ * encode the UTF-16 string as Latin1
+ */
+simdutf_warn_unused size_t latin1_length_from_utf16(size_t length) noexcept;
+
+/**
+ * Compute the number of code units that this Latin1 string would require in
+ * UTF-16 format.
+ *
+ * @param length        the length of the string in Latin1 code units (char)
+ * @return the length of the string in 2-byte code units (char16_t) required to
+ * encode the Latin1 string as UTF-16
+ */
+simdutf_warn_unused size_t utf16_length_from_latin1(size_t length) noexcept;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert Latin1 string into UTF-32 string.
  *
@@ -1496,7 +1567,7 @@ convert_latin1_to_utf16be(const detail::input_span_of_byte_like auto &input,
  */
 simdutf_warn_unused size_t convert_latin1_to_utf32(
     const char *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf32(
     const detail::input_span_of_byte_like auto &latin1_input,
     std::span<char32_t> utf32_output) noexcept {
@@ -1504,8 +1575,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf32(
       reinterpret_cast<const char *>(latin1_input.data()), latin1_input.size(),
       utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert possibly broken UTF-8 string into latin1 string.
  *
@@ -1521,7 +1594,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_latin1_to_utf32(
 simdutf_warn_unused size_t convert_utf8_to_latin1(const char *input,
                                                   size_t length,
                                                   char *latin1_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf8_to_latin1(
     const detail::input_span_of_byte_like auto &input,
     detail::output_span_of_byte_like auto &&output) noexcept {
@@ -1529,8 +1602,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf8_to_latin1(
                                 input.size(),
                                 reinterpret_cast<char *>(output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness, convert possibly broken UTF-8 string into a UTF-16
  * string.
@@ -1546,15 +1621,17 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf8_to_latin1(
  */
 simdutf_warn_unused size_t convert_utf8_to_utf16(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf8_to_utf16(const detail::input_span_of_byte_like auto &input,
                       std::span<char16_t> output) noexcept {
   return convert_utf8_to_utf16(reinterpret_cast<const char *>(input.data()),
                                input.size(), output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Using native endianness, convert a Latin1 string into a UTF-16 string.
  *
@@ -1565,15 +1642,17 @@ convert_utf8_to_utf16(const detail::input_span_of_byte_like auto &input,
  */
 simdutf_warn_unused size_t convert_latin1_to_utf16(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_latin1_to_utf16(const detail::input_span_of_byte_like auto &input,
                         std::span<char16_t> output) noexcept {
   return convert_latin1_to_utf16(reinterpret_cast<const char *>(input.data()),
                                  input.size(), output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Convert possibly broken UTF-8 string into UTF-16LE string.
  *
@@ -1588,7 +1667,7 @@ convert_latin1_to_utf16(const detail::input_span_of_byte_like auto &input,
  */
 simdutf_warn_unused size_t convert_utf8_to_utf16le(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf8_to_utf16le(const detail::input_span_of_byte_like auto &utf8_input,
                         std::span<char16_t> utf16_output) noexcept {
@@ -1596,7 +1675,7 @@ convert_utf8_to_utf16le(const detail::input_span_of_byte_like auto &utf8_input,
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-8 string into UTF-16BE string.
@@ -1612,7 +1691,7 @@ convert_utf8_to_utf16le(const detail::input_span_of_byte_like auto &utf8_input,
  */
 simdutf_warn_unused size_t convert_utf8_to_utf16be(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf8_to_utf16be(const detail::input_span_of_byte_like auto &utf8_input,
                         std::span<char16_t> utf16_output) noexcept {
@@ -1620,8 +1699,10 @@ convert_utf8_to_utf16be(const detail::input_span_of_byte_like auto &utf8_input,
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert possibly broken UTF-8 string into latin1 string with errors.
  * If the string cannot be represented as Latin1, an error
@@ -1640,7 +1721,7 @@ convert_utf8_to_utf16be(const detail::input_span_of_byte_like auto &utf8_input,
  */
 simdutf_warn_unused result convert_utf8_to_latin1_with_errors(
     const char *input, size_t length, char *latin1_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf8_to_latin1_with_errors(
     const detail::input_span_of_byte_like auto &utf8_input,
@@ -1649,8 +1730,10 @@ convert_utf8_to_latin1_with_errors(
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness, convert possibly broken UTF-8 string into UTF-16
  * string and stop on error.
@@ -1668,7 +1751,7 @@ convert_utf8_to_latin1_with_errors(
  */
 simdutf_warn_unused result convert_utf8_to_utf16_with_errors(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf8_to_utf16_with_errors(
     const detail::input_span_of_byte_like auto &utf8_input,
@@ -1677,7 +1760,7 @@ convert_utf8_to_utf16_with_errors(
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-8 string into UTF-16LE string and stop on error.
@@ -1695,7 +1778,7 @@ convert_utf8_to_utf16_with_errors(
  */
 simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf8_to_utf16le_with_errors(
     const detail::input_span_of_byte_like auto &utf8_input,
@@ -1704,7 +1787,7 @@ convert_utf8_to_utf16le_with_errors(
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-8 string into UTF-16BE string and stop on error.
@@ -1722,7 +1805,7 @@ convert_utf8_to_utf16le_with_errors(
  */
 simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(
     const char *input, size_t length, char16_t *utf16_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf8_to_utf16be_with_errors(
     const detail::input_span_of_byte_like auto &utf8_input,
@@ -1731,8 +1814,10 @@ convert_utf8_to_utf16be_with_errors(
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /**
  * Convert possibly broken UTF-8 string into UTF-32 string.
  *
@@ -1747,7 +1832,7 @@ convert_utf8_to_utf16be_with_errors(
  */
 simdutf_warn_unused size_t convert_utf8_to_utf32(
     const char *input, size_t length, char32_t *utf32_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf8_to_utf32(const detail::input_span_of_byte_like auto &utf8_input,
                       std::span<char32_t> utf32_output) noexcept {
@@ -1755,7 +1840,7 @@ convert_utf8_to_utf32(const detail::input_span_of_byte_like auto &utf8_input,
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-8 string into UTF-32 string and stop on error.
@@ -1773,7 +1858,7 @@ convert_utf8_to_utf32(const detail::input_span_of_byte_like auto &utf8_input,
  */
 simdutf_warn_unused result convert_utf8_to_utf32_with_errors(
     const char *input, size_t length, char32_t *utf32_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf8_to_utf32_with_errors(
     const detail::input_span_of_byte_like auto &utf8_input,
@@ -1782,8 +1867,10 @@ convert_utf8_to_utf32_with_errors(
       reinterpret_cast<const char *>(utf8_input.data()), utf8_input.size(),
       utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert valid UTF-8 string into latin1 string.
  *
@@ -1805,7 +1892,7 @@ convert_utf8_to_utf32_with_errors(
  */
 simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
     const char *input, size_t length, char *latin1_output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
     const detail::input_span_of_byte_like auto &valid_utf8_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -1813,8 +1900,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size(), latin1_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness, convert valid UTF-8 string into a UTF-16 string.
  *
@@ -1827,7 +1916,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_latin1(
  */
 simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
     const char *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
     const detail::input_span_of_byte_like auto &valid_utf8_input,
     std::span<char16_t> utf16_output) noexcept {
@@ -1835,7 +1924,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-8 string into UTF-16LE string.
@@ -1849,7 +1938,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16(
  */
 simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
     const char *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
     const detail::input_span_of_byte_like auto &valid_utf8_input,
     std::span<char16_t> utf16_output) noexcept {
@@ -1857,7 +1946,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-8 string into UTF-16BE string.
@@ -1871,7 +1960,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(
  */
 simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
     const char *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
     const detail::input_span_of_byte_like auto &valid_utf8_input,
     std::span<char16_t> utf16_output) noexcept {
@@ -1879,8 +1968,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /**
  * Convert valid UTF-8 string into UTF-32 string.
  *
@@ -1893,7 +1984,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(
  */
 simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
     const char *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
     const detail::input_span_of_byte_like auto &valid_utf8_input,
     std::span<char32_t> utf32_output) noexcept {
@@ -1901,8 +1992,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 /**
  * Return the number of bytes that this Latin1 string would require in UTF-8
  * format.
@@ -1913,13 +2006,13 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf8_to_utf32(
  */
 simdutf_warn_unused size_t utf8_length_from_latin1(const char *input,
                                                    size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t utf8_length_from_latin1(
     const detail::input_span_of_byte_like auto &latin1_input) noexcept {
   return utf8_length_from_latin1(
       reinterpret_cast<const char *>(latin1_input.data()), latin1_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Compute the number of bytes that this UTF-8 string would require in Latin1
@@ -1936,15 +2029,17 @@ simdutf_really_inline simdutf_warn_unused size_t utf8_length_from_latin1(
  */
 simdutf_warn_unused size_t latin1_length_from_utf8(const char *input,
                                                    size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t latin1_length_from_utf8(
     const detail::input_span_of_byte_like auto &valid_utf8_input) noexcept {
   return latin1_length_from_utf8(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Compute the number of 2-byte code units that this UTF-8 string would require
  * in UTF-16LE format.
@@ -1961,15 +2056,17 @@ simdutf_really_inline simdutf_warn_unused size_t latin1_length_from_utf8(
  */
 simdutf_warn_unused size_t utf16_length_from_utf8(const char *input,
                                                   size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t utf16_length_from_utf8(
     const detail::input_span_of_byte_like auto &valid_utf8_input) noexcept {
   return utf16_length_from_utf8(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /**
  * Compute the number of 4-byte code units that this UTF-8 string would require
  * in UTF-32 format.
@@ -1988,15 +2085,17 @@ simdutf_really_inline simdutf_warn_unused size_t utf16_length_from_utf8(
  */
 simdutf_warn_unused size_t utf32_length_from_utf8(const char *input,
                                                   size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf8(
     const detail::input_span_of_byte_like auto &valid_utf8_input) noexcept {
   return utf32_length_from_utf8(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness, convert possibly broken UTF-16 string into UTF-8
  * string.
@@ -2015,15 +2114,17 @@ simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf8(
 simdutf_warn_unused size_t convert_utf16_to_utf8(const char16_t *input,
                                                  size_t length,
                                                  char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16_to_utf8(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
   return convert_utf16_to_utf8(utf16_input.data(), utf16_input.size(),
                                reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Using native endianness, convert possibly broken UTF-16 string into Latin1
  * string.
@@ -2041,7 +2142,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16_to_utf8(
  */
 simdutf_warn_unused size_t convert_utf16_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16_to_latin1(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -2049,7 +2150,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16_to_latin1(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16LE string into Latin1 string.
@@ -2069,7 +2170,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16_to_latin1(
  */
 simdutf_warn_unused size_t convert_utf16le_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16le_to_latin1(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -2077,7 +2178,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16le_to_latin1(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into Latin1 string.
@@ -2095,7 +2196,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16le_to_latin1(
  */
 simdutf_warn_unused size_t convert_utf16be_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16be_to_latin1(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -2103,8 +2204,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16be_to_latin1(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Convert possibly broken UTF-16LE string into UTF-8 string.
  *
@@ -2122,14 +2225,14 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16be_to_latin1(
 simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t *input,
                                                    size_t length,
                                                    char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16le_to_utf8(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
   return convert_utf16le_to_utf8(utf16_input.data(), utf16_input.size(),
                                  reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into UTF-8 string.
@@ -2148,15 +2251,17 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16le_to_utf8(
 simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t *input,
                                                    size_t length,
                                                    char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf16be_to_utf8(
     std::span<char16_t> utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
   return convert_utf16be_to_utf8(utf16_input.data(), utf16_input.size(),
                                  reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Using native endianness, convert possibly broken UTF-16 string into Latin1
  * string.
@@ -2175,7 +2280,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf16be_to_utf8(
  */
 simdutf_warn_unused result convert_utf16_to_latin1_with_errors(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16_to_latin1_with_errors(
     std::span<char16_t> utf16_input,
@@ -2184,7 +2289,7 @@ convert_utf16_to_latin1_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16LE string into Latin1 string.
@@ -2203,7 +2308,7 @@ convert_utf16_to_latin1_with_errors(
  */
 simdutf_warn_unused result convert_utf16le_to_latin1_with_errors(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16le_to_latin1_with_errors(
     std::span<char16_t> utf16_input,
@@ -2212,7 +2317,7 @@ convert_utf16le_to_latin1_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into Latin1 string.
@@ -2233,7 +2338,7 @@ convert_utf16le_to_latin1_with_errors(
  */
 simdutf_warn_unused result convert_utf16be_to_latin1_with_errors(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16be_to_latin1_with_errors(
     std::span<char16_t> utf16_input,
@@ -2242,8 +2347,10 @@ convert_utf16be_to_latin1_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Using native endianness, convert possibly broken UTF-16 string into UTF-8
  * string and stop on error.
@@ -2263,7 +2370,7 @@ convert_utf16be_to_latin1_with_errors(
  */
 simdutf_warn_unused result convert_utf16_to_utf8_with_errors(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16_to_utf8_with_errors(
     std::span<char16_t> utf16_input,
@@ -2272,7 +2379,7 @@ convert_utf16_to_utf8_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16LE string into UTF-8 string and stop on error.
@@ -2292,7 +2399,7 @@ convert_utf16_to_utf8_with_errors(
  */
 simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16le_to_utf8_with_errors(
     std::span<char16_t> utf16_input,
@@ -2301,7 +2408,7 @@ convert_utf16le_to_utf8_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into UTF-8 string and stop on error.
@@ -2321,7 +2428,7 @@ convert_utf16le_to_utf8_with_errors(
  */
 simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16be_to_utf8_with_errors(
     std::span<char16_t> utf16_input,
@@ -2330,7 +2437,7 @@ convert_utf16be_to_utf8_with_errors(
       utf16_input.data(), utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness, convert valid UTF-16 string into UTF-8 string.
@@ -2341,13 +2448,13 @@ convert_utf16be_to_utf8_with_errors(
  *
  * @param input         the UTF-16 string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf8_buffer   the pointer to buffer that can hold the conversion
+ * @param utf8_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
     std::span<char16_t> valid_utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -2355,8 +2462,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Using native endianness, convert UTF-16 string into Latin1 string.
  *
@@ -2378,7 +2487,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_utf8(
  */
 simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
     std::span<char16_t> valid_utf16_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -2386,7 +2495,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-16LE string into Latin1 string.
@@ -2409,7 +2518,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16_to_latin1(
  */
 simdutf_warn_unused size_t convert_valid_utf16le_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf16le_to_latin1(
     std::span<char16_t> valid_utf16_input,
@@ -2418,7 +2527,7 @@ convert_valid_utf16le_to_latin1(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-16BE string into Latin1 string.
@@ -2441,7 +2550,7 @@ convert_valid_utf16le_to_latin1(
  */
 simdutf_warn_unused size_t convert_valid_utf16be_to_latin1(
     const char16_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf16be_to_latin1(
     std::span<char16_t> valid_utf16_input,
@@ -2450,8 +2559,10 @@ convert_valid_utf16be_to_latin1(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Convert valid UTF-16LE string into UTF-8 string.
  *
@@ -2462,13 +2573,13 @@ convert_valid_utf16be_to_latin1(
  *
  * @param input         the UTF-16LE string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf8_buffer   the pointer to buffer that can hold the conversion
+ * @param utf8_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
     std::span<char16_t> valid_utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -2476,7 +2587,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-16BE string into UTF-8 string.
@@ -2487,13 +2598,13 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(
  *
  * @param input         the UTF-16BE string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf8_buffer   the pointer to buffer that can hold the conversion
+ * @param utf8_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
     const char16_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
     std::span<char16_t> valid_utf16_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -2501,8 +2612,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
       valid_utf16_input.data(), valid_utf16_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /**
  * Using native endianness, convert possibly broken UTF-16 string into UTF-32
  * string.
@@ -2520,14 +2633,14 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(
  */
 simdutf_warn_unused size_t convert_utf16_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf16_to_utf32(std::span<const char16_t> utf16_input,
                        std::span<char32_t> utf32_output) noexcept {
   return convert_utf16_to_utf32(utf16_input.data(), utf16_input.size(),
                                 utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16LE string into UTF-32 string.
@@ -2545,14 +2658,14 @@ convert_utf16_to_utf32(std::span<const char16_t> utf16_input,
  */
 simdutf_warn_unused size_t convert_utf16le_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf16le_to_utf32(std::span<const char16_t> utf16_input,
                          std::span<char32_t> utf32_output) noexcept {
   return convert_utf16le_to_utf32(utf16_input.data(), utf16_input.size(),
                                   utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into UTF-32 string.
@@ -2570,14 +2683,14 @@ convert_utf16le_to_utf32(std::span<const char16_t> utf16_input,
  */
 simdutf_warn_unused size_t convert_utf16be_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf16be_to_utf32(std::span<const char16_t> utf16_input,
                          std::span<char32_t> utf32_output) noexcept {
   return convert_utf16be_to_utf32(utf16_input.data(), utf16_input.size(),
                                   utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness, convert possibly broken UTF-16 string into
@@ -2598,14 +2711,14 @@ convert_utf16be_to_utf32(std::span<const char16_t> utf16_input,
  */
 simdutf_warn_unused result convert_utf16_to_utf32_with_errors(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16_to_utf32_with_errors(std::span<const char16_t> utf16_input,
                                    std::span<char32_t> utf32_output) noexcept {
   return convert_utf16_to_utf32_with_errors(
       utf16_input.data(), utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16LE string into UTF-32 string and stop on error.
@@ -2625,7 +2738,7 @@ convert_utf16_to_utf32_with_errors(std::span<const char16_t> utf16_input,
  */
 simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16le_to_utf32_with_errors(
     std::span<const char16_t> utf16_input,
@@ -2633,7 +2746,7 @@ convert_utf16le_to_utf32_with_errors(
   return convert_utf16le_to_utf32_with_errors(
       utf16_input.data(), utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-16BE string into UTF-32 string and stop on error.
@@ -2653,7 +2766,7 @@ convert_utf16le_to_utf32_with_errors(
  */
 simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf16be_to_utf32_with_errors(
     std::span<const char16_t> utf16_input,
@@ -2661,7 +2774,7 @@ convert_utf16be_to_utf32_with_errors(
   return convert_utf16be_to_utf32_with_errors(
       utf16_input.data(), utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness, convert valid UTF-16 string into UTF-32 string.
@@ -2673,20 +2786,20 @@ convert_utf16be_to_utf32_with_errors(
  *
  * @param input         the UTF-16 string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf32_buffer   the pointer to buffer that can hold the conversion
+ * @param utf32_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf16_to_utf32(std::span<const char16_t> valid_utf16_input,
                              std::span<char32_t> utf32_output) noexcept {
   return convert_valid_utf16_to_utf32(
       valid_utf16_input.data(), valid_utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-16LE string into UTF-32 string.
@@ -2697,20 +2810,20 @@ convert_valid_utf16_to_utf32(std::span<const char16_t> valid_utf16_input,
  *
  * @param input         the UTF-16LE string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf32_buffer   the pointer to buffer that can hold the conversion
+ * @param utf32_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf16le_to_utf32(std::span<const char16_t> valid_utf16_input,
                                std::span<char32_t> utf32_output) noexcept {
   return convert_valid_utf16le_to_utf32(
       valid_utf16_input.data(), valid_utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-16BE string into UTF-32 string.
@@ -2721,21 +2834,23 @@ convert_valid_utf16le_to_utf32(std::span<const char16_t> valid_utf16_input,
  *
  * @param input         the UTF-16BE string to convert
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param utf32_buffer   the pointer to buffer that can hold the conversion
+ * @param utf32_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(
     const char16_t *input, size_t length, char32_t *utf32_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf16be_to_utf32(std::span<const char16_t> valid_utf16_input,
                                std::span<char32_t> utf32_output) noexcept {
   return convert_valid_utf16be_to_utf32(
       valid_utf16_input.data(), valid_utf16_input.size(), utf32_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 /**
  * Compute the number of bytes that this UTF-16LE/BE string would require in
  * Latin1 format.
@@ -2763,14 +2878,16 @@ simdutf_warn_unused size_t latin1_length_from_utf16(size_t length) noexcept;
  */
 simdutf_warn_unused size_t utf8_length_from_utf16(const char16_t *input,
                                                   size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf8_length_from_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
   return utf8_length_from_utf16(valid_utf16_input.data(),
                                 valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 /**
  * Compute the number of bytes that this UTF-16LE string would require in UTF-8
  * format.
@@ -2784,13 +2901,13 @@ utf8_length_from_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t *input,
                                                     size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf8_length_from_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
   return utf8_length_from_utf16le(valid_utf16_input.data(),
                                   valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Compute the number of bytes that this UTF-16BE string would require in UTF-8
@@ -2805,14 +2922,16 @@ utf8_length_from_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t *input,
                                                     size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf8_length_from_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
   return utf8_length_from_utf16be(valid_utf16_input.data(),
                                   valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /**
  * Convert possibly broken UTF-32 string into UTF-8 string.
  *
@@ -2829,14 +2948,14 @@ utf8_length_from_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
 simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t *input,
                                                  size_t length,
                                                  char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf32_to_utf8(
     std::span<const char32_t> utf32_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
   return convert_utf32_to_utf8(utf32_input.data(), utf32_input.size(),
                                reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-32 string into UTF-8 string and stop on error.
@@ -2856,7 +2975,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf32_to_utf8(
  */
 simdutf_warn_unused result convert_utf32_to_utf8_with_errors(
     const char32_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf32_to_utf8_with_errors(
     std::span<const char32_t> utf32_input,
@@ -2865,7 +2984,7 @@ convert_utf32_to_utf8_with_errors(
       utf32_input.data(), utf32_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-32 string into UTF-8 string.
@@ -2876,13 +2995,13 @@ convert_utf32_to_utf8_with_errors(
  *
  * @param input         the UTF-32 string to convert
  * @param length        the length of the string in 4-byte code units (char32_t)
- * @param utf8_buffer   the pointer to buffer that can hold the conversion
+ * @param utf8_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
     const char32_t *input, size_t length, char *utf8_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
     std::span<const char32_t> valid_utf32_input,
     detail::output_span_of_byte_like auto &&utf8_output) noexcept {
@@ -2890,8 +3009,10 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
       valid_utf32_input.data(), valid_utf32_input.size(),
       reinterpret_cast<char *>(utf8_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /**
  * Using native endianness, convert possibly broken UTF-32 string into a UTF-16
  * string.
@@ -2908,14 +3029,14 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_utf8(
  */
 simdutf_warn_unused size_t convert_utf32_to_utf16(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf32_to_utf16(std::span<const char32_t> utf32_input,
                        std::span<char16_t> utf16_output) noexcept {
   return convert_utf32_to_utf16(utf32_input.data(), utf32_input.size(),
                                 utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-32 string into UTF-16LE string.
@@ -2932,15 +3053,17 @@ convert_utf32_to_utf16(std::span<const char32_t> utf32_input,
  */
 simdutf_warn_unused size_t convert_utf32_to_utf16le(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf32_to_utf16le(std::span<const char32_t> utf32_input,
                          std::span<char16_t> utf16_output) noexcept {
   return convert_utf32_to_utf16le(utf32_input.data(), utf32_input.size(),
                                   utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 /**
  * Convert possibly broken UTF-32 string into Latin1 string.
  *
@@ -2957,7 +3080,7 @@ convert_utf32_to_utf16le(std::span<const char32_t> utf32_input,
  */
 simdutf_warn_unused size_t convert_utf32_to_latin1(
     const char32_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_utf32_to_latin1(
     std::span<char32_t> utf32_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -2965,7 +3088,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf32_to_latin1(
       utf32_input.data(), utf32_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-32 string into Latin1 string and stop on error.
@@ -2986,7 +3109,7 @@ simdutf_really_inline simdutf_warn_unused size_t convert_utf32_to_latin1(
  */
 simdutf_warn_unused result convert_utf32_to_latin1_with_errors(
     const char32_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf32_to_latin1_with_errors(
     std::span<char32_t> utf32_input,
@@ -2995,7 +3118,7 @@ convert_utf32_to_latin1_with_errors(
       utf32_input.data(), utf32_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-32 string into Latin1 string.
@@ -3013,13 +3136,13 @@ convert_utf32_to_latin1_with_errors(
  *
  * @param input         the UTF-32 string to convert
  * @param length        the length of the string in 4-byte code units (char32_t)
- * @param latin1_buffer   the pointer to buffer that can hold the conversion
+ * @param latin1_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
     const char32_t *input, size_t length, char *latin1_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
     std::span<char32_t> valid_utf32_input,
     detail::output_span_of_byte_like auto &&latin1_output) noexcept {
@@ -3027,8 +3150,34 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
       valid_utf32_input.data(), valid_utf32_input.size(),
       reinterpret_cast<char *>(latin1_output.data()));
 }
-#endif
+  #endif // SIMDUTF_SPAN
+
+/**
+ * Compute the number of bytes that this UTF-32 string would require in Latin1
+ * format.
+ *
+ * This function does not validate the input. It is acceptable to pass invalid
+ * UTF-32 strings but in such cases the result is implementation defined.
+ *
+ * This function is not BOM-aware.
+ *
+ * @param length        the length of the string in 4-byte code units (char32_t)
+ * @return the number of bytes required to encode the UTF-32 string as Latin1
+ */
+simdutf_warn_unused size_t latin1_length_from_utf32(size_t length) noexcept;
+
+/**
+ * Compute the number of bytes that this Latin1 string would require in UTF-32
+ * format.
+ *
+ * @param length        the length of the string in Latin1 code units (char)
+ * @return the length of the string in 4-byte code units (char32_t) required to
+ * encode the Latin1 string as UTF-32
+ */
+simdutf_warn_unused size_t utf32_length_from_latin1(size_t length) noexcept;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /**
  * Convert possibly broken UTF-32 string into UTF-16BE string.
  *
@@ -3044,14 +3193,14 @@ simdutf_really_inline simdutf_warn_unused size_t convert_valid_utf32_to_latin1(
  */
 simdutf_warn_unused size_t convert_utf32_to_utf16be(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_utf32_to_utf16be(std::span<const char32_t> utf32_input,
                          std::span<char16_t> utf16_output) noexcept {
   return convert_utf32_to_utf16be(utf32_input.data(), utf32_input.size(),
                                   utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness, convert possibly broken UTF-32 string into UTF-16
@@ -3072,14 +3221,14 @@ convert_utf32_to_utf16be(std::span<const char32_t> utf32_input,
  */
 simdutf_warn_unused result convert_utf32_to_utf16_with_errors(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf32_to_utf16_with_errors(std::span<const char32_t> utf32_input,
                                    std::span<char16_t> utf16_output) noexcept {
   return convert_utf32_to_utf16_with_errors(
       utf32_input.data(), utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-32 string into UTF-16LE string and stop on error.
@@ -3099,7 +3248,7 @@ convert_utf32_to_utf16_with_errors(std::span<const char32_t> utf32_input,
  */
 simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf32_to_utf16le_with_errors(
     std::span<const char32_t> utf32_input,
@@ -3107,7 +3256,7 @@ convert_utf32_to_utf16le_with_errors(
   return convert_utf32_to_utf16le_with_errors(
       utf32_input.data(), utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert possibly broken UTF-32 string into UTF-16BE string and stop on error.
@@ -3127,7 +3276,7 @@ convert_utf32_to_utf16le_with_errors(
  */
 simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result
 convert_utf32_to_utf16be_with_errors(
     std::span<const char32_t> utf32_input,
@@ -3135,7 +3284,7 @@ convert_utf32_to_utf16be_with_errors(
   return convert_utf32_to_utf16be_with_errors(
       utf32_input.data(), utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness, convert valid UTF-32 string into a UTF-16 string.
@@ -3146,20 +3295,20 @@ convert_utf32_to_utf16be_with_errors(
  *
  * @param input         the UTF-32 string to convert
  * @param length        the length of the string in 4-byte code units (char32_t)
- * @param utf16_buffer   the pointer to buffer that can hold the conversion
+ * @param utf16_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf32_to_utf16(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf32_to_utf16(std::span<const char32_t> valid_utf32_input,
                              std::span<char16_t> utf16_output) noexcept {
   return convert_valid_utf32_to_utf16(
       valid_utf32_input.data(), valid_utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-32 string into UTF-16LE string.
@@ -3170,20 +3319,20 @@ convert_valid_utf32_to_utf16(std::span<const char32_t> valid_utf32_input,
  *
  * @param input         the UTF-32 string to convert
  * @param length        the length of the string in 4-byte code units (char32_t)
- * @param utf16_buffer   the pointer to buffer that can hold the conversion
+ * @param utf16_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf32_to_utf16le(std::span<const char32_t> valid_utf32_input,
                                std::span<char16_t> utf16_output) noexcept {
   return convert_valid_utf32_to_utf16le(
       valid_utf32_input.data(), valid_utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert valid UTF-32 string into UTF-16BE string.
@@ -3194,21 +3343,23 @@ convert_valid_utf32_to_utf16le(std::span<const char32_t> valid_utf32_input,
  *
  * @param input         the UTF-32 string to convert
  * @param length        the length of the string in 4-byte code units (char32_t)
- * @param utf16_buffer   the pointer to buffer that can hold the conversion
+ * @param utf16_buffer   the pointer to a buffer that can hold the conversion
  * result
  * @return number of written code units; 0 if conversion is not possible
  */
 simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(
     const char32_t *input, size_t length, char16_t *utf16_buffer) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 convert_valid_utf32_to_utf16be(std::span<const char32_t> valid_utf32_input,
                                std::span<char16_t> utf16_output) noexcept {
   return convert_valid_utf32_to_utf16be(
       valid_utf32_input.data(), valid_utf32_input.size(), utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 /**
  * Change the endianness of the input. Can be used to go from UTF-16LE to
  * UTF-16BE or from UTF-16BE to UTF-16LE.
@@ -3219,20 +3370,22 @@ convert_valid_utf32_to_utf16be(std::span<const char32_t> valid_utf32_input,
  *
  * @param input         the UTF-16 string to process
  * @param length        the length of the string in 2-byte code units (char16_t)
- * @param output        the pointer to buffer that can hold the conversion
+ * @param output        the pointer to a buffer that can hold the conversion
  * result
  */
 void change_endianness_utf16(const char16_t *input, size_t length,
                              char16_t *output) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline void
 change_endianness_utf16(std::span<const char16_t> utf16_input,
                         std::span<char16_t> utf16_output) noexcept {
   return change_endianness_utf16(utf16_input.data(), utf16_input.size(),
                                  utf16_output.data());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 /**
  * Compute the number of bytes that this UTF-32 string would require in UTF-8
  * format.
@@ -3246,14 +3399,16 @@ change_endianness_utf16(std::span<const char16_t> utf16_input,
  */
 simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t *input,
                                                   size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf8_length_from_utf32(std::span<const char32_t> valid_utf32_input) noexcept {
   return utf8_length_from_utf32(valid_utf32_input.data(),
                                 valid_utf32_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 /**
  * Compute the number of two-byte code units that this UTF-32 string would
  * require in UTF-16 format.
@@ -3267,13 +3422,13 @@ utf8_length_from_utf32(std::span<const char32_t> valid_utf32_input) noexcept {
  */
 simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t *input,
                                                    size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf16_length_from_utf32(std::span<const char32_t> valid_utf32_input) noexcept {
   return utf16_length_from_utf32(valid_utf32_input.data(),
                                  valid_utf32_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Using native endianness; Compute the number of bytes that this UTF-16
@@ -3292,13 +3447,13 @@ utf16_length_from_utf32(std::span<const char32_t> valid_utf32_input) noexcept {
  */
 simdutf_warn_unused size_t utf32_length_from_utf16(const char16_t *input,
                                                    size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 utf32_length_from_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
   return utf32_length_from_utf16(valid_utf16_input.data(),
                                  valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Compute the number of bytes that this UTF-16LE string would require in UTF-32
@@ -3317,13 +3472,13 @@ utf32_length_from_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t *input,
                                                      size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf16le(
     std::span<const char16_t> valid_utf16_input) noexcept {
   return utf32_length_from_utf16le(valid_utf16_input.data(),
                                    valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Compute the number of bytes that this UTF-16BE string would require in UTF-32
@@ -3342,14 +3497,16 @@ simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf16le(
  */
 simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t *input,
                                                      size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf16be(
     std::span<const char16_t> valid_utf16_input) noexcept {
   return utf32_length_from_utf16be(valid_utf16_input.data(),
                                    valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
 /**
  * Count the number of code points (characters) in the string assuming that
  * it is valid.
@@ -3366,12 +3523,12 @@ simdutf_really_inline simdutf_warn_unused size_t utf32_length_from_utf16be(
  */
 simdutf_warn_unused size_t count_utf16(const char16_t *input,
                                        size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 count_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
   return count_utf16(valid_utf16_input.data(), valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Count the number of code points (characters) in the string assuming that
@@ -3389,12 +3546,12 @@ count_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t count_utf16le(const char16_t *input,
                                          size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 count_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
   return count_utf16le(valid_utf16_input.data(), valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Count the number of code points (characters) in the string assuming that
@@ -3412,13 +3569,15 @@ count_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t count_utf16be(const char16_t *input,
                                          size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 count_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
   return count_utf16be(valid_utf16_input.data(), valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
 /**
  * Count the number of code points (characters) in the string assuming that
  * it is valid.
@@ -3433,13 +3592,13 @@ count_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t count_utf8(const char *input,
                                       size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t count_utf8(
     const detail::input_span_of_byte_like auto &valid_utf8_input) noexcept {
   return count_utf8(reinterpret_cast<const char *>(valid_utf8_input.data()),
                     valid_utf8_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Given a valid UTF-8 string having a possibly truncated last character,
@@ -3456,15 +3615,17 @@ simdutf_really_inline simdutf_warn_unused size_t count_utf8(
  * @return the length of the string in bytes, possibly shorter by 1 to 3 bytes
  */
 simdutf_warn_unused size_t trim_partial_utf8(const char *input, size_t length);
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t trim_partial_utf8(
     const detail::input_span_of_byte_like auto &valid_utf8_input) noexcept {
   return trim_partial_utf8(
       reinterpret_cast<const char *>(valid_utf8_input.data()),
       valid_utf8_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_UTF16
 /**
  * Given a valid UTF-16BE string having a possibly truncated last character,
  * this function checks the end of string. If the last character is truncated
@@ -3481,13 +3642,13 @@ simdutf_really_inline simdutf_warn_unused size_t trim_partial_utf8(
  */
 simdutf_warn_unused size_t trim_partial_utf16be(const char16_t *input,
                                                 size_t length);
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 trim_partial_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
   return trim_partial_utf16be(valid_utf16_input.data(),
                               valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Given a valid UTF-16LE string having a possibly truncated last character,
@@ -3505,13 +3666,13 @@ trim_partial_utf16be(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t trim_partial_utf16le(const char16_t *input,
                                                 size_t length);
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 trim_partial_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
   return trim_partial_utf16le(valid_utf16_input.data(),
                               valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Given a valid UTF-16 string having a possibly truncated last character,
@@ -3529,13 +3690,18 @@ trim_partial_utf16le(std::span<const char16_t> valid_utf16_input) noexcept {
  */
 simdutf_warn_unused size_t trim_partial_utf16(const char16_t *input,
                                               size_t length);
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 trim_partial_utf16(std::span<const char16_t> valid_utf16_input) noexcept {
   return trim_partial_utf16(valid_utf16_input.data(), valid_utf16_input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_BASE64
+  #ifndef SIMDUTF_NEED_TRAILING_ZEROES
+    #define SIMDUTF_NEED_TRAILING_ZEROES 1
+  #endif
 // base64_options are used to specify the base64 encoding options.
 // ASCII spaces are ' ', '\t', '\n', '\r', '\f'
 // garbage characters are characters that are not part of the base64 alphabet
@@ -3577,14 +3743,14 @@ enum last_chunk_handling_options : uint64_t {
  */
 simdutf_warn_unused size_t
 maximal_binary_length_from_base64(const char *input, size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 maximal_binary_length_from_base64(
     const detail::input_span_of_byte_like auto &input) noexcept {
   return maximal_binary_length_from_base64(
       reinterpret_cast<const char *>(input.data()), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Provide the maximal binary length in bytes given the base64 input.
@@ -3598,12 +3764,12 @@ maximal_binary_length_from_base64(
  */
 simdutf_warn_unused size_t maximal_binary_length_from_base64(
     const char16_t *input, size_t length) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 maximal_binary_length_from_base64(std::span<const char16_t> input) noexcept {
   return maximal_binary_length_from_base64(input.data(), input.size());
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert a base64 input to a binary output.
@@ -3646,7 +3812,7 @@ maximal_binary_length_from_base64(std::span<const char16_t> input) noexcept {
  *
  * @param input         the base64 string to process
  * @param length        the length of the string in bytes
- * @param output        the pointer to buffer that can hold the conversion
+ * @param output        the pointer to a buffer that can hold the conversion
  * result (should be at least maximal_binary_length_from_base64(input, length)
  * bytes long).
  * @param options       the base64 options to use, usually base64_default or
@@ -3663,7 +3829,7 @@ simdutf_warn_unused result base64_to_binary(
     const char *input, size_t length, char *output,
     base64_options options = base64_default,
     last_chunk_handling_options last_chunk_options = loose) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result base64_to_binary(
     const detail::input_span_of_byte_like auto &input,
     detail::output_span_of_byte_like auto &&binary_output,
@@ -3674,7 +3840,7 @@ simdutf_really_inline simdutf_warn_unused result base64_to_binary(
                           reinterpret_cast<char *>(binary_output.data()),
                           options, last_chunk_options);
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Provide the base64 length in bytes given the length of a binary input.
@@ -3699,7 +3865,7 @@ simdutf_warn_unused size_t base64_length_from_binary(
  *
  * @param input         the binary to process
  * @param length        the length of the input in bytes
- * @param output        the pointer to buffer that can hold the conversion
+ * @param output        the pointer to a buffer that can hold the conversion
  * result (should be at least base64_length_from_binary(length) bytes long)
  * @param options       the base64 options to use, can be base64_default or
  * base64_url, is base64_default by default.
@@ -3708,7 +3874,7 @@ simdutf_warn_unused size_t base64_length_from_binary(
  */
 size_t binary_to_base64(const char *input, size_t length, char *output,
                         base64_options options = base64_default) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused size_t
 binary_to_base64(const detail::input_span_of_byte_like auto &input,
                  detail::output_span_of_byte_like auto &&binary_output,
@@ -3717,7 +3883,58 @@ binary_to_base64(const detail::input_span_of_byte_like auto &input,
       reinterpret_cast<const char *>(input.data()), input.size(),
       reinterpret_cast<char *>(binary_output.data()), options);
 }
-#endif
+  #endif // SIMDUTF_SPAN
+
+  #if SIMDUTF_ATOMIC_REF
+/**
+ * Convert a binary input to a base64 output, using atomic accesses.
+ * This function comes with a potentially significant performance
+ * penalty, but it may be useful in some cases where the input and
+ * output buffers are shared between threads, to avoid undefined
+ * behavior in case of data races.
+ *
+ * The function is for advanced users. Its main use case is when
+ * to silence sanitizer warnings. We have no documented use case
+ * where this function is actually necessary in terms of practical correctness.
+ *
+ * This function is only available when simdutf is compiled with
+ * C++20 support and __cpp_lib_atomic_ref >= 201806L. You may check
+ * the availability of this function by checking the macro
+ * SIMDUTF_ATOMIC_REF.
+ *
+ * The default option (simdutf::base64_default) uses the characters `+` and `/`
+ * as part of its alphabet. Further, it adds padding (`=`) at the end of the
+ * output to ensure that the output length is a multiple of four.
+ *
+ * The URL option (simdutf::base64_url) uses the characters `-` and `_` as part
+ * of its alphabet. No padding is added at the end of the output.
+ *
+ * This function always succeeds.
+ *
+ * @brief atomic_binary_to_base64
+ * @param input         the binary to process
+ * @param length        the length of the input in bytes
+ * @param output        the pointer to a buffer that can hold the conversion
+ * result (should be at least base64_length_from_binary(length) bytes long)
+ * @param options       the base64 options to use, can be base64_default or
+ * base64_url, is base64_default by default.
+ * @return number of written bytes, will be equal to
+ * base64_length_from_binary(length, options)
+ */
+size_t
+atomic_binary_to_base64(const char *input, size_t length, char *output,
+                        base64_options options = base64_default) noexcept;
+    #if SIMDUTF_SPAN
+simdutf_really_inline simdutf_warn_unused size_t
+atomic_binary_to_base64(const detail::input_span_of_byte_like auto &input,
+                        detail::output_span_of_byte_like auto &&binary_output,
+                        base64_options options = base64_default) noexcept {
+  return atomic_binary_to_base64(
+      reinterpret_cast<const char *>(input.data()), input.size(),
+      reinterpret_cast<char *>(binary_output.data()), options);
+}
+    #endif // SIMDUTF_SPAN
+  #endif   // SIMDUTF_ATOMIC_REF
 
 /**
  * Convert a base64 input to a binary output.
@@ -3749,7 +3966,7 @@ binary_to_base64(const detail::input_span_of_byte_like auto &input,
  * characters) must be divisible by four.
  *
  * You should call this function with a buffer that is at least
- * maximal_binary_length_from_utf6_base64(input, length) bytes long. If you fail
+ * maximal_binary_length_from_base64(input, length) bytes long. If you fail
  * to provide that much space, the function may cause a buffer overflow.
  *
  * Advanced users may want to taylor how the last chunk is handled. By default,
@@ -3761,7 +3978,7 @@ binary_to_base64(const detail::input_span_of_byte_like auto &input,
  * @param input         the base64 string to process, in ASCII stored as 16-bit
  * units
  * @param length        the length of the string in 16-bit units
- * @param output        the pointer to buffer that can hold the conversion
+ * @param output        the pointer to a buffer that can hold the conversion
  * result (should be at least maximal_binary_length_from_base64(input, length)
  * bytes long).
  * @param options       the base64 options to use, can be base64_default or
@@ -3780,7 +3997,7 @@ base64_to_binary(const char16_t *input, size_t length, char *output,
                  base64_options options = base64_default,
                  last_chunk_handling_options last_chunk_options =
                      last_chunk_handling_options::loose) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result base64_to_binary(
     std::span<const char16_t> input,
     detail::output_span_of_byte_like auto &&binary_output,
@@ -3790,7 +4007,7 @@ simdutf_really_inline simdutf_warn_unused result base64_to_binary(
                           reinterpret_cast<char *>(binary_output.data()),
                           options, last_chunk_options);
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 /**
  * Convert a base64 input to a binary output.
@@ -3838,7 +4055,7 @@ simdutf_really_inline simdutf_warn_unused result base64_to_binary(
  * @param input         the base64 string to process, in ASCII stored as 8-bit
  * or 16-bit units
  * @param length        the length of the string in 8-bit or 16-bit units.
- * @param output        the pointer to buffer that can hold the conversion
+ * @param output        the pointer to a buffer that can hold the conversion
  * result.
  * @param outlen        the number of bytes that can be written in the output
  * buffer. Upon return, it is modified to reflect how many bytes were written.
@@ -3858,7 +4075,7 @@ base64_to_binary_safe(const char *input, size_t length, char *output,
                       size_t &outlen, base64_options options = base64_default,
                       last_chunk_handling_options last_chunk_options =
                           last_chunk_handling_options::loose) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result base64_to_binary_safe(
     const detail::input_span_of_byte_like auto &input,
     detail::output_span_of_byte_like auto &&binary_output,
@@ -3873,14 +4090,14 @@ simdutf_really_inline simdutf_warn_unused result base64_to_binary_safe(
                                reinterpret_cast<char *>(binary_output.data()),
                                outlen, options, last_chunk_options);
 }
-#endif
+  #endif // SIMDUTF_SPAN
 
 simdutf_warn_unused result
 base64_to_binary_safe(const char16_t *input, size_t length, char *output,
                       size_t &outlen, base64_options options = base64_default,
                       last_chunk_handling_options last_chunk_options =
                           last_chunk_handling_options::loose) noexcept;
-#if SIMDUTF_SPAN
+  #if SIMDUTF_SPAN
 simdutf_really_inline simdutf_warn_unused result base64_to_binary_safe(
     std::span<const char16_t> input,
     detail::output_span_of_byte_like auto &&binary_output,
@@ -3894,7 +4111,8 @@ simdutf_really_inline simdutf_warn_unused result base64_to_binary_safe(
                                reinterpret_cast<char *>(binary_output.data()),
                                outlen, options, last_chunk_options);
 }
-#endif
+  #endif // SIMDUTF_SPAN
+#endif   // SIMDUTF_FEATURE_BASE64
 
 /**
  * An implementation of simdutf for a particular CPU architecture.
@@ -3938,6 +4156,7 @@ class implementation {
    */
   bool supported_by_runtime_system() const;
 
+#if SIMDUTF_FEATURE_DETECT_ENCODING
   /**
    * This function will try to detect the encoding
    * @param input the string to identify
@@ -3955,6 +4174,7 @@ class implementation {
    */
   virtual int detect_encodings(const char *input,
                                size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_DETECT_ENCODING
 
   /**
    * @private For internal implementation use
@@ -3967,6 +4187,7 @@ class implementation {
     return _required_instruction_sets;
   }
 
+#if SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
   /**
    * Validate the UTF-8 string.
    *
@@ -3978,7 +4199,9 @@ class implementation {
    */
   simdutf_warn_unused virtual bool validate_utf8(const char *buf,
                                                  size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF8
   /**
    * Validate the UTF-8 string and stop on errors.
    *
@@ -3993,7 +4216,9 @@ class implementation {
    */
   simdutf_warn_unused virtual result
   validate_utf8_with_errors(const char *buf, size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_ASCII
   /**
    * Validate the ASCII string.
    *
@@ -4020,7 +4245,9 @@ class implementation {
    */
   simdutf_warn_unused virtual result
   validate_ascii_with_errors(const char *buf, size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_ASCII
 
+#if SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
   /**
    * Validate the UTF-16LE string.This function may be best when you expect
    * the input to be almost always valid. Otherwise, consider using
@@ -4037,7 +4264,9 @@ class implementation {
    */
   simdutf_warn_unused virtual bool
   validate_utf16le(const char16_t *buf, size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF16
   /**
    * Validate the UTF-16BE string. This function may be best when you expect
    * the input to be almost always valid. Otherwise, consider using
@@ -4094,7 +4323,9 @@ class implementation {
   simdutf_warn_unused virtual result
   validate_utf16be_with_errors(const char16_t *buf,
                                size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
   /**
    * Validate the UTF-32 string.
    *
@@ -4109,7 +4340,9 @@ class implementation {
    */
   simdutf_warn_unused virtual bool
   validate_utf32(const char32_t *buf, size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF32 || SIMDUTF_FEATURE_DETECT_ENCODING
 
+#if SIMDUTF_FEATURE_UTF32
   /**
    * Validate the UTF-32 string and stop on error.
    *
@@ -4128,7 +4361,9 @@ class implementation {
   simdutf_warn_unused virtual result
   validate_utf32_with_errors(const char32_t *buf,
                              size_t len) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert Latin1 string into UTF8 string.
    *
@@ -4142,7 +4377,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_latin1_to_utf8(const char *input, size_t length,
                          char *utf8_output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert possibly Latin1 string into UTF-16LE string.
    *
@@ -4170,7 +4407,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_latin1_to_utf16be(const char *input, size_t length,
                             char16_t *utf16_output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert Latin1 string into UTF-32 string.
    *
@@ -4184,7 +4423,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_latin1_to_utf32(const char *input, size_t length,
                           char32_t *utf32_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert possibly broken UTF-8 string into latin1 string.
    *
@@ -4243,7 +4484,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_valid_utf8_to_latin1(const char *input, size_t length,
                                char *latin1_output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
   /**
    * Convert possibly broken UTF-8 string into UTF-16LE string.
    *
@@ -4313,7 +4556,9 @@ class implementation {
   simdutf_warn_unused virtual result convert_utf8_to_utf16be_with_errors(
       const char *input, size_t length,
       char16_t *utf16_output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
   /**
    * Convert possibly broken UTF-8 string into UTF-32 string.
    *
@@ -4347,7 +4592,9 @@ class implementation {
   simdutf_warn_unused virtual result
   convert_utf8_to_utf32_with_errors(const char *input, size_t length,
                                     char32_t *utf32_output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
   /**
    * Convert valid UTF-8 string into UTF-16LE string.
    *
@@ -4375,7 +4622,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_valid_utf8_to_utf16be(const char *input, size_t length,
                                 char16_t *utf16_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
   /**
    * Convert valid UTF-8 string into UTF-32 string.
    *
@@ -4389,7 +4638,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_valid_utf8_to_utf32(const char *input, size_t length,
                               char32_t *utf32_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
   /**
    * Compute the number of 2-byte code units that this UTF-8 string would
    * require in UTF-16LE format.
@@ -4404,7 +4655,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   utf16_length_from_utf8(const char *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
   /**
    * Compute the number of 4-byte code units that this UTF-8 string would
    * require in UTF-32 format.
@@ -4421,7 +4674,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   utf32_length_from_utf8(const char *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert possibly broken UTF-16LE string into Latin1 string.
    *
@@ -4555,7 +4810,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   convert_valid_utf16be_to_latin1(const char16_t *input, size_t length,
                                   char *latin1_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
   /**
    * Convert possibly broken UTF-16LE string into UTF-8 string.
    *
@@ -4648,7 +4905,7 @@ class implementation {
    * @param input         the UTF-16LE string to convert
    * @param length        the length of the string in 2-byte code units
    * (char16_t)
-   * @param utf8_buffer   the pointer to buffer that can hold the conversion
+   * @param utf8_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
@@ -4666,14 +4923,16 @@ class implementation {
    * @param input         the UTF-16BE string to convert
    * @param length        the length of the string in 2-byte code units
    * (char16_t)
-   * @param utf8_buffer   the pointer to buffer that can hold the conversion
+   * @param utf8_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
   simdutf_warn_unused virtual size_t
   convert_valid_utf16be_to_utf8(const char16_t *input, size_t length,
                                 char *utf8_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
   /**
    * Convert possibly broken UTF-16LE string into UTF-32 string.
    *
@@ -4766,7 +5025,7 @@ class implementation {
    * @param input         the UTF-16LE string to convert
    * @param length        the length of the string in 2-byte code units
    * (char16_t)
-   * @param utf32_buffer   the pointer to buffer that can hold the conversion
+   * @param utf32_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
@@ -4784,14 +5043,16 @@ class implementation {
    * @param input         the UTF-16BE string to convert
    * @param length        the length of the string in 2-byte code units
    * (char16_t)
-   * @param utf32_buffer   the pointer to buffer that can hold the conversion
+   * @param utf32_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
   simdutf_warn_unused virtual size_t
   convert_valid_utf16be_to_utf32(const char16_t *input, size_t length,
                                  char32_t *utf32_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
   /**
    * Compute the number of bytes that this UTF-16LE string would require in
    * UTF-8 format.
@@ -4827,7 +5088,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   utf8_length_from_utf16be(const char16_t *input,
                            size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert possibly broken UTF-32 string into Latin1 string.
    *
@@ -4844,11 +5107,12 @@ class implementation {
    * @return number of written code units; 0 if input is not a valid UTF-32
    * string
    */
-
   simdutf_warn_unused virtual size_t
   convert_utf32_to_latin1(const char32_t *input, size_t length,
                           char *latin1_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
   /**
    * Convert possibly broken UTF-32 string into Latin1 string and stop on error.
    * If the string cannot be represented as Latin1, an error is returned.
@@ -4888,14 +5152,16 @@ class implementation {
    * @param input         the UTF-32 string to convert
    * @param length        the length of the string in 4-byte code units
    * (char32_t)
-   * @param latin1_buffer   the pointer to buffer that can hold the conversion
+   * @param latin1_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
   simdutf_warn_unused virtual size_t
   convert_valid_utf32_to_latin1(const char32_t *input, size_t length,
                                 char *latin1_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
   /**
    * Convert possibly broken UTF-32 string into UTF-8 string.
    *
@@ -4946,14 +5212,16 @@ class implementation {
    * @param input         the UTF-32 string to convert
    * @param length        the length of the string in 4-byte code units
    * (char32_t)
-   * @param utf8_buffer   the pointer to buffer that can hold the conversion
+   * @param utf8_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
   simdutf_warn_unused virtual size_t
   convert_valid_utf32_to_utf8(const char32_t *input, size_t length,
                               char *utf8_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
   /**
    * Return the number of bytes that this UTF-16 string would require in Latin1
    * format.
@@ -4965,8 +5233,12 @@ class implementation {
    * @return the number of bytes required to encode the UTF-16 string as Latin1
    */
   simdutf_warn_unused virtual size_t
-  utf16_length_from_latin1(size_t length) const noexcept = 0;
+  utf16_length_from_latin1(size_t length) const noexcept {
+    return length;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
   /**
    * Convert possibly broken UTF-32 string into UTF-16LE string.
    *
@@ -5059,7 +5331,7 @@ class implementation {
    * @param input         the UTF-32 string to convert
    * @param length        the length of the string in 4-byte code units
    * (char32_t)
-   * @param utf16_buffer   the pointer to buffer that can hold the conversion
+   * @param utf16_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
@@ -5077,14 +5349,16 @@ class implementation {
    * @param input         the UTF-32 string to convert
    * @param length        the length of the string in 4-byte code units
    * (char32_t)
-   * @param utf16_buffer   the pointer to buffer that can hold the conversion
+   * @param utf16_buffer   the pointer to a buffer that can hold the conversion
    * result
    * @return number of written code units; 0 if conversion is not possible
    */
   simdutf_warn_unused virtual size_t
   convert_valid_utf32_to_utf16be(const char32_t *input, size_t length,
                                  char16_t *utf16_buffer) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
   /**
    * Change the endianness of the input. Can be used to go from UTF-16LE to
    * UTF-16BE or from UTF-16BE to UTF-16LE.
@@ -5096,12 +5370,14 @@ class implementation {
    * @param input         the UTF-16 string to process
    * @param length        the length of the string in 2-byte code units
    * (char16_t)
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result
    */
   virtual void change_endianness_utf16(const char16_t *input, size_t length,
                                        char16_t *output) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
   /**
    * Return the number of bytes that this Latin1 string would require in UTF-8
    * format.
@@ -5112,7 +5388,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   utf8_length_from_latin1(const char *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
   /**
    * Compute the number of bytes that this UTF-32 string would require in UTF-8
    * format.
@@ -5128,7 +5406,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   utf8_length_from_utf32(const char32_t *input,
                          size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
   /**
    * Compute the number of bytes that this UTF-32 string would require in Latin1
    * format.
@@ -5141,8 +5421,12 @@ class implementation {
    * @return the number of bytes required to encode the UTF-32 string as Latin1
    */
   simdutf_warn_unused virtual size_t
-  latin1_length_from_utf32(size_t length) const noexcept = 0;
+  latin1_length_from_utf32(size_t length) const noexcept {
+    return length;
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
   /**
    * Compute the number of bytes that this UTF-8 string would require in Latin1
    * format.
@@ -5156,7 +5440,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   latin1_length_from_utf8(const char *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
   /**
    * Compute the number of bytes that this UTF-16LE/BE string would require in
    * Latin1 format.
@@ -5173,8 +5459,12 @@ class implementation {
    * Latin1
    */
   simdutf_warn_unused virtual size_t
-  latin1_length_from_utf16(size_t length) const noexcept = 0;
+  latin1_length_from_utf16(size_t length) const noexcept {
+    return length;
+  }
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
   /**
    * Compute the number of two-byte code units that this UTF-32 string would
    * require in UTF-16 format.
@@ -5190,19 +5480,24 @@ class implementation {
   simdutf_warn_unused virtual size_t
   utf16_length_from_utf32(const char32_t *input,
                           size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
   /**
    * Return the number of bytes that this UTF-32 string would require in Latin1
    * format.
    *
-   * @param input         the UTF-32 string to convert
    * @param length        the length of the string in 4-byte code units
    * (char32_t)
    * @return the number of bytes required to encode the UTF-32 string as Latin1
    */
   simdutf_warn_unused virtual size_t
-  utf32_length_from_latin1(size_t length) const noexcept = 0;
+  utf32_length_from_latin1(size_t length) const noexcept {
+    return length;
+  }
+#endif // SIMDUTF_FEATURE_UTF32 && SIMDUTF_FEATURE_LATIN1
 
+#if SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
   /**
    * Compute the number of bytes that this UTF-16LE string would require in
    * UTF-32 format.
@@ -5244,7 +5539,9 @@ class implementation {
   simdutf_warn_unused virtual size_t
   utf32_length_from_utf16be(const char16_t *input,
                             size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16 && SIMDUTF_FEATURE_UTF32
 
+#if SIMDUTF_FEATURE_UTF16
   /**
    * Count the number of code points (characters) in the string assuming that
    * it is valid.
@@ -5280,7 +5577,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   count_utf16be(const char16_t *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF16
 
+#if SIMDUTF_FEATURE_UTF8
   /**
    * Count the number of code points (characters) in the string assuming that
    * it is valid.
@@ -5295,7 +5594,9 @@ class implementation {
    */
   simdutf_warn_unused virtual size_t
   count_utf8(const char *input, size_t length) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_UTF8
 
+#if SIMDUTF_FEATURE_BASE64
   /**
    * Provide the maximal binary length in bytes given the base64 input.
    * In general, if the input contains ASCII spaces, the result will be less
@@ -5306,9 +5607,8 @@ class implementation {
    * @param length        the length of the base64 input in bytes
    * @return maximal number of binary bytes
    */
-  simdutf_warn_unused virtual size_t
-  maximal_binary_length_from_base64(const char *input,
-                                    size_t length) const noexcept = 0;
+  simdutf_warn_unused size_t maximal_binary_length_from_base64(
+      const char *input, size_t length) const noexcept;
 
   /**
    * Provide the maximal binary length in bytes given the base64 input.
@@ -5321,9 +5621,8 @@ class implementation {
    * @param length        the length of the base64 input in 16-bit units
    * @return maximal number of binary bytes
    */
-  simdutf_warn_unused virtual size_t
-  maximal_binary_length_from_base64(const char16_t *input,
-                                    size_t length) const noexcept = 0;
+  simdutf_warn_unused size_t maximal_binary_length_from_base64(
+      const char16_t *input, size_t length) const noexcept;
 
   /**
    * Convert a base64 input to a binary output.
@@ -5347,7 +5646,7 @@ class implementation {
    *
    * @param input         the base64 string to process
    * @param length        the length of the string in bytes
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result (should be at least maximal_binary_length_from_base64(input, length)
    * bytes long).
    * @param options       the base64 options to use, can be base64_default or
@@ -5386,7 +5685,7 @@ class implementation {
    *
    * @param input         the base64 string to process
    * @param length        the length of the string in bytes
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result (should be at least maximal_binary_length_from_base64(input, length)
    * bytes long).
    * @param options       the base64 options to use, can be base64_default or
@@ -5416,13 +5715,13 @@ class implementation {
    * character that is not a valid base64 character (INVALID_BASE64_CHARACTER).
    *
    * You should call this function with a buffer that is at least
-   * maximal_binary_length_from_utf6_base64(input, length) bytes long. If you
+   * maximal_binary_length_from_base64(input, length) bytes long. If you
    * fail to provide that much space, the function may cause a buffer overflow.
    *
    * @param input         the base64 string to process, in ASCII stored as
    * 16-bit units
    * @param length        the length of the string in 16-bit units
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result (should be at least maximal_binary_length_from_base64(input, length)
    * bytes long).
    * @param options       the base64 options to use, can be base64_default or
@@ -5461,7 +5760,7 @@ class implementation {
    *
    * @param input         the base64 string to process
    * @param length        the length of the string in bytes
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result (should be at least maximal_binary_length_from_base64(input, length)
    * bytes long).
    * @param options       the base64 options to use, can be base64_default or
@@ -5482,9 +5781,8 @@ class implementation {
    * base64_url, is base64_default by default.
    * @return number of base64 bytes
    */
-  simdutf_warn_unused virtual size_t base64_length_from_binary(
-      size_t length,
-      base64_options options = base64_default) const noexcept = 0;
+  simdutf_warn_unused size_t base64_length_from_binary(
+      size_t length, base64_options options = base64_default) const noexcept;
 
   /**
    * Convert a binary input to a base64 output.
@@ -5500,7 +5798,7 @@ class implementation {
    *
    * @param input         the binary to process
    * @param length        the length of the input in bytes
-   * @param output        the pointer to buffer that can hold the conversion
+   * @param output        the pointer to a buffer that can hold the conversion
    * result (should be at least base64_length_from_binary(length) bytes long)
    * @param options       the base64 options to use, can be base64_default or
    * base64_url, is base64_default by default.
@@ -5510,6 +5808,27 @@ class implementation {
   virtual size_t
   binary_to_base64(const char *input, size_t length, char *output,
                    base64_options options = base64_default) const noexcept = 0;
+#endif // SIMDUTF_FEATURE_BASE64
+
+#ifdef SIMDUTF_INTERNAL_TESTS
+  // This method is exported only in developer mode, its purpose
+  // is to expose some internal test procedures from the given
+  // implementation and then use them through our standard test
+  // framework.
+  //
+  // Regular users should not use it, the tests of the public
+  // API are enough.
+
+  struct TestProcedure {
+    // display name
+    std::string name;
+
+    // procedure should return whether given test pass or not
+    void (*procedure)(const implementation &);
+  };
+
+  virtual std::vector<TestProcedure> internal_tests() const;
+#endif
 
 protected:
   /** @private Construct an implementation with the given name and description.
diff --git a/deps/v8/AUTHORS b/deps/v8/AUTHORS
index 7b43882edca7ea..eeac21260b74dd 100644
--- a/deps/v8/AUTHORS
+++ b/deps/v8/AUTHORS
@@ -188,6 +188,7 @@ Matt Hanselman <mjhanselman@gmail.com>
 Matthew Sporleder <msporleder@gmail.com>
 Maxim Mazurok <maxim@mazurok.com>
 Maxim Mossienko <maxim.mossienko@gmail.com>
+Meir Shpilraien <meir@redis.com>
 Michael Lutz <michi@icosahedron.de>
 Michael Mclaughlin <m8ch88l@gmail.com>
 Michael Smith <mike@w3.org>
diff --git a/deps/v8/src/inspector/string-16.cc b/deps/v8/src/inspector/string-16.cc
index a8b786a8166d54..6df9963e970e51 100644
--- a/deps/v8/src/inspector/string-16.cc
+++ b/deps/v8/src/inspector/string-16.cc
@@ -27,7 +27,7 @@ bool isSpaceOrNewLine(UChar c) {
   return isASCII(c) && c <= ' ' && (c == ' ' || (c <= 0xD && c >= 0x9));
 }
 
-int64_t charactersToInteger(const UChar* characters, size_t length,
+int64_t charactersToInteger(const uint16_t* characters, size_t length,
                             bool* ok = nullptr) {
   std::vector<char> buffer;
   buffer.reserve(length + 1);
@@ -50,6 +50,8 @@ int64_t charactersToInteger(const UChar* characters, size_t length,
 
 String16::String16(const UChar* characters, size_t size)
     : m_impl(characters, size) {}
+String16::String16(const uint16_t* characters, size_t size)
+    : m_impl(reinterpret_cast<const UChar*>(characters), size) {}
 
 String16::String16(const UChar* characters) : m_impl(characters) {}
 
@@ -241,6 +243,10 @@ String16 String16::fromUTF16LE(const UChar* stringStart, size_t length) {
 #endif  // V8_TARGET_BIG_ENDIAN
 }
 
+String16 String16::fromUTF16LE(const uint16_t* stringStart, size_t length) {
+  return fromUTF16LE(reinterpret_cast<const UChar*>(stringStart), length);
+}
+
 std::string String16::utf8() const {
   return UTF16ToUTF8(m_impl.data(), m_impl.size());
 }
diff --git a/deps/v8/src/inspector/string-16.h b/deps/v8/src/inspector/string-16.h
index 29651ac9592922..feb091433b9687 100644
--- a/deps/v8/src/inspector/string-16.h
+++ b/deps/v8/src/inspector/string-16.h
@@ -17,7 +17,7 @@
 
 namespace v8_inspector {
 
-using UChar = uint16_t;
+using UChar = char16_t;
 
 class String16 {
  public:
@@ -27,6 +27,7 @@ class String16 {
   String16(const String16&) V8_NOEXCEPT = default;
   String16(String16&&) V8_NOEXCEPT = default;
   String16(const UChar* characters, size_t size);
+  String16(const uint16_t* characters, size_t size);
   V8_EXPORT String16(const UChar* characters);
   V8_EXPORT String16(const char* characters);
   String16(const char* characters, size_t size);
@@ -48,7 +49,9 @@ class String16 {
   int toInteger(bool* ok = nullptr) const;
   std::pair<size_t, size_t> getTrimmedOffsetAndLength() const;
   String16 stripWhiteSpace() const;
-  const UChar* characters16() const { return m_impl.c_str(); }
+  const uint16_t* characters16() const {
+    return reinterpret_cast<const uint16_t*>(m_impl.c_str());
+  }
   size_t length() const { return m_impl.length(); }
   bool isEmpty() const { return !m_impl.length(); }
   UChar operator[](size_t index) const { return m_impl[index]; }
@@ -78,6 +81,8 @@ class String16 {
   // On Big endian architectures, byte order needs to be flipped.
   V8_EXPORT static String16 fromUTF16LE(const UChar* stringStart,
                                         size_t length);
+  V8_EXPORT static String16 fromUTF16LE(const uint16_t* stringStart,
+                                        size_t length);
 
   std::size_t hash() const {
     if (!hash_code) {
diff --git a/deps/v8/src/inspector/v8-string-conversions.cc b/deps/v8/src/inspector/v8-string-conversions.cc
index 0c75e66b972274..8cf19be816c240 100644
--- a/deps/v8/src/inspector/v8-string-conversions.cc
+++ b/deps/v8/src/inspector/v8-string-conversions.cc
@@ -12,7 +12,7 @@
 
 namespace v8_inspector {
 namespace {
-using UChar = uint16_t;
+using UChar = char16_t;
 using UChar32 = uint32_t;
 
 bool isASCII(UChar c) { return !(c & ~0x7F); }
@@ -386,7 +386,7 @@ std::string UTF16ToUTF8(const UChar* stringStart, size_t length) {
 
 std::basic_string<UChar> UTF8ToUTF16(const char* stringStart, size_t length) {
   if (!stringStart || !length) return std::basic_string<UChar>();
-  std::vector<uint16_t> buffer(length);
+  std::vector<UChar> buffer(length);
   UChar* bufferStart = buffer.data();
 
   UChar* bufferCurrent = bufferStart;
@@ -395,7 +395,7 @@ std::basic_string<UChar> UTF8ToUTF16(const char* stringStart, size_t length) {
                          reinterpret_cast<const char*>(stringStart + length),
                          &bufferCurrent, bufferCurrent + buffer.size(), nullptr,
                          true) != conversionOK)
-    return std::basic_string<uint16_t>();
+    return std::basic_string<UChar>();
   size_t utf16Length = bufferCurrent - bufferStart;
   return std::basic_string<UChar>(bufferStart, bufferStart + utf16Length);
 }
diff --git a/deps/v8/src/inspector/v8-string-conversions.h b/deps/v8/src/inspector/v8-string-conversions.h
index eb33c6816a5882..0f674f0bc81947 100644
--- a/deps/v8/src/inspector/v8-string-conversions.h
+++ b/deps/v8/src/inspector/v8-string-conversions.h
@@ -5,14 +5,15 @@
 #ifndef V8_INSPECTOR_V8_STRING_CONVERSIONS_H_
 #define V8_INSPECTOR_V8_STRING_CONVERSIONS_H_
 
+
 #include <cstdint>
 #include <string>
 
 // Conversion routines between UT8 and UTF16, used by string-16.{h,cc}. You may
 // want to use string-16.h directly rather than these.
 namespace v8_inspector {
-std::basic_string<uint16_t> UTF8ToUTF16(const char* stringStart, size_t length);
-std::string UTF16ToUTF8(const uint16_t* stringStart, size_t length);
+std::basic_string<char16_t> UTF8ToUTF16(const char* stringStart, size_t length);
+std::string UTF16ToUTF8(const char16_t* stringStart, size_t length);
 }  // namespace v8_inspector
 
 #endif  // V8_INSPECTOR_V8_STRING_CONVERSIONS_H_
diff --git a/deps/v8/third_party/inspector_protocol/crdtp/test_platform_v8.cc b/deps/v8/third_party/inspector_protocol/crdtp/test_platform_v8.cc
index c9d89eaa42f1d8..1a46d781b89d47 100644
--- a/deps/v8/third_party/inspector_protocol/crdtp/test_platform_v8.cc
+++ b/deps/v8/third_party/inspector_protocol/crdtp/test_platform_v8.cc
@@ -11,13 +11,16 @@
 namespace v8_crdtp {
 
 std::string UTF16ToUTF8(span<uint16_t> in) {
-  return v8_inspector::UTF16ToUTF8(in.data(), in.size());
+  return v8_inspector::UTF16ToUTF8(reinterpret_cast<const char16_t*>(in.data()),
+                                   in.size());
 }
 
 std::vector<uint16_t> UTF8ToUTF16(span<uint8_t> in) {
-  std::basic_string<uint16_t> utf16 = v8_inspector::UTF8ToUTF16(
+  std::basic_string<char16_t> utf16 = v8_inspector::UTF8ToUTF16(
       reinterpret_cast<const char*>(in.data()), in.size());
-  return std::vector<uint16_t>(utf16.begin(), utf16.end());
+  return std::vector<uint16_t>(
+      reinterpret_cast<const uint16_t*>(utf16.data()),
+      reinterpret_cast<const uint16_t*>(utf16.data()) + utf16.size());
 }
 
 }  // namespace v8_crdtp
diff --git a/deps/v8/third_party/zlib/zutil.h b/deps/v8/third_party/zlib/zutil.h
index e0466922244edc..a3f671c94f8fd2 100644
--- a/deps/v8/third_party/zlib/zutil.h
+++ b/deps/v8/third_party/zlib/zutil.h
@@ -152,17 +152,8 @@ extern z_const char * const z_errmsg[10]; /* indexed by 2-zlib_error */
 #  endif
 #endif
 
-#if defined(MACOS) || defined(TARGET_OS_MAC)
+#if defined(MACOS)
 #  define OS_CODE  7
-#  ifndef Z_SOLO
-#    if defined(__MWERKS__) && __dest_os != __be_os && __dest_os != __win32_os
-#      include <unix.h> /* for fdopen */
-#    else
-#      ifndef fdopen
-#        define fdopen(fd,mode) NULL /* No fdopen() */
-#      endif
-#    endif
-#  endif
 #endif
 
 #ifdef __acorn
@@ -185,18 +176,6 @@ extern z_const char * const z_errmsg[10]; /* indexed by 2-zlib_error */
 #  define OS_CODE 19
 #endif
 
-#if defined(_BEOS_) || defined(RISCOS)
-#  define fdopen(fd,mode) NULL /* No fdopen() */
-#endif
-
-#if (defined(_MSC_VER) && (_MSC_VER > 600)) && !defined __INTERIX
-#  if defined(_WIN32_WCE)
-#    define fdopen(fd,mode) NULL /* No fdopen() */
-#  else
-#    define fdopen(fd,type)  _fdopen(fd,type)
-#  endif
-#endif
-
 #if defined(__BORLANDC__) && !defined(MSDOS)
   #pragma warn -8004
   #pragma warn -8008
diff --git a/doc/api/async_context.md b/doc/api/async_context.md
index 6e983a1986772f..6ed6a970dd70af 100644
--- a/doc/api/async_context.md
+++ b/doc/api/async_context.md
@@ -75,8 +75,8 @@ http.get('http://localhost:8080');
 http.get('http://localhost:8080');
 // Prints:
 //   0: start
-//   1: start
 //   0: finish
+//   1: start
 //   1: finish
 ```
 
@@ -107,8 +107,8 @@ http.get('http://localhost:8080');
 http.get('http://localhost:8080');
 // Prints:
 //   0: start
-//   1: start
 //   0: finish
+//   1: start
 //   1: finish
 ```
 
diff --git a/doc/api/cli.md b/doc/api/cli.md
index cefda053007e1c..d39b090ba6cc29 100644
--- a/doc/api/cli.md
+++ b/doc/api/cli.md
@@ -2003,6 +2003,11 @@ this is not set and the network interfaces are included.
 
 <!-- YAML
 added: v1.6.0
+changes:
+  - version:
+      - v20.19.0
+    pr-url: https://github.com/nodejs/node/pull/51977
+    description: This option also supports ECMAScript module.
 -->
 
 Preload the specified module at startup.
@@ -2010,8 +2015,6 @@ Preload the specified module at startup.
 Follows `require()`'s module resolution
 rules. `module` may be either a path to a file, or a node module name.
 
-Only CommonJS modules are supported.
-Use [`--import`][] to preload an [ECMAScript module][].
 Modules preloaded with `--require` will run before modules preloaded with `--import`.
 
 ### `--secure-heap=n`
diff --git a/doc/api/deprecations.md b/doc/api/deprecations.md
index c04c15e013e7d1..21f2ba7d5e1a32 100644
--- a/doc/api/deprecations.md
+++ b/doc/api/deprecations.md
@@ -2886,11 +2886,11 @@ changes:
     description: Documentation-only deprecation.
 -->
 
-Type: Runtime
+Type: End-of-Life
 
 Using a trailing `"/"` to define subpath folder mappings in the
-[subpath exports][] or [subpath imports][] fields is deprecated. Use
-[subpath patterns][] instead.
+[subpath exports][] or [subpath imports][] fields is no longer supported.
+Use [subpath patterns][] instead.
 
 ### DEP0149: `http.IncomingMessage#connection`
 
@@ -3377,10 +3377,10 @@ be added when a function is bound to an `AsyncResource`.
 changes:
   - version: v20.1.0
     pr-url: https://github.com/nodejs/node/pull/47740
-    description: Documentation-only deprecation.
+    description: Runtime deprecation.
 -->
 
-Type: Documentation-only
+Type: Runtime
 
 In a future version of Node.js, [`assert.CallTracker`][],
 will be removed.
@@ -3573,6 +3573,22 @@ Instantiating classes without the `new` qualifier exported by the `node:repl` mo
 It is recommended to use the `new` qualifier instead. This applies to all REPL classes, including
 `REPLServer` and `Recoverable`.
 
+<!-- md-lint skip-deprecation DEP0186 -->
+
+### DEP0187: Passing invalid argument types to `fs.existsSync`
+
+<!-- YAML
+changes:
+  - version: v20.19.3
+    pr-url: https://github.com/nodejs/node/pull/55892
+    description: Documentation-only.
+-->
+
+Type: Documentation-only
+
+Passing non-supported argument types is deprecated and, instead of returning `false`,
+will throw an error in a future version.
+
 [NIST SP 800-38D]: https://nvlpubs.nist.gov/nistpubs/Legacy/SP/nistspecialpublication800-38d.pdf
 [RFC 6066]: https://tools.ietf.org/html/rfc6066#section-3
 [RFC 8247 Section 2.4]: https://www.rfc-editor.org/rfc/rfc8247#section-2.4
diff --git a/doc/api/fs.md b/doc/api/fs.md
index d2ee714c0460a1..889154f9bc66fd 100644
--- a/doc/api/fs.md
+++ b/doc/api/fs.md
@@ -721,7 +721,7 @@ added:
 * `options` {Object}
   * `offset` {integer} **Default:** `0`
   * `length` {integer} **Default:** `buffer.byteLength - offset`
-  * `position` {integer} **Default:** `null`
+  * `position` {integer|null} **Default:** `null`
 * Returns: {Promise}
 
 Write `buffer` to the file.
@@ -1048,6 +1048,8 @@ behavior is similar to `cp dir1/ dir2/`.
 deprecated: v10.0.0
 -->
 
+> Stability: 0 - Deprecated
+
 * `path` {string|Buffer|URL}
 * `mode` {integer}
 * Returns: {Promise} Fulfills with `undefined` upon success.
@@ -3059,6 +3061,8 @@ changes:
                  it will emit a deprecation warning with id DEP0013.
 -->
 
+> Stability: 0 - Deprecated
+
 * `path` {string|Buffer|URL}
 * `mode` {integer}
 * `callback` {Function}
@@ -4836,7 +4840,7 @@ added:
 * `options` {Object}
   * `offset` {integer} **Default:** `0`
   * `length` {integer} **Default:** `buffer.byteLength - offset`
-  * `position` {integer} **Default:** `null`
+  * `position` {integer|null} **Default:** `null`
 * `callback` {Function}
   * `err` {Error}
   * `bytesWritten` {integer}
@@ -5520,6 +5524,8 @@ Synchronous version of [`fs.futimes()`][]. Returns `undefined`.
 deprecated: v0.4.7
 -->
 
+> Stability: 0 - Deprecated
+
 * `path` {string|Buffer|URL}
 * `mode` {integer}
 
@@ -6313,7 +6319,7 @@ added:
 * `options` {Object}
   * `offset` {integer} **Default:** `0`
   * `length` {integer} **Default:** `buffer.byteLength - offset`
-  * `position` {integer} **Default:** `null`
+  * `position` {integer|null} **Default:** `null`
 * Returns: {number} The number of bytes written.
 
 For detailed information, see the documentation of the asynchronous version of
diff --git a/doc/api/globals.md b/doc/api/globals.md
index d25b50d2b12942..558e008f05564a 100644
--- a/doc/api/globals.md
+++ b/doc/api/globals.md
@@ -498,6 +498,46 @@ changes:
 
 A browser-compatible implementation of the [`fetch()`][] function.
 
+```mjs
+const res = await fetch('https://nodejs.org/api/documentation.json');
+if (res.ok) {
+  const data = await res.json();
+  console.log(data);
+}
+```
+
+The implementation is based upon [undici](https://undici.nodejs.org), an HTTP/1.1 client
+written from scratch for Node.js. You can figure out which version of `undici` is bundled
+in your Node.js process reading the `process.versions.undici` property.
+
+## Custom dispatcher
+
+You can use a custom dispatcher to dispatch requests passing it in fetch's options object.
+The dispatcher must be compatible with `undici`'s
+[`Dispatcher` class](https://undici.nodejs.org/#/docs/api/Dispatcher.md).
+
+```js
+fetch(url, { dispatcher: new MyAgent() });
+```
+
+It is possible to change the global dispatcher in Node.js installing `undici` and using
+the `setGlobalDispatcher()` method. Calling this method will affect both `undici` and
+Node.js.
+
+```mjs
+import { setGlobalDispatcher } from 'undici';
+setGlobalDispatcher(new MyAgent());
+```
+
+## Related classes
+
+The following globals are available to use with `fetch`:
+
+* [`FormData`](https://nodejs.org/api/globals.html#class-formdata)
+* [`Headers`](https://nodejs.org/api/globals.html#class-headers)
+* [`Request`](https://nodejs.org/api/globals.html#request)
+* [`Response`](https://nodejs.org/api/globals.html#response).
+
 ## Class: `File`
 
 <!-- YAML
diff --git a/doc/api/https.md b/doc/api/https.md
index a168addacf8e64..dcd36096ea4393 100644
--- a/doc/api/https.md
+++ b/doc/api/https.md
@@ -341,6 +341,7 @@ changes:
 * `options` {Object | string | URL} Accepts the same `options` as
   [`https.request()`][], with the method set to GET by default.
 * `callback` {Function}
+* Returns: {http.ClientRequest}
 
 Like [`http.get()`][] but for HTTPS.
 
diff --git a/doc/api/process.md b/doc/api/process.md
index b1542e0be1e23c..9f72b88bf703d9 100644
--- a/doc/api/process.md
+++ b/doc/api/process.md
@@ -578,6 +578,10 @@ address such failures, a non-operational
 `resource.loaded`, which would prevent the `'unhandledRejection'` event from
 being emitted.
 
+If an `'unhandledRejection'` event is emitted but not handled it will
+be raised as an uncaught exception. This alongside other behaviors of
+`'unhandledRejection'` events can changed via the [`--unhandled-rejections`][] flag.
+
 ### Event: `'warning'`
 
 <!-- YAML
diff --git a/doc/api/readline.md b/doc/api/readline.md
index cb1aa52605dc4a..352857e90a1646 100644
--- a/doc/api/readline.md
+++ b/doc/api/readline.md
@@ -698,6 +698,7 @@ added: v17.0.0
     **Default:** `500`.
   * `tabSize` {integer} The number of spaces a tab is equal to (minimum 1).
     **Default:** `8`.
+  * `signal` {AbortSignal} Allows closing the interface using an AbortSignal.
 * Returns: {readlinePromises.Interface}
 
 The `readlinePromises.createInterface()` method creates a new `readlinePromises.Interface`
diff --git a/doc/api/repl.md b/doc/api/repl.md
index a134c493a54812..00c95dac4165a6 100644
--- a/doc/api/repl.md
+++ b/doc/api/repl.md
@@ -303,7 +303,19 @@ When a new [`repl.REPLServer`][] is created, a custom evaluation function may be
 provided. This can be used, for instance, to implement fully customized REPL
 applications.
 
-The following illustrates an example of a REPL that squares a given number:
+An evaluation function accepts the following four arguments:
+
+* `code` {string} The code to be executed (e.g. `1 + 1`).
+* `context` {Object} The context in which the code is executed. This can either be the JavaScript `global`
+  context or a context specific to the REPL instance, depending on the `useGlobal` option.
+* `replResourceName` {string} An identifier for the REPL resource associated with the current code
+  evaluation. This can be useful for debugging purposes.
+* `callback` {Function} A function to invoke once the code evaluation is complete. The callback takes two parameters:
+  * An error object to provide if an error occurred during evaluation, or `null`/`undefined` if no error occurred.
+  * The result of the code evaluation (this is not relevant if an error is provided).
+
+The following illustrates an example of a REPL that squares a given number, an error is instead printed
+if the provided input is not actually a number:
 
 ```mjs
 import repl from 'node:repl';
@@ -312,8 +324,12 @@ function byThePowerOfTwo(number) {
   return number * number;
 }
 
-function myEval(cmd, context, filename, callback) {
-  callback(null, byThePowerOfTwo(cmd));
+function myEval(code, context, replResourceName, callback) {
+  if (Number.isNaN(code)) {
+    callback(new Error(`${code.trim()} is not a number`));
+  } else {
+    callback(null, byThePowerOfTwo(code));
+  }
 }
 
 repl.start({ prompt: 'Enter a number: ', eval: myEval });
@@ -326,8 +342,12 @@ function byThePowerOfTwo(number) {
   return number * number;
 }
 
-function myEval(cmd, context, filename, callback) {
-  callback(null, byThePowerOfTwo(cmd));
+function myEval(code, context, replResourceName, callback) {
+  if (Number.isNaN(code)) {
+    callback(new Error(`${code.trim()} is not a number`));
+  } else {
+    callback(null, byThePowerOfTwo(code));
+  }
 }
 
 repl.start({ prompt: 'Enter a number: ', eval: myEval });
@@ -691,7 +711,8 @@ changes:
   * `eval` {Function} The function to be used when evaluating each given line
     of input. **Default:** an async wrapper for the JavaScript `eval()`
     function. An `eval` function can error with `repl.Recoverable` to indicate
-    the input was incomplete and prompt for additional lines.
+    the input was incomplete and prompt for additional lines. See the
+    [custom evaluation functions][] section for more details.
   * `useColors` {boolean} If `true`, specifies that the default `writer`
     function should include ANSI color styling to REPL output. If a custom
     `writer` function is provided then this has no effect. **Default:** checking
@@ -914,4 +935,5 @@ avoiding open network interfaces.
 [`repl.start()`]: #replstartoptions
 [`reverse-i-search`]: #reverse-i-search
 [`util.inspect()`]: util.md#utilinspectobject-options
+[custom evaluation functions]: #custom-evaluation-functions
 [stream]: stream.md
diff --git a/doc/api/util.md b/doc/api/util.md
index 61b4e68aef9db2..d57230966cbf4b 100644
--- a/doc/api/util.md
+++ b/doc/api/util.md
@@ -1809,7 +1809,7 @@ changes:
   - version: v20.18.0
     pr-url: https://github.com/nodejs/node/pull/54389
     description: Respect isTTY and environment variables
-      such as NO_COLORS, NODE_DISABLE_COLORS, and FORCE_COLOR.
+      such as NO_COLOR, NODE_DISABLE_COLORS, and FORCE_COLOR.
 -->
 
 * `format` {string | Array} A text format or an Array
@@ -1821,7 +1821,7 @@ changes:
 
 This function returns a formatted text considering the `format` passed
 for printing in a terminal. It is aware of the terminal's capabilities
-and acts according to the configuration set via `NO_COLORS`,
+and acts according to the configuration set via `NO_COLOR`,
 `NODE_DISABLE_COLORS` and `FORCE_COLOR` environment variables.
 
 ```mjs
diff --git a/doc/api/vm.md b/doc/api/vm.md
index a82df8a95ac825..a8a14732bdf855 100644
--- a/doc/api/vm.md
+++ b/doc/api/vm.md
@@ -1037,13 +1037,13 @@ changes:
   * `contextExtensions` {Object\[]} An array containing a collection of context
     extensions (objects wrapping the current scope) to be applied while
     compiling. **Default:** `[]`.
-* `importModuleDynamically`
-  {Function|vm.constants.USE\_MAIN\_CONTEXT\_DEFAULT\_LOADER}
-  Used to specify the how the modules should be loaded during the evaluation of
-  this function when `import()` is called. This option is part of the
-  experimental modules API. We do not recommend using it in a production
-  environment. For detailed information, see
-  [Support of dynamic `import()` in compilation APIs][].
+  * `importModuleDynamically`
+    {Function|vm.constants.USE\_MAIN\_CONTEXT\_DEFAULT\_LOADER}
+    Used to specify the how the modules should be loaded during the evaluation of
+    this function when `import()` is called. This option is part of the
+    experimental modules API. We do not recommend using it in a production
+    environment. For detailed information, see
+    [Support of dynamic `import()` in compilation APIs][].
 * Returns: {Function}
 
 Compiles the given code into the provided context (if no context is
diff --git a/doc/api/webcrypto.md b/doc/api/webcrypto.md
index fa7e248c982efa..e0425bc7782866 100644
--- a/doc/api/webcrypto.md
+++ b/doc/api/webcrypto.md
@@ -2,7 +2,12 @@
 
 <!-- YAML
 changes:
-  - version: v20.0.0
+  - version: v20.19.3
+    pr-url: https://github.com/nodejs/node/pull/56142
+    description: Algorithms `Ed25519` and `X25519` are now stable.
+  - version:
+    - v20.0.0
+    - v18.17.0
     pr-url: https://github.com/nodejs/node/pull/46067
     description: Arguments are now coerced and validated as per their WebIDL
       definitions like in other Web Crypto API implementations.
@@ -111,9 +116,7 @@ async function generateEcKey(namedCurve = 'P-521') {
 }
 ```
 
-#### Ed25519/Ed448/X25519/X448 key pairs
-
-> Stability: 1 - Experimental
+#### Ed25519/X25519 key pairs
 
 ```js
 const { subtle } = globalThis.crypto;
@@ -351,28 +354,28 @@ async function digest(data, algorithm = 'SHA-512') {
 The table details the algorithms supported by the Node.js Web Crypto API
 implementation and the APIs supported for each:
 
-| Algorithm                                                 | `generateKey` | `exportKey` | `importKey` | `encrypt` | `decrypt` | `wrapKey` | `unwrapKey` | `deriveBits` | `deriveKey` | `sign` | `verify` | `digest` |
-| --------------------------------------------------------- | ------------- | ----------- | ----------- | --------- | --------- | --------- | ----------- | ------------ | ----------- | ------ | -------- | -------- |
-| `'RSASSA-PKCS1-v1_5'`                                     | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'RSA-PSS'`                                               | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'RSA-OAEP'`                                              | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
-| `'ECDSA'`                                                 | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'Ed25519'` <span class="experimental-inline"></span>[^1] | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'Ed448'` <span class="experimental-inline"></span>[^1]   | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'ECDH'`                                                  | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
-| `'X25519'` <span class="experimental-inline"></span>[^1]  | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
-| `'X448'` <span class="experimental-inline"></span>[^1]    | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
-| `'AES-CTR'`                                               | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
-| `'AES-CBC'`                                               | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
-| `'AES-GCM'`                                               | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
-| `'AES-KW'`                                                | ✔             | ✔           | ✔           |           |           | ✔         | ✔           |              |             |        |          |          |
-| `'HMAC'`                                                  | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
-| `'HKDF'`                                                  |               | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
-| `'PBKDF2'`                                                |               | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
-| `'SHA-1'`                                                 |               |             |             |           |           |           |             |              |             |        |          | ✔        |
-| `'SHA-256'`                                               |               |             |             |           |           |           |             |              |             |        |          | ✔        |
-| `'SHA-384'`                                               |               |             |             |           |           |           |             |              |             |        |          | ✔        |
-| `'SHA-512'`                                               |               |             |             |           |           |           |             |              |             |        |          | ✔        |
+| Algorithm                                               | `generateKey` | `exportKey` | `importKey` | `encrypt` | `decrypt` | `wrapKey` | `unwrapKey` | `deriveBits` | `deriveKey` | `sign` | `verify` | `digest` |
+| ------------------------------------------------------- | ------------- | ----------- | ----------- | --------- | --------- | --------- | ----------- | ------------ | ----------- | ------ | -------- | -------- |
+| `'RSASSA-PKCS1-v1_5'`                                   | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'RSA-PSS'`                                             | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'RSA-OAEP'`                                            | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
+| `'ECDSA'`                                               | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'Ed25519'`                                             | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'Ed448'` <span class="experimental-inline"></span>[^1] | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'ECDH'`                                                | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
+| `'X25519'`                                              | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
+| `'X448'` <span class="experimental-inline"></span>[^1]  | ✔             | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
+| `'AES-CTR'`                                             | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
+| `'AES-CBC'`                                             | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
+| `'AES-GCM'`                                             | ✔             | ✔           | ✔           | ✔         | ✔         | ✔         | ✔           |              |             |        |          |          |
+| `'AES-KW'`                                              | ✔             | ✔           | ✔           |           |           | ✔         | ✔           |              |             |        |          |          |
+| `'HMAC'`                                                | ✔             | ✔           | ✔           |           |           |           |             |              |             | ✔      | ✔        |          |
+| `'HKDF'`                                                |               | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
+| `'PBKDF2'`                                              |               | ✔           | ✔           |           |           |           |             | ✔            | ✔           |        |          |          |
+| `'SHA-1'`                                               |               |             |             |           |           |           |             |              |             |        |          | ✔        |
+| `'SHA-256'`                                             |               |             |             |           |           |           |             |              |             |        |          | ✔        |
+| `'SHA-384'`                                             |               |             |             |           |           |           |             |              |             |        |          | ✔        |
+| `'SHA-512'`                                             |               |             |             |           |           |           |             |              |             |        |          | ✔        |
 
 ## Class: `Crypto`
 
@@ -494,24 +497,24 @@ The possible usages are:
 Valid key usages depend on the key algorithm (identified by
 `cryptokey.algorithm.name`).
 
-| Key Type                                                  | `'encrypt'` | `'decrypt'` | `'sign'` | `'verify'` | `'deriveKey'` | `'deriveBits'` | `'wrapKey'` | `'unwrapKey'` |
-| --------------------------------------------------------- | ----------- | ----------- | -------- | ---------- | ------------- | -------------- | ----------- | ------------- |
-| `'AES-CBC'`                                               | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
-| `'AES-CTR'`                                               | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
-| `'AES-GCM'`                                               | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
-| `'AES-KW'`                                                |             |             |          |            |               |                | ✔           | ✔             |
-| `'ECDH'`                                                  |             |             |          |            | ✔             | ✔              |             |               |
-| `'X25519'` <span class="experimental-inline"></span>[^1]  |             |             |          |            | ✔             | ✔              |             |               |
-| `'X448'` <span class="experimental-inline"></span>[^1]    |             |             |          |            | ✔             | ✔              |             |               |
-| `'ECDSA'`                                                 |             |             | ✔        | ✔          |               |                |             |               |
-| `'Ed25519'` <span class="experimental-inline"></span>[^1] |             |             | ✔        | ✔          |               |                |             |               |
-| `'Ed448'` <span class="experimental-inline"></span>[^1]   |             |             | ✔        | ✔          |               |                |             |               |
-| `'HDKF'`                                                  |             |             |          |            | ✔             | ✔              |             |               |
-| `'HMAC'`                                                  |             |             | ✔        | ✔          |               |                |             |               |
-| `'PBKDF2'`                                                |             |             |          |            | ✔             | ✔              |             |               |
-| `'RSA-OAEP'`                                              | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
-| `'RSA-PSS'`                                               |             |             | ✔        | ✔          |               |                |             |               |
-| `'RSASSA-PKCS1-v1_5'`                                     |             |             | ✔        | ✔          |               |                |             |               |
+| Key Type                                                | `'encrypt'` | `'decrypt'` | `'sign'` | `'verify'` | `'deriveKey'` | `'deriveBits'` | `'wrapKey'` | `'unwrapKey'` |
+| ------------------------------------------------------- | ----------- | ----------- | -------- | ---------- | ------------- | -------------- | ----------- | ------------- |
+| `'AES-CBC'`                                             | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
+| `'AES-CTR'`                                             | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
+| `'AES-GCM'`                                             | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
+| `'AES-KW'`                                              |             |             |          |            |               |                | ✔           | ✔             |
+| `'ECDH'`                                                |             |             |          |            | ✔             | ✔              |             |               |
+| `'X25519'`                                              |             |             |          |            | ✔             | ✔              |             |               |
+| `'X448'` <span class="experimental-inline"></span>[^1]  |             |             |          |            | ✔             | ✔              |             |               |
+| `'ECDSA'`                                               |             |             | ✔        | ✔          |               |                |             |               |
+| `'Ed25519'`                                             |             |             | ✔        | ✔          |               |                |             |               |
+| `'Ed448'` <span class="experimental-inline"></span>[^1] |             |             | ✔        | ✔          |               |                |             |               |
+| `'HDKF'`                                                |             |             |          |            | ✔             | ✔              |             |               |
+| `'HMAC'`                                                |             |             | ✔        | ✔          |               |                |             |               |
+| `'PBKDF2'`                                              |             |             |          |            | ✔             | ✔              |             |               |
+| `'RSA-OAEP'`                                            | ✔           | ✔           |          |            |               |                | ✔           | ✔             |
+| `'RSA-PSS'`                                             |             |             | ✔        | ✔          |               |                |             |               |
+| `'RSASSA-PKCS1-v1_5'`                                   |             |             | ✔        | ✔          |               |                |             |               |
 
 ## Class: `CryptoKeyPair`
 
@@ -596,9 +599,6 @@ Using the method and parameters specified in `algorithm` and the keying
 material provided by `baseKey`, `subtle.deriveBits()` attempts to generate
 `length` bits.
 
-The Node.js implementation requires that `length`, when a number, is a multiple
-of `8`.
-
 When `length` is not provided or `null` the maximum number of bits for a given
 algorithm is generated. This is allowed for the `'ECDH'`, `'X25519'`, and `'X448'`
 algorithms, for other algorithms `length` is required to be a number.
@@ -609,7 +609,7 @@ containing the generated data.
 The algorithms currently supported include:
 
 * `'ECDH'`
-* `'X25519'` <span class="experimental-inline"></span>[^1]
+* `'X25519'`
 * `'X448'` <span class="experimental-inline"></span>[^1]
 * `'HKDF'`
 * `'PBKDF2'`
@@ -649,7 +649,7 @@ generate raw keying material, then passing the result into the
 The algorithms currently supported include:
 
 * `'ECDH'`
-* `'X25519'` <span class="experimental-inline"></span>[^1]
+* `'X25519'`
 * `'X448'` <span class="experimental-inline"></span>[^1]
 * `'HKDF'`
 * `'PBKDF2'`
@@ -733,22 +733,22 @@ When `format` is `'jwk'` and the export is successful, the returned promise
 will be resolved with a JavaScript object conforming to the [JSON Web Key][]
 specification.
 
-| Key Type                                                  | `'spki'` | `'pkcs8'` | `'jwk'` | `'raw'` |
-| --------------------------------------------------------- | -------- | --------- | ------- | ------- |
-| `'AES-CBC'`                                               |          |           | ✔       | ✔       |
-| `'AES-CTR'`                                               |          |           | ✔       | ✔       |
-| `'AES-GCM'`                                               |          |           | ✔       | ✔       |
-| `'AES-KW'`                                                |          |           | ✔       | ✔       |
-| `'ECDH'`                                                  | ✔        | ✔         | ✔       | ✔       |
-| `'ECDSA'`                                                 | ✔        | ✔         | ✔       | ✔       |
-| `'Ed25519'` <span class="experimental-inline"></span>[^1] | ✔        | ✔         | ✔       | ✔       |
-| `'Ed448'` <span class="experimental-inline"></span>[^1]   | ✔        | ✔         | ✔       | ✔       |
-| `'HDKF'`                                                  |          |           |         |         |
-| `'HMAC'`                                                  |          |           | ✔       | ✔       |
-| `'PBKDF2'`                                                |          |           |         |         |
-| `'RSA-OAEP'`                                              | ✔        | ✔         | ✔       |         |
-| `'RSA-PSS'`                                               | ✔        | ✔         | ✔       |         |
-| `'RSASSA-PKCS1-v1_5'`                                     | ✔        | ✔         | ✔       |         |
+| Key Type                                                | `'spki'` | `'pkcs8'` | `'jwk'` | `'raw'` |
+| ------------------------------------------------------- | -------- | --------- | ------- | ------- |
+| `'AES-CBC'`                                             |          |           | ✔       | ✔       |
+| `'AES-CTR'`                                             |          |           | ✔       | ✔       |
+| `'AES-GCM'`                                             |          |           | ✔       | ✔       |
+| `'AES-KW'`                                              |          |           | ✔       | ✔       |
+| `'ECDH'`                                                | ✔        | ✔         | ✔       | ✔       |
+| `'ECDSA'`                                               | ✔        | ✔         | ✔       | ✔       |
+| `'Ed25519'`                                             | ✔        | ✔         | ✔       | ✔       |
+| `'Ed448'` <span class="experimental-inline"></span>[^1] | ✔        | ✔         | ✔       | ✔       |
+| `'HDKF'`                                                |          |           |         |         |
+| `'HMAC'`                                                |          |           | ✔       | ✔       |
+| `'PBKDF2'`                                              |          |           |         |         |
+| `'RSA-OAEP'`                                            | ✔        | ✔         | ✔       |         |
+| `'RSA-PSS'`                                             | ✔        | ✔         | ✔       |         |
+| `'RSASSA-PKCS1-v1_5'`                                   | ✔        | ✔         | ✔       |         |
 
 ### `subtle.generateKey(algorithm, extractable, keyUsages)`
 
@@ -777,10 +777,10 @@ include:
 * `'RSA-PSS'`
 * `'RSA-OAEP'`
 * `'ECDSA'`
-* `'Ed25519'` <span class="experimental-inline"></span>[^1]
+* `'Ed25519'`
 * `'Ed448'` <span class="experimental-inline"></span>[^1]
 * `'ECDH'`
-* `'X25519'` <span class="experimental-inline"></span>[^1]
+* `'X25519'`
 * `'X448'` <span class="experimental-inline"></span>[^1]
 
 The {CryptoKey} (secret key) generating algorithms supported include:
@@ -829,24 +829,24 @@ If importing a `'PBKDF2'` key, `extractable` must be `false`.
 
 The algorithms currently supported include:
 
-| Key Type                                                  | `'spki'` | `'pkcs8'` | `'jwk'` | `'raw'` |
-| --------------------------------------------------------- | -------- | --------- | ------- | ------- |
-| `'AES-CBC'`                                               |          |           | ✔       | ✔       |
-| `'AES-CTR'`                                               |          |           | ✔       | ✔       |
-| `'AES-GCM'`                                               |          |           | ✔       | ✔       |
-| `'AES-KW'`                                                |          |           | ✔       | ✔       |
-| `'ECDH'`                                                  | ✔        | ✔         | ✔       | ✔       |
-| `'X25519'` <span class="experimental-inline"></span>[^1]  | ✔        | ✔         | ✔       | ✔       |
-| `'X448'` <span class="experimental-inline"></span>[^1]    | ✔        | ✔         | ✔       | ✔       |
-| `'ECDSA'`                                                 | ✔        | ✔         | ✔       | ✔       |
-| `'Ed25519'` <span class="experimental-inline"></span>[^1] | ✔        | ✔         | ✔       | ✔       |
-| `'Ed448'` <span class="experimental-inline"></span>[^1]   | ✔        | ✔         | ✔       | ✔       |
-| `'HDKF'`                                                  |          |           |         | ✔       |
-| `'HMAC'`                                                  |          |           | ✔       | ✔       |
-| `'PBKDF2'`                                                |          |           |         | ✔       |
-| `'RSA-OAEP'`                                              | ✔        | ✔         | ✔       |         |
-| `'RSA-PSS'`                                               | ✔        | ✔         | ✔       |         |
-| `'RSASSA-PKCS1-v1_5'`                                     | ✔        | ✔         | ✔       |         |
+| Key Type                                                | `'spki'` | `'pkcs8'` | `'jwk'` | `'raw'` |
+| ------------------------------------------------------- | -------- | --------- | ------- | ------- |
+| `'AES-CBC'`                                             |          |           | ✔       | ✔       |
+| `'AES-CTR'`                                             |          |           | ✔       | ✔       |
+| `'AES-GCM'`                                             |          |           | ✔       | ✔       |
+| `'AES-KW'`                                              |          |           | ✔       | ✔       |
+| `'ECDH'`                                                | ✔        | ✔         | ✔       | ✔       |
+| `'X25519'`                                              | ✔        | ✔         | ✔       | ✔       |
+| `'X448'` <span class="experimental-inline"></span>[^1]  | ✔        | ✔         | ✔       | ✔       |
+| `'ECDSA'`                                               | ✔        | ✔         | ✔       | ✔       |
+| `'Ed25519'`                                             | ✔        | ✔         | ✔       | ✔       |
+| `'Ed448'` <span class="experimental-inline"></span>[^1] | ✔        | ✔         | ✔       | ✔       |
+| `'HDKF'`                                                |          |           |         | ✔       |
+| `'HMAC'`                                                |          |           | ✔       | ✔       |
+| `'PBKDF2'`                                              |          |           |         | ✔       |
+| `'RSA-OAEP'`                                            | ✔        | ✔         | ✔       |         |
+| `'RSA-PSS'`                                             | ✔        | ✔         | ✔       |         |
+| `'RSASSA-PKCS1-v1_5'`                                   | ✔        | ✔         | ✔       |         |
 
 ### `subtle.sign(algorithm, key, data)`
 
@@ -879,7 +879,7 @@ The algorithms currently supported include:
 * `'RSASSA-PKCS1-v1_5'`
 * `'RSA-PSS'`
 * `'ECDSA'`
-* `'Ed25519'` <span class="experimental-inline"></span>[^1]
+* `'Ed25519'`
 * `'Ed448'` <span class="experimental-inline"></span>[^1]
 * `'HMAC'`
 
@@ -927,10 +927,10 @@ The unwrapped key algorithms supported include:
 * `'RSA-PSS'`
 * `'RSA-OAEP'`
 * `'ECDSA'`
-* `'Ed25519'` <span class="experimental-inline"></span>[^1]
+* `'Ed25519'`
 * `'Ed448'` <span class="experimental-inline"></span>[^1]
 * `'ECDH'`
-* `'X25519'` <span class="experimental-inline"></span>[^1]
+* `'X25519'`
 * `'X448'` <span class="experimental-inline"></span>[^1]
 * `'HMAC'`
 * `'AES-CTR'`
@@ -970,7 +970,7 @@ The algorithms currently supported include:
 * `'RSASSA-PKCS1-v1_5'`
 * `'RSA-PSS'`
 * `'ECDSA'`
-* `'Ed25519'` <span class="experimental-inline"></span>[^1]
+* `'Ed25519'`
 * `'Ed448'` <span class="experimental-inline"></span>[^1]
 * `'HMAC'`
 
@@ -1643,8 +1643,8 @@ added: v15.0.0
 
 The length (in bytes) of the random salt to use.
 
-[^1]: An experimental implementation of
-    [Secure Curves in the Web Cryptography API][] as of 30 August 2023
+[^1]: An experimental implementation of Ed448 and X448 algorithms from
+    [Secure Curves in the Web Cryptography API][] as of 21 October 2024
 
 [JSON Web Key]: https://tools.ietf.org/html/rfc7517
 [Key usages]: #cryptokeyusages
diff --git a/doc/api/zlib.md b/doc/api/zlib.md
index 7fb73d387f5410..b4c763a130a372 100644
--- a/doc/api/zlib.md
+++ b/doc/api/zlib.md
@@ -12,7 +12,7 @@ Gzip, Deflate/Inflate, and Brotli.
 To access it:
 
 ```mjs
-import os from 'node:zlib';
+import zlib from 'node:zlib';
 ```
 
 ```cjs
diff --git a/doc/api_assets/style.css b/doc/api_assets/style.css
index 9f56028d09162c..1288b1a5956390 100644
--- a/doc/api_assets/style.css
+++ b/doc/api_assets/style.css
@@ -220,7 +220,6 @@ li.picker-header.expanded > .picker,
 :root:not(.has-js) li.picker-header:focus-within > .picker,
 :root:not(.has-js) li.picker-header:hover > .picker {
   display: block;
-  z-index: 1;
 }
 
 li.picker-header a span {
@@ -817,7 +816,7 @@ kbd {
 .header {
   position: sticky;
   top: -1px;
-  z-index: 1;
+  z-index: 10;
   padding-top: 1rem;
   background-color: var(--color-fill-app);
 }
diff --git a/doc/changelogs/CHANGELOG_ARCHIVE.md b/doc/changelogs/CHANGELOG_ARCHIVE.md
index 14fe32c7f5a48e..7b119389facd78 100644
--- a/doc/changelogs/CHANGELOG_ARCHIVE.md
+++ b/doc/changelogs/CHANGELOG_ARCHIVE.md
@@ -14,6 +14,8 @@
 </tr>
 <tr>
 <td valign="top">
+<a href="#0.11.16">0.11.16</a><br/>
+<a href="#0.11.15">0.11.15</a><br/>
 <a href="#0.11.14">0.11.14</a><br/>
 <a href="#0.11.13">0.11.13</a><br/>
 <a href="#0.11.12">0.11.12</a><br/>
@@ -175,6 +177,90 @@
   * [0.10.x](CHANGELOG_V010.md)
   * [io.js](CHANGELOG_IOJS.md)
 
+<a id="0.11.16"></a>
+
+## 2015.01.29, Version 0.11.16 (Unstable)
+
+* openssl: Upgrade to 1.0.1l
+* npm: Upgrade to 2.3.0
+* url: revert support of `path` for url.format" (Julien Gilli)
+* assert: use `util.inspect()` to create error messages (cjihrig)
+* net: throw on invalid socket timeouts (cjihrig)
+* url: fix parsing of ssh urls (Evan Lucas)
+
+<a id="0.11.15"></a>
+
+## 2015.01.20, Version 0.11.15 (Unstable)
+
+* v8: Upgrade to 3.28.73
+* uv: Upgrade to 1.0.2
+* npm: Upgrade to v2.1.6
+* uv: float patch to revert tty breakage (Trevor Norris)
+* v8: re-implement debugger-agent (Fedor Indutny)
+* v8: apply floating irhydra patch (Fedor Indutny)
+* v8: fix postmortem-metadata generator (Refael Ackermann)
+* debugger: fix unhandled error in setBreakpoint (Miroslav Bajtoš)
+* async-wrap: add event hooks (Trevor Norris)
+* async-wrap: expose async-wrap as binding (Trevor Norris)
+* buffer, doc: misc. fix and cleanup (Trevor Norris)
+* buffer: add generic functions for (u)int ops (Yazhong Liu)
+* buffer: fix and cleanup fill() (Trevor Norris)
+* buffer: mv floating point read/write checks to JS (Trevor Norris)
+* build, i18n: improve Intl build, add "--with-intl" (Steven R. Loomis)
+* build: add small-icu support for binary packages (Julien Gilli)
+* build: do not generate support for libuv's probes (Julien Gilli)
+* build: i18n: add icu config options (Steven R. Loomis)
+* build: i18n: support little-endian machines (Steven Loomis)
+* build: vcbuild fix "The input line is too long." (Alexis Campailla)
+* child\_process: improve `spawn()` argument handling (cjihrig)
+* cluster: avoid race enabling debugger in worker (Timothy J Fontaine)
+* cluster: `cluster.disconnect()` should check status (Sam Roberts)
+* cluster: do not signal children in debug mode (Fedor Indutny)
+* cluster: don't assert if worker has no handles (Sam Roberts)
+* core: fix usage of `uv_cwd` (Saúl Ibarra Corretgé)
+* core: replace `uv_fs_readdir` with `uv_fs_scandir` (Saúl Ibarra Corretgé)
+* crypto: `createDiffieHellman` throw for bad args (Trevor Norris)
+* crypto: lower RSS usage for TLSCallbacks (Fedor Indutny)
+* crypto: store thread id as pointer-sized (Alexis Campailla)
+* dns: propagate domain for c-ares methods (Chris Dickinson)
+* fs: fix symlink error message (Vladimir Kurchatkin)
+* http: Improve `_addHeaderLines` method (Jackson Tian)
+* http: cleanup `setHeader()` (Trevor Norris)
+* http: rename `flush` to `flushHeaders` (Timothy J Fontaine)
+* lib,src: fix `spawnSync` ignoring its 'env' option (Juanjo)
+* modules: adding load linked modules feature (Thorsten Lorenz)
+* net: Make `server.connections` un-enumerable (Patrick Mooney)
+* net: add `pauseOnConnect` option to `createServer()` (cjihrig)
+* net: make `connect()` input validation synchronous (cjihrig)
+* node: avoid automatic microtask runs (Vladimir Kurchatkin)
+* node: fix throws before timer module is loaded (Trevor Norris)
+* openssl: fix keypress requirement in apps on win32 (Fedor Indutny)
+* path: added `parse()` and `format()` functions (Rory Bradford)
+* path: allow calling platform specific methods (Timothy J Fontaine)
+* path: don't lower-cases drive letters (Bert Belder)
+* path: refactor `normalizeArray()` (Nathan Woltman)
+* process: pid can be a string in process.kill() (Sam Roberts)
+* readline: fix performance issue when large line (Jicheng Li)
+* readline: should not require an output stream. (Julien Gilli)
+* smalloc: check if obj has external data (Vladimir Kurchatkin)
+* smalloc: don't allow to dispose typed arrays (Vladimir Kurchatkin)
+* smalloc: fix bad assert for zero length data (Trevor Norris)
+* smalloc: fix copyOnto optimization (Vladimir Kurchatkin)
+* src: all wrap's now use actual FunctionTemplate (Trevor Norris)
+* src: fix VC++ warning C4244 (Rasmus Christian Pedersen)
+* src: remove Async Listener (Trevor Norris)
+* stream: switch `_writableState.buffer` to queue (Chris Dickinson)
+* streams: make setDefaultEncoding() throw (Brian White)
+* streams: set default encoding for writable streams (Johnny Ray)
+* tls: remove `tls.createSecurePair` code deprecation (Jackson Tian)
+* tls\_wrap: ignore `ZERO_RETURN` after `close_notify` (Fedor Indutny)
+* url: change hostname regex to negate invalid chars (Jonathan Johnson)
+* url: fixed encoding for slash switching emulation. (Evan Rutledge Borden)
+* url: improve parsing speed (CGavrila)
+* url: make query() consistent (Gabriel Wicke)
+* url: support `path` for `url.format` (Yazhong Liu)
+* util: add es6 Symbol support for `util.inspect` (gyson)
+
 <a id="0.11.14"></a>
 
 ## 2014.09.24, Version 0.11.14 (Unstable)
diff --git a/doc/changelogs/CHANGELOG_V012.md b/doc/changelogs/CHANGELOG_V012.md
index 00be56d2d14f3e..f0cd61f5056941 100644
--- a/doc/changelogs/CHANGELOG_V012.md
+++ b/doc/changelogs/CHANGELOG_V012.md
@@ -279,6 +279,8 @@ Security Update
 
 ## 2015.11.25, Version 0.12.8 (LTS), @rvagg
 
+### Commits
+
 * \[[`d9399569bd`](https://github.com/nodejs/node/commit/d9399569bd)] - build: backport tools/release.sh (Rod Vagg) <https://github.com/nodejs/node/pull/3642>
 * \[[`78c5b4c8bd`](https://github.com/nodejs/node/commit/78c5b4c8bd)] - build: backport config for new CI infrastructure (Rod Vagg) <https://github.com/nodejs/node/pull/3642>
 * \[[`83441616a5`](https://github.com/nodejs/node/commit/83441616a5)] - build: fix --without-ssl compile time error (Ben Noordhuis) <https://github.com/nodejs/node/pull/3825>
@@ -310,6 +312,10 @@ Security Update
 * \[[`881d9bea01`](https://github.com/nodejs/node/commit/881d9bea01)] - doc: readable event clarification (James M Snell) <https://github.com/nodejs/node-v0.x-archive/pull/25591>
 * \[[`b6378f0c75`](https://github.com/nodejs/node/commit/b6378f0c75)] - doc: stream.unshift does not reset reading state (James M Snell) <https://github.com/nodejs/node-v0.x-archive/pull/25591>
 
+<a id="0.12.7"></a>
+
+## 2015-07-09, Version 0.12.7 (Stable)
+
 ### Commits
 
 * \[[`0cf9f27703`](https://github.com/nodejs/node/commit/0cf9f27703)] - **deps**: upgrade openssl sources to 1.0.1p [#25654](https://github.com/joyent/node/pull/25654)
diff --git a/doc/changelogs/CHANGELOG_V20.md b/doc/changelogs/CHANGELOG_V20.md
index 4cdf93ecbf3a6f..736684726072ca 100644
--- a/doc/changelogs/CHANGELOG_V20.md
+++ b/doc/changelogs/CHANGELOG_V20.md
@@ -9,6 +9,7 @@
 </tr>
 <tr>
 <td>
+<a href="#20.19.3">20.19.3</a><br/>
 <a href="#20.19.2">20.19.2</a><br/>
 <a href="#20.19.1">20.19.1</a><br/>
 <a href="#20.19.0">20.19.0</a><br/>
@@ -71,6 +72,82 @@
   * [io.js](CHANGELOG_IOJS.md)
   * [Archive](CHANGELOG_ARCHIVE.md)
 
+<a id="20.19.3"></a>
+
+## 2025-06-23, Version 20.19.3 'Iron' (LTS), @marco-ippolito
+
+### Notable Changes
+
+* \[[`c535a3c483`](https://github.com/nodejs/node/commit/c535a3c483)] - **crypto**: graduate WebCryptoAPI Ed25519 and X25519 algorithms as stable (Filip Skokan) [#56142](https://github.com/nodejs/node/pull/56142)
+* \[[`af1dc63815`](https://github.com/nodejs/node/commit/af1dc63815)] - **crypto**: update root certificates to NSS 3.108 (Node.js GitHub Bot) [#57381](https://github.com/nodejs/node/pull/57381)
+* \[[`01d63a4ddf`](https://github.com/nodejs/node/commit/01d63a4ddf)] - **deps**: update timezone to 2025b (Node.js GitHub Bot) [#57857](https://github.com/nodejs/node/pull/57857)
+* \[[`b6daa344eb`](https://github.com/nodejs/node/commit/b6daa344eb)] - **doc**: add dario-piotrowicz to collaborators (Dario Piotrowicz) [#58102](https://github.com/nodejs/node/pull/58102)
+
+### Commits
+
+* \[[`fc1fa7a357`](https://github.com/nodejs/node/commit/fc1fa7a357)] - **build**: use FILE\_OFFSET\_BITS=64 esp. on 32-bit arch (RafaelGSS) [#58090](https://github.com/nodejs/node/pull/58090)
+* \[[`79e0812181`](https://github.com/nodejs/node/commit/79e0812181)] - **build**: use glob for dependencies of out/Makefile (Richard Lau) [#55789](https://github.com/nodejs/node/pull/55789)
+* \[[`f56e62851a`](https://github.com/nodejs/node/commit/f56e62851a)] - **crypto**: allow length=0 for HKDF and PBKDF2 in SubtleCrypto.deriveBits (Filip Skokan) [#55866](https://github.com/nodejs/node/pull/55866)
+* \[[`c535a3c483`](https://github.com/nodejs/node/commit/c535a3c483)] - **crypto**: graduate WebCryptoAPI Ed25519 and X25519 algorithms as stable (Filip Skokan) [#56142](https://github.com/nodejs/node/pull/56142)
+* \[[`39925de8b1`](https://github.com/nodejs/node/commit/39925de8b1)] - **crypto**: allow non-multiple of 8 in SubtleCrypto.deriveBits (Filip Skokan) [#55296](https://github.com/nodejs/node/pull/55296)
+* \[[`af1dc63815`](https://github.com/nodejs/node/commit/af1dc63815)] - **crypto**: update root certificates to NSS 3.108 (Node.js GitHub Bot) [#57381](https://github.com/nodejs/node/pull/57381)
+* \[[`d09008add3`](https://github.com/nodejs/node/commit/d09008add3)] - **deps**: V8: cherry-pick 1a3ecc2483b2 (Michaël Zasso) [#58342](https://github.com/nodejs/node/pull/58342)
+* \[[`fd56652425`](https://github.com/nodejs/node/commit/fd56652425)] - **deps**: V8: cherry-pick 182d9c05e78b (Andrey Kosyakov) [#58342](https://github.com/nodejs/node/pull/58342)
+* \[[`447481e829`](https://github.com/nodejs/node/commit/447481e829)] - **deps**: V8: cherry-pick third\_party/zlib\@646b7f569718 (Hans Wennborg) [#58342](https://github.com/nodejs/node/pull/58342)
+* \[[`eb447168df`](https://github.com/nodejs/node/commit/eb447168df)] - **deps**: update simdutf to 6.4.2 (Node.js GitHub Bot) [#57855](https://github.com/nodejs/node/pull/57855)
+* \[[`01d63a4ddf`](https://github.com/nodejs/node/commit/01d63a4ddf)] - **deps**: update timezone to 2025b (Node.js GitHub Bot) [#57857](https://github.com/nodejs/node/pull/57857)
+* \[[`10fb49f2a9`](https://github.com/nodejs/node/commit/10fb49f2a9)] - **deps**: update icu to 77.1 (Node.js GitHub Bot) [#57455](https://github.com/nodejs/node/pull/57455)
+* \[[`f1dc7d0205`](https://github.com/nodejs/node/commit/f1dc7d0205)] - **deps**: update corepack to 0.32.0 (Node.js GitHub Bot) [#57265](https://github.com/nodejs/node/pull/57265)
+* \[[`7a2e64bb8a`](https://github.com/nodejs/node/commit/7a2e64bb8a)] - **deps**: update simdutf to 6.4.0 (Node.js GitHub Bot) [#56764](https://github.com/nodejs/node/pull/56764)
+* \[[`e80669be0d`](https://github.com/nodejs/node/commit/e80669be0d)] - **doc**: mention reports should align with Node.js CoC (Rafael Gonzaga) [#57607](https://github.com/nodejs/node/pull/57607)
+* \[[`7b2c0bc92e`](https://github.com/nodejs/node/commit/7b2c0bc92e)] - **doc**: add gurgunday as triager (Gürgün Dayıoğlu) [#57594](https://github.com/nodejs/node/pull/57594)
+* \[[`791e4879de`](https://github.com/nodejs/node/commit/791e4879de)] - **doc**: document REPL custom eval arguments (Dario Piotrowicz) [#57690](https://github.com/nodejs/node/pull/57690)
+* \[[`2917f09876`](https://github.com/nodejs/node/commit/2917f09876)] - **doc**: improved fetch docs (Alessandro Miliucci) [#57296](https://github.com/nodejs/node/pull/57296)
+* \[[`d940b15843`](https://github.com/nodejs/node/commit/d940b15843)] - **doc**: clarify `unhandledRejection` events behaviors in process doc (Dario Piotrowicz) [#57654](https://github.com/nodejs/node/pull/57654)
+* \[[`71c664fab7`](https://github.com/nodejs/node/commit/71c664fab7)] - **doc**: update position type to integer | null in fs (Yukihiro Hasegawa) [#57745](https://github.com/nodejs/node/pull/57745)
+* \[[`0c0fbfa9c6`](https://github.com/nodejs/node/commit/0c0fbfa9c6)] - **doc**: add missing v0.x changelog entries (Antoine du Hamel) [#57779](https://github.com/nodejs/node/pull/57779)
+* \[[`e99462c9fc`](https://github.com/nodejs/node/commit/e99462c9fc)] - **doc**: correct deprecation type of `assert.CallTracker` (René) [#57997](https://github.com/nodejs/node/pull/57997)
+* \[[`c7e92696ef`](https://github.com/nodejs/node/commit/c7e92696ef)] - **doc**: add returns for https.get (Eng Zer Jun) [#58025](https://github.com/nodejs/node/pull/58025)
+* \[[`ccc42b69ce`](https://github.com/nodejs/node/commit/ccc42b69ce)] - **doc**: fix env variable name in `util.styleText` (Antoine du Hamel) [#58072](https://github.com/nodejs/node/pull/58072)
+* \[[`b6daa344eb`](https://github.com/nodejs/node/commit/b6daa344eb)] - **doc**: add dario-piotrowicz to collaborators (Dario Piotrowicz) [#58102](https://github.com/nodejs/node/pull/58102)
+* \[[`e5d6a3df16`](https://github.com/nodejs/node/commit/e5d6a3df16)] - **doc**: fix `AsyncLocalStorage` example response changes after node v18 (Naor Tedgi (Abu Emma)) [#57969](https://github.com/nodejs/node/pull/57969)
+* \[[`f006411998`](https://github.com/nodejs/node/commit/f006411998)] - **doc**: fix typo of file `zlib.md` (yusheng chen) [#58093](https://github.com/nodejs/node/pull/58093)
+* \[[`5193735df4`](https://github.com/nodejs/node/commit/5193735df4)] - **doc**: add missing options.signal to readlinePromises.createInterface() (Jimmy Leung) [#55456](https://github.com/nodejs/node/pull/55456)
+* \[[`fd44af730f`](https://github.com/nodejs/node/commit/fd44af730f)] - **doc**: fix misaligned options in vm.compileFunction() (Jimmy Leung) [#58145](https://github.com/nodejs/node/pull/58145)
+* \[[`0fdcc0ddcd`](https://github.com/nodejs/node/commit/0fdcc0ddcd)] - **doc**: add ambassaor message (Brian Muenzenmeyer) [#57600](https://github.com/nodejs/node/pull/57600)
+* \[[`5ca9616bd3`](https://github.com/nodejs/node/commit/5ca9616bd3)] - **doc**: increase z-index of header element (Dario Piotrowicz) [#57851](https://github.com/nodejs/node/pull/57851)
+* \[[`81342d10f0`](https://github.com/nodejs/node/commit/81342d10f0)] - **doc**: fix deprecation type for `DEP0148` (Livia Medeiros) [#57785](https://github.com/nodejs/node/pull/57785)
+* \[[`776becfe01`](https://github.com/nodejs/node/commit/776becfe01)] - **doc**: remove mention of `--require` not supporting ES modules (Huáng Jùnliàng) [#57620](https://github.com/nodejs/node/pull/57620)
+* \[[`3140a8f133`](https://github.com/nodejs/node/commit/3140a8f133)] - **doc**: add missing `deprecated` badges in `fs.md` (Yukihiro Hasegawa) [#57384](https://github.com/nodejs/node/pull/57384)
+* \[[`441ce24ae3`](https://github.com/nodejs/node/commit/441ce24ae3)] - **doc**: deprecate passing invalid types in `fs.existsSync` (Carlos Espa) [#55892](https://github.com/nodejs/node/pull/55892)
+* \[[`0556f54544`](https://github.com/nodejs/node/commit/0556f54544)] - **http**: correctly translate HTTP method (Paolo Insogna) [#52701](https://github.com/nodejs/node/pull/52701)
+* \[[`c2c6d2b035`](https://github.com/nodejs/node/commit/c2c6d2b035)] - **http**: be more generational GC friendly (ywave620) [#56767](https://github.com/nodejs/node/pull/56767)
+* \[[`cdf3fa241c`](https://github.com/nodejs/node/commit/cdf3fa241c)] - **http2**: skip writeHead if stream is closed (Shima Ryuhei) [#57686](https://github.com/nodejs/node/pull/57686)
+* \[[`bbd5aec785`](https://github.com/nodejs/node/commit/bbd5aec785)] - **http2**: fix graceful session close (Kushagra Pandey) [#57808](https://github.com/nodejs/node/pull/57808)
+* \[[`b427ae4f34`](https://github.com/nodejs/node/commit/b427ae4f34)] - **meta**: remove `build-windows.yml` (Aviv Keller) [#54662](https://github.com/nodejs/node/pull/54662)
+* \[[`49e624f554`](https://github.com/nodejs/node/commit/49e624f554)] - **os**: fix netmask format check condition in getCIDR function (Wiyeong Seo) [#57324](https://github.com/nodejs/node/pull/57324)
+* \[[`d582954434`](https://github.com/nodejs/node/commit/d582954434)] - **src**: remove unused variable in crypto\_x509.cc (Michaël Zasso) [#57754](https://github.com/nodejs/node/pull/57754)
+* \[[`234a505e96`](https://github.com/nodejs/node/commit/234a505e96)] - **src**: allow embedder customization of OOMErrorHandler (Shelley Vohr) [#57325](https://github.com/nodejs/node/pull/57325)
+* \[[`c0252cd380`](https://github.com/nodejs/node/commit/c0252cd380)] - **src**: fix -Wunreachable-code-return in node\_sea (Shelley Vohr) [#57664](https://github.com/nodejs/node/pull/57664)
+* \[[`fcd1622fc1`](https://github.com/nodejs/node/commit/fcd1622fc1)] - **src**: fix kill signal 0 on Windows (Stefan Stojanovic) [#57695](https://github.com/nodejs/node/pull/57695)
+* \[[`850192b06b`](https://github.com/nodejs/node/commit/850192b06b)] - **test**: skip broken sea on rhel8 (Marco Ippolito) [#58761](https://github.com/nodejs/node/pull/58761)
+* \[[`3cf7cfb695`](https://github.com/nodejs/node/commit/3cf7cfb695)] - **test**: update WPT for WebCryptoAPI to edd42c005c (Node.js GitHub Bot) [#57365](https://github.com/nodejs/node/pull/57365)
+* \[[`f57765bdcf`](https://github.com/nodejs/node/commit/f57765bdcf)] - **test**: mark test-without-async-context-frame flaky on windows (James M Snell) [#56753](https://github.com/nodejs/node/pull/56753)
+* \[[`275ea8e7ef`](https://github.com/nodejs/node/commit/275ea8e7ef)] - **test**: force GC in test-file-write-stream4 (Luigi Pinca) [#57930](https://github.com/nodejs/node/pull/57930)
+* \[[`da6a13c338`](https://github.com/nodejs/node/commit/da6a13c338)] - **test**: deflake test-http2-options-max-headers-block-length (Luigi Pinca) [#57959](https://github.com/nodejs/node/pull/57959)
+* \[[`56fce6691e`](https://github.com/nodejs/node/commit/56fce6691e)] - **test**: prevent extraneous HOSTNAME substitution in test-runner-output (René) [#58076](https://github.com/nodejs/node/pull/58076)
+* \[[`c9c0be5596`](https://github.com/nodejs/node/commit/c9c0be5596)] - **test**: update expected error message for macOS (Antoine du Hamel) [#57742](https://github.com/nodejs/node/pull/57742)
+* \[[`3cbf5f93d2`](https://github.com/nodejs/node/commit/3cbf5f93d2)] - **test**: fix missing edge case in test-blob-slice-with-large-size (Joyee Cheung) [#58414](https://github.com/nodejs/node/pull/58414)
+* \[[`bffd4ec379`](https://github.com/nodejs/node/commit/bffd4ec379)] - **test**: skip in test-buffer-tostring-rangeerror on allocation failure (Joyee Cheung) [#58415](https://github.com/nodejs/node/pull/58415)
+* \[[`8237346fb7`](https://github.com/nodejs/node/commit/8237346fb7)] - **test,crypto**: update WebCryptoAPI WPT (Filip Skokan) [#54593](https://github.com/nodejs/node/pull/54593)
+* \[[`b90c4ab937`](https://github.com/nodejs/node/commit/b90c4ab937)] - **tools**: remove unused `osx-pkg-postinstall.sh` (Antoine du Hamel) [#57667](https://github.com/nodejs/node/pull/57667)
+* \[[`414013dcfb`](https://github.com/nodejs/node/commit/414013dcfb)] - **tools**: edit create-release-proposal workflow to handle pr body length (Elves Vieira) [#57841](https://github.com/nodejs/node/pull/57841)
+* \[[`7c449ed6b3`](https://github.com/nodejs/node/commit/7c449ed6b3)] - **tools**: fix tarball testing directory (Marco Ippolito) [#57994](https://github.com/nodejs/node/pull/57994)
+* \[[`d164dc2d38`](https://github.com/nodejs/node/commit/d164dc2d38)] - **tools**: update sccache version to v0.10.0 (Marco Ippolito) [#57994](https://github.com/nodejs/node/pull/57994)
+* \[[`debd3c2cc0`](https://github.com/nodejs/node/commit/debd3c2cc0)] - **tools**: disable failing test envs in `test-linux` CI (Antoine du Hamel) [#58351](https://github.com/nodejs/node/pull/58351)
+* \[[`152112505a`](https://github.com/nodejs/node/commit/152112505a)] - **typings**: fix `ImportModuleDynamicallyCallback` return type (Chengzhong Wu) [#57160](https://github.com/nodejs/node/pull/57160)
+* \[[`363bf744ab`](https://github.com/nodejs/node/commit/363bf744ab)] - **worker**: flush stdout and stderr on exit (Matteo Collina) [#56428](https://github.com/nodejs/node/pull/56428)
+
 <a id="20.19.2"></a>
 
 ## 2025-05-14, Version 20.19.2 'Iron' (LTS), @RafaelGSS
diff --git a/doc/contributing/advocacy-ambassador-program.md b/doc/contributing/advocacy-ambassador-program.md
index 76f29b73586691..e8c7936c9412b4 100644
--- a/doc/contributing/advocacy-ambassador-program.md
+++ b/doc/contributing/advocacy-ambassador-program.md
@@ -123,29 +123,29 @@ when no longer relevant.
 
 The goal is to raise awareness of XYZ in the JavaScript ecosystem.
 
-#### Related Links
+##### Related Links
 
-List of links with more information about the topic to provide brackground
+List of links with more information about the topic to provide background
 or the information to be shared.
 
-#### Project contacts
+##### Project contacts
 
 Add a list of GitHub handles for those within the project that
-have volunteered to be contacated when necessary by ambassadors
+have volunteered to be contacted when necessary by ambassadors
 to get more info about the message to be promoted.
 
 #### Node.js is a great choice for a JavaScript runtime
 
 ##### Goal
 
-Highlight the benefits of chosing Node.js as your backend JavaScript runtime. Focus on what is great
+Highlight the benefits of choosing Node.js as your backend JavaScript runtime. Focus on what is great
 about Node.js without drawing comparisons to alternatives. We don't want to say negative things about
 other options, only highlight what is great about Node.js as a choice.
 
 Some of the things to highlight include:
 
-* How widely it is used (you never get fired for chosing Node.js).
-* The openess of the project. It is part of the OpenJS Foundation and it's governance is set up to avoid
+* How widely it is used (you never get fired for choosing Node.js).
+* The openness of the project. It is part of the OpenJS Foundation and it's governance is set up to avoid
   any one company from dominating the project. Decisions are made by the collaborators (of which there are quite
   a few) versus a small number of people.
 * It has predictable and stable releases and has delivered on the release schedule since 2015.
@@ -157,7 +157,7 @@ Some of the things to highlight include:
 * Single threaded programming model enables very low resource consumption, making it ideal for containerised workloads.
 * Highly vibrant ecosystem with enterprise support from many vendors.
 
-#### Related Links
+##### Related Links
 
 * <https://github.com/nodejs/release>
 * <https://github.com/nodejs/node/blob/main/doc/contributing/security-release-process.md>
@@ -165,13 +165,13 @@ Some of the things to highlight include:
 * <https://github.com/mhdawson/presentations/blob/main/2024/NodeConfEU_2024-Node.js_whats_next.pdf>
   for slide  usage and topping recent surveys.
 
-#### Project contacts
+##### Project contacts
 
 * @mhdawson
 
-### How things get done in the Node.js project
+#### How things get done in the Node.js project
 
-#### Goal
+##### Goal
 
 Help people understand that no people are paid to answer their issues or
 implement their pull requests. Things get done based on what volunteers work on
@@ -182,7 +182,7 @@ Some of the things to highlight include:
 * Nobody is paid specifically to answer issues, fix bugs or implement new features.
 * No company owns/supports Node.js. Most contributions are from individuals
   as opposed to organizations. When an individual becomes a collaborator
-  the access and priviledges are granted to the indidual, their employer does
+  the access and privileges are granted to the individual, their employer does
   not gain any additional rights in the project.
 * The governance of the project is specifically designed to prevent one or
   a small number of companies from dominating the project.
@@ -198,10 +198,10 @@ Some of the things to highlight include:
 * People are volunteering their time to review your PRs and answer questions in
   the issues you open. Be mindfull of your asks for their time and acknowledge
   the gift of their time. Too many issues/PRs in a short period of time may
-  overwelm maintainers leading to less progress versus more, try to pace your
+  overwhelm maintainers leading to less progress versus more, try to pace your
   issues and PRs so that you don't have too many open at the same time. The
-  same goes for comments in discussions, try to avoid overwelming a discussion
-  with too many responses, even too much useful data can overwelm a discussion
+  same goes for comments in discussions, try to avoid overwhelming a discussion
+  with too many responses, even too much useful data can overwhelm a discussion
   leading to lower engagement.
 * While volunteers work to do the right thing for the community, the project
   does not owe anybody anything and does not tolerate abusive or
@@ -210,11 +210,11 @@ Some of the things to highlight include:
 * If you depend on timely support or an SLA, contract with a company that provides
   paid support and will prioritize your issues.
 
-#### Related Links
+##### Related Links
 
 * <https://github.com/nodejs/TSC/blob/main/TSC-Charter.md#section-3-establishment-of-the-tsc>
 
-#### Project contacts
+##### Project contacts
 
 * @mcollina
 * @mhdawson
@@ -233,7 +233,7 @@ Some of the things to highlight include:
 * The `tsconfig.json` configuration options to use in combination for type checking.
 * Updates on the implementation advancements.
 
-#### Related Links
+##### Related Links
 
 <!-- lint disable prohibited-strings remark-lint-->
 
@@ -245,6 +245,49 @@ Some of the things to highlight include:
 
 <!-- lint enable prohibited-strings remark-lint-->
 
-#### Project contacts
+##### Project contacts
 
 * @marco-ippolito
+
+#### Do I still need this dependency for my Node.js app?
+
+##### Goal
+
+Advancements over time in Node.js are improving the out of the box experience.
+New versions are released all the time across Active LTS and Current development lines.
+It's easy to miss something between the release notes and our busy work schedules.
+
+Each of these on its own is respectable, but together they make a more cohesive narrative.
+This also shows a [healthy ecosystem at work](https://brianmuenzenmeyer.com/posts/2024-do-i-need-this-node-dependency/#oss-pace-layers),
+with projects learning from one another and their users.
+
+"Recent" new or newish features, ordered by availability:
+
+| Feature                                                                                               | Introduced                                             | Release Status                                                     |
+| ----------------------------------------------------------------------------------------------------- | ------------------------------------------------------ | ------------------------------------------------------------------ |
+| [testing source code](https://nodejs.org/api/test.html)                                               | [16.17.0](https://nodejs.org/en/blog/release/v16.17.0) | Stable as of 20.0.0                                                |
+| [watching source code](https://nodejs.org/api/cli.html#--watch)                                       | [16.19.0](https://nodejs.org/en/blog/release/v16.19.0) | Stable as of 20.13.0                                               |
+| [parsing arguments](https://nodejs.org/api/util.html#utilparseargsconfig)                             | [18.3.0](https://nodejs.org/en/blog/release/v18.3.0)   | Stable as of 20.0.0                                                |
+| [reading environment](https://nodejs.org/api/cli.html#--env-fileconfig)                               | [20.6.0](https://nodejs.org/en/blog/release/v20.6.0)   | Active Development                                                 |
+| [styling output](https://nodejs.org/docs/latest-v22.x/api/util.html#utilstyletextformat-text-options) | [20.12.0](https://nodejs.org/en/blog/release/v20.12.0) | Stable, as of [22.13.0](https://github.com/nodejs/node/pull/56329) |
+| [run scripts](https://nodejs.org/docs/latest/api/cli.html#--run)                                      | [22.0.0](https://nodejs.org/en/blog/release/v22.0.0)   | Stable, as of 22.0.0                                               |
+| [run TypeScript](https://nodejs.org/api/cli.html#--experimental-strip-types)                          | [22.6.0](https://nodejs.org/en/blog/release/v22.6.0)   | Active Development                                                 |
+
+##### Related Links
+
+<!-- lint disable prohibited-strings remark-lint-->
+
+* <https://brianmuenzenmeyer.com/posts/2024-do-i-need-this-node-dependency/>
+* <https://nodejs.org/api/test.html>
+* <https://nodejs.org/api/cli.html#--watch>
+* <https://nodejs.org/api/util.html#utilparseargsconfig>
+* <https://nodejs.org/api/cli.html#--env-fileconfig>
+* <https://nodejs.org/docs/latest-v22.x/api/util.html#utilstyletextformat-text-options>
+* <https://nodejs.org/api/cli.html#--run>
+* <https://nodejs.org/api/cli.html#--experimental-strip-types>
+
+<!-- lint enable prohibited-strings remark-lint-->
+
+##### Project contacts
+
+* @bmuenzenmeyer
diff --git a/lib/_http_common.js b/lib/_http_common.js
index 8a78bafac72b5e..5ccf43e1d66668 100644
--- a/lib/_http_common.js
+++ b/lib/_http_common.js
@@ -27,7 +27,7 @@ const {
 } = primordials;
 const { setImmediate } = require('timers');
 
-const { methods, HTTPParser } = internalBinding('http_parser');
+const { methods, allMethods, HTTPParser } = internalBinding('http_parser');
 const { getOptionValue } = require('internal/options');
 const insecureHTTPParser = getOptionValue('--insecure-http-parser');
 
@@ -108,7 +108,7 @@ function parserOnHeadersComplete(versionMajor, versionMinor, headers, method,
 
   if (typeof method === 'number') {
     // server only
-    incoming.method = methods[method];
+    incoming.method = allMethods[method];
   } else {
     // client only
     incoming.statusCode = statusCode;
diff --git a/lib/_http_outgoing.js b/lib/_http_outgoing.js
index 56676b4d73e9f1..7c6639201889ea 100644
--- a/lib/_http_outgoing.js
+++ b/lib/_http_outgoing.js
@@ -1172,6 +1172,8 @@ OutgoingMessage.prototype._flushOutput = function _flushOutput(socket) {
   // Refs: https://github.com/nodejs/node/pull/30958
   for (let i = 0; i < outputLength; i++) {
     const { data, encoding, callback } = outputData[i];
+    // Avoid any potential ref to Buffer in new generation from old generation
+    outputData[i].data = null;
     ret = socket.write(data, encoding, callback);
   }
   socket.uncork();
diff --git a/lib/internal/bootstrap/switches/is_not_main_thread.js b/lib/internal/bootstrap/switches/is_not_main_thread.js
index c2cb0106f09ae0..6577a7a97ea789 100644
--- a/lib/internal/bootstrap/switches/is_not_main_thread.js
+++ b/lib/internal/bootstrap/switches/is_not_main_thread.js
@@ -31,14 +31,24 @@ process.removeListener('removeListener', stopListeningIfSignal);
 
 const {
   createWorkerStdio,
+  kStdioWantsMoreDataCallback,
 } = require('internal/worker/io');
 
 let workerStdio;
 function lazyWorkerStdio() {
-  if (!workerStdio) workerStdio = createWorkerStdio();
+  if (workerStdio === undefined) {
+    workerStdio = createWorkerStdio();
+    process.on('exit', flushSync);
+  }
+
   return workerStdio;
 }
 
+function flushSync() {
+  workerStdio.stdout[kStdioWantsMoreDataCallback]();
+  workerStdio.stderr[kStdioWantsMoreDataCallback]();
+}
+
 function getStdout() { return lazyWorkerStdio().stdout; }
 
 function getStderr() { return lazyWorkerStdio().stderr; }
diff --git a/lib/internal/crypto/cfrg.js b/lib/internal/crypto/cfrg.js
index 9c5e59ebb3bf86..b80bbf0f3bd23c 100644
--- a/lib/internal/crypto/cfrg.js
+++ b/lib/internal/crypto/cfrg.js
@@ -274,6 +274,14 @@ async function cfrgImportKey(
           'DataError');
       }
 
+      if (keyData.alg !== undefined && (name === 'Ed25519' || name === 'Ed448')) {
+        if (keyData.alg !== name && keyData.alg !== 'EdDSA') {
+          throw lazyDOMException(
+            'JWK "alg" does not match the requested algorithm',
+            'DataError');
+        }
+      }
+
       if (!isPublic && typeof keyData.x !== 'string') {
         throw lazyDOMException('Invalid JWK', 'DataError');
       }
diff --git a/lib/internal/crypto/diffiehellman.js b/lib/internal/crypto/diffiehellman.js
index 59bbf8ff71233c..5962df0ecd17fb 100644
--- a/lib/internal/crypto/diffiehellman.js
+++ b/lib/internal/crypto/diffiehellman.js
@@ -5,6 +5,7 @@ const {
   MathCeil,
   ObjectDefineProperty,
   SafeSet,
+  Uint8Array,
 } = primordials;
 
 const { Buffer } = require('buffer');
@@ -295,6 +296,8 @@ function diffieHellman(options) {
   return statelessDH(privateKey[kHandle], publicKey[kHandle]);
 }
 
+let masks;
+
 // The ecdhDeriveBits function is part of the Web Crypto API and serves both
 // deriveKeys and deriveBits functions.
 async function ecdhDeriveBits(algorithm, baseKey, length) {
@@ -342,18 +345,25 @@ async function ecdhDeriveBits(algorithm, baseKey, length) {
 
   // If the length is not a multiple of 8 the nearest ceiled
   // multiple of 8 is sliced.
-  length = MathCeil(length / 8);
-  const { byteLength } = bits;
+  const sliceLength = MathCeil(length / 8);
 
+  const { byteLength } = bits;
   // If the length is larger than the derived secret, throw.
-  // Otherwise, we either return the secret or a truncated
-  // slice.
-  if (byteLength < length)
+  if (byteLength < sliceLength)
     throw lazyDOMException('derived bit length is too small', 'OperationError');
 
-  return length === byteLength ?
-    bits :
-    ArrayBufferPrototypeSlice(bits, 0, length);
+  const slice = ArrayBufferPrototypeSlice(bits, 0, sliceLength);
+
+  const mod = length % 8;
+  if (mod === 0)
+    return slice;
+
+  // eslint-disable-next-line no-sparse-arrays
+  masks ||= [, 0b10000000, 0b11000000, 0b11100000, 0b11110000, 0b11111000, 0b11111100, 0b11111110];
+
+  const masked = new Uint8Array(slice);
+  masked[sliceLength - 1] = masked[sliceLength - 1] & masks[mod];
+  return masked.buffer;
 }
 
 module.exports = {
diff --git a/lib/internal/crypto/hkdf.js b/lib/internal/crypto/hkdf.js
index 757a2391a0167f..849e593e440a04 100644
--- a/lib/internal/crypto/hkdf.js
+++ b/lib/internal/crypto/hkdf.js
@@ -1,6 +1,7 @@
 'use strict';
 
 const {
+  ArrayBuffer,
   FunctionPrototypeCall,
 } = primordials;
 
@@ -141,7 +142,7 @@ async function hkdfDeriveBits(algorithm, baseKey, length) {
   const { hash, salt, info } = algorithm;
 
   if (length === 0)
-    throw lazyDOMException('length cannot be zero', 'OperationError');
+    return new ArrayBuffer(0);
   if (length === null)
     throw lazyDOMException('length cannot be null', 'OperationError');
   if (length % 8) {
diff --git a/lib/internal/crypto/pbkdf2.js b/lib/internal/crypto/pbkdf2.js
index 697ceffa542aa7..4148725d0340db 100644
--- a/lib/internal/crypto/pbkdf2.js
+++ b/lib/internal/crypto/pbkdf2.js
@@ -1,6 +1,7 @@
 'use strict';
 
 const {
+  ArrayBuffer,
   FunctionPrototypeCall,
 } = primordials;
 
@@ -98,10 +99,8 @@ async function pbkdf2DeriveBits(algorithm, baseKey, length) {
       'iterations cannot be zero',
       'OperationError');
 
-  const raw = baseKey[kKeyObject].export();
-
   if (length === 0)
-    throw lazyDOMException('length cannot be zero', 'OperationError');
+    return new ArrayBuffer(0);
   if (length === null)
     throw lazyDOMException('length cannot be null', 'OperationError');
   if (length % 8) {
@@ -113,7 +112,7 @@ async function pbkdf2DeriveBits(algorithm, baseKey, length) {
   let result;
   try {
     result = await pbkdf2Promise(
-      raw, salt, iterations, length / 8, normalizeHashName(hash.name),
+      baseKey[kKeyObject].export(), salt, iterations, length / 8, normalizeHashName(hash.name),
     );
   } catch (err) {
     throw lazyDOMException(
diff --git a/lib/internal/crypto/util.js b/lib/internal/crypto/util.js
index c208517b2cde9e..396acee091e8c5 100644
--- a/lib/internal/crypto/util.js
+++ b/lib/internal/crypto/util.js
@@ -189,18 +189,22 @@ const kSupportedAlgorithms = {
     'AES-GCM': 'AesKeyGenParams',
     'AES-KW': 'AesKeyGenParams',
     'HMAC': 'HmacKeyGenParams',
+    'Ed25519': null,
+    'X25519': null,
   },
   'sign': {
     'RSASSA-PKCS1-v1_5': null,
     'RSA-PSS': 'RsaPssParams',
     'ECDSA': 'EcdsaParams',
     'HMAC': null,
+    'Ed25519': null,
   },
   'verify': {
     'RSASSA-PKCS1-v1_5': null,
     'RSA-PSS': 'RsaPssParams',
     'ECDSA': 'EcdsaParams',
     'HMAC': null,
+    'Ed25519': null,
   },
   'importKey': {
     'RSASSA-PKCS1-v1_5': 'RsaHashedImportParams',
@@ -215,11 +219,14 @@ const kSupportedAlgorithms = {
     'AES-CBC': null,
     'AES-GCM': null,
     'AES-KW': null,
+    'Ed25519': null,
+    'X25519': null,
   },
   'deriveBits': {
     'HKDF': 'HkdfParams',
     'PBKDF2': 'Pbkdf2Params',
     'ECDH': 'EcdhKeyDeriveParams',
+    'X25519': 'EcdhKeyDeriveParams',
   },
   'encrypt': {
     'RSA-OAEP': 'RsaOaepParams',
@@ -251,17 +258,6 @@ const kSupportedAlgorithms = {
 };
 
 const experimentalAlgorithms = ObjectEntries({
-  'X25519': {
-    generateKey: null,
-    importKey: null,
-    deriveBits: 'EcdhKeyDeriveParams',
-  },
-  'Ed25519': {
-    generateKey: null,
-    sign: null,
-    verify: null,
-    importKey: null,
-  },
   'X448': {
     generateKey: null,
     importKey: null,
diff --git a/lib/internal/crypto/webcrypto.js b/lib/internal/crypto/webcrypto.js
index 8cd27717532344..004c5a59acd302 100644
--- a/lib/internal/crypto/webcrypto.js
+++ b/lib/internal/crypto/webcrypto.js
@@ -475,6 +475,7 @@ async function exportKeyJWK(key) {
       // Fall through
     case 'Ed448':
       jwk.crv ||= key.algorithm.name;
+      jwk.alg = key.algorithm.name;
       return jwk;
     case 'AES-CTR':
       // Fall through
diff --git a/lib/internal/http2/compat.js b/lib/internal/http2/compat.js
index 2437009486668c..a02d3b0c5844fa 100644
--- a/lib/internal/http2/compat.js
+++ b/lib/internal/http2/compat.js
@@ -706,7 +706,7 @@ class Http2ServerResponse extends Stream {
   writeHead(statusCode, statusMessage, headers) {
     const state = this[kState];
 
-    if (state.closed || this.stream.destroyed)
+    if (state.closed || this.stream.destroyed || this.stream.closed)
       return this;
     if (this[kStream].headersSent)
       throw new ERR_HTTP2_HEADERS_SENT();
diff --git a/lib/internal/http2/core.js b/lib/internal/http2/core.js
index d0602acb0a9f09..3e2c5e7b74525b 100644
--- a/lib/internal/http2/core.js
+++ b/lib/internal/http2/core.js
@@ -1068,6 +1068,7 @@ function setupHandle(socket, type, options) {
   if (typeof options.selectPadding === 'function')
     this[kSelectPadding] = options.selectPadding;
   handle.consume(socket._handle);
+  handle.ongracefulclosecomplete = this[kMaybeDestroy].bind(this, null);
 
   this[kHandle] = handle;
   if (this[kNativeFields]) {
@@ -1592,6 +1593,10 @@ class Http2Session extends EventEmitter {
     if (typeof callback === 'function')
       this.once('close', callback);
     this.goaway();
+    const handle = this[kHandle];
+    if (handle) {
+      handle.setGracefulClose();
+    }
     this[kMaybeDestroy]();
   }
 
@@ -1612,11 +1617,13 @@ class Http2Session extends EventEmitter {
   // * session is closed and there are no more pending or open streams
   [kMaybeDestroy](error) {
     if (error == null) {
+      const handle = this[kHandle];
+      const hasPendingData = !!handle && handle.hasPendingData();
       const state = this[kState];
       // Do not destroy if we're not closed and there are pending/open streams
       if (!this.closed ||
           state.streams.size > 0 ||
-          state.pendingStreams.size > 0) {
+          state.pendingStreams.size > 0 || hasPendingData) {
         return;
       }
     }
@@ -3310,7 +3317,7 @@ function socketOnClose() {
     state.streams.forEach((stream) => stream.close(NGHTTP2_CANCEL));
     state.pendingStreams.forEach((stream) => stream.close(NGHTTP2_CANCEL));
     session.close();
-    session[kMaybeDestroy](err);
+    closeSession(session, NGHTTP2_NO_ERROR, err);
   }
 }
 
diff --git a/lib/internal/modules/esm/loader.js b/lib/internal/modules/esm/loader.js
index 5712472c114bb8..d9b2c75dd978d0 100644
--- a/lib/internal/modules/esm/loader.js
+++ b/lib/internal/modules/esm/loader.js
@@ -280,7 +280,7 @@ class ModuleLoader {
    * @param {string} source Source code. TODO(joyeecheung): pass the raw buffer.
    * @param {string} isMain Whether this module is a main module.
    * @param {CJSModule|undefined} parent Parent module, if any.
-   * @returns {{wrap: ModuleWrap, namespace: ModuleNamespaceObject}}
+   * @returns {{wrap: ModuleWrap, namespace: import('internal/modules/esm/utils').ModuleNamespaceObject}}
    */
   importSyncForRequire(mod, filename, source, isMain, parent) {
     const url = pathToFileURL(filename).href;
diff --git a/lib/internal/modules/esm/utils.js b/lib/internal/modules/esm/utils.js
index 38b5eaf1d7e0f2..d7d3928bc2a947 100644
--- a/lib/internal/modules/esm/utils.js
+++ b/lib/internal/modules/esm/utils.js
@@ -101,12 +101,18 @@ function getConditionsSet(conditions) {
   return getDefaultConditionsSet();
 }
 
+/**
+ * @typedef {{
+ *   [Symbol.toStringTag]: 'Module',
+ * }} ModuleNamespaceObject
+ */
+
 /**
  * @callback ImportModuleDynamicallyCallback
  * @param {string} specifier
  * @param {ModuleWrap|ContextifyScript|Function|vm.Module} callbackReferrer
  * @param {Record<string, string>} attributes
- * @returns { Promise<void> }
+ * @returns {Promise<ModuleNamespaceObject>}
  */
 
 /**
diff --git a/lib/internal/util.js b/lib/internal/util.js
index 26566cd1f65602..0c34a5ea4e564b 100644
--- a/lib/internal/util.js
+++ b/lib/internal/util.js
@@ -9,6 +9,7 @@ const {
   Error,
   ErrorCaptureStackTrace,
   FunctionPrototypeCall,
+  NumberParseInt,
   ObjectDefineProperties,
   ObjectDefineProperty,
   ObjectGetOwnPropertyDescriptor,
@@ -34,7 +35,9 @@ const {
   SafeSet,
   SafeWeakMap,
   SafeWeakRef,
+  StringPrototypeIncludes,
   StringPrototypeReplace,
+  StringPrototypeSlice,
   StringPrototypeToLowerCase,
   StringPrototypeToUpperCase,
   Symbol,
@@ -825,6 +828,59 @@ function setupCoverageHooks(dir) {
   return coverageDirectory;
 }
 
+// Returns the number of ones in the binary representation of the decimal
+// number.
+function countBinaryOnes(n) {
+  // Count the number of bits set in parallel, which is faster than looping
+  n = n - ((n >>> 1) & 0x55555555);
+  n = (n & 0x33333333) + ((n >>> 2) & 0x33333333);
+  return ((n + (n >>> 4) & 0xF0F0F0F) * 0x1010101) >>> 24;
+}
+
+function getCIDR(address, netmask, family) {
+  let ones = 0;
+  let split = '.';
+  let range = 10;
+  let groupLength = 8;
+  let hasZeros = false;
+  let lastPos = 0;
+
+  if (family === 'IPv6') {
+    split = ':';
+    range = 16;
+    groupLength = 16;
+  }
+
+  for (let i = 0; i < netmask.length; i++) {
+    if (netmask[i] !== split) {
+      if (i + 1 < netmask.length) {
+        continue;
+      }
+      i++;
+    }
+    const part = StringPrototypeSlice(netmask, lastPos, i);
+    lastPos = i + 1;
+    if (part !== '') {
+      if (hasZeros) {
+        if (part !== '0') {
+          return null;
+        }
+      } else {
+        const binary = NumberParseInt(part, range);
+        const binaryOnes = countBinaryOnes(binary);
+        ones += binaryOnes;
+        if (binaryOnes !== groupLength) {
+          if (StringPrototypeIncludes(binary.toString(2), '01')) {
+            return null;
+          }
+          hasZeros = true;
+        }
+      }
+    }
+  }
+
+  return `${address}/${ones}`;
+}
 
 const handleTypes = ['TCP', 'TTY', 'UDP', 'FILE', 'PIPE', 'UNKNOWN'];
 function guessHandleType(fd) {
@@ -889,6 +945,7 @@ module.exports = {
   filterDuplicateStrings,
   filterOwnProperties,
   getConstructorOf,
+  getCIDR,
   getCWDURL,
   getInternalGlobal,
   getStructuredStack,
diff --git a/lib/internal/worker/io.js b/lib/internal/worker/io.js
index a7bc2b0d20c2e2..2af57235d9f065 100644
--- a/lib/internal/worker/io.js
+++ b/lib/internal/worker/io.js
@@ -373,9 +373,13 @@ class WritableWorkerStdio extends Writable {
       chunks: ArrayPrototypeMap(chunks,
                                 ({ chunk, encoding }) => ({ chunk, encoding })),
     });
-    ArrayPrototypePush(this[kWritableCallbacks], cb);
-    if (this[kPort][kWaitingStreams]++ === 0)
-      this[kPort].ref();
+    if (process._exiting) {
+      cb();
+    } else {
+      ArrayPrototypePush(this[kWritableCallbacks], cb);
+      if (this[kPort][kWaitingStreams]++ === 0)
+        this[kPort].ref();
+    }
   }
 
   _final(cb) {
diff --git a/lib/os.js b/lib/os.js
index f39e8319ca6b80..e6a7f07032f922 100644
--- a/lib/os.js
+++ b/lib/os.js
@@ -24,7 +24,6 @@
 const {
   ArrayPrototypePush,
   Float64Array,
-  NumberParseInt,
   ObjectDefineProperties,
   StringPrototypeEndsWith,
   StringPrototypeSlice,
@@ -41,6 +40,7 @@ const {
   },
   hideStackFrames,
 } = require('internal/errors');
+const { getCIDR } = require('internal/util');
 const { validateInt32 } = require('internal/validators');
 
 const {
@@ -208,60 +208,6 @@ function endianness() {
 }
 endianness[SymbolToPrimitive] = () => kEndianness;
 
-// Returns the number of ones in the binary representation of the decimal
-// number.
-function countBinaryOnes(n) {
-  // Count the number of bits set in parallel, which is faster than looping
-  n = n - ((n >>> 1) & 0x55555555);
-  n = (n & 0x33333333) + ((n >>> 2) & 0x33333333);
-  return ((n + (n >>> 4) & 0xF0F0F0F) * 0x1010101) >>> 24;
-}
-
-function getCIDR(address, netmask, family) {
-  let ones = 0;
-  let split = '.';
-  let range = 10;
-  let groupLength = 8;
-  let hasZeros = false;
-  let lastPos = 0;
-
-  if (family === 'IPv6') {
-    split = ':';
-    range = 16;
-    groupLength = 16;
-  }
-
-  for (let i = 0; i < netmask.length; i++) {
-    if (netmask[i] !== split) {
-      if (i + 1 < netmask.length) {
-        continue;
-      }
-      i++;
-    }
-    const part = StringPrototypeSlice(netmask, lastPos, i);
-    lastPos = i + 1;
-    if (part !== '') {
-      if (hasZeros) {
-        if (part !== '0') {
-          return null;
-        }
-      } else {
-        const binary = NumberParseInt(part, range);
-        const binaryOnes = countBinaryOnes(binary);
-        ones += binaryOnes;
-        if (binaryOnes !== groupLength) {
-          if ((binary & 1) !== 0) {
-            return null;
-          }
-          hasZeros = true;
-        }
-      }
-    }
-  }
-
-  return `${address}/${ones}`;
-}
-
 /**
  * @returns {Record<string, Array<{
  *  address: string,
diff --git a/src/api/environment.cc b/src/api/environment.cc
index 46106fa94b3055..87f632cdd0ec71 100644
--- a/src/api/environment.cc
+++ b/src/api/environment.cc
@@ -230,7 +230,10 @@ void SetIsolateErrorHandlers(v8::Isolate* isolate, const IsolateSettings& s) {
   auto* fatal_error_cb = s.fatal_error_callback ?
       s.fatal_error_callback : OnFatalError;
   isolate->SetFatalErrorHandler(fatal_error_cb);
-  isolate->SetOOMErrorHandler(OOMErrorHandler);
+
+  auto* oom_error_cb =
+      s.oom_error_callback ? s.oom_error_callback : OOMErrorHandler;
+  isolate->SetOOMErrorHandler(oom_error_cb);
 
   if ((s.flags & SHOULD_NOT_SET_PREPARE_STACK_TRACE_CALLBACK) == 0) {
     auto* prepare_stack_trace_cb = s.prepare_stack_trace_callback ?
diff --git a/src/crypto/crypto_x509.cc b/src/crypto/crypto_x509.cc
index a2fc10f3b324cd..31d52f2cf90450 100644
--- a/src/crypto/crypto_x509.cc
+++ b/src/crypto/crypto_x509.cc
@@ -147,7 +147,6 @@ MaybeLocal<Object> X509Certificate::GetPeerCert(
     const SSLPointer& ssl,
     GetPeerCertificateFlag flag) {
   ClearErrorOnReturn clear_error_on_return;
-  MaybeLocal<Object> maybe_cert;
 
   bool is_server =
       static_cast<int>(flag) & static_cast<int>(GetPeerCertificateFlag::SERVER);
diff --git a/src/env_properties.h b/src/env_properties.h
index 555c8fd09111ea..94cc1f2b92e0df 100644
--- a/src/env_properties.h
+++ b/src/env_properties.h
@@ -249,6 +249,7 @@
   V(onsignal_string, "onsignal")                                               \
   V(onunpipe_string, "onunpipe")                                               \
   V(onwrite_string, "onwrite")                                                 \
+  V(ongracefulclosecomplete_string, "ongracefulclosecomplete")                 \
   V(openssl_error_stack, "opensslErrorStack")                                  \
   V(options_string, "options")                                                 \
   V(order_string, "order")                                                     \
diff --git a/src/node.h b/src/node.h
index ec5f6d0d25731d..1a879cfb3ad366 100644
--- a/src/node.h
+++ b/src/node.h
@@ -483,6 +483,7 @@ struct IsolateSettings {
   v8::Isolate::AbortOnUncaughtExceptionCallback
       should_abort_on_uncaught_exception_callback = nullptr;
   v8::FatalErrorCallback fatal_error_callback = nullptr;
+  v8::OOMErrorCallback oom_error_callback = nullptr;
   v8::PrepareStackTraceCallback prepare_stack_trace_callback = nullptr;
 
   // Miscellaneous callbacks
diff --git a/src/node_http2.cc b/src/node_http2.cc
index 73a3836cfeff1b..c190252e5a5499 100644
--- a/src/node_http2.cc
+++ b/src/node_http2.cc
@@ -545,7 +545,8 @@ Http2Session::Http2Session(Http2State* http2_state,
     : AsyncWrap(http2_state->env(), wrap, AsyncWrap::PROVIDER_HTTP2SESSION),
       js_fields_(http2_state->env()->isolate()),
       session_type_(type),
-      http2_state_(http2_state) {
+      http2_state_(http2_state),
+      graceful_close_initiated_(false) {
   MakeWeak();
   statistics_.session_type = type;
   statistics_.start_time = uv_hrtime();
@@ -749,6 +750,24 @@ void Http2Stream::EmitStatistics() {
   });
 }
 
+void Http2Session::HasPendingData(const FunctionCallbackInfo<Value>& args) {
+  Http2Session* session;
+  ASSIGN_OR_RETURN_UNWRAP(&session, args.Holder());
+  args.GetReturnValue().Set(session->HasPendingData());
+}
+
+bool Http2Session::HasPendingData() const {
+  nghttp2_session* session = session_.get();
+  int want_write = nghttp2_session_want_write(session);
+  // It is expected that want_read will alway be 0 if graceful
+  // session close is initiated and goaway frame is sent.
+  int want_read = nghttp2_session_want_read(session);
+  if (want_write == 0 && want_read == 0) {
+    return false;
+  }
+  return true;
+}
+
 void Http2Session::EmitStatistics() {
   if (LIKELY(!HasHttp2Observer(env())))
     return;
@@ -1725,6 +1744,7 @@ void Http2Session::HandleSettingsFrame(const nghttp2_frame* frame) {
 void Http2Session::OnStreamAfterWrite(WriteWrap* w, int status) {
   Debug(this, "write finished with status %d", status);
 
+  MaybeNotifyGracefulCloseComplete();
   CHECK(is_write_in_progress());
   set_write_in_progress(false);
 
@@ -1945,6 +1965,7 @@ uint8_t Http2Session::SendPendingData() {
   if (!res.async) {
     set_write_in_progress(false);
     ClearOutgoing(res.err);
+    MaybeNotifyGracefulCloseComplete();
   }
 
   MaybeStopReading();
@@ -3418,6 +3439,8 @@ void Initialize(Local<Object> target,
   SetProtoMethod(isolate, session, "receive", Http2Session::Receive);
   SetProtoMethod(isolate, session, "destroy", Http2Session::Destroy);
   SetProtoMethod(isolate, session, "goaway", Http2Session::Goaway);
+  SetProtoMethod(
+      isolate, session, "hasPendingData", Http2Session::HasPendingData);
   SetProtoMethod(isolate, session, "settings", Http2Session::Settings);
   SetProtoMethod(isolate, session, "request", Http2Session::Request);
   SetProtoMethod(
@@ -3438,6 +3461,8 @@ void Initialize(Local<Object> target,
       "remoteSettings",
       Http2Session::RefreshSettings<nghttp2_session_get_remote_settings,
                                     false>);
+  SetProtoMethod(
+      isolate, session, "setGracefulClose", Http2Session::SetGracefulClose);
   SetConstructorFunction(context, target, "Http2Session", session);
 
   Local<Object> constants = Object::New(isolate);
@@ -3492,6 +3517,38 @@ void Initialize(Local<Object> target,
   nghttp2_set_debug_vprintf_callback(NgHttp2Debug);
 #endif
 }
+
+void Http2Session::SetGracefulClose(const FunctionCallbackInfo<Value>& args) {
+  Http2Session* session;
+  ASSIGN_OR_RETURN_UNWRAP(&session, args.Holder());
+  CHECK_NOT_NULL(session);
+  // Set the graceful close flag
+  session->SetGracefulCloseInitiated(true);
+
+  Debug(session, "Setting graceful close initiated flag");
+}
+
+void Http2Session::MaybeNotifyGracefulCloseComplete() {
+  nghttp2_session* session = session_.get();
+
+  if (!IsGracefulCloseInitiated()) {
+    return;
+  }
+
+  int want_write = nghttp2_session_want_write(session);
+  int want_read = nghttp2_session_want_read(session);
+  bool should_notify = (want_write == 0 && want_read == 0);
+
+  if (should_notify) {
+    Debug(this, "Notifying JS after write in graceful close mode");
+
+    // Make the callback to JavaScript
+    HandleScope scope(env()->isolate());
+    MakeCallback(env()->ongracefulclosecomplete_string(), 0, nullptr);
+  }
+
+  return;
+}
 }  // namespace http2
 }  // namespace node
 
diff --git a/src/node_http2.h b/src/node_http2.h
index 3ba05cbe7f9ce6..a60a7ba029db3e 100644
--- a/src/node_http2.h
+++ b/src/node_http2.h
@@ -712,6 +712,7 @@ class Http2Session : public AsyncWrap,
   static void Consume(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void Receive(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void Destroy(const v8::FunctionCallbackInfo<v8::Value>& args);
+  static void HasPendingData(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void Settings(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void Request(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void SetNextStreamID(const v8::FunctionCallbackInfo<v8::Value>& args);
@@ -723,6 +724,7 @@ class Http2Session : public AsyncWrap,
   static void Ping(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void AltSvc(const v8::FunctionCallbackInfo<v8::Value>& args);
   static void Origin(const v8::FunctionCallbackInfo<v8::Value>& args);
+  static void SetGracefulClose(const v8::FunctionCallbackInfo<v8::Value>& args);
 
   template <get_setting fn, bool local>
   static void RefreshSettings(const v8::FunctionCallbackInfo<v8::Value>& args);
@@ -735,6 +737,7 @@ class Http2Session : public AsyncWrap,
 
   BaseObjectPtr<Http2Ping> PopPing();
   bool AddPing(const uint8_t* data, v8::Local<v8::Function> callback);
+  bool HasPendingData() const;
 
   BaseObjectPtr<Http2Settings> PopSettings();
   bool AddSettings(v8::Local<v8::Function> callback);
@@ -785,6 +788,13 @@ class Http2Session : public AsyncWrap,
 
   Statistics statistics_ = {};
 
+  bool IsGracefulCloseInitiated() const {
+    return graceful_close_initiated_;
+  }
+  void SetGracefulCloseInitiated(bool value) {
+    graceful_close_initiated_ = value;
+  }
+
  private:
   void EmitStatistics();
 
@@ -951,8 +961,13 @@ class Http2Session : public AsyncWrap,
   void CopyDataIntoOutgoing(const uint8_t* src, size_t src_length);
   void ClearOutgoing(int status);
 
+  void MaybeNotifyGracefulCloseComplete();
+
   friend class Http2Scope;
   friend class Http2StreamListener;
+
+  // Flag to indicate that JavaScript has initiated a graceful closure
+  bool graceful_close_initiated_ = false;
 };
 
 struct Http2SessionPerformanceEntryTraits {
diff --git a/src/node_http_parser.cc b/src/node_http_parser.cc
index bdde10dbef6241..663f7da5b901a7 100644
--- a/src/node_http_parser.cc
+++ b/src/node_http_parser.cc
@@ -1305,7 +1305,9 @@ void InitializeHttpParser(Local<Object> target,
          Integer::NewFromUnsigned(env->isolate(), kLenientAll));
 
   Local<Array> methods = Array::New(env->isolate());
+  Local<Array> all_methods = Array::New(env->isolate());
   size_t method_index = -1;
+  size_t all_method_index = -1;
 #define V(num, name, string)                                                   \
   methods                                                                      \
       ->Set(env->context(),                                                    \
@@ -1314,9 +1316,23 @@ void InitializeHttpParser(Local<Object> target,
       .Check();
   HTTP_METHOD_MAP(V)
 #undef V
+#define V(num, name, string)                                                   \
+  all_methods                                                                  \
+      ->Set(env->context(),                                                    \
+            ++all_method_index,                                                \
+            FIXED_ONE_BYTE_STRING(env->isolate(), #string))                    \
+      .Check();
+  HTTP_ALL_METHOD_MAP(V)
+#undef V
+
   target->Set(env->context(),
               FIXED_ONE_BYTE_STRING(env->isolate(), "methods"),
               methods).Check();
+  target
+      ->Set(env->context(),
+            FIXED_ONE_BYTE_STRING(env->isolate(), "allMethods"),
+            all_methods)
+      .Check();
 
   t->Inherit(AsyncWrap::GetConstructorTemplate(env));
   SetProtoMethod(isolate, t, "close", Parser::Close);
diff --git a/src/node_root_certs.h b/src/node_root_certs.h
index ee229fc7740627..69634bca1db986 100644
--- a/src/node_root_certs.h
+++ b/src/node_root_certs.h
@@ -354,38 +354,6 @@
 "W8mw0FfB+j564ZfJ\n"
 "-----END CERTIFICATE-----",
 
-/* SwissSign Silver CA - G2 */
-"-----BEGIN CERTIFICATE-----\n"
-"MIIFvTCCA6WgAwIBAgIITxvUL1S7L0swDQYJKoZIhvcNAQEFBQAwRzELMAkGA1UEBhMCQ0gx\n"
-"FTATBgNVBAoTDFN3aXNzU2lnbiBBRzEhMB8GA1UEAxMYU3dpc3NTaWduIFNpbHZlciBDQSAt\n"
-"IEcyMB4XDTA2MTAyNTA4MzI0NloXDTM2MTAyNTA4MzI0NlowRzELMAkGA1UEBhMCQ0gxFTAT\n"
-"BgNVBAoTDFN3aXNzU2lnbiBBRzEhMB8GA1UEAxMYU3dpc3NTaWduIFNpbHZlciBDQSAtIEcy\n"
-"MIICIjANBgkqhkiG9w0BAQEFAAOCAg8AMIICCgKCAgEAxPGHf9N4Mfc4yfjDmUO8x/e8N+dO\n"
-"cbpLj6VzHVxumK4DV644N0MvFz0fyM5oEMF4rhkDKxD6LHmD9ui5aLlV8gREpzn5/ASLHvGi\n"
-"TSf5YXu6t+WiE7brYT7QbNHm+/pe7R20nqA1W6GSy/BJkv6FCgU+5tkL4k+73JU3/JHpMjUi\n"
-"0R86TieFnbAVlDLaYQ1HTWBCrpJH6INaUFjpiou5XaHc3ZlKHzZnu0jkg7Y360g6rw9njxcH\n"
-"6ATK72oxh9TAtvmUcXtnZLi2kUpCe2UuMGoM9ZDulebyzYLs2aFK7PayS+VFheZteJMELpyC\n"
-"bTapxDFkH4aDCyr0NQp4yVXPQbBH6TCfmb5hqAaEuSh6XzjZG6k4sIN/c8HDO0gqgg8hm7jM\n"
-"qDXDhBuDsz6+pJVpATqJAHgE2cn0mRmrVn5bi4Y5FZGkECwJMoBgs5PAKrYYC51+jUnyEEp/\n"
-"+dVGLxmSo5mnJqy7jDzmDrxHB9xzUfFwZC8I+bRHHTBsROopN4WSaGa8gzj+ezku01DwH/te\n"
-"YLappvonQfGbGHLy9YR0SslnxFSuSGTfjNFusB3hB48IHpmccelM2KX3RxIfdNFRnobzwqIj\n"
-"QAtz20um53MGjMGg6cFZrEb65i/4z3GcRm25xBWNOHkDRUjvxF3XCO6HOSKGsg0PWEP3calI\n"
-"Lv3q1h8CAwEAAaOBrDCBqTAOBgNVHQ8BAf8EBAMCAQYwDwYDVR0TAQH/BAUwAwEB/zAdBgNV\n"
-"HQ4EFgQUF6DNweRBtjpbO8tFnb0cwpj6hlgwHwYDVR0jBBgwFoAUF6DNweRBtjpbO8tFnb0c\n"
-"wpj6hlgwRgYDVR0gBD8wPTA7BglghXQBWQEDAQEwLjAsBggrBgEFBQcCARYgaHR0cDovL3Jl\n"
-"cG9zaXRvcnkuc3dpc3NzaWduLmNvbS8wDQYJKoZIhvcNAQEFBQADggIBAHPGgeAn0i0P4JUw\n"
-"4ppBf1AsX19iYamGamkYDHRJ1l2E6kFSGG9YrVBWIGrGvShpWJHckRE1qTodvBqlYJ7YH39F\n"
-"kWnZfrt4csEGDyrOj4VwYaygzQu4OSlWhDJOhrs9xCrZ1x9y7v5RoSJBsXECYxqCsGKrXlcS\n"
-"H9/L3XWgwF15kIwb4FDm3jH+mHtwX6WQ2K34ArZv02DdQEsixT2tOnqfGhpHkXkzuoLcMmkD\n"
-"lm4fS/Bx/uNncqCxv1yL5PqZIseEuRuNI5c/7SXgz2W79WEE790eslpBIlqhn10s6FvJbakM\n"
-"DHiqYMZWjwFaDGi8aRl5xB9+lwW/xekkUV7U1UtT7dkjWjYDZaPBA61BMPNGG4WQr2W11bHk\n"
-"Flt4dR2Xem1ZqSqPe97Dh4kQmUlzeMg9vVE1dCrV8X5pGyq7O70luJpaPXJhkGaH7gzWTdQR\n"
-"dAtq/gsD/KNVV4n+SsuuWxcFyPKNIzFTONItaj+CuY0IavdeQXRuwxF+B6wpYJE/OMpXEA29\n"
-"MC/HpeZBoNquBYeaoKRlbEwJDIm6uNO5wJOKMPqN5ZprFQFOZ6raYlY+hAhm0sQ2fac+EPyI\n"
-"4NSA5QC9qvNOBqN6avlicuMJT+ubDgEj8Z+7fNzcbBGXJbLytGMU0gYqZ4yD9c7qB9iaah7s\n"
-"5Aq7KkzrCWA5zspi2C5u\n"
-"-----END CERTIFICATE-----",
-
 /* SecureTrust CA */
 "-----BEGIN CERTIFICATE-----\n"
 "MIIDuDCCAqCgAwIBAgIQDPCOXAgWpa1Cf/DrJxhZ0DANBgkqhkiG9w0BAQUFADBIMQswCQYD\n"
@@ -3598,4 +3566,66 @@
 "4E/0BdGF9jVg3PVys0Z9AjBEmEYagoUeYWmJSwdLZrWeqrqgHkHZAXQ6bkU6iYAZezKYVWOr\n"
 "62Nuk22rGwlgMU4=\n"
 "-----END CERTIFICATE-----",
+
+/* D-TRUST BR Root CA 2 2023 */
+"-----BEGIN CERTIFICATE-----\n"
+"MIIFqTCCA5GgAwIBAgIQczswBEhb2U14LnNLyaHcZjANBgkqhkiG9w0BAQ0FADBIMQswCQYD\n"
+"VQQGEwJERTEVMBMGA1UEChMMRC1UcnVzdCBHbWJIMSIwIAYDVQQDExlELVRSVVNUIEJSIFJv\n"
+"b3QgQ0EgMiAyMDIzMB4XDTIzMDUwOTA4NTYzMVoXDTM4MDUwOTA4NTYzMFowSDELMAkGA1UE\n"
+"BhMCREUxFTATBgNVBAoTDEQtVHJ1c3QgR21iSDEiMCAGA1UEAxMZRC1UUlVTVCBCUiBSb290\n"
+"IENBIDIgMjAyMzCCAiIwDQYJKoZIhvcNAQEBBQADggIPADCCAgoCggIBAK7/CVmRgApKaOYk\n"
+"P7in5Mg6CjoWzckjYaCTcfKri3OPoGdlYNJUa2NRb0kz4HIHE304zQaSBylSa053bATTlfrd\n"
+"TIzZXcFhfUvnKLNEgXtRr90zsWh81k5M/itoucpmacTsXld/9w3HnDY25QdgrMBM6ghs7wZ8\n"
+"T1soegj8k12b9py0i4a6Ibn08OhZWiihNIQaJZG2tY/vsvmA+vk9PBFy2OMvhnbFeSzBqZCT\n"
+"Rphny4NqoFAjpzv2gTng7fC5v2Xx2Mt6++9zA84A9H3X4F07ZrjcjrqDy4d2A/wl2ecjbwb9\n"
+"Z/Pg/4S8R7+1FhhGaRTMBffb00msa8yr5LULQyReS2tNZ9/WtT5PeB+UcSTq3nD88ZP+npNa\n"
+"5JRal1QMNXtfbO4AHyTsA7oC9Xb0n9Sa7YUsOCIvx9gvdhFP/Wxc6PWOJ4d/GUohR5AdeY0c\n"
+"W/jPSoXk7bNbjb7EZChdQcRurDhaTyN0dKkSw/bSuREVMweR2Ds3OmMwBtHFIjYoYiMQ4EbM\n"
+"l6zWK11kJNXuHA7e+whadSr2Y23OC0K+0bpwHJwh5Q8xaRfX/Aq03u2AnMuStIv13lmiWAml\n"
+"Y0cL4UEyNEHZmrHZqLAbWt4NDfTisl01gLmB1IRpkQLLddCNxbU9CZEJjxShFHR5PtbJFR2k\n"
+"WVki3PaKRT08EtY+XTIvAgMBAAGjgY4wgYswDwYDVR0TAQH/BAUwAwEB/zAdBgNVHQ4EFgQU\n"
+"Z5Dw1t61GNVGKX5cq/ieCLxklRAwDgYDVR0PAQH/BAQDAgEGMEkGA1UdHwRCMEAwPqA8oDqG\n"
+"OGh0dHA6Ly9jcmwuZC10cnVzdC5uZXQvY3JsL2QtdHJ1c3RfYnJfcm9vdF9jYV8yXzIwMjMu\n"
+"Y3JsMA0GCSqGSIb3DQEBDQUAA4ICAQA097N3U9swFrktpSHxQCF16+tIFoE9c+CeJyrrd6kT\n"
+"pGoKWloUMz1oH4Guaf2Mn2VsNELZLdB/eBaxOqwjMa1ef67nriv6uvw8l5VAk1/DLQOj7aRv\n"
+"U9f6QA4w9QAgLABMjDu0ox+2v5Eyq6+SmNMW5tTRVFxDWy6u71cqqLRvpO8NVhTaIasgdp4D\n"
+"/Ca4nj8+AybmTNudX0KEPUUDAxxZiMrcLmEkWqTqJwtzEr5SswrPMhfiHocaFpVIbVrg0M8J\n"
+"kiZmkdijYQ6qgYF/6FKC0ULn4B0Y+qSFNueG4A3rvNTJ1jxD8V1Jbn6Bm2m1iWKPiFLY1/4n\n"
+"wSPFyysCu7Ff/vtDhQNGvl3GyiEm/9cCnnRK3PgTFbGBVzbLZVzRHTF36SXDw7IyN9XxmAnk\n"
+"bWOACKsGkoHU6XCPpz+y7YaMgmo1yEJagtFSGkUPFaUA8JR7ZSdXOUPPfH/mvTWze/EZTN46\n"
+"ls/pdu4D58JDUjxqgejBWoC9EV2Ta/vH5mQ/u2kc6d0li690yVRAysuTEwrt+2aSEcr1wPrY\n"
+"g1UDfNPFIkZ1cGt5SAYqgpq/5usWDiJFAbzdNpQ0qTUmiteXue4Icr80knCDgKs4qllo3UCk\n"
+"GJCy89UDyibK79XH4I9TjvAA46jtn/mtd+ArY0+ew+43u3gJhJ65bvspmZDogNOfJA==\n"
+"-----END CERTIFICATE-----",
+
+/* D-TRUST EV Root CA 2 2023 */
+"-----BEGIN CERTIFICATE-----\n"
+"MIIFqTCCA5GgAwIBAgIQaSYJfoBLTKCnjHhiU19abzANBgkqhkiG9w0BAQ0FADBIMQswCQYD\n"
+"VQQGEwJERTEVMBMGA1UEChMMRC1UcnVzdCBHbWJIMSIwIAYDVQQDExlELVRSVVNUIEVWIFJv\n"
+"b3QgQ0EgMiAyMDIzMB4XDTIzMDUwOTA5MTAzM1oXDTM4MDUwOTA5MTAzMlowSDELMAkGA1UE\n"
+"BhMCREUxFTATBgNVBAoTDEQtVHJ1c3QgR21iSDEiMCAGA1UEAxMZRC1UUlVTVCBFViBSb290\n"
+"IENBIDIgMjAyMzCCAiIwDQYJKoZIhvcNAQEBBQADggIPADCCAgoCggIBANiOo4mAC7JXUtyp\n"
+"U0w3uX9jFxPvp1sjW2l1sJkKF8GLxNuo4MwxusLyzV3pt/gdr2rElYfXR8mV2IIEUD2BCP/k\n"
+"PbOx1sWy/YgJ25yE7CUXFId/MHibaljJtnMoPDT3mfd/06b4HEV8rSyMlD/YZxBTfiLNTiVR\n"
+"8CUkNRFeEMbsh2aJgWi6zCudR3Mfvc2RpHJqnKIbGKBv7FD0fUDCqDDPvXPIEysQEx6Lmqg6\n"
+"lHPTGGkKSv/BAQP/eX+1SH977ugpbzZMlWGG2Pmic4ruri+W7mjNPU0oQvlFKzIbRlUWaqZL\n"
+"Kfm7lVa/Rh3sHZMdwGWyH6FDrlaeoLGPaxK3YG14C8qKXO0elg6DpkiVjTujIcSuWMYAsoS0\n"
+"I6SWhjW42J7YrDRJmGOVxcttSEfi8i4YHtAxq9107PncjLgcjmgjutDzUNzPZY9zOjLHfP7K\n"
+"giJPvo5iR2blzYfi6NUPGJ/lBHJLRjwQ8kTCZFZxTnXonMkmdMV9WdEKWw9t/p51HBjGGjp8\n"
+"2A0EzM23RWV6sY+4roRIPrN6TagD4uJ+ARZZaBhDM7DS3LAaQzXupdqpRlyuhoFBAUp0Juyf\n"
+"Br/CBTdkdXgpaP3F9ev+R/nkhbDhezGdpn9yo7nELC7MmVcOIQxFAZRl62UJxmMiCzNJkkg8\n"
+"/M3OsD6Onov4/knFNXJHAgMBAAGjgY4wgYswDwYDVR0TAQH/BAUwAwEB/zAdBgNVHQ4EFgQU\n"
+"qvyREBuHkV8Wub9PS5FeAByxMoAwDgYDVR0PAQH/BAQDAgEGMEkGA1UdHwRCMEAwPqA8oDqG\n"
+"OGh0dHA6Ly9jcmwuZC10cnVzdC5uZXQvY3JsL2QtdHJ1c3RfZXZfcm9vdF9jYV8yXzIwMjMu\n"
+"Y3JsMA0GCSqGSIb3DQEBDQUAA4ICAQCTy6UfmRHsmg1fLBWTxj++EI14QvBukEdHjqOSMo1w\n"
+"j/Zbjb6JzkcBahsgIIlbyIIQbODnmaprxiqgYzWRaoUlrRc4pZt+UPJ26oUFKidBK7GB0aL2\n"
+"QHWpDsvxVUjY7NHss+jOFKE17MJeNRqrphYBBo7q3C+jisosketSjl8MmxfPy3MHGcRqwnNU\n"
+"73xDUmPBEcrCRbH0O1P1aa4846XerOhUt7KR/aypH/KH5BfGSah82ApB9PI+53c0BFLd6IHy\n"
+"TS9URZ0V4U/M5d40VxDJI3IXcI1QcB9WbMy5/zpaT2N6w25lBx2Eof+pDGOJbbJAiDnXH3do\n"
+"tfyc1dZnaVuodNv8ifYbMvekJKZ2t0dT741Jj6m2g1qllpBFYfXeA08mD6iL8AOWsKwV0HFa\n"
+"anuU5nCT2vFp4LJiTZ6P/4mdm13NRemUAiKN4DV/6PEEeXFsVIP4M7kFMhtYVRFP0OUnR3Hs\n"
+"7dpn1mKmS00PaaLJvOwiS5THaJQXfuKOKD62xur1NGyfN4gHONuGcfrNlUhDbqNPgofXNJhu\n"
+"S5N5YHVpD/Aa1VP6IQzCP+k/HxiMkl14p3ZnGbuy6n/pcAlWVqOwDAstNl7F6cTVg8uGF5cs\n"
+"bBNvh1qvSaYd2804BC5f4ko1Di1L+KIkBI3Y4WNeApI02phhXBxvWHZks/wCuPWdCg==\n"
+"-----END CERTIFICATE-----",
 #endif  // defined(NODE_WANT_INTERNALS) && NODE_WANT_INTERNALS
diff --git a/src/node_sea.cc b/src/node_sea.cc
index fb9f933a19fa70..d70401f4eccdc1 100644
--- a/src/node_sea.cc
+++ b/src/node_sea.cc
@@ -637,8 +637,9 @@ bool MaybeLoadSingleExecutableApplication(Environment* env) {
 
   LoadEnvironment(env, LoadSingleExecutableApplication);
   return true;
-#endif
+#else
   return false;
+#endif
 }
 
 void Initialize(Local<Object> target,
diff --git a/src/node_version.h b/src/node_version.h
index a756ad717e3c0b..cae2f453260259 100644
--- a/src/node_version.h
+++ b/src/node_version.h
@@ -29,7 +29,7 @@
 #define NODE_VERSION_IS_LTS 1
 #define NODE_VERSION_LTS_CODENAME "Iron"
 
-#define NODE_VERSION_IS_RELEASE 0
+#define NODE_VERSION_IS_RELEASE 1
 
 #ifndef NODE_STRINGIFY
 #define NODE_STRINGIFY(n) NODE_STRINGIFY_HELPER(n)
diff --git a/src/process_wrap.cc b/src/process_wrap.cc
index 74f46bcb651bef..f0ce0b55f07843 100644
--- a/src/process_wrap.cc
+++ b/src/process_wrap.cc
@@ -314,7 +314,7 @@ class ProcessWrap : public HandleWrap {
     int signal = args[0]->Int32Value(env->context()).FromJust();
 #ifdef _WIN32
     if (signal != SIGKILL && signal != SIGTERM && signal != SIGINT &&
-        signal != SIGQUIT) {
+        signal != SIGQUIT && signal != 0) {
       signal = SIGKILL;
     }
 #endif
diff --git a/test/common/wpt.js b/test/common/wpt.js
index 0673d4bc8d721e..98e1d12216fbef 100644
--- a/test/common/wpt.js
+++ b/test/common/wpt.js
@@ -451,8 +451,16 @@ class StatusLoader {
 
   load() {
     const dir = path.join(__dirname, '..', 'wpt');
-    const statusFile = path.join(dir, 'status', `${this.path}.json`);
-    const result = JSON.parse(fs.readFileSync(statusFile, 'utf8'));
+    let statusFile = path.join(dir, 'status', `${this.path}.json`);
+    let result;
+
+    if (fs.existsSync(statusFile)) {
+      result = JSON.parse(fs.readFileSync(statusFile, 'utf8'));
+    } else {
+      statusFile = path.join(dir, 'status', `${this.path}.cjs`);
+      result = require(statusFile);
+    }
+
     this.rules.addRules(result);
 
     const subDir = fixtures.path('wpt', this.path);
diff --git a/test/fixtures/tz-version.txt b/test/fixtures/tz-version.txt
index 0846b7f265fa58..ef468adcecf9e2 100644
--- a/test/fixtures/tz-version.txt
+++ b/test/fixtures/tz-version.txt
@@ -1 +1 @@
-2025a
+2025b
diff --git a/test/fixtures/wpt/README.md b/test/fixtures/wpt/README.md
index fa23c5f61e2708..58673011a736a3 100644
--- a/test/fixtures/wpt/README.md
+++ b/test/fixtures/wpt/README.md
@@ -25,15 +25,15 @@ Last update:
 - interfaces: https://github.com/web-platform-tests/wpt/tree/df731dab88/interfaces
 - performance-timeline: https://github.com/web-platform-tests/wpt/tree/17ebc3aea0/performance-timeline
 - resource-timing: https://github.com/web-platform-tests/wpt/tree/22d38586d0/resource-timing
-- resources: https://github.com/web-platform-tests/wpt/tree/1e140d63ec/resources
+- resources: https://github.com/web-platform-tests/wpt/tree/919874f84f/resources
 - streams: https://github.com/web-platform-tests/wpt/tree/2bd26e124c/streams
 - url: https://github.com/web-platform-tests/wpt/tree/67880a4eb8/url
 - user-timing: https://github.com/web-platform-tests/wpt/tree/5ae85bf826/user-timing
 - wasm/jsapi: https://github.com/web-platform-tests/wpt/tree/cde25e7e3c/wasm/jsapi
 - wasm/webapi: https://github.com/web-platform-tests/wpt/tree/fd1b23eeaa/wasm/webapi
-- WebCryptoAPI: https://github.com/web-platform-tests/wpt/tree/5e042cbc4e/WebCryptoAPI
+- WebCryptoAPI: https://github.com/web-platform-tests/wpt/tree/edd42c005c/WebCryptoAPI
 - webidl/ecmascript-binding/es-exceptions: https://github.com/web-platform-tests/wpt/tree/a370aad338/webidl/ecmascript-binding/es-exceptions
 - webmessaging/broadcastchannel: https://github.com/web-platform-tests/wpt/tree/e97fac4791/webmessaging/broadcastchannel
 
 [Web Platform Tests]: https://github.com/web-platform-tests/wpt
-[`git node wpt`]: https://github.com/nodejs/node-core-utils/blob/main/docs/git-node.md#git-node-wpt
\ No newline at end of file
+[`git node wpt`]: https://github.com/nodejs/node-core-utils/blob/main/docs/git-node.md#git-node-wpt
diff --git a/test/fixtures/wpt/WebCryptoAPI/crypto_key_cached_slots.https.any.js b/test/fixtures/wpt/WebCryptoAPI/crypto_key_cached_slots.https.any.js
new file mode 100644
index 00000000000000..f573fac1c96606
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/crypto_key_cached_slots.https.any.js
@@ -0,0 +1,44 @@
+// META: title=WebCryptoAPI: CryptoKey cached ECMAScript objects
+
+// https://w3c.github.io/webcrypto/#dom-cryptokey-algorithm
+// https://github.com/servo/servo/issues/33908
+
+promise_test(function() {
+    return self.crypto.subtle.generateKey(
+        {
+          name: "AES-CTR",
+          length: 256,
+        },
+        true,
+        ["encrypt"],
+      ).then(
+        function(key) {
+          let a = key.algorithm;
+          let b = key.algorithm;
+          assert_true(a === b);
+        },
+        function(err) {
+            assert_unreached("generateKey threw an unexpected error: " + err.toString());
+        }
+    );
+}, "CryptoKey.algorithm getter returns cached object");
+
+promise_test(function() {
+    return self.crypto.subtle.generateKey(
+        {
+          name: "AES-CTR",
+          length: 256,
+        },
+        true,
+        ["encrypt"],
+      ).then(
+        function(key) {
+          let a = key.usages;
+          let b = key.usages;
+          assert_true(a === b);
+        },
+        function(err) {
+            assert_unreached("generateKey threw an unexpected error: " + err.toString());
+        }
+    );
+}, "CryptoKey.usages getter returns cached object");
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.js
index ef6905e574c158..8ab9db7bf71318 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.js
@@ -1,14 +1,19 @@
+function define_tests_25519() {
+    return define_tests("X25519");
+}
+
+function define_tests_448() {
+    return define_tests("X448");
+}
 
-function define_tests() {
+function define_tests(algorithmName) {
   // May want to test prefixed implementations.
   var subtle = self.crypto.subtle;
 
   // Verify the derive functions perform checks against the all-zero value results,
   // ensuring small-order points are rejected.
   // https://www.rfc-editor.org/rfc/rfc7748#section-6.1
-  // TODO: The spec states that the check must be done on use, but there is discussion about doing it on import.
-  // https://github.com/WICG/webcrypto-secure-curves/pull/13
-  Object.keys(kSmallOrderPoint).forEach(function(algorithmName) {
+  {
       kSmallOrderPoint[algorithmName].forEach(function(test) {
           promise_test(async() => {
               let derived;
@@ -23,22 +28,23 @@ function define_tests() {
                                                  false, [])
                   derived = await subtle.deriveBits({name: algorithmName, public: publicKey}, privateKey, 8 * sizes[algorithmName]);
               } catch (err) {
-                  assert_false(privateKey === undefined, "Private key should be valid.");
-                  assert_false(publicKey === undefined, "Public key should be valid.");
+                  assert_true(privateKey !== undefined, "Private key should be valid.");
+                  assert_true(publicKey !== undefined, "Public key should be valid.");
                   assert_equals(err.name, "OperationError", "Should throw correct error, not " + err.name + ": " + err.message + ".");
               }
               assert_equals(derived, undefined, "Operation succeeded, but should not have.");
           }, algorithmName + " key derivation checks for all-zero value result with a key of order " + test.order);
       });
-  });
+  }
 
   return importKeys(pkcs8, spki, sizes)
   .then(function(results) {
       publicKeys = results.publicKeys;
       privateKeys = results.privateKeys;
       noDeriveBitsKeys = results.noDeriveBitsKeys;
+      ecdhKeys = results.ecdhKeys;
 
-      Object.keys(sizes).forEach(function(algorithmName) {
+      {
           // Basic success case
           promise_test(function(test) {
               return subtle.deriveBits({name: algorithmName, public: publicKeys[algorithmName]}, privateKeys[algorithmName], 8 * sizes[algorithmName])
@@ -59,25 +65,6 @@ function define_tests() {
               });
           }, algorithmName + " mixed case parameters");
 
-          // Null length
-          // "Null" is not valid per the current spec
-          //   - https://github.com/w3c/webcrypto/issues/322
-          //   - https://github.com/w3c/webcrypto/issues/329
-          //
-          // Proposal for a spec change:
-          //   - https://github.com/w3c/webcrypto/pull/345
-          //
-          // This test case may be replaced by these new tests:
-          //   - https://github.com/web-platform-tests/wpt/pull/43400
-          promise_test(function(test) {
-              return subtle.deriveBits({name: algorithmName, public: publicKeys[algorithmName]}, privateKeys[algorithmName], null)
-              .then(function(derivation) {
-                  assert_true(equalBuffers(derivation, derivations[algorithmName]), "Derived correct bits");
-              }, function(err) {
-                  assert_unreached("deriveBits failed with error " + err.name + ": " + err.message);
-              });
-          }, algorithmName + " with null length");
-
           // Shorter than entire derivation per algorithm
           promise_test(function(test) {
               return subtle.deriveBits({name: algorithmName, public: publicKeys[algorithmName]}, privateKeys[algorithmName], 8 * sizes[algorithmName] - 32)
@@ -122,11 +109,7 @@ function define_tests() {
 
           // - wrong algorithm
           promise_test(function(test) {
-              publicKey = publicKeys["X25519"];
-              if (algorithmName === "X25519") {
-                  publicKey = publicKeys["X448"];
-              }
-              return subtle.deriveBits({name: algorithmName, public: publicKey}, privateKeys[algorithmName], 8 * sizes[algorithmName])
+              return subtle.deriveBits({name: algorithmName, public: ecdhKeys[algorithmName]}, privateKeys[algorithmName], 8 * sizes[algorithmName])
               .then(function(derivation) {
                   assert_unreached("deriveBits succeeded but should have failed with InvalidAccessError");
               }, function(err) {
@@ -186,16 +169,17 @@ function define_tests() {
                   assert_equals(err.name, "OperationError", "Should throw correct error, not " + err.name + ": " + err.message);
               });
           }, algorithmName + " asking for too many bits");
-      });
+      }
   });
 
   function importKeys(pkcs8, spki, sizes) {
       var privateKeys = {};
       var publicKeys = {};
       var noDeriveBitsKeys = {};
+      var ecdhPublicKeys = {};
 
       var promises = [];
-      Object.keys(pkcs8).forEach(function(algorithmName) {
+      {
           var operation = subtle.importKey("pkcs8", pkcs8[algorithmName],
                                           {name: algorithmName},
                                           false, ["deriveBits", "deriveKey"])
@@ -205,8 +189,8 @@ function define_tests() {
                               privateKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
-      Object.keys(pkcs8).forEach(function(algorithmName) {
+      }
+      {
           var operation = subtle.importKey("pkcs8", pkcs8[algorithmName],
                                           {name: algorithmName},
                                           false, ["deriveKey"])
@@ -216,8 +200,8 @@ function define_tests() {
                               noDeriveBitsKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
-      Object.keys(spki).forEach(function(algorithmName) {
+      }
+      {
           var operation = subtle.importKey("spki", spki[algorithmName],
                                           {name: algorithmName},
                                           false, [])
@@ -227,10 +211,17 @@ function define_tests() {
                               publicKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
-
+      }
+      {
+          var operation = subtle.importKey("spki", ecSPKI,
+                                           {name: "ECDH", namedCurve: "P-256"},
+                                           false, [])
+                          .then(function(key) {
+                              ecdhPublicKeys[algorithmName] = key;
+                          });
+      }
       return Promise.all(promises)
-             .then(function(results) {return {privateKeys: privateKeys, publicKeys: publicKeys, noDeriveBitsKeys: noDeriveBitsKeys}});
+             .then(function(results) {return {privateKeys: privateKeys, publicKeys: publicKeys, noDeriveBitsKeys: noDeriveBitsKeys, ecdhKeys: ecdhPublicKeys}});
   }
 
   // Compares two ArrayBuffer or ArrayBufferView objects. If bitCount is
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve25519.https.any.js
new file mode 100644
index 00000000000000..866192e0193bc1
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve25519.https.any.js
@@ -0,0 +1,10 @@
+// META: title=WebCryptoAPI: deriveKey() Using ECDH with CFRG Elliptic Curves
+// META: script=cfrg_curves_bits_fixtures.js
+// META: script=cfrg_curves_bits.js
+
+// Define subtests from a `promise_test` to ensure the harness does not
+// complete before the subtests are available. `explicit_done` cannot be used
+// for this purpose because the global `done` function is automatically invoked
+// by the WPT infrastructure in dedicated worker tests defined using the
+// "multi-global" pattern.
+promise_test(define_tests_25519, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve448.https.any.js
similarity index 75%
rename from test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.https.any.js
rename to test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve448.https.any.js
index c1837591ee85d4..32485c68107e5c 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_curve448.https.any.js
@@ -1,4 +1,4 @@
-// META: title=WebCryptoAPI: deriveBits() Using ECDH with CFRG Elliptic Curves
+// META: title=WebCryptoAPI: deriveKey() Using ECDH with CFRG Elliptic Curves
 // META: script=cfrg_curves_bits_fixtures.js
 // META: script=cfrg_curves_bits.js
 
@@ -7,4 +7,4 @@
 // for this purpose because the global `done` function is automatically invoked
 // by the WPT infrastructure in dedicated worker tests defined using the
 // "multi-global" pattern.
-promise_test(define_tests, 'setup - define tests');
+promise_test(define_tests_448, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_fixtures.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_fixtures.js
index ffdeb51eab9700..c376c75bfe6cf8 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_fixtures.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_bits_fixtures.js
@@ -35,3 +35,6 @@ var kSmallOrderPoint = {
         { order: "p+1 (=1, order 1)", vector : new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255]) },
     ]
 };
+
+// "P-256":
+var ecSPKI = new Uint8Array([48, 89, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 3, 66, 0, 4, 154, 116, 32, 120, 126, 95, 77, 105, 211, 232, 34, 114, 115, 1, 109, 56, 224, 71, 129, 133, 223, 127, 238, 156, 142, 103, 60, 202, 211, 79, 126, 128, 254, 49, 141, 182, 221, 107, 119, 218, 99, 32, 165, 246, 151, 89, 9, 68, 23, 177, 52, 239, 138, 139, 116, 193, 101, 4, 57, 198, 115, 0, 90, 61]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.js
index 81244ba05a8766..62f9e00aa33846 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.js
@@ -1,5 +1,12 @@
+function define_tests_25519() {
+    return define_tests("X25519");
+}
+
+function define_tests_448() {
+    return define_tests("X448");
+}
 
-function define_tests() {
+function define_tests(algorithmName) {
   // May want to test prefixed implementations.
   var subtle = self.crypto.subtle;
 
@@ -8,7 +15,7 @@ function define_tests() {
   // https://www.rfc-editor.org/rfc/rfc7748#section-6.1
   // TODO: The spec states that the check must be done on use, but there is discussion about doing it on import.
   // https://github.com/WICG/webcrypto-secure-curves/pull/13
-  Object.keys(kSmallOrderPoint).forEach(function(algorithmName) {
+  {
       kSmallOrderPoint[algorithmName].forEach(function(test) {
           promise_test(async() => {
               let derived;
@@ -32,10 +39,10 @@ function define_tests() {
               assert_equals(derived, undefined, "Operation succeeded, but should not have.");
           }, algorithmName + " deriveBits checks for all-zero value result with a key of order " + test.order);
       });
-  });
+  }
 
   // Ensure the keys generated by each algorithm are valid for key derivation.
-  Object.keys(sizes).forEach(function(algorithmName) {
+  {
       promise_test(async() => {
           let derived;
           try {
@@ -46,15 +53,16 @@ function define_tests() {
           }
           assert_false (derived === undefined, "Key derivation failed.");
       }, "Key derivation using a " + algorithmName + " generated keys.");
-  });
+  }
 
   return importKeys(pkcs8, spki, sizes)
   .then(function(results) {
       publicKeys = results.publicKeys;
       privateKeys = results.privateKeys;
       noDeriveKeyKeys = results.noDeriveKeyKeys;
+      ecdhKeys = results.ecdhKeys;
 
-      Object.keys(sizes).forEach(function(algorithmName) {
+      {
           // Basic success case
           promise_test(function(test) {
               return subtle.deriveKey({name: algorithmName, public: publicKeys[algorithmName]}, privateKeys[algorithmName], {name: "HMAC", hash: "SHA-256", length: 256}, true, ["sign", "verify"])
@@ -102,11 +110,7 @@ function define_tests() {
 
           // - wrong algorithm
           promise_test(function(test) {
-              publicKey = publicKeys["X25519"];
-              if (algorithmName === "X25519") {
-                  publicKey = publicKeys["X448"];
-              }
-              return subtle.deriveKey({name: algorithmName, public: publicKey}, privateKeys[algorithmName], {name: "HMAC", hash: "SHA-256", length: 256}, true, ["sign", "verify"])
+              return subtle.deriveKey({name: algorithmName, public: ecdhKeys[algorithmName]}, privateKeys[algorithmName], {name: "HMAC", hash: "SHA-256", length: 256}, true, ["sign", "verify"])
               .then(function(key) {return crypto.subtle.exportKey("raw", key);})
               .then(function(exportedKey) {
                   assert_unreached("deriveKey succeeded but should have failed with InvalidAccessError");
@@ -161,16 +165,17 @@ function define_tests() {
                   });
               });
           }, algorithmName + " public property value is a secret key");
-      });
+      }
   });
 
   function importKeys(pkcs8, spki, sizes) {
       var privateKeys = {};
       var publicKeys = {};
       var noDeriveKeyKeys = {};
+      var ecdhPublicKeys = {};
 
       var promises = [];
-      Object.keys(pkcs8).forEach(function(algorithmName) {
+      {
           var operation = subtle.importKey("pkcs8", pkcs8[algorithmName],
                                           {name: algorithmName},
                                           false, ["deriveBits", "deriveKey"])
@@ -180,8 +185,8 @@ function define_tests() {
                             privateKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
-      Object.keys(pkcs8).forEach(function(algorithmName) {
+      }
+      {
           var operation = subtle.importKey("pkcs8", pkcs8[algorithmName],
                                           {name: algorithmName},
                                           false, ["deriveBits"])
@@ -191,8 +196,8 @@ function define_tests() {
                             noDeriveKeyKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
-      Object.keys(spki).forEach(function(algorithmName) {
+      }
+      {
           var operation = subtle.importKey("spki", spki[algorithmName],
                                           {name: algorithmName},
                                           false, [])
@@ -202,10 +207,18 @@ function define_tests() {
                             publicKeys[algorithmName] = null;
                           });
           promises.push(operation);
-      });
+      }
+      {
+          var operation = subtle.importKey("spki", ecSPKI,
+                                           {name: "ECDH", namedCurve: "P-256"},
+                                           false, [])
+                          .then(function(key) {
+                              ecdhPublicKeys[algorithmName] = key;
+                          });
+      }
 
       return Promise.all(promises)
-             .then(function(results) {return {privateKeys: privateKeys, publicKeys: publicKeys, noDeriveKeyKeys: noDeriveKeyKeys}});
+             .then(function(results) {return {privateKeys: privateKeys, publicKeys: publicKeys, noDeriveKeyKeys: noDeriveKeyKeys, ecdhKeys: ecdhPublicKeys}});
   }
 
   // Compares two ArrayBuffer or ArrayBufferView objects. If bitCount is
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve25519.https.any.js
new file mode 100644
index 00000000000000..91390ba5c2a17a
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve25519.https.any.js
@@ -0,0 +1,10 @@
+// META: title=WebCryptoAPI: deriveKey() Using ECDH with CFRG Elliptic Curves
+// META: script=cfrg_curves_bits_fixtures.js
+// META: script=cfrg_curves_keys.js
+
+// Define subtests from a `promise_test` to ensure the harness does not
+// complete before the subtests are available. `explicit_done` cannot be used
+// for this purpose because the global `done` function is automatically invoked
+// by the WPT infrastructure in dedicated worker tests defined using the
+// "multi-global" pattern.
+promise_test(define_tests_25519, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve448.https.any.js
similarity index 89%
rename from test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.https.any.js
rename to test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve448.https.any.js
index 96658a56e81da9..b34e366376a70f 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/cfrg_curves_keys_curve448.https.any.js
@@ -7,4 +7,4 @@
 // for this purpose because the global `done` function is automatically invoked
 // by the WPT infrastructure in dedicated worker tests defined using the
 // "multi-global" pattern.
-promise_test(define_tests, 'setup - define tests');
+promise_test(define_tests_448, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.https.any.js
new file mode 100644
index 00000000000000..5edc832b6163bd
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.https.any.js
@@ -0,0 +1,9 @@
+// META: title=WebCryptoAPI: deriveKey() Using HKDF and PBKDF2 from an ECDH key
+// META: script=derive_key_and_encrypt.js
+// META: script=../util/helpers.js
+
+// Test imported from WebKit's source, defined to check the impact of the
+// 'Get Key Length' behavior of HKDF and PBKDF2, which should return 'null'
+// in both cases, in the 'deriveKey' operation.
+// https://bugs.webkit.org/show_bug.cgi?id=282096
+promise_test(define_tests, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.js
new file mode 100644
index 00000000000000..5963a852fcfbe1
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derive_key_and_encrypt.js
@@ -0,0 +1,49 @@
+let iv = new Uint8Array(Array(12).keys());
+let salt = new Uint8Array(Array(10).keys());
+let plaintext = new Uint8Array(Array(100).keys());
+
+function define_tests() {
+    importKeys().then((keys) => {
+        // Make sure that ecdh produces the same shared secret and the same encryption results using a key derived from that secret.
+        keys.forEach(keyData => {
+            promise_test(async() => {
+                let hkdfKey = await crypto.subtle.deriveKey({name: "ECDH", public: keyData.publicKey }, keyData.privateKey, { name: "HKDF", hash: "" , salt: new Uint8Array(), info: new Uint8Array() }, false, ["deriveKey"]);
+                let aesKey = await crypto.subtle.deriveKey({name: "HKDF", hash: "SHA-256", salt: salt, info: plaintext}, hkdfKey, {name:"AES-GCM", length: 256}, true, ["encrypt", "decrypt"]);
+                let result = await crypto.subtle.encrypt({ name: "AES-GCM", iv: iv }, aesKey, plaintext);
+                assert_equals(bytesToHexString(result), "a6280c522670eaf82f6564afbeb20a5b3f2d4e13c5596f6df3dcff8c34cb2118d2770fb24d83cfac5079c323118485bb01170292ee41eb82b07208f4840478fea3771d8922785c476ba06c2a0b933fc1661431419530a916ad4468545d1af5004a1149fea241c2ff1582ee58a8b7d79935de5def");
+            }, "HKDF derivation of a ECDH key " + keyData.test);
+            promise_test(async() => {
+                let pkdf2Key = await crypto.subtle.deriveKey({name: "ECDH", public: keyData.publicKey }, keyData.privateKey, { name: "PBKDF2", hash: "" , salt: new Uint8Array(), iterations: 32 }, false, ["deriveKey"]);
+                let aesKey = await crypto.subtle.deriveKey({name: "PBKDF2", hash: "SHA-256", salt: salt, iterations: 32 }, pkdf2Key, { name:"AES-GCM", length: 256 }, true, ["encrypt", "decrypt"]);
+                let result = await crypto.subtle.encrypt({ name: "AES-GCM", iv: iv }, aesKey, plaintext);
+                assert_equals(bytesToHexString(result), "c6201dfbb6fa92c1c246f6ce52f8f1c037f087efde41bac7f6485a2a8207623d2d3825b9cbe8ef864a90378667ed25544ce44cd2904bd96c19f0eeb611d626185165a8afb4e52f95700d7880f83939a42712fc4e377f198c01a61b397b76c3a4b93d932c321084bbef33332169dea09458b27df3");
+            }, "PBKDF2 derivation of a ECDH key " + keyData.test);
+        });
+    }, (e) => {
+        assert_unreached("Setup failed: " + e.message);
+    });
+
+    return Promise.resolve("define_tests");
+}
+
+async function importKeys() {
+    // "ECDSA" with a 'P-256' curve
+    let keyData = [
+        hexStringToUint8Array("308187020100301306072a8648ce3d020106082a8648ce3d030107046d306b0201010420fe77a808a7109ba5ceb93ebebad2c84a714d864ad29b62d6537e1969035c0079a144034200042684c752eef1c927a80c74e8b02ce459f848b5977f37fd878b36dae632be9a6cadd56126e404a4f75c535e5769d95b49fb1106f784f3d231b776d1f4d57927ce"),
+        hexStringToUint8Array("042684c752eef1c927a80c74e8b02ce459f848b5977f37fd878b36dae632be9a6cadd56126e404a4f75c535e5769d95b49fb1106f784f3d231b776d1f4d57927ce"),
+        hexStringToUint8Array("308187020100301306072a8648ce3d020106082a8648ce3d030107046d306b020101042067521ccd1f85516118182bca3394c273bab9ce5cd6265105559e325e01f2df1ca144034200043042d8698882f2b59de972390d3fc9277e2e677a6c560148017c9475218fda1b38f76f7645fbcaf3d03e6259d080204fbafb04731b6ad53cb25c3d35d95b7c73"),
+        hexStringToUint8Array("043042d8698882f2b59de972390d3fc9277e2e677a6c560148017c9475218fda1b38f76f7645fbcaf3d03e6259d080204fbafb04731b6ad53cb25c3d35d95b7c73"),
+    ];
+    let extractable = true;
+    var allKeys = await Promise.all([
+        crypto.subtle.importKey("pkcs8", keyData[0], {name: "ECDH", namedCurve: "P-256"}, extractable, ["deriveKey", 'deriveBits']),
+        crypto.subtle.importKey("raw", keyData[1], {name: "ECDH", namedCurve: "P-256"}, extractable, []),
+        crypto.subtle.importKey("pkcs8", keyData[2], {name: "ECDH", namedCurve: "P-256"}, extractable, ["deriveKey", 'deriveBits']),
+        crypto.subtle.importKey("raw", keyData[3], {name: "ECDH", namedCurve: "P-256"}, extractable, []),
+    ]);
+    // Test cases defined combining public and private keys of each key-pair.
+    return [
+        { test: 1, publicKey: allKeys[3], privateKey: allKeys[0] },
+        { test: 2, publicKey: allKeys[1], privateKey: allKeys[2] }
+    ];
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.https.any.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.https.any.js
new file mode 100644
index 00000000000000..0aee2e3c172d30
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.https.any.js
@@ -0,0 +1,11 @@
+// META: title=WebCryptoAPI: deriveBits() tests for the 'length' parameter
+// META: script=derived_bits_length.js
+// META: script=derived_bits_length_vectors.js
+// META: script=derived_bits_length_testcases.js
+
+// Define subtests from a `promise_test` to ensure the harness does not
+// complete before the subtests are available. `explicit_done` cannot be used
+// for this purpose because the global `done` function is automatically invoked
+// by the WPT infrastructure in dedicated worker tests defined using the
+// "multi-global" pattern.
+promise_test(define_tests, 'setup - define tests');
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.js
new file mode 100644
index 00000000000000..5a7ed7eb50a0a0
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length.js
@@ -0,0 +1,36 @@
+function define_tests() {
+    // May want to test prefixed implementations.
+    var subtle = self.crypto.subtle;
+
+    Object.keys(testCases).forEach(algorithm => {
+        let testData = algorithms[algorithm];
+        testCases[algorithm].forEach(testParam => {
+            promise_test(async() => {
+                let derivedBits, privateKey, publicKey;
+                try {
+                    privateKey = await subtle.importKey(testData.privateKey.format, testData.privateKey.data, testData.importAlg, false, ["deriveBits"]);
+                    if (testData.deriveAlg.public !== undefined) {
+                        publicKey = await subtle.importKey(testData.publicKey.format, testData.publicKey.data, testData.importAlg, false, []);
+                        testData.deriveAlg.public = publicKey;
+                    }
+                    if (testParam.length === "omitted")
+                        derivedBits = await subtle.deriveBits(testData.deriveAlg, privateKey);
+                    else
+                        derivedBits = await subtle.deriveBits(testData.deriveAlg, privateKey, testParam.length);
+                    if (testParam.expected === undefined) {
+                        assert_unreached("deriveBits should have thrown an OperationError exception.");
+                    }
+                    assert_array_equals(new Uint8Array(derivedBits), testParam.expected, "Derived bits do not match the expected result.");
+                } catch (err) {
+                    if (err instanceof AssertionError || testParam.expected !== undefined) {
+                        throw err;
+                    }
+                    assert_true(privateKey !== undefined, "Key should be valid.");
+                    assert_equals(err.name, "OperationError", "deriveBits correctly threw OperationError: " + err.message);
+                }
+            }, algorithm + " derivation with " + testParam.length + " as 'length' parameter");
+        });
+    });
+
+    return Promise.resolve("define_tests");
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_testcases.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_testcases.js
new file mode 100644
index 00000000000000..2679fa79e2a044
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_testcases.js
@@ -0,0 +1,38 @@
+var testCases = {
+    "HKDF": [
+        {length: 256, expected: algorithms["HKDF"].derivation},
+        {length: 384, expected: algorithms["HKDF"].derivation384},
+        {length: 230, expected: undefined}, // should throw an exception, not multiple of 8
+        {length: 0, expected: emptyArray},
+        {length: null, expected: undefined }, // should throw an exception
+        {length: undefined, expected: undefined }, // should throw an exception
+        {length: "omitted", expected: undefined }, // default value is null, so should throw
+    ],
+    "PBKDF2": [
+        {length: 256, expected: algorithms["PBKDF2"].derivation},
+        {length: 384, expected: algorithms["PBKDF2"].derivation384},
+        {length: 230, expected: undefined}, // should throw an exception, not multiple of 8
+        {length: 0, expected: emptyArray},
+        {length: null, expected: undefined }, // should throw an exception
+        {length: undefined, expected: undefined }, // should throw an exception
+        {length: "omitted", expected: undefined }, // default value is null, so should throw
+    ],
+    "ECDH": [
+        {length: 256, expected: algorithms["ECDH"].derivation},
+        {length: 384, expected: undefined}, // should throw an exception, bigger than the output size
+        {length: 230, expected: algorithms["ECDH"].derivation230},
+        {length: 0, expected: emptyArray},
+        {length: null, expected: algorithms["ECDH"].derivation},
+        {length: undefined, expected: algorithms["ECDH"].derivation},
+        {length: "omitted", expected: algorithms["ECDH"].derivation }, // default value is null
+    ],
+    "X25519": [
+        {length: 256, expected: algorithms["X25519"].derivation},
+        {length: 384, expected: undefined}, // should throw an exception, bigger than the output size
+        {length: 230, expected: algorithms["X25519"].derivation230},
+        {length: 0, expected: emptyArray},
+        {length: null, expected: algorithms["X25519"].derivation},
+        {length: undefined, expected: algorithms["X25519"].derivation},
+        {length: "omitted", expected: algorithms["X25519"].derivation }, // default value is null
+    ],
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_vectors.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_vectors.js
new file mode 100644
index 00000000000000..391f81d1871d76
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/derived_bits_length_vectors.js
@@ -0,0 +1,41 @@
+const emptyArray = new Uint8Array([]);
+const rawKey = new Uint8Array([85, 115, 101, 114, 115, 32, 115, 104, 111, 117, 108, 100, 32, 112, 105, 99, 107, 32, 108, 111, 110, 103, 32, 112, 97, 115, 115, 112, 104, 114, 97, 115, 101, 115, 32, 40, 110, 111, 116, 32, 117, 115, 101, 32, 115, 104, 111, 114, 116, 32, 112, 97, 115, 115, 119, 111, 114, 100, 115, 41, 33]);
+const salt = new Uint8Array([83, 111, 100, 105, 117, 109, 32, 67, 104, 108, 111, 114, 105, 100, 101, 32, 99, 111, 109, 112, 111, 117, 110, 100]);
+const info = new Uint8Array([72, 75, 68, 70, 32, 101, 120, 116, 114, 97, 32, 105, 110, 102, 111]);
+
+var algorithms = {
+    "HKDF": {
+        importAlg: {name: "HKDF"},
+        privateKey: {format: "raw", data: rawKey},
+        deriveAlg: {name: "HKDF", salt: salt, hash: "SHA-256", info: info},
+        derivation: new Uint8Array([49, 183, 214, 133, 48, 168, 99, 231, 23, 192, 129, 202, 105, 23, 182, 134, 80, 179, 221, 154, 41, 243, 6, 6, 226, 202, 209, 153, 190, 193, 77, 19]),
+        derivation384: new Uint8Array([49, 183, 214, 133, 48, 168, 99, 231, 23, 192, 129, 202, 105, 23, 182, 134, 80, 179, 221, 154, 41, 243, 6, 6, 226, 202, 209, 153, 190, 193, 77, 19, 165, 50, 181, 8, 254, 59, 122, 199, 25, 224,146, 248, 105, 105, 75, 84]),
+        derivation230: undefined,
+    },
+    "PBKDF2": {
+        importAlg: {name: "PBKDF2"},
+        privateKey: {format: "raw", data: rawKey},
+        deriveAlg: {name: "PBKDF2", salt: salt, hash: "SHA-256", iterations: 100000},
+        derivation: new Uint8Array([17, 153, 45, 139, 129, 51, 17, 36, 76, 84, 75, 98, 41, 41, 69, 226, 8, 212, 3, 206, 189, 107, 149, 82, 161, 165, 98, 6, 93, 153, 88, 234]),
+        derivation384: new Uint8Array([17, 153, 45, 139, 129, 51, 17, 36, 76, 84, 75, 98, 41, 41, 69, 226, 8, 212, 3, 206, 189, 107, 149, 82, 161, 165, 98, 6, 93, 153, 88, 234, 39, 104, 8, 112, 222, 57, 166, 47, 102, 146, 195, 59, 219, 239, 238, 47]),
+        derivation230: undefined,
+    },
+    "ECDH": {
+        importAlg: {name: "ECDH",  namedCurve: "P-256"},
+        privateKey: {format: "pkcs8", data: new Uint8Array([48, 129, 135, 2, 1, 0, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 4, 109, 48, 107, 2, 1, 1, 4, 32, 15, 247, 79, 232, 241, 202, 175, 97, 92, 206, 241, 29, 217, 53, 114, 87, 98, 217, 216, 65, 236, 186, 185, 94, 170, 38, 68, 123, 52, 100, 245, 113, 161, 68, 3, 66, 0, 4, 140, 96, 11, 44, 102, 25, 45, 97, 158, 39, 210, 37, 107, 59, 151, 118, 178, 141, 30, 5, 246, 13, 234, 189, 98, 174, 123, 154, 211, 157, 224, 217, 59, 4, 102, 109, 199, 119, 14, 126, 207, 13, 211, 203, 203, 211, 110, 221, 107, 94, 220, 153, 81, 7, 55, 161, 237, 104, 46, 205, 112, 244, 10, 47])},
+        publicKey: {format: "spki", data: new Uint8Array([48, 89, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 3, 66, 0, 4, 154, 116, 32, 120, 126, 95, 77, 105, 211, 232, 34, 114, 115, 1, 109, 56, 224, 71, 129, 133, 223, 127, 238, 156, 142, 103, 60, 202, 211, 79, 126, 128, 254, 49, 141, 182, 221, 107, 119, 218, 99, 32, 165, 246, 151, 89, 9, 68, 23, 177, 52, 239, 138, 139, 116, 193, 101, 4, 57, 198, 115, 0, 90, 61])},
+        deriveAlg: {name: "ECDH", public: new Uint8Array ([])},
+        derivation: new Uint8Array([14, 143, 60, 77, 177, 178, 162, 131, 115, 90, 0, 220, 87, 31, 26, 232, 151, 28, 227, 35, 250, 17, 131, 137, 203, 95, 65, 196, 59, 61, 181, 161]),
+        derivation384: undefined,
+        derivation230: new Uint8Array([14, 143, 60, 77, 177, 178, 162, 131, 115, 90, 0, 220, 87, 31, 26, 232, 151, 28, 227, 35, 250, 17, 131, 137, 203, 95, 65, 196, 56]),
+    },
+    "X25519": {
+        importAlg: {name: "X25519"},
+        privateKey: {format: "pkcs8", data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97])},
+        publicKey: {format: "spki", data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6])},
+        deriveAlg: {name: "X25519", public: new Uint8Array ([])},
+        derivation: new Uint8Array([39, 104, 64, 157, 250, 185, 158, 194, 59, 140, 137, 185, 63, 245, 136, 2, 149, 247, 97, 118, 8, 143, 137, 228, 61, 254, 190, 126, 161, 149, 0, 8]),
+        derivation384: undefined,
+        derivation230: new Uint8Array([39, 104, 64, 157, 250, 185, 158, 194, 59, 140, 137, 185, 63, 245, 136, 2, 149, 247, 97, 118, 8, 143, 137, 228, 61, 254, 190, 126, 160]),
+    }
+};
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/ecdh_bits.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/ecdh_bits.js
index cb9747a529fd53..36b29c20a282ab 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/ecdh_bits.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/ecdh_bits.js
@@ -55,25 +55,6 @@ function define_tests() {
                 });
             }, namedCurve + " mixed case parameters");
 
-            // Null length
-            // "Null" is not valid per the current spec
-            //   - https://github.com/w3c/webcrypto/issues/322
-            //   - https://github.com/w3c/webcrypto/issues/329
-            //
-            // Proposal for a spec change:
-            //   - https://github.com/w3c/webcrypto/pull/345
-            //
-            // This test case may be replaced by these new tests:
-            //   - https://github.com/web-platform-tests/wpt/pull/43400
-            promise_test(function(test) {
-                return subtle.deriveBits({name: "ECDH", public: publicKeys[namedCurve]}, privateKeys[namedCurve], null)
-                .then(function(derivation) {
-                    assert_true(equalBuffers(derivation, derivations[namedCurve]), "Derived correct bits");
-                }, function(err) {
-                    assert_unreached("deriveBits failed with error " + err.name + ": " + err.message);
-                });
-            }, namedCurve + " with null length");
-
             // Shorter than entire derivation per algorithm
             promise_test(function(test) {
                 return subtle.deriveBits({name: "ECDH", public: publicKeys[namedCurve]}, privateKeys[namedCurve], 8 * sizes[namedCurve] - 32)
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/hkdf.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/hkdf.js
index 3903da5cddff94..0384f88ec73e43 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/hkdf.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/hkdf.js
@@ -45,13 +45,13 @@ function define_tests() {
                             });
                         }, testName);
 
-                        // 0 length (OperationError)
+                        // 0 length
                         subsetTest(promise_test, function(test) {
                             return subtle.deriveBits(algorithm, baseKeys[derivedKeySize], 0)
                             .then(function(derivation) {
                                 assert_equals(derivation.byteLength, 0, "Derived correctly empty key");
                             }, function(err) {
-                                assert_equals(err.name, "OperationError", "deriveBits with 0 length correctly threw OperationError: " + err.message);
+                                assert_unreached("deriveBits failed with error " + err.name + ": " + err.message);
                             });
                         }, testName + " with 0 length");
 
@@ -139,25 +139,6 @@ function define_tests() {
                             });
                         }, testName + " with missing info");
 
-                        // length null (OperationError)
-                        // "Null" is not valid per the current spec
-                        //   - https://github.com/w3c/webcrypto/issues/322
-                        //   - https://github.com/w3c/webcrypto/issues/329
-                        //
-                        // Proposal for a spec change:
-                        //   - https://github.com/w3c/webcrypto/pull/345
-                        //
-                        // This test case may be replaced by these new tests:
-                        //   - https://github.com/web-platform-tests/wpt/pull/43400
-                        subsetTest(promise_test, function(test) {
-                            return subtle.deriveBits(algorithm, baseKeys[derivedKeySize], null)
-                            .then(function(derivation) {
-                                assert_unreached("null length should have thrown an OperationError");
-                            }, function(err) {
-                                assert_equals(err.name, "OperationError", "deriveBits with null length correctly threw OperationError: " + err.message);
-                            });
-                        }, testName + " with null length");
-
                         // length not multiple of 8 (OperationError)
                         subsetTest(promise_test, function(test) {
                             return subtle.deriveBits(algorithm, baseKeys[derivedKeySize], 44)
diff --git a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/pbkdf2.js b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/pbkdf2.js
index 4e4ae79d800a40..38cf3b1bfe952c 100644
--- a/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/pbkdf2.js
+++ b/test/fixtures/wpt/WebCryptoAPI/derive_bits_keys/pbkdf2.js
@@ -42,6 +42,16 @@ function define_tests() {
                             });
                         }, testName);
 
+                        // 0 length
+                        subsetTest(promise_test, function(test) {
+                            return subtle.deriveBits({name: "PBKDF2", salt: salts[saltSize], hash: hashName, iterations: parseInt(iterations)}, baseKeys[passwordSize], 0)
+                            .then(function(derivation) {
+                                assert_true(equalBuffers(derivation.byteLength, 0, "Derived correctly empty key"));
+                            }, function(err) {
+                                assert_unreached("deriveBits failed with error " + err.name + ": " + err.message);
+                            });
+                        }, testName + " with 0 length");
+
                         // Check for correct deriveKey results for every kind of
                         // key that can be created by the deriveKeys operation.
                         derivedKeyTypes.forEach(function(derivedKeyType) {
@@ -103,36 +113,6 @@ function define_tests() {
 
                         });
 
-                        // Test various error conditions for deriveBits below:
-                        // length null (OperationError)
-                        // "Null" is not valid per the current spec
-                        //   - https://github.com/w3c/webcrypto/issues/322
-                        //   - https://github.com/w3c/webcrypto/issues/329
-                        //
-                        // Proposal for a spec change:
-                        //   - https://github.com/w3c/webcrypto/pull/345
-                        //
-                        // This test case may be replaced by these new tests:
-                        //   - https://github.com/web-platform-tests/wpt/pull/43400
-                        subsetTest(promise_test, function(test) {
-                            return subtle.deriveBits({name: "PBKDF2", salt: salts[saltSize], hash: hashName, iterations: parseInt(iterations)}, baseKeys[passwordSize], null)
-                            .then(function(derivation) {
-                                assert_unreached("null length should have thrown an OperationError");
-                            }, function(err) {
-                                assert_equals(err.name, "OperationError", "deriveBits with null length correctly threw OperationError: " + err.message);
-                            });
-                        }, testName + " with null length");
-
-                        // 0 length (OperationError)
-                        subsetTest(promise_test, function(test) {
-                            return subtle.deriveBits({name: "PBKDF2", salt: salts[saltSize], hash: hashName, iterations: parseInt(iterations)}, baseKeys[passwordSize], 0)
-                            .then(function(derivation) {
-                                assert_unreached("0 length should have thrown an OperationError");
-                            }, function(err) {
-                                assert_equals(err.name, "OperationError", "deriveBits with 0 length correctly threw OperationError: " + err.message);
-                            });
-                        }, testName + " with 0 length");
-
                         // length not multiple of 8 (OperationError)
                         subsetTest(promise_test, function(test) {
                             return subtle.deriveBits({name: "PBKDF2", salt: salts[saltSize], hash: hashName, iterations: parseInt(iterations)}, baseKeys[passwordSize], 44)
diff --git a/test/fixtures/wpt/WebCryptoAPI/digest/digest.https.any.js b/test/fixtures/wpt/WebCryptoAPI/digest/digest.https.any.js
index 379d9311f30247..3b0972b1f2bf7d 100644
--- a/test/fixtures/wpt/WebCryptoAPI/digest/digest.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/digest/digest.https.any.js
@@ -118,6 +118,20 @@
         });
     });
 
+    // Call digest() with empty algorithm object
+    Object.keys(sourceData).forEach(function(size) {
+        promise_test(function(test) {
+            var promise = subtle.digest({}, sourceData[size])
+            .then(function(result) {
+                assert_unreached("digest() with missing algorithm name should have thrown a TypeError");
+            }, function(err) {
+                assert_equals(err.name, "TypeError", "Missing algorithm name should cause TypeError")
+            });
+
+            return promise;
+        }, "empty algorithm object with " + size);
+    });
+
 
     done();
 
diff --git a/test/fixtures/wpt/WebCryptoAPI/generateKey/failures.js b/test/fixtures/wpt/WebCryptoAPI/generateKey/failures.js
index e0f0279a69bb88..deaac636a99be5 100644
--- a/test/fixtures/wpt/WebCryptoAPI/generateKey/failures.js
+++ b/test/fixtures/wpt/WebCryptoAPI/generateKey/failures.js
@@ -166,6 +166,14 @@ function run_test(algorithmNames) {
         });
     });
 
+    // Empty algorithm should fail with TypeError
+    allValidUsages(["decrypt", "sign", "deriveBits"], true, []) // Small search space, shouldn't matter because should fail before used
+        .forEach(function(usages) {
+            [false, true, "RED", 7].forEach(function(extractable){
+                testError({}, extractable, usages, "TypeError", "Empty algorithm");
+            });
+        });
+
 
     // Algorithms normalize okay, but usages bad (though not empty).
     // It shouldn't matter what other extractable is. Should fail
diff --git a/test/fixtures/wpt/WebCryptoAPI/generateKey/successes.js b/test/fixtures/wpt/WebCryptoAPI/generateKey/successes.js
index e1c1665b511030..a9a168e1adbf72 100644
--- a/test/fixtures/wpt/WebCryptoAPI/generateKey/successes.js
+++ b/test/fixtures/wpt/WebCryptoAPI/generateKey/successes.js
@@ -5,7 +5,8 @@ function run_test(algorithmNames, slowTest) {
     setup({explicit_timeout: true});
 
 // These tests check that generateKey successfully creates keys
-// when provided any of a wide set of correct parameters.
+// when provided any of a wide set of correct parameters
+// and that they can be exported afterwards.
 //
 // There are a lot of combinations of possible parameters,
 // resulting in a very large number of tests
@@ -68,9 +69,32 @@ function run_test(algorithmNames, slowTest) {
                 } else {
                     assert_goodCryptoKey(result, algorithm, extractable, usages, "secret");
                 }
+                return result;
             }, function(err) {
-                assert_unreached("Threw an unexpected error: " + err.toString());
-            });
+                assert_unreached("generateKey threw an unexpected error: " + err.toString());
+            })
+            .then(async function (result) {
+                if (resultType === "CryptoKeyPair") {
+                    await Promise.all([
+                        subtle.exportKey('jwk', result.publicKey),
+                        subtle.exportKey('spki', result.publicKey),
+                        result.publicKey.algorithm.name.startsWith('RSA') ? undefined : subtle.exportKey('raw', result.publicKey),
+                        ...(extractable ? [
+                            subtle.exportKey('jwk', result.privateKey),
+                            subtle.exportKey('pkcs8', result.privateKey),
+                        ] : [])
+                    ]);
+                } else {
+                    if (extractable) {
+                        await Promise.all([
+                            subtle.exportKey('raw', result),
+                            subtle.exportKey('jwk', result),
+                        ]);
+                    }
+                }
+            }, function(err) {
+                assert_unreached("exportKey threw an unexpected error: " + err.toString());
+            })
         }, testTag + ": generateKey" + parameterString(algorithm, extractable, usages));
     }
 
diff --git a/test/fixtures/wpt/WebCryptoAPI/getRandomValues.any.js b/test/fixtures/wpt/WebCryptoAPI/getRandomValues.any.js
index 1a3370ea13d2c0..574134eb76dcd8 100644
--- a/test/fixtures/wpt/WebCryptoAPI/getRandomValues.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/getRandomValues.any.js
@@ -1,4 +1,15 @@
 // Step 1.
+test(function() {
+    assert_throws_dom("TypeMismatchError", function() {
+        self.crypto.getRandomValues(new Float16Array(6))
+    }, "Float16Array")
+
+    assert_throws_dom("TypeMismatchError", function() {
+        const len = 65536 / Float16Array.BYTES_PER_ELEMENT + 1;
+        self.crypto.getRandomValues(new Float16Array(len));
+    }, "Float16Array (too long)")
+}, "Float16 arrays");
+
 test(function() {
     assert_throws_dom("TypeMismatchError", function() {
         self.crypto.getRandomValues(new Float32Array(6))
@@ -57,4 +68,10 @@ for (const array of arrays) {
     test(function() {
         assert_true(self.crypto.getRandomValues(new ctor(0)).length == 0)
     }, "Null arrays: " + array);
+
+    test(function() {
+        class Buffer extends ctor {}
+        // Must not throw for the test to pass
+        self.crypto.getRandomValues(new Buffer(256));
+    }, "Subclass of " + array);
 }
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/crashtests/importKey-unsettled-promise.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/crashtests/importKey-unsettled-promise.https.any.js
new file mode 100644
index 00000000000000..0ceeea390ebf97
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/crashtests/importKey-unsettled-promise.https.any.js
@@ -0,0 +1,17 @@
+// META: title=WebCryptoAPI: Assure promise returned by importKey is settled.
+// META: timeout=long
+// META: script=/common/gc.js
+
+'use strict';
+
+promise_test(async () => {
+  const jwkKey = {};
+  const extractable = true;
+  crypto.subtle.importKey("jwk", jwkKey, {name: "UNSUPPORTED", hash: "SHA-224"}, extractable, []).then(
+      () => { assert_unreached("Unsupported algorithm should cause promise rejection")},
+      (err) => {
+        assert_equals(err.name, "NotSupportedError");
+      });
+  await garbageCollect();
+})
+
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDH.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDH.https.any.js
new file mode 100644
index 00000000000000..423d399f19def4
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDH.https.any.js
@@ -0,0 +1,10 @@
+// META: title=WebCryptoAPI: importKey() for Failures
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=ec_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
+
+// Setup: define the correct behaviors that should be sought, and create
+// helper functions that generate all possible test parameters for
+// different situations.
+run_test(["ECDH"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDSA.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDSA.https.any.js
new file mode 100644
index 00000000000000..527940798a42a5
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_ECDSA.https.any.js
@@ -0,0 +1,10 @@
+// META: title=WebCryptoAPI: importKey() for Failures
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=ec_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
+
+// Setup: define the correct behaviors that should be sought, and create
+// helper functions that generate all possible test parameters for
+// different situations.
+run_test(["ECDSA"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_fixtures.js b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_fixtures.js
new file mode 100644
index 00000000000000..a2d25e816cbd73
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/ec_importKey_failures_fixtures.js
@@ -0,0 +1,225 @@
+// Setup: define the correct behaviors that should be sought, and create
+// helper functions that generate all possible test parameters for
+// different situations.
+function getValidKeyData(algorithm) {
+    return validKeyData[algorithm.namedCurve];
+}
+
+function getBadKeyLengthData(algorithm) {
+    return badKeyLengthData[algorithm.namedCurve];
+}
+
+function getMissingJWKFieldKeyData(algorithm) {
+    // The curve doesn't affect when testing for missing JWK fields.
+    return missingJWKFieldKeyData["P-521"];
+}
+
+function getMismatchedJWKKeyData(algorithm) {
+    // TODO: Implement test cases where the public key doesn't match the private key.
+    return [];
+}
+
+function getMismatchedKtyField(algorithm) {
+    return mismatchedKtyField[algorithm.name];
+}
+
+function getMismatchedCrvField(algorithm) {
+    return mismatchedCrvField[algorithm.name];
+}
+
+var validKeyData = {
+    "P-521": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 129, 155, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 35, 3, 129, 134, 0, 4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56, 251]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56, 251]),
+        },
+        {
+            format:"pkcs8",
+            data: new Uint8Array([48, 129, 238, 2, 1, 0, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 35, 4, 129, 214, 48, 129, 211, 2, 1, 1, 4, 66, 0, 244, 8, 117, 131, 104, 186, 147, 15, 48, 247, 106, 224, 84, 254, 92, 210, 206, 127, 218, 44, 159, 118, 166, 212, 54, 207, 117, 214, 108, 68, 11, 254, 99, 49, 199, 193, 114, 161, 36, 120, 25, 60, 130, 81, 72, 123, 201, 18, 99, 250, 80, 33, 127, 133, 255, 99, 111, 89, 205, 84, 110, 58, 180, 131, 180, 161, 129, 137, 3, 129, 134, 0, 4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56, 251]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-521",
+                x: "AVb0efjfHiCn_8BM5CDD4VSuJRmWvuQvA0uE1Bt0PzTkXzEbgTqc3sjNpZu7vTHUYLMpJSHnwbci5WZ8A9svrnU_",
+                y: "AVAXNs_iRzlDINjkr8L9ObWpMxBhuB4iQSgrnheJGCK1t54FL0WXtZZD_Tk3nFG9USXE9IvD8CXOPNNpUyhsyzj7",
+                d: "APQIdYNoupMPMPdq4FT-XNLOf9osn3am1DbPddZsRAv-YzHHwXKhJHgZPIJRSHvJEmP6UCF_hf9jb1nNVG46tIO0"
+            }
+        }
+    ],
+    "P-256": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 89, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97, 232]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97, 232]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 129, 135, 2, 1, 0, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 4, 109, 48, 107, 2, 1, 1, 4, 32, 19, 211, 58, 45, 90, 191, 156, 249, 235, 178, 31, 248, 96, 212, 174, 254, 110, 86, 231, 119, 144, 244, 222, 233, 180, 8, 132, 235, 211, 53, 68, 234, 161, 68, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97, 232]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-256",
+                x: "0hCwpvnZ8BKGgFi0P6T0cQGFQ7ugDJJQ35JXwqyuXdE",
+                y: "zgN1UtSBRQzjm00QlXAbF1v6s0uObAmeGPHBmDWDYeg",
+                d: "E9M6LVq_nPnrsh_4YNSu_m5W53eQ9N7ptAiE69M1ROo"
+            }
+        },
+    ],
+    "P-384": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 118, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 34, 3, 98, 0, 4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172, 232]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172, 232]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 129, 182, 2, 1, 0, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 34, 4, 129, 158, 48, 129, 155, 2, 1, 1, 4, 48, 69, 55, 181, 153, 7, 132, 211, 194, 210, 46, 150, 168, 249, 47, 161, 170, 73, 46, 232, 115, 229, 118, 164, 21, 130, 225, 68, 24, 60, 152, 136, 209, 14, 107, 158, 180, 206, 212, 178, 204, 64, 18, 228, 172, 94, 168, 64, 115, 161, 100, 3, 98, 0, 4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172, 232]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-384",
+                x: "IZwU1mYXs27G2IVrOFtzp000T9iude8EZDXdpU47RL1fvevR0I3Wni19wdwhjLQ1",
+                y: "vSgTjMd4M3qEL2vWGyQOdCSfJGZ8KlgQp2v8KOAzX4imUB3sAZdtqFr7AIactqzo",
+                d: "RTe1mQeE08LSLpao-S-hqkku6HPldqQVguFEGDyYiNEOa560ztSyzEAS5KxeqEBz"
+            }
+        }
+    ]
+};
+
+// Removed just the last byte.
+var badKeyLengthData = {
+    "P-521": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 129, 155, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 35, 3, 129, 134, 0, 4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56]),
+        },
+        {
+            format:"pkcs8",
+            data: new Uint8Array([48, 129, 238, 2, 1, 0, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 35, 4, 129, 214, 48, 129, 211, 2, 1, 1, 4, 66, 0, 244, 8, 117, 131, 104, 186, 147, 15, 48, 247, 106, 224, 84, 254, 92, 210, 206, 127, 218, 44, 159, 118, 166, 212, 54, 207, 117, 214, 108, 68, 11, 254, 99, 49, 199, 193, 114, 161, 36, 120, 25, 60, 130, 81, 72, 123, 201, 18, 99, 250, 80, 33, 127, 133, 255, 99, 111, 89, 205, 84, 110, 58, 180, 131, 180, 161, 129, 137, 3, 129, 134, 0, 4, 1, 86, 244, 121, 248, 223, 30, 32, 167, 255, 192, 76, 228, 32, 195, 225, 84, 174, 37, 25, 150, 190, 228, 47, 3, 75, 132, 212, 27, 116, 63, 52, 228, 95, 49, 27, 129, 58, 156, 222, 200, 205, 165, 155, 187, 189, 49, 212, 96, 179, 41, 37, 33, 231, 193, 183, 34, 229, 102, 124, 3, 219, 47, 174, 117, 63, 1, 80, 23, 54, 207, 226, 71, 57, 67, 32, 216, 228, 175, 194, 253, 57, 181, 169, 51, 16, 97, 184, 30, 34, 65, 40, 43, 158, 23, 137, 24, 34, 181, 183, 158, 5, 47, 69, 151, 181, 150, 67, 253, 57, 55, 156, 81, 189, 81, 37, 196, 244, 139, 195, 240, 37, 206, 60, 211, 105, 83, 40, 108, 203, 56]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-521",
+                x: "AVb0efjfHiCn_8BM5CDD4VSuJRmWvuQvA0uE1Bt0PzTkXzEbgTqc3sjNpZu7vTHUYLMpJSHnwbci5WZ8A9svrnU",
+                y: "AVAXNs_iRzlDINjkr8L9ObWpMxBhuB4iQSgrnheJGCK1t54FL0WXtZZD_Tk3nFG9USXE9IvD8CXOPNNpUyhsyzj7",
+                d: "APQIdYNoupMPMPdq4FT-XNLOf9osn3am1DbPddZsRAv-YzHHwXKhJHgZPIJRSHvJEmP6UCF_hf9jb1nNVG46tIO0"
+            }
+        }
+    ],
+    "P-256": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 89, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 129, 135, 2, 1, 0, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 4, 109, 48, 107, 2, 1, 1, 4, 32, 19, 211, 58, 45, 90, 191, 156, 249, 235, 178, 31, 248, 96, 212, 174, 254, 110, 86, 231, 119, 144, 244, 222, 233, 180, 8, 132, 235, 211, 53, 68, 234, 161, 68, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-256",
+                x: "0hCwpvnZ8BKGgFi0P6T0cQGFQ7ugDJJQ35JXwqyuXd",
+                y: "zgN1UtSBRQzjm00QlXAbF1v6s0uObAmeGPHBmDWDYeg",
+                d: "E9M6LVq_nPnrsh_4YNSu_m5W53eQ9N7ptAiE69M1ROo"
+            }
+        },
+    ],
+    "P-384": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 118, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 34, 3, 98, 0, 4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 129, 182, 2, 1, 0, 48, 16, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 5, 43, 129, 4, 0, 34, 4, 129, 158, 48, 129, 155, 2, 1, 1, 4, 48, 69, 55, 181, 153, 7, 132, 211, 194, 210, 46, 150, 168, 249, 47, 161, 170, 73, 46, 232, 115, 229, 118, 164, 21, 130, 225, 68, 24, 60, 152, 136, 209, 14, 107, 158, 180, 206, 212, 178, 204, 64, 18, 228, 172, 94, 168, 64, 115, 161, 100, 3, 98, 0, 4, 33, 156, 20, 214, 102, 23, 179, 110, 198, 216, 133, 107, 56, 91, 115, 167, 77, 52, 79, 216, 174, 117, 239, 4, 100, 53, 221, 165, 78, 59, 68, 189, 95, 189, 235, 209, 208, 141, 214, 158, 45, 125, 193, 220, 33, 140, 180, 53, 189, 40, 19, 140, 199, 120, 51, 122, 132, 47, 107, 214, 27, 36, 14, 116, 36, 159, 36, 102, 124, 42, 88, 16, 167, 107, 252, 40, 224, 51, 95, 136, 166, 80, 29, 236, 1, 151, 109, 168, 90, 251, 0, 134, 156, 182, 172]),
+        },
+        {
+            format: "jwk",
+            data: {
+                kty: "EC",
+                crv: "P-384",
+                x: "IZwU1mYXs27G2IVrOFtzp000T9iude8EZDXdpU47RL1fvevR0I3Wni19wdwhjLQ",
+                y: "vSgTjMd4M3qEL2vWGyQOdCSfJGZ8KlgQp2v8KOAzX4imUB3sAZdtqFr7AIactqzo",
+                d: "RTe1mQeE08LSLpao-S-hqkku6HPldqQVguFEGDyYiNEOa560ztSyzEAS5KxeqEBz"
+            }
+        }
+    ]
+};
+
+var missingJWKFieldKeyData = {
+    "P-521": [
+        {
+            param: "x",
+            data: {
+                kty: "EC",
+                crv: "P-521",
+                y: "AVAXNs_iRzlDINjkr8L9ObWpMxBhuB4iQSgrnheJGCK1t54FL0WXtZZD_Tk3nFG9USXE9IvD8CXOPNNpUyhsyzj7",
+                d: "APQIdYNoupMPMPdq4FT-XNLOf9osn3am1DbPddZsRAv-YzHHwXKhJHgZPIJRSHvJEmP6UCF_hf9jb1nNVG46tIO0"
+            }
+        },
+        {
+            param: "kty",
+            data: {
+                crv: "P-521",
+                x: "AVb0efjfHiCn_8BM5CDD4VSuJRmWvuQvA0uE1Bt0PzTkXzEbgTqc3sjNpZu7vTHUYLMpJSHnwbci5WZ8A9svrnU_",
+                y: "AVAXNs_iRzlDINjkr8L9ObWpMxBhuB4iQSgrnheJGCK1t54FL0WXtZZD_Tk3nFG9USXE9IvD8CXOPNNpUyhsyzj7",
+                d: "APQIdYNoupMPMPdq4FT-XNLOf9osn3am1DbPddZsRAv-YzHHwXKhJHgZPIJRSHvJEmP6UCF_hf9jb1nNVG46tIO0"
+            }
+        },
+        {
+            param: "crv",
+            data: {
+                kty: "EC",
+                x: "AVb0efjfHiCn_8BM5CDD4VSuJRmWvuQvA0uE1Bt0PzTkXzEbgTqc3sjNpZu7vTHUYLMpJSHnwbci5WZ8A9svrnU_",
+                y: "AVAXNs_iRzlDINjkr8L9ObWpMxBhuB4iQSgrnheJGCK1t54FL0WXtZZD_Tk3nFG9USXE9IvD8CXOPNNpUyhsyzj7",
+                d: "APQIdYNoupMPMPdq4FT-XNLOf9osn3am1DbPddZsRAv-YzHHwXKhJHgZPIJRSHvJEmP6UCF_hf9jb1nNVG46tIO0"
+            }
+        }
+    ]
+};
+
+// The 'kty' field doesn't match the key algorithm.
+var mismatchedKtyField =  {
+    "P-521": "OKP",
+    "P-256": "OKP",
+    "P-384": "OKP",
+}
+
+// The 'kty' field doesn't match the key algorithm.
+var mismatchedCrvField =  {
+    "P-521": "P-256",
+    "P-256": "P-384",
+    "P-384": "P-521",
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures.js b/test/fixtures/wpt/WebCryptoAPI/import_export/importKey_failures.js
similarity index 61%
rename from test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures.js
rename to test/fixtures/wpt/WebCryptoAPI/import_export/importKey_failures.js
index ebdb73616d6581..453461a8771f51 100644
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures.js
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/importKey_failures.js
@@ -20,8 +20,10 @@ function run_test(algorithmNames) {
     var allTestVectors = [ // Parameters that should work for importKey / exportKey
         {name: "Ed25519", privateUsages: ["sign"], publicUsages: ["verify"]},
         {name: "Ed448", privateUsages: ["sign"], publicUsages: ["verify"]},
+        {name: "ECDSA", privateUsages: ["sign"], publicUsages: ["verify"]},
         {name: "X25519",  privateUsages: ["deriveKey", "deriveBits"], publicUsages: []},
         {name: "X448",  privateUsages: ["deriveKey", "deriveBits"], publicUsages: []},
+        {name: "ECDH",  privateUsages: ["deriveKey", "deriveBits"], publicUsages: []}
     ];
 
     var testVectors = [];
@@ -109,6 +111,10 @@ function run_test(algorithmNames) {
         return [];
     }
 
+    function isPrivateKey(data) {
+        return data.d !== undefined;
+    }
+
 // Now test for properly handling errors
 // - Unsupported algorithm
 // - Bad usages for algorithm
@@ -121,8 +127,8 @@ function run_test(algorithmNames) {
     // due to SyntaxError
     testVectors.forEach(function(vector) {
         var name = vector.name;
-        validKeyData.forEach(function(test) {
-            allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).forEach(function(test) {
                 invalidUsages(validUsages(vector, test.format, test.data)).forEach(function(usages) {
                     [true, false].forEach(function(extractable) {
                         testError(test.format, algorithm, test.data, name, usages, extractable, "SyntaxError", "Bad usages");
@@ -136,8 +142,8 @@ function run_test(algorithmNames) {
     // Should fail due to SyntaxError
     testVectors.forEach(function(vector) {
         var name = vector.name;
-        validKeyData.filter((test) => test.format === 'pkcs8' || (test.format === 'jwk' && test.data.d)).forEach(function(test) {
-            allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).filter((test) => test.format === 'pkcs8' || (test.format === 'jwk' && isPrivateKey(test.data))).forEach(function(test) {
                 [true, false].forEach(function(extractable) {
                     testError(test.format, algorithm, test.data, name, [/* Empty usages */], extractable, "SyntaxError", "Empty usages");
                 });
@@ -145,11 +151,11 @@ function run_test(algorithmNames) {
         });
     });
 
-    // Algorithms normalize okay, usages ok. The length of the key must thouw a DataError exception.
+    // Algorithms normalize okay, usages ok. The length of the key must throw a DataError exception.
     testVectors.forEach(function(vector) {
         var name = vector.name;
-        badKeyLengthData.forEach(function(test) {
-            allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getBadKeyLengthData(algorithm).forEach(function(test) {
                 allValidUsages(validUsages(vector, test.format, test.data)).forEach(function(usages) {
                     [true, false].forEach(function(extractable) {
                         testError(test.format, algorithm, test.data, name, usages, extractable, "DataError", "Bad key length");
@@ -159,11 +165,11 @@ function run_test(algorithmNames) {
         });
     });
 
-    // Algorithms normalize okay, usages ok and valid key. The lack of the mandatory JWK parameter must throw a syntax error.
+    // Algorithms normalize okay, usages ok and valid key. The lack of the mandatory JWK parameter must throw a DataError exception.
     testVectors.forEach(function(vector) {
         var name = vector.name;
-        missingJWKFieldKeyData.forEach(function(test) {
-            allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+           getMissingJWKFieldKeyData(algorithm).forEach(function(test) {
                 allValidUsages(validUsages(vector, 'jwk', test.data)).forEach(function(usages) {
                     [true, false].forEach(function(extractable) {
                         testError('jwk', algorithm, test.data, name, usages, extractable, "DataError", "Missing JWK '" + test.param + "' parameter");
@@ -176,8 +182,8 @@ function run_test(algorithmNames) {
     // Algorithms normalize okay, usages ok and valid key. The public key is not compatible with the private key.
     testVectors.forEach(function(vector) {
         var name = vector.name;
-        invalidJWKKeyData.forEach(function(data) {
-            allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getMismatchedJWKKeyData(algorithm).forEach(function(data) {
                 allValidUsages(vector.privateUsages).forEach(function(usages) {
                     [true].forEach(function(extractable) {
                         testError('jwk', algorithm, data, name, usages, extractable, "DataError", "Invalid key pair");
@@ -186,4 +192,79 @@ function run_test(algorithmNames) {
             });
         });
     });
+
+    // Missing mandatory "name" field on algorithm
+    testVectors.forEach(function(vector) {
+        var name = vector.name;
+        // We just need *some* valid keydata, so pick the first available algorithm.
+        var algorithm = allAlgorithmSpecifiersFor(name)[0];
+        getValidKeyData(algorithm).forEach(function(test) {
+            validUsages(vector, test.format, test.data).forEach(function(usages) {
+                [true, false].forEach(function(extractable) {
+                    testError(test.format, {}, test.data, name, usages, extractable, "TypeError", "Missing algorithm name");
+                });
+            });
+        });
+    });
+
+    // The 'kty' field is not correct.
+    testVectors.forEach(function(vector) {
+        var name = vector.name;
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).forEach(function(test) {
+                if (test.format === "jwk") {
+                    var data = {crv: test.data.crv, kty: test.data.kty, d: test.data.d, x: test.data.x, d: test.data.d};
+                    data.kty = getMismatchedKtyField(algorithm);
+                    var usages =  validUsages(vector, 'jwk', test.data);
+                    testError('jwk', algorithm, data, name, usages, true, "DataError", "Invalid 'kty' field");
+                }
+            });
+        });
+    });
+
+    // The 'ext' field is not correct.
+    testVectors.forEach(function(vector) {
+        var name = vector.name;
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).forEach(function(test) {
+                if (test.format === "jwk") {
+                    var data = {crv: test.data.crv, kty: test.data.kty, d: test.data.d, x: test.data.x, d: test.data.d};
+                    data.ext = false;
+                    var usages =  validUsages(vector, 'jwk', test.data);
+                    testError('jwk', algorithm, data, name, usages, true, "DataError", "Import from a non-extractable");
+                }
+            });
+        });
+    });
+
+    // The 'use' field is incorrect.
+    testVectors.forEach(function(vector) {
+        var name = vector.name;
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).forEach(function(test) {
+                if (test.format === "jwk") {
+                    var data = {crv: test.data.crv, kty: test.data.kty, d: test.data.d, x: test.data.x, d: test.data.d};
+                    data.use = "invalid";
+                    var usages =  validUsages(vector, 'jwk', test.data);
+                    if (usages.length !== 0)
+                        testError('jwk', algorithm, data, name, usages, true, "DataError", "Invalid 'use' field");
+                }
+            });
+        });
+    });
+
+    // The 'crv' field is incorrect.
+    testVectors.forEach(function(vector) {
+        var name = vector.name;
+        allAlgorithmSpecifiersFor(name).forEach(function(algorithm) {
+            getValidKeyData(algorithm).forEach(function(test) {
+                if (test.format === "jwk") {
+                    var data = {crv: test.data.crv, kty: test.data.kty, d: test.data.d, x: test.data.x, d: test.data.d};
+                    data.crv = getMismatchedCrvField(algorithm)
+                    var usages =  validUsages(vector, 'jwk', test.data);
+                    testError('jwk', algorithm, data, name, usages, true, "DataError", "Invalid 'crv' field");
+                }
+            });
+        });
+    });
 }
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.https.any.js
deleted file mode 100644
index a56bd31cbe14b1..00000000000000
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.https.any.js
+++ /dev/null
@@ -1,280 +0,0 @@
-// META: title=WebCryptoAPI: importKey() for OKP keys
-// META: timeout=long
-// META: script=../util/helpers.js
-
-// Test importKey and exportKey for OKP algorithms. Only "happy paths" are
-// currently tested - those where the operation should succeed.
-
-    var subtle = crypto.subtle;
-
-    var keyData = {
-        "Ed25519": {
-            spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204]),
-            raw: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204]),
-            pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31]),
-            jwk: {
-                crv: "Ed25519",
-                d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
-                x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
-                kty: "OKP"
-            }
-        },
-
-        "Ed448": {
-            spki: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
-            raw: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
-            pkcs8: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
-            jwk: {
-                crv: "Ed448",
-                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-                kty: "OKP"
-            }
-        },
-
-        "X25519": {
-            spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
-            raw: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
-            pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97]),
-            jwk: {
-                crv: "X25519",
-                d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
-                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-                kty: "OKP"
-            }
-        },
-
-        "X448": {
-            spki: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
-            raw: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
-            pkcs8: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146, 158]),
-            jwk: {
-                crv: "X448",
-                d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
-                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-                kty: "OKP"
-            }
-        },
-
-    };
-
-    // combinations to test
-    var testVectors = [
-        {name: "Ed25519", privateUsages: ["sign"], publicUsages: ["verify"]},
-        {name: "Ed448", privateUsages: ["sign"], publicUsages: ["verify"]},
-        {name: "X25519",  privateUsages: ["deriveKey", "deriveBits"], publicUsages: []},
-        {name: "X448",  privateUsages: ["deriveKey", "deriveBits"], publicUsages: []},
-    ];
-
-    // TESTS ARE HERE:
-    // Test every test vector, along with all available key data
-    testVectors.forEach(function(vector) {
-        [true, false].forEach(function(extractable) {
-
-            // Test public keys first
-            allValidUsages(vector.publicUsages, true).forEach(function(usages) {
-                ['spki', 'jwk', 'raw'].forEach(function(format) {
-                    var algorithm = {name: vector.name};
-                    var data = keyData[vector.name];
-                    if (format === "jwk") { // Not all fields used for public keys
-                        data = {jwk: {kty: keyData[vector.name].jwk.kty, crv: keyData[vector.name].jwk.crv, x: keyData[vector.name].jwk.x}};
-                    }
-
-                    testFormat(format, algorithm, data, vector.name, usages, extractable);
-
-                    // Test for https://github.com/WICG/webcrypto-secure-curves/pull/24
-                    if (format === "jwk" && extractable) {
-                        testJwkAlgBehaviours(algorithm, data.jwk, vector.name, usages);
-                    }
-                });
-
-            });
-
-            // Next, test private keys
-            allValidUsages(vector.privateUsages).forEach(function(usages) {
-                ['pkcs8', 'jwk'].forEach(function(format) {
-                    var algorithm = {name: vector.name};
-                    var data = keyData[vector.name];
-
-                    testFormat(format, algorithm, data, vector.name, usages, extractable);
-
-                    // Test for https://github.com/WICG/webcrypto-secure-curves/pull/24
-                    if (format === "jwk" && extractable) {
-                        testJwkAlgBehaviours(algorithm, data.jwk, vector.name, usages);
-                    }
-                });
-            });
-        });
-    });
-
-
-    // Test importKey with a given key format and other parameters. If
-    // extrable is true, export the key and verify that it matches the input.
-    function testFormat(format, algorithm, keyData, keySize, usages, extractable) {
-        promise_test(function(test) {
-            return subtle.importKey(format, keyData[format], algorithm, extractable, usages).
-            then(function(key) {
-                assert_equals(key.constructor, CryptoKey, "Imported a CryptoKey object");
-                assert_goodCryptoKey(key, algorithm, extractable, usages, (format === 'pkcs8' || (format === 'jwk' && keyData[format].d)) ? 'private' : 'public');
-                if (!extractable) {
-                    return;
-                }
-
-                return subtle.exportKey(format, key).
-                then(function(result) {
-                    if (format !== "jwk") {
-                        assert_true(equalBuffers(keyData[format], result), "Round trip works");
-                    } else {
-                        assert_true(equalJwk(keyData[format], result), "Round trip works");
-                    }
-                }, function(err) {
-                    assert_unreached("Threw an unexpected error: " + err.toString());
-                });
-            }, function(err) {
-                assert_unreached("Threw an unexpected error: " + err.toString());
-            });
-        }, "Good parameters: " + keySize.toString() + " bits " + parameterString(format, keyData[format], algorithm, extractable, usages));
-    }
-
-    // Test importKey/exportKey "alg" behaviours, alg is ignored upon import and alg is missing for Ed25519 and Ed448 JWK export
-    // https://github.com/WICG/webcrypto-secure-curves/pull/24
-    function testJwkAlgBehaviours(algorithm, keyData, crv, usages) {
-        promise_test(function(test) {
-            return subtle.importKey('jwk', { ...keyData, alg: 'this is ignored' }, algorithm, true, usages).
-            then(function(key) {
-                assert_equals(key.constructor, CryptoKey, "Imported a CryptoKey object");
-
-                return subtle.exportKey('jwk', key).
-                then(function(result) {
-                    assert_equals(Object.keys(result).length, keyData.d ? 6 : 5, "Correct number of JWK members");
-                    assert_equals(result.alg, undefined, 'No JWK "alg" member is present');
-                    assert_true(equalJwk(keyData, result), "Round trip works");
-                }, function(err) {
-                    assert_unreached("Threw an unexpected error: " + err.toString());
-                });
-            }, function(err) {
-                assert_unreached("Threw an unexpected error: " + err.toString());
-            });
-        }, "Good parameters with ignored JWK alg: " + crv.toString() + " " + parameterString('jwk', keyData, algorithm, true, usages));
-    }
-
-
-
-    // Helper methods follow:
-
-    // Are two array buffers the same?
-    function equalBuffers(a, b) {
-        if (a.byteLength !== b.byteLength) {
-            return false;
-        }
-
-        var aBytes = new Uint8Array(a);
-        var bBytes = new Uint8Array(b);
-
-        for (var i=0; i<a.byteLength; i++) {
-            if (aBytes[i] !== bBytes[i]) {
-                return false;
-            }
-        }
-
-        return true;
-    }
-
-    // Are two Jwk objects "the same"? That is, does the object returned include
-    // matching values for each property that was expected? It's okay if the
-    // returned object has extra methods; they aren't checked.
-    function equalJwk(expected, got) {
-        var fields = Object.keys(expected);
-        var fieldName;
-
-        for(var i=0; i<fields.length; i++) {
-            fieldName = fields[i];
-            if (!(fieldName in got)) {
-                return false;
-            }
-            if (expected[fieldName] !== got[fieldName]) {
-                return false;
-            }
-        }
-
-        return true;
-    }
-
-    // Build minimal Jwk objects from raw key data and algorithm specifications
-    function jwkData(keyData, algorithm) {
-        var result = {
-            kty: "oct",
-            k: byteArrayToUnpaddedBase64(keyData)
-        };
-
-        if (algorithm.name.substring(0, 3) === "AES") {
-            result.alg = "A" + (8 * keyData.byteLength).toString() + algorithm.name.substring(4);
-        } else if (algorithm.name === "HMAC") {
-            result.alg = "HS" + algorithm.hash.substring(4);
-        }
-        return result;
-    }
-
-    // Jwk format wants Base 64 without the typical padding at the end.
-    function byteArrayToUnpaddedBase64(byteArray){
-        var binaryString = "";
-        for (var i=0; i<byteArray.byteLength; i++){
-            binaryString += String.fromCharCode(byteArray[i]);
-        }
-        var base64String = btoa(binaryString);
-
-        return base64String.replace(/=/g, "");
-    }
-
-    // Convert method parameters to a string to uniquely name each test
-    function parameterString(format, data, algorithm, extractable, usages) {
-        if ("byteLength" in data) {
-            data = "buffer(" + data.byteLength.toString() + ")";
-        } else {
-            data = "object(" + Object.keys(data).join(", ") + ")";
-        }
-        var result = "(" +
-                        objectToString(format) + ", " +
-                        objectToString(data) + ", " +
-                        objectToString(algorithm) + ", " +
-                        objectToString(extractable) + ", " +
-                        objectToString(usages) +
-                     ")";
-
-        return result;
-    }
-
-    // Character representation of any object we may use as a parameter.
-    function objectToString(obj) {
-        var keyValuePairs = [];
-
-        if (Array.isArray(obj)) {
-            return "[" + obj.map(function(elem){return objectToString(elem);}).join(", ") + "]";
-        } else if (typeof obj === "object") {
-            Object.keys(obj).sort().forEach(function(keyName) {
-                keyValuePairs.push(keyName + ": " + objectToString(obj[keyName]));
-            });
-            return "{" + keyValuePairs.join(", ") + "}";
-        } else if (typeof obj === "undefined") {
-            return "undefined";
-        } else {
-            return obj.toString();
-        }
-
-        var keyValuePairs = [];
-
-        Object.keys(obj).sort().forEach(function(keyName) {
-            var value = obj[keyName];
-            if (typeof value === "object") {
-                value = objectToString(value);
-            } else if (typeof value === "array") {
-                value = "[" + value.map(function(elem){return objectToString(elem);}).join(", ") + "]";
-            } else {
-                value = value.toString();
-            }
-
-            keyValuePairs.push(keyName + ": " + value);
-        });
-
-        return "{" + keyValuePairs.join(", ") + "}";
-    }
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.js
new file mode 100644
index 00000000000000..8627f85e466199
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey.js
@@ -0,0 +1,220 @@
+var subtle = crypto.subtle;
+
+function runTests(algorithmName) {
+    var algorithm = {name: algorithmName};
+    var data = keyData[algorithmName];
+    var jwkData = {jwk: {kty: data.jwk.kty, crv: data.jwk.crv, x: data.jwk.x}};
+
+    [true, false].forEach(function(extractable) {
+        // Test public keys first
+        allValidUsages(data.publicUsages, true).forEach(function(usages) {
+            ['spki', 'jwk', 'raw'].forEach(function(format) {
+                if (format === "jwk") { // Not all fields used for public keys
+                    testFormat(format, algorithm, jwkData, algorithmName, usages, extractable);
+                    // Test for https://github.com/w3c/webcrypto/pull/401
+                    if (extractable) {
+                        testJwkAlgBehaviours(algorithm, jwkData.jwk, algorithmName, usages);
+                    }
+                } else {
+                    testFormat(format, algorithm, data, algorithmName, usages, extractable);
+                }
+            });
+
+        });
+
+        // Next, test private keys
+        allValidUsages(data.privateUsages).forEach(function(usages) {
+            ['pkcs8', 'jwk'].forEach(function(format) {
+                testFormat(format, algorithm, data, algorithmName, usages, extractable);
+
+                // Test for https://github.com/w3c/webcrypto/pull/401
+                if (format === "jwk" && extractable) {
+                    testJwkAlgBehaviours(algorithm, data.jwk, algorithmName, usages);
+                }
+            });
+        });
+    });
+}
+
+
+// Test importKey with a given key format and other parameters. If
+// extrable is true, export the key and verify that it matches the input.
+function testFormat(format, algorithm, keyData, keySize, usages, extractable) {
+    [algorithm, algorithm.name].forEach((alg) => {
+        promise_test(function(test) {
+            return subtle.importKey(format, keyData[format], alg, extractable, usages).
+                then(function(key) {
+                    assert_equals(key.constructor, CryptoKey, "Imported a CryptoKey object");
+                    assert_goodCryptoKey(key, algorithm, extractable, usages, (format === 'pkcs8' || (format === 'jwk' && keyData[format].d)) ? 'private' : 'public');
+                    if (!extractable) {
+                        return;
+                    }
+
+                    return subtle.exportKey(format, key).
+                        then(function(result) {
+                            if (format !== "jwk") {
+                                assert_true(equalBuffers(keyData[format], result), "Round trip works");
+                            } else {
+                                assert_true(equalJwk(keyData[format], result), "Round trip works");
+                            }
+                        }, function(err) {
+                            assert_unreached("Threw an unexpected error: " + err.toString());
+                        });
+                }, function(err) {
+                    assert_unreached("Threw an unexpected error: " + err.toString());
+                });
+        }, "Good parameters: " + keySize.toString() + " bits " + parameterString(format, keyData[format], alg, extractable, usages));
+    });
+}
+
+// Test importKey/exportKey "alg" behaviours (https://github.com/w3c/webcrypto/pull/401)
+// - alg is ignored for ECDH import
+// - TODO: alg is checked to be the algorithm.name or EdDSA for Ed25519 and Ed448 import
+// - alg is missing for ECDH export
+// - alg is the algorithm name for Ed25519 and Ed448 export
+function testJwkAlgBehaviours(algorithm, keyData, crv, usages) {
+    [algorithm, algorithm.name].forEach((alg) => {
+        (crv.startsWith('Ed') ? [algorithm.name, 'EdDSA'] : ['this is ignored']).forEach((jwkAlg) => {
+            promise_test(function(test) {
+                return subtle.importKey('jwk', { ...keyData, alg: jwkAlg }, alg, true, usages).
+                    then(function(key) {
+                        assert_equals(key.constructor, CryptoKey, "Imported a CryptoKey object");
+
+                        return subtle.exportKey('jwk', key).
+                            then(function(result) {
+                                let expectedKeys = crv.startsWith('Ed') ? 6 : 5
+                                if (keyData.d) expectedKeys++
+                                assert_equals(Object.keys(result).length, expectedKeys, "Correct number of JWK members");
+                                assert_equals(result.alg, crv.startsWith('Ed') ? algorithm.name : undefined, 'Expected JWK "alg" member');
+                                assert_true(equalJwk(keyData, result), "Round trip works");
+                            }, function(err) {
+                            assert_unreached("Threw an unexpected error: " + err.toString());
+                            });
+                    }, function(err) {
+                        assert_unreached("Threw an unexpected error: " + err.toString());
+                    });
+            }, 'Good parameters with JWK alg' + (crv.startsWith('Ed') ? ` ${jwkAlg}: ` : ': ') + crv.toString() + " " + parameterString('jwk', keyData, alg, true, usages, jwkAlg));
+        });
+    });
+}
+
+
+
+// Helper methods follow:
+
+// Are two array buffers the same?
+function equalBuffers(a, b) {
+    if (a.byteLength !== b.byteLength) {
+        return false;
+    }
+
+    var aBytes = new Uint8Array(a);
+    var bBytes = new Uint8Array(b);
+
+    for (var i=0; i<a.byteLength; i++) {
+        if (aBytes[i] !== bBytes[i]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+// Are two Jwk objects "the same"? That is, does the object returned include
+// matching values for each property that was expected? It's okay if the
+// returned object has extra methods; they aren't checked.
+function equalJwk(expected, got) {
+    var fields = Object.keys(expected);
+    var fieldName;
+
+    for(var i=0; i<fields.length; i++) {
+        fieldName = fields[i];
+        if (!(fieldName in got)) {
+            return false;
+        }
+        if (expected[fieldName] !== got[fieldName]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+// Build minimal Jwk objects from raw key data and algorithm specifications
+function jwkData(keyData, algorithm) {
+    var result = {
+        kty: "oct",
+        k: byteArrayToUnpaddedBase64(keyData)
+    };
+
+    if (algorithm.name.substring(0, 3) === "AES") {
+        result.alg = "A" + (8 * keyData.byteLength).toString() + algorithm.name.substring(4);
+    } else if (algorithm.name === "HMAC") {
+        result.alg = "HS" + algorithm.hash.substring(4);
+    }
+    return result;
+}
+
+// Jwk format wants Base 64 without the typical padding at the end.
+function byteArrayToUnpaddedBase64(byteArray){
+    var binaryString = "";
+    for (var i=0; i<byteArray.byteLength; i++){
+        binaryString += String.fromCharCode(byteArray[i]);
+    }
+    var base64String = btoa(binaryString);
+
+    return base64String.replace(/=/g, "");
+}
+
+// Convert method parameters to a string to uniquely name each test
+function parameterString(format, data, algorithm, extractable, usages) {
+    if ("byteLength" in data) {
+        data = "buffer(" + data.byteLength.toString() + ")";
+    } else {
+        data = "object(" + Object.keys(data).join(", ") + ")";
+    }
+    var result = "(" +
+        objectToString(format) + ", " +
+        objectToString(data) + ", " +
+        objectToString(algorithm) + ", " +
+        objectToString(extractable) + ", " +
+        objectToString(usages) +
+        ")";
+
+    return result;
+}
+
+// Character representation of any object we may use as a parameter.
+function objectToString(obj) {
+    var keyValuePairs = [];
+
+    if (Array.isArray(obj)) {
+        return "[" + obj.map(function(elem){return objectToString(elem);}).join(", ") + "]";
+    } else if (typeof obj === "object") {
+        Object.keys(obj).sort().forEach(function(keyName) {
+            keyValuePairs.push(keyName + ": " + objectToString(obj[keyName]));
+        });
+        return "{" + keyValuePairs.join(", ") + "}";
+    } else if (typeof obj === "undefined") {
+        return "undefined";
+    } else {
+        return obj.toString();
+    }
+
+    var keyValuePairs = [];
+
+    Object.keys(obj).sort().forEach(function(keyName) {
+        var value = obj[keyName];
+        if (typeof value === "object") {
+            value = objectToString(value);
+        } else if (typeof value === "array") {
+            value = "[" + value.map(function(elem){return objectToString(elem);}).join(", ") + "]";
+        } else {
+            value = value.toString();
+        }
+
+        keyValuePairs.push(keyName + ": " + value);
+    });
+
+    return "{" + keyValuePairs.join(", ") + "}";
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed25519.https.any.js
new file mode 100644
index 00000000000000..49656489d425dd
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed25519.https.any.js
@@ -0,0 +1,9 @@
+// META: title=WebCryptoAPI: importKey() for OKP keys
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=okp_importKey_fixtures.js
+// META: script=okp_importKey.js
+
+
+// Test importKey and exportKey for OKP algorithms.
+runTests("Ed25519");
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed448.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed448.https.any.js
new file mode 100644
index 00000000000000..5bb7460c1fbc93
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_Ed448.https.any.js
@@ -0,0 +1,9 @@
+// META: title=WebCryptoAPI: importKey() for OKP keys
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=okp_importKey_fixtures.js
+// META: script=okp_importKey.js
+
+
+// Test importKey and exportKey for OKP algorithms.
+runTests("Ed448");
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X25519.https.any.js
new file mode 100644
index 00000000000000..de1431d6cc2467
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X25519.https.any.js
@@ -0,0 +1,9 @@
+// META: title=WebCryptoAPI: importKey() for OKP keys
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=okp_importKey_fixtures.js
+// META: script=okp_importKey.js
+
+
+// Test importKey and exportKey for OKP algorithms.
+runTests("X25519");
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X448.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X448.https.any.js
new file mode 100644
index 00000000000000..f8552be3c826dc
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_X448.https.any.js
@@ -0,0 +1,9 @@
+// META: title=WebCryptoAPI: importKey() for OKP keys
+// META: timeout=long
+// META: script=../util/helpers.js
+// META: script=okp_importKey_fixtures.js
+// META: script=okp_importKey.js
+
+
+// Test importKey and exportKey for OKP algorithms.
+runTests("X448");
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed25519.https.any.js
index 7453a3a55052b3..e07b796df83d19 100644
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed25519.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed25519.https.any.js
@@ -1,110 +1,7 @@
 // META: title=WebCryptoAPI: importKey() for Failures
 // META: timeout=long
 // META: script=../util/helpers.js
-// META: script=okp_importKey_failures.js
-
-// Setup: define the correct behaviors that should be sought, and create
-// helper functions that generate all possible test parameters for
-// different situations.
-var validKeyData = [
-    {
-      format: "spki",
-      data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204])
-    },
-    {
-      format: "pkcs8",
-      data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31])
-    },
-    {
-      format: "raw",
-      data: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204])
-    },
-    {
-      format: "jwk",
-      data: {
-          crv: "Ed25519",
-          d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
-          x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
-          kty: "OKP"
-      },
-    },
-    {
-      format: "jwk",
-      data: {
-          crv: "Ed25519",
-          x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
-          kty: "OKP"
-      },
-    },
-];
-
-// Removed just the last byte.
-var badKeyLengthData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61])
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168])
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61])
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "Ed25519",
-            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB",
-            x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
-            kty: "OKP"
-        }
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "Ed25519",
-            x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPc",
-            kty: "OKP"
-        }
-    },
-];
-
-var missingJWKFieldKeyData = [
-    {
-        param: "x",
-        data: {
-            crv: "Ed25519",
-            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
-            kty: "OKP"
-        },
-    },
-    {
-        param: "kty",
-        data: {
-            crv: "Ed25519",
-            x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
-            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
-        },
-    },
-    {
-        param: "crv",
-        data: {
-            x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
-            kty: "OKP"
-        },
-    }
-];
-
-// The public key doesn't match the private key.
-var invalidJWKKeyData = [
-    {
-        crv: "Ed25519",
-        d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
-        x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
-        kty: "OKP"
-    },
-];
+// META: script=okp_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
 
 run_test(["Ed25519"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed448.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed448.https.any.js
index db2d47827af329..8ff3de5c79d3ab 100644
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed448.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_Ed448.https.any.js
@@ -1,111 +1,8 @@
 // META: title=WebCryptoAPI: importKey() for Failures
 // META: timeout=long
 // META: script=../util/helpers.js
-// META: script=okp_importKey_failures.js
+// META: script=okp_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
 
-// Setup: define the correct behaviors that should be sought, and create
-// helper functions that generate all possible test parameters for
-// different situations.
-var validKeyData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "Ed448",
-            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-            kty: "OKP"
-        },
-    },
-    {
-        format: "jwk",
-        data: {
-          crv: "Ed448",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-            kty: "OKP"
-        },
-    },
-];
-
-// Removed just the last byte.
-var badKeyLengthData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90]),
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29]),
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90]),
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "Ed448",
-            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-            kty: "OKP"
-        },
-    },
-    {
-        format: "jwk",
-        data: {
-          crv: "Ed448",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalq",
-            kty: "OKP"
-        },
-    },
-];
-
-var missingJWKFieldKeyData = [
-    {
-        param: "x",
-        data: {
-            crv: "Ed448",
-            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-            kty: "OKP"
-        }
-    },
-    {
-        param: "kty",
-        data: {
-            crv: "Ed448",
-            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-        }
-    },
-    {
-        param: "crv",
-        data: {
-            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
-            kty: "OKP"
-        }
-    }
-];
-
-// The public key doesn't match the private key.
-var invalidJWKKeyData = [
-    {
-        crv: "Ed448",
-        d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
-        x: "X9dEm1m0Yf0s54fsYWrUah2hNCSFpw4fig6nXYDpZ3jt8SR2m0bHBhvWeD3x5Q9s0foavq_oJWGA",
-        kty: "OKP"
-    },
-];
 
 run_test(["Ed448"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X25519.https.any.js
index d4d099f7656033..a0116bee888512 100644
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X25519.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X25519.https.any.js
@@ -1,110 +1,8 @@
 // META: title=WebCryptoAPI: importKey() for Failures
 // META: timeout=long
 // META: script=../util/helpers.js
-// META: script=okp_importKey_failures.js
+// META: script=okp_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
 
-// Setup: define the correct behaviors that should be sought, and create
-// helper functions that generate all possible test parameters for
-// different situations.
-var validKeyData = [
-    {
-      format: "spki",
-      data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
-    },
-    {
-      format: "pkcs8",
-      data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97]),
-    },
-    {
-      format: "raw",
-      data: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
-    },
-    {
-      format: "jwk",
-      data: {
-          crv: "X25519",
-          d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
-          x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-          kty: "OKP"
-      },
-    },
-    {
-      format: "jwk",
-      data: {
-          crv: "X25519",
-          x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-          kty: "OKP"
-      },
-    },
-];
-
-// Removed just the last byte.
-var badKeyLengthData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151]),
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255]),
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151]),
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "X25519",
-            x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lw",
-            kty: "OKP"
-        }
-    },
-    {
-      format: "jwk",
-      data: {
-          crv: "X25519",
-          d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2",
-          x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-          kty: "OKP"
-      },
-    },
-];
-
-var missingJWKFieldKeyData = [
-    {
-        param: "x",
-        data: {
-            crv: "X25519",
-            d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
-            kty: "OKP"
-        },
-    },
-    {
-        param: "kty",
-        data: {
-            crv: "X25519",
-            d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
-            x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-        },
-    },
-    {
-        param: "crv",
-        data: {
-            x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
-            kty: "OKP"
-        },
-    }
-];
-
-// The public key doesn't match the private key.
-var invalidJWKKeyData = [
-    {
-        crv: "X25519",
-        d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
-        x: "hSDwCYkwp1R0i33ctD73Wg2_Og0mOBr066SpjqqbTmo",
-        kty: "OKP"
-    },
-];
 
 run_test(["X25519"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X448.https.any.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X448.https.any.js
index d8ac902e672566..eccce68fac731c 100644
--- a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X448.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_X448.https.any.js
@@ -1,111 +1,8 @@
 // META: title=WebCryptoAPI: importKey() for Failures
 // META: timeout=long
 // META: script=../util/helpers.js
-// META: script=okp_importKey_failures.js
+// META: script=okp_importKey_failures_fixtures.js
+// META: script=importKey_failures.js
 
-// Setup: define the correct behaviors that should be sought, and create
-// helper functions that generate all possible test parameters for
-// different situations.
-var validKeyData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146, 158]),
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "X448",
-            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-            kty: "OKP"
-        },
-    },
-    {
-        format: "jwk",
-        data: {
-          crv: "X448",
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-            kty: "OKP"
-        },
-    },
-];
-
-// Removed just the last byte.
-var badKeyLengthData = [
-    {
-        format: "spki",
-        data: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206]),
-    },
-    {
-        format: "pkcs8",
-        data: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146]),
-    },
-    {
-        format: "raw",
-        data: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206]),
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "X448",
-            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp",
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-            kty: "OKP"
-        },
-    },
-    {
-        format: "jwk",
-        data: {
-            crv: "X448",
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm",
-            kty: "OKP"
-        },
-    },
-];
-
-var missingJWKFieldKeyData = [
-    {
-        param: "x",
-        data: {
-            crv: "X448",
-            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
-            kty: "OKP"
-        }
-    },
-    {
-        param: "kty",
-        data: {
-            crv: "X448",
-            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-        }
-    },
-    {
-        param: "crv",
-        data: {
-            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
-            kty: "OKP"
-        }
-    }
-];
-
-// The public key doesn't match the private key.
-var invalidJWKKeyData = [
-    {
-
-        crv: "X448",
-        kty: "OKP",
-        d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
-        x: "mwj3zDG34+Z9ItWuoSEHSic70rg94Jxj+qc9LCLF2bvINmRyQdlT1AxbEtqIEg1TF3+A5TLEH6A",
-    },
-];
 
 run_test(["X448"]);
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_fixtures.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_fixtures.js
new file mode 100644
index 00000000000000..9bedddc5c5944a
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_failures_fixtures.js
@@ -0,0 +1,438 @@
+// Setup: define the correct behaviors that should be sought, and create
+// helper functions that generate all possible test parameters for
+// different situations.
+function getValidKeyData(algorithm) {
+    return validKeyData[algorithm.name];
+}
+
+function getBadKeyLengthData(algorithm) {
+    return badKeyLengthData[algorithm.name];
+}
+
+function getMissingJWKFieldKeyData(algorithm) {
+    return missingJWKFieldKeyData[algorithm.name];
+}
+
+function getMismatchedJWKKeyData(algorithm) {
+    return mismatchedJWKKeyData[algorithm.name];
+}
+
+function getMismatchedKtyField(algorithm) {
+    return mismatchedKtyField[algorithm.name];
+}
+
+function getMismatchedCrvField(algorithm) {
+    return mismatchedCrvField[algorithm.name];
+}
+
+var validKeyData = {
+    "Ed25519": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204])
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31])
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204])
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed25519",
+                d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+                x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed25519",
+                x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
+                kty: "OKP"
+            },
+        }
+    ],
+    "Ed448": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed448",
+                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed448",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+                kty: "OKP"
+            },
+        },
+    ],
+    "X25519": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X25519",
+                d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X25519",
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+                kty: "OKP"
+            },
+        },
+    ],
+    "X448": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146, 158]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X448",
+                d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X448",
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+                kty: "OKP"
+            },
+        },
+    ],
+};
+
+// Removed just the last byte.
+var badKeyLengthData = {
+    "Ed25519": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61])
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168])
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61])
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed25519",
+                d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB",
+                x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
+                kty: "OKP"
+            }
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed25519",
+                x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPc",
+                kty: "OKP"
+            }
+        }
+    ],
+    "Ed448": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed448",
+                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "Ed448",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalq",
+                kty: "OKP"
+            },
+        },
+    ],
+    "X25519": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X25519",
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lw",
+                kty: "OKP"
+            }
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X25519",
+                d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2",
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+                kty: "OKP"
+            },
+        },
+    ],
+    "X448": [
+        {
+            format: "spki",
+            data: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206]),
+        },
+        {
+            format: "pkcs8",
+            data: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146]),
+        },
+        {
+            format: "raw",
+            data: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206]),
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X448",
+                d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp",
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+                kty: "OKP"
+            },
+        },
+        {
+            format: "jwk",
+            data: {
+                crv: "X448",
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm",
+                kty: "OKP"
+            },
+        },
+    ],
+};
+
+var missingJWKFieldKeyData =  {
+    "Ed25519": [
+        {
+            param: "x",
+            data: {
+                crv: "Ed25519",
+                d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+                kty: "OKP"
+            },
+        },
+        {
+            param: "kty",
+            data: {
+                crv: "Ed25519",
+                x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
+                d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+            },
+        },
+        {
+            param: "crv",
+            data: {
+                x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
+                kty: "OKP"
+            },
+        }
+    ],
+    "Ed448": [
+        {
+            param: "x",
+            data: {
+                crv: "Ed448",
+                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+                kty: "OKP"
+            }
+        },
+        {
+            param: "kty",
+            data: {
+                crv: "Ed448",
+                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+            }
+        },
+        {
+            param: "crv",
+            data: {
+                d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+                x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+                kty: "OKP"
+            }
+        }
+    ],
+    "X25519": [
+        {
+            param: "x",
+            data: {
+                crv: "X25519",
+                d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+                kty: "OKP"
+            },
+        },
+        {
+            param: "kty",
+            data: {
+                crv: "X25519",
+                d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+            },
+        },
+        {
+            param: "crv",
+            data: {
+                x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+                kty: "OKP"
+            },
+        }
+    ],
+    "X448": [
+        {
+            param: "x",
+            data: {
+                crv: "X448",
+                d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+                kty: "OKP"
+            }
+        },
+        {
+            param: "kty",
+            data: {
+                crv: "X448",
+                d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+            }
+        },
+        {
+            param: "crv",
+            data: {
+                x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+                kty: "OKP"
+            }
+        }
+    ],
+};
+
+// The public key doesn't match the private key.
+var mismatchedJWKKeyData =  {
+    "Ed25519": [
+        {
+            crv: "Ed25519",
+            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+            x: "11qYAYKxCrfVS_7TyWQHOg7hcvPapiMlrwIaaPcHURo",
+            kty: "OKP"
+        },
+    ],
+    "Ed448": [
+        {
+            crv: "Ed448",
+            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+            x: "X9dEm1m0Yf0s54fsYWrUah2hNCSFpw4fig6nXYDpZ3jt8SR2m0bHBhvWeD3x5Q9s0foavq_oJWGA",
+            kty: "OKP"
+        },
+    ],
+    "X25519": [
+        {
+            crv: "X25519",
+            d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+            x: "hSDwCYkwp1R0i33ctD73Wg2_Og0mOBr066SpjqqbTmo",
+            kty: "OKP"
+        },
+    ],
+    "X448": [
+        {
+
+            crv: "X448",
+            kty: "OKP",
+            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+            x: "mwj3zDG34+Z9ItWuoSEHSic70rg94Jxj+qc9LCLF2bvINmRyQdlT1AxbEtqIEg1TF3+A5TLEH6A",
+        },
+    ],
+}
+
+// The 'kty' field doesn't match the key algorithm.
+var mismatchedKtyField =  {
+    "Ed25519": "EC",
+    "X25519": "EC",
+    "Ed448": "EC",
+    "X448": "EC",
+}
+
+// The 'kty' field doesn't match the key algorithm.
+var mismatchedCrvField =  {
+    "Ed25519": "X25519",
+    "X25519": "Ed448",
+    "Ed448": "X25519",
+    "X448": "Ed25519",
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_fixtures.js b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_fixtures.js
new file mode 100644
index 00000000000000..58b41d52601eb9
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/import_export/okp_importKey_fixtures.js
@@ -0,0 +1,58 @@
+var keyData = {
+    "Ed25519": {
+        privateUsages: ["sign"],
+        publicUsages: ["verify"],
+        spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204]),
+        raw: new Uint8Array([216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204]),
+        pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31]),
+        jwk: {
+            crv: "Ed25519",
+            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+            x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
+            kty: "OKP"
+        }
+    },
+
+    "Ed448": {
+        privateUsages: ["sign"],
+        publicUsages: ["verify"],
+        spki: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
+        raw: new Uint8Array([171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
+        pkcs8: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
+        jwk: {
+            crv: "Ed448",
+            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+            kty: "OKP"
+        }
+    },
+
+    "X25519": {
+        privateUsages: ["deriveKey", "deriveBits"],
+        publicUsages: [],
+        spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
+        raw: new Uint8Array([28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
+        pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97]),
+        jwk: {
+            crv: "X25519",
+            d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+            x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+            kty: "OKP"
+        }
+    },
+
+    "X448": {
+        privateUsages: ["deriveKey", "deriveBits"],
+        publicUsages: [],
+        spki: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
+        raw: new Uint8Array([182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
+        pkcs8: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146, 158]),
+        jwk: {
+            crv: "X448",
+            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+            kty: "OKP"
+        }
+    },
+
+};
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.js
index d77a8808831176..4024674e3c3d7f 100644
--- a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.js
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.js
@@ -1,363 +1,214 @@
 
-function run_test() {
-  setup({explicit_done: true});
-
+function run_test(algorithmName) {
   var subtle = self.crypto.subtle; // Change to test prefixed implementations
 
-  // When are all these tests really done? When all the promises they use have resolved.
-  var all_promises = [];
-
   // Source file [algorithm_name]_vectors.js provides the getTestVectors method
   // for the algorithm that drives these tests.
-  var testVectors = getTestVectors();
+  var testVectors = getTestVectors(algorithmName);
 
-  // Test verification first, because signing tests rely on that working
   testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
-              var operation = subtle.verify(algorithm, vector.publicKey, vector.signature, vector.data)
-              .then(function(is_verified) {
-                  assert_true(is_verified, "Signature verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              return operation;
-          }, vector.name + " verification");
-
-      }, function(err) {
-          // We need a failed test if the importVectorKey operation fails, so
-          // we know we never tested verification.
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " verification");
-      });
-
-      all_promises.push(promise);
-  });
+      var algorithm = {name: vector.algorithmName};
 
-  // Test verification with an altered buffer after call
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
+      // Test verification first, because signing tests rely on that working
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              isVerified = await subtle.verify(algorithm, key, vector.signature, vector.data)
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_true(isVerified, "Signature verified");
+      }, vector.name + " verification");
+
+      // Test verification with an altered buffer after call
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
               var signature = copyBuffer(vector.signature);
-              var operation = subtle.verify(algorithm, vector.publicKey, signature, vector.data)
-              .then(function(is_verified) {
-                  assert_true(is_verified, "Signature verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              signature[0] = 255 - signature[0];
-              return operation;
-          }, vector.name + " verification with altered signature after call");
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " verification with altered signature after call");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Check for successful verification even if data is altered after call.
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
+              [isVerified] = await Promise.all([
+                  subtle.verify(algorithm, key, signature, vector.data),
+                  signature[0] = 255 - signature[0]
+              ]);
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_true(isVerified, "Signature verified");
+      }, vector.name + " verification with altered signature after call");
+
+      // Check for successful verification even if data is altered after call.
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
               var data = copyBuffer(vector.data);
-              var operation = subtle.verify(algorithm, vector.publicKey, vector.signature, data)
-              .then(function(is_verified) {
-                  assert_true(is_verified, "Signature verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              data[0] = 255 - data[0];
-              return operation;
-          }, vector.name + " with altered data after call");
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " with altered data after call");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Check for failures due to using privateKey to verify.
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
-              return subtle.verify(algorithm, vector.privateKey, vector.signature, vector.data)
-              .then(function(data) {
-                  assert_unreached("Should have thrown error for using privateKey to verify in " + vector.name + ": " + err.message + "'");
-              }, function(err) {
-                  assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
-              });
-          }, vector.name + " using privateKey to verify");
-
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " using privateKey to verify");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Check for failures due to using publicKey to sign.
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
-              return subtle.sign(algorithm, vector.publicKey, vector.data)
-              .then(function(signature) {
-                  assert_unreached("Should have thrown error for using publicKey to sign in " + vector.name + ": " + err.message + "'");
-              }, function(err) {
-                  assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
-              });
-          }, vector.name + " using publicKey to sign");
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " using publicKey to sign");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Check for failures due to no "verify" usage.
-  testVectors.forEach(function(originalVector) {
-      var vector = Object.assign({}, originalVector);
-
-      var promise = importVectorKeys(vector, [], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
-              return subtle.verify(algorithm, vector.publicKey, vector.signature, vector.data)
-              .then(function(data) {
-                  assert_unreached("Should have thrown error for no verify usage in " + vector.name + ": " + err.message + "'");
-              }, function(err) {
-                  assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
-              });
-          }, vector.name + " no verify usage");
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " no verify usage");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Check for successful signing and verification.
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
-          promise_test(function(test) {
-              return subtle.sign(algorithm, vector.privateKey, vector.data)
-              .then(function(signature) {
-                  assert_true(equalBuffers(signature, vector.signature), "Signing did not give the expected output");
-                  // Can we verify the signature?
-                  return subtle.verify(algorithm, vector.publicKey, signature, vector.data)
-                  .then(function(is_verified) {
-                      assert_true(is_verified, "Round trip verification works");
-                      return signature;
-                  }, function(err) {
-                      assert_unreached("verify error for test " + vector.name + ": " + err.message + "'");
-                  });
-              }, function(err) {
-                  assert_unreached("sign error for test " + vector.name + ": '" + err.message + "'");
-              });
-          }, vector.name + " round trip");
-
-      }, function(err) {
-          // We need a failed test if the importVectorKey operation fails, so
-          // we know we never tested signing or verifying
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " round trip");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Test signing with the wrong algorithm
-  testVectors.forEach(function(vector) {
-      // Want to get the key for the wrong algorithm
-      var promise = subtle.generateKey({name: "HMAC", hash: "SHA-1"}, false, ["sign", "verify"])
-      .then(function(wrongKey) {
-          var algorithm = {name: vector.algorithmName};
-          return importVectorKeys(vector, ["verify"], ["sign"])
-          .then(function(vectors) {
-              promise_test(function(test) {
-                  var operation = subtle.sign(algorithm, wrongKey, vector.data)
-                  .then(function(signature) {
-                      assert_unreached("Signing should not have succeeded for " + vector.name);
-                  }, function(err) {
-                      assert_equals(err.name, "InvalidAccessError", "Should have thrown InvalidAccessError instead of '" + err.message + "'");
-                  });
-
-                  return operation;
-              }, vector.name + " signing with wrong algorithm name");
-
-          }, function(err) {
-              // We need a failed test if the importVectorKey operation fails, so
-              // we know we never tested verification.
-              promise_test(function(test) {
-                  assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-              }, "importVectorKeys step: " + vector.name + " signing with wrong algorithm name");
-          });
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("Generate wrong key for test " + vector.name + " failed: '" + err.message + "'");
-          }, "generate wrong key step: " + vector.name + " signing with wrong algorithm name");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Test verification with the wrong algorithm
-  testVectors.forEach(function(vector) {
-      // Want to get the key for the wrong algorithm
-      var promise = subtle.generateKey({name: "HMAC", hash: "SHA-1"}, false, ["sign", "verify"])
-      .then(function(wrongKey) {
-          return importVectorKeys(vector, ["verify"], ["sign"])
-          .then(function(vectors) {
-              var algorithm = {name: vector.algorithmName};
-              promise_test(function(test) {
-                  var operation = subtle.verify(algorithm, wrongKey, vector.signature, vector.data)
-                  .then(function(signature) {
-                      assert_unreached("Verifying should not have succeeded for " + vector.name);
-                  }, function(err) {
-                      assert_equals(err.name, "InvalidAccessError", "Should have thrown InvalidAccessError instead of '" + err.message + "'");
-                  });
-
-                  return operation;
-              }, vector.name + " verifying with wrong algorithm name");
-
-          }, function(err) {
-              // We need a failed test if the importVectorKey operation fails, so
-              // we know we never tested verification.
-              promise_test(function(test) {
-                  assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-              }, "importVectorKeys step: " + vector.name + " verifying with wrong algorithm name");
-          });
-      }, function(err) {
-          promise_test(function(test) {
-              assert_unreached("Generate wrong key for test " + vector.name + " failed: '" + err.message + "'");
-          }, "generate wrong key step: " + vector.name + " verifying with wrong algorithm name");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Test verification fails with wrong signature
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
+              [isVerified] = await Promise.all([
+                  subtle.verify(algorithm, key, vector.signature, data),
+                  data[0] = 255 - data[0]
+              ]);
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_true(isVerified, "Signature verified");
+      }, vector.name + " with altered data after call");
+
+      // Check for failures due to using privateKey to verify.
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("pkcs8", vector.privateKeyBuffer, algorithm, false, ["sign"]);
+              isVerified = await subtle.verify(algorithm, key, vector.signature, vector.data)
+              assert_unreached("Should have thrown error for using privateKey to verify in " + vector.name);
+          } catch (err) {
+              if (err instanceof AssertionError)
+                  throw err;
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
+          };
+          assert_false(isVerified, "Signature verified");
+      }, vector.name + " using privateKey to verify");
+
+      // Check for failures due to using publicKey to sign.
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              let signature = await subtle.sign(algorithm, key, vector.data);
+              assert_unreached("Should have thrown error for using publicKey to sign in " + vector.name);
+          } catch (err) {
+              if (err instanceof AssertionError)
+                  throw err;
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
+          };
+      }, vector.name + " using publicKey to sign");
+
+      // Check for failures due to no "verify" usage.
+      promise_test(async() => {
+          let isVerified = false;
+          let key;
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, []);
+              isVerified = await subtle.verify(algorithm, key, vector.signature, vector.data)
+              assert_unreached("Should have thrown error for no verify usage in " + vector.name);
+          } catch (err) {
+              if (err instanceof AssertionError)
+                  throw err;
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_equals(err.name, "InvalidAccessError", "Should throw InvalidAccessError instead of '" + err.message + "'");
+          };
+          assert_false(isVerified, "Signature verified");
+      }, vector.name + " no verify usage");
+
+      // Check for successful signing and verification.
+      var algorithm = {name: vector.algorithmName};
+      promise_test(async() => {
+          let isVerified = false;
+          let privateKey, publicKey;
+          let signature;
+          try {
+              privateKey = await subtle.importKey("pkcs8", vector.privateKeyBuffer, algorithm, false, ["sign"]);
+              publicKey = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              signature = await subtle.sign(algorithm, privateKey, vector.data);
+              isVerified = await subtle.verify(algorithm, publicKey, vector.signature, vector.data)
+          } catch (err) {
+              assert_false(publicKey === undefined || privateKey === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_false(signature === undefined, "sign error for test " + vector.name + ": '" + err.message + "'");
+              assert_unreached("verify error for test " + vector.name + ": " + err.message + "'");
+          };
+          assert_true(isVerified, "Round trip verification works");
+      }, vector.name + " round trip");
+
+      // Test signing with the wrong algorithm
+      var algorithm = {name: vector.algorithmName};
+      promise_test(async() => {
+          let wrongKey;
+          try {
+              wrongKey = await subtle.generateKey({name: "HMAC", hash: "SHA-1"}, false, ["sign", "verify"])
+              let signature = await subtle.sign(algorithm, wrongKey, vector.data);
+              assert_unreached("Signing should not have succeeded for " + vector.name);
+          } catch (err) {
+              if (err instanceof AssertionError)
+                  throw err;
+              assert_false(wrongKey === undefined, "Generate wrong key for test " + vector.name + " failed: '" + err.message + "'");
+              assert_equals(err.name, "InvalidAccessError", "Should have thrown InvalidAccessError instead of '" + err.message + "'");
+          };
+      }, vector.name + " signing with wrong algorithm name");
+
+      // Test verification with the wrong algorithm
+      var algorithm = {name: vector.algorithmName};
+      promise_test(async() => {
+          let wrongKey;
+          try {
+              wrongKey = await subtle.generateKey({name: "HMAC", hash: "SHA-1"}, false, ["sign", "verify"])
+              let isVerified = await subtle.verify(algorithm, wrongKey, vector.signature, vector.data)
+              assert_unreached("Verifying should not have succeeded for " + vector.name);
+          } catch (err) {
+              if (err instanceof AssertionError)
+                  throw err;
+              assert_false(wrongKey === undefined, "Generate wrong key for test " + vector.name + " failed: '" + err.message + "'");
+              assert_equals(err.name, "InvalidAccessError", "Should have thrown InvalidAccessError instead of '" + err.message + "'");
+          };
+      }, vector.name + " verifying with wrong algorithm name");
+
+      // Test verification fails with wrong signature
+      var algorithm = {name: vector.algorithmName};
+      promise_test(async() => {
+          let key;
+          let isVerified = true;
           var signature = copyBuffer(vector.signature);
           signature[0] = 255 - signature[0];
-          promise_test(function(test) {
-              var operation = subtle.verify(algorithm, vector.publicKey, signature, vector.data)
-              .then(function(is_verified) {
-                  assert_false(is_verified, "Signature NOT verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              return operation;
-          }, vector.name + " verification failure due to altered signature");
-
-      }, function(err) {
-          // We need a failed test if the importVectorKey operation fails, so
-          // we know we never tested verification.
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " verification failure due to altered signature");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Test verification fails with short (odd length) signature
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              isVerified = await subtle.verify(algorithm, key, signature, vector.data)
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_false(isVerified, "Signature verified");
+      }, vector.name + " verification failure due to altered signature");
+
+      // Test verification fails with short (odd length) signature
+      promise_test(async() => {
+          let key;
+          let isVerified = true;
           var signature = vector.signature.slice(1); // Skip the first byte
-          promise_test(function(test) {
-              var operation = subtle.verify(algorithm, vector.publicKey, signature, vector.data)
-              .then(function(is_verified) {
-                  assert_false(is_verified, "Signature NOT verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              return operation;
-          }, vector.name + " verification failure due to shortened signature");
-
-      }, function(err) {
-          // We need a failed test if the importVectorKey operation fails, so
-          // we know we never tested verification.
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " verification failure due to shortened signature");
-      });
-
-      all_promises.push(promise);
-  });
-
-  // Test verification fails with wrong data
-  testVectors.forEach(function(vector) {
-      var promise = importVectorKeys(vector, ["verify"], ["sign"])
-      .then(function(vectors) {
-          var algorithm = {name: vector.algorithmName};
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              isVerified = await subtle.verify(algorithm, key, signature, vector.data)
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_false(isVerified, "Signature verified");
+      }, vector.name + " verification failure due to shortened signature");
+
+      // Test verification fails with wrong data
+      promise_test(async() => {
+          let key;
+          let isVerified = true;
           var data = copyBuffer(vector.data);
           data[0] = 255 - data[0];
-          promise_test(function(test) {
-              var operation = subtle.verify(algorithm, vector.publicKey, vector.signature, data)
-              .then(function(is_verified) {
-                  assert_false(is_verified, "Signature NOT verified");
-              }, function(err) {
-                  assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
-              });
-
-              return operation;
-          }, vector.name + " verification failure due to altered data");
-
-      }, function(err) {
-          // We need a failed test if the importVectorKey operation fails, so
-          // we know we never tested verification.
-          promise_test(function(test) {
-              assert_unreached("importVectorKeys failed for " + vector.name + ". Message: ''" + err.message + "''");
-          }, "importVectorKeys step: " + vector.name + " verification failure due to altered data");
-      });
-
-      all_promises.push(promise);
-  });
-
-
-  promise_test(function() {
-      return Promise.all(all_promises)
-          .then(function() {done();})
-          .catch(function() {done();})
-  }, "setup");
-
-  // Test that generated keys are valid for signing and verifying.
-  testVectors.forEach(function(vector) {
-      var algorithm = {name: vector.algorithmName};
+          try {
+              key = await subtle.importKey("spki", vector.publicKeyBuffer, algorithm, false, ["verify"]);
+              isVerified = await subtle.verify(algorithm, key, vector.signature, data)
+          } catch (err) {
+              assert_false(key === undefined, "importKey failed for " + vector.name + ". Message: ''" + err.message + "''");
+              assert_unreached("Verification should not throw error " + vector.name + ": " + err.message + "'");
+          };
+          assert_false(isVerified, "Signature verified");
+      }, vector.name + " verification failure due to altered data");
+
+      // Test that generated keys are valid for signing and verifying.
       promise_test(async() => {
           let key = await subtle.generateKey(algorithm, false, ["sign", "verify"]);
           let signature = await subtle.sign(algorithm, key.privateKey, vector.data);
@@ -366,42 +217,6 @@ function run_test() {
       }, "Sign and verify using generated " + vector.algorithmName + " keys.");
   });
 
-
-  // A test vector has all needed fields for signing and verifying, EXCEPT that the
-  // key field may be null. This function replaces that null with the Correct
-  // CryptoKey object.
-  //
-  // Returns a Promise that yields an updated vector on success.
-  function importVectorKeys(vector, publicKeyUsages, privateKeyUsages) {
-      var publicPromise, privatePromise;
-
-      if (vector.publicKey !== null) {
-          publicPromise = new Promise(function(resolve, reject) {
-              resolve(vector);
-          });
-      } else {
-          publicPromise = subtle.importKey(vector.publicKeyFormat, vector.publicKeyBuffer, {name: vector.algorithmName}, false, publicKeyUsages)
-          .then(function(key) {
-              vector.publicKey = key;
-              return vector;
-          });        // Returns a copy of the sourceBuffer it is sent.
-      }
-
-      if (vector.privateKey !== null) {
-          privatePromise = new Promise(function(resolve, reject) {
-              resolve(vector);
-          });
-      } else {
-          privatePromise = subtle.importKey(vector.privateKeyFormat, vector.privateKeyBuffer, {name: vector.algorithmName}, false, privateKeyUsages)
-          .then(function(key) {
-              vector.privateKey = key;
-              return vector;
-          });
-      }
-
-      return Promise.all([publicPromise, privatePromise]);
-  }
-
   // Returns a copy of the sourceBuffer it is sent.
   function copyBuffer(sourceBuffer) {
       var source = new Uint8Array(sourceBuffer);
@@ -414,22 +229,5 @@ function run_test() {
       return copy;
   }
 
-  function equalBuffers(a, b) {
-      if (a.byteLength !== b.byteLength) {
-          return false;
-      }
-
-      var aBytes = new Uint8Array(a);
-      var bBytes = new Uint8Array(b);
-
-      for (var i=0; i<a.byteLength; i++) {
-          if (aBytes[i] !== bBytes[i]) {
-              return false;
-          }
-      }
-
-      return true;
-  }
-
   return;
 }
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve25519.https.any.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve25519.https.any.js
new file mode 100644
index 00000000000000..3be9eef41a9a0f
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve25519.https.any.js
@@ -0,0 +1,6 @@
+// META: title=WebCryptoAPI: sign() and verify() Using EdDSA
+// META: script=eddsa_vectors.js
+// META: script=eddsa.js
+// META: timeout=long
+
+run_test("Ed25519");
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.https.any.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve448.https.any.js
similarity index 88%
rename from test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.https.any.js
rename to test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve448.https.any.js
index 85993a48b91a8c..6960c296bae6af 100644
--- a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_curve448.https.any.js
@@ -3,4 +3,4 @@
 // META: script=eddsa.js
 // META: timeout=long
 
-run_test();
+run_test("Ed448");
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.https.any.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.https.any.js
new file mode 100644
index 00000000000000..be025557fc20ab
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.https.any.js
@@ -0,0 +1,6 @@
+// META: title=WebCryptoAPI: verify() Using EdDSA with small-order points
+// META: script=eddsa_vectors.js
+// META: script=eddsa_small_order_points.js
+// META: timeout=long
+
+run_test();
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.js
new file mode 100644
index 00000000000000..ae973155992c87
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_small_order_points.js
@@ -0,0 +1,26 @@
+
+function run_test() {
+  var subtle = self.crypto.subtle; // Change to test prefixed implementations
+
+  // When verifying an Ed25519 or Ed448 signature, if the public key or the first half of the signature (R) is
+  // an invalid or small-order element, return false.
+  Object.keys(kSmallOrderTestCases).forEach(function (algorithmName) {
+      var algorithm = {name: algorithmName};
+      kSmallOrderTestCases[algorithmName].forEach(function(test) {
+          promise_test(async() => {
+              let isVerified = true;
+              let publicKey;
+              try {
+                  publicKey = await subtle.importKey("raw", test.keyData, algorithm, false, ["verify"])
+                  isVerified = await subtle.verify(algorithm, publicKey, test.signature, test.message);
+              } catch (err) {
+                  assert_true(publicKey !== undefined, "Public key should be valid.");
+                  assert_unreached("The operation shouldn't fail, but it thown this error: " + err.name + ": " + err.message + ".");
+              }
+              assert_equals(isVerified, test.verified, "Signature verification result.");
+          }, algorithmName + " Verification checks with small-order key of order - Test " + test.id);
+      });
+  });
+
+  return;
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_vectors.js b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_vectors.js
index 96ec2b01af96f2..78240586c2e34c 100644
--- a/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_vectors.js
+++ b/test/fixtures/wpt/WebCryptoAPI/sign_verify/eddsa_vectors.js
@@ -16,7 +16,7 @@
 //     algorithmName - the name of the AlgorithmIdentifier parameter to provide to sign
 //     data - the text to sign
 //     signature - the expected signature
-function getTestVectors() {
+function getTestVectors(algorithmName) {
   var pkcs8 = {
       "Ed25519": new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31]),
       "Ed448": new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
@@ -37,7 +37,7 @@ function getTestVectors() {
   }
 
   var vectors = [];
-  ["Ed25519", "Ed448"].forEach(function(algorithmName) {
+  {
     var vector = {
       name: "EdDSA " + algorithmName,
       publicKeyBuffer: spki[algorithmName],
@@ -52,7 +52,146 @@ function getTestVectors() {
     };
 
     vectors.push(vector);
-  });
-
+  }
   return vectors;
 }
+
+// https://eprint.iacr.org/2020/1244.pdf#table.caption.3
+var kSmallOrderPoints = [
+    // Canonical serializations
+    [0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], // #0 - Order 1
+    [0xEC, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x7F], // #1 - Order 2
+    [0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80], // #2 - Order 4
+    [0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00], // #3 - Order 4
+    [0xC7, 0x17, 0x6A, 0x70, 0x3D, 0x4D, 0xD8, 0x4F, 0xBA, 0x3C, 0x0B, 0x76, 0x0D, 0x10, 0x67, 0x0F, 0x2A, 0x20, 0x53, 0xFA, 0x2C, 0x39, 0xCC, 0xC6, 0x4E, 0xC7, 0xFD, 0x77, 0x92, 0xAC, 0x03, 0x7A], // #4 - Order 8
+    [0xC7, 0x17, 0x6A, 0x70, 0x3D, 0x4D, 0xD8, 0x4F, 0xBA, 0x3C, 0x0B, 0x76, 0x0D, 0x10, 0x67, 0x0F, 0x2A, 0x20, 0x53, 0xFA, 0x2C, 0x39, 0xCC, 0xC6, 0x4E, 0xC7, 0xFD, 0x77, 0x92, 0xAC, 0x03, 0xFA], // #5 - Order 8
+    [0x26, 0xE8, 0x95, 0x8F, 0xC2, 0xB2, 0x27, 0xB0, 0x45, 0xC3, 0xF4, 0x89, 0xF2, 0xEF, 0x98, 0xF0, 0xD5, 0xDF, 0xAC, 0x05, 0xD3, 0xC6, 0x33, 0x39, 0xB1, 0x38, 0x02, 0x88, 0x6D, 0x53, 0xFC, 0x05], // #6 - Order 8
+    [0x26, 0xE8, 0x95, 0x8F, 0xC2, 0xB2, 0x27, 0xB0, 0x45, 0xC3, 0xF4, 0x89, 0xF2, 0xEF, 0x98, 0xF0, 0xD5, 0xDF, 0xAC, 0x05, 0xD3, 0xC6, 0x33, 0x39, 0xB1, 0x38, 0x02, 0x88, 0x6D, 0x53, 0xFC, 0x85], // #7 - Order 8
+
+    // Non-canonical serializatons
+    [0x01, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x80], // #8  - Order 1
+    [0xEC, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF], // #9  - Order 2
+    [0xEE, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x7F], // #10 - Order 1
+    [0xEE, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF], // #11 - Order 1
+    [0xED, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF], // #12 - Order 4
+    [0xED, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x7F], // #13 - Order 4
+];
+
+
+var pubKeys = [
+    [0xc7, 0x17, 0x6a, 0x70, 0x3d, 0x4d, 0xd8, 0x4f, 0xba, 0x3c, 0x0b, 0x76, 0x0d, 0x10, 0x67, 0x0f, 0x2a, 0x20, 0x53, 0xfa, 0x2c, 0x39, 0xcc, 0xc6, 0x4e, 0xc7, 0xfd, 0x77, 0x92, 0xac, 0x03, 0xfa], // kSmallOrderPoints #5
+    [0xf7, 0xba, 0xde, 0xc5, 0xb8, 0xab, 0xea, 0xf6, 0x99, 0x58, 0x39, 0x92, 0x21, 0x9b, 0x7b, 0x22, 0x3f, 0x1d, 0xf3, 0xfb, 0xbe, 0xa9, 0x19, 0x84, 0x4e, 0x3f, 0x7c, 0x55, 0x4a, 0x43, 0xdd, 0x43], // highest 32 bytes of case "1" signature
+    [0xcd, 0xb2, 0x67, 0xce, 0x40, 0xc5, 0xcd, 0x45, 0x30, 0x6f, 0xa5, 0xd2, 0xf2, 0x97, 0x31, 0x45, 0x93, 0x87, 0xdb, 0xf9, 0xeb, 0x93, 0x3b, 0x7b, 0xd5, 0xae, 0xd9, 0xa7, 0x65, 0xb8, 0x8d, 0x4d],
+    [0x44, 0x2a, 0xad, 0x9f, 0x08, 0x9a, 0xd9, 0xe1, 0x46, 0x47, 0xb1, 0xef, 0x90, 0x99, 0xa1, 0xff, 0x47, 0x98, 0xd7, 0x85, 0x89, 0xe6, 0x6f, 0x28, 0xec, 0xa6, 0x9c, 0x11, 0xf5, 0x82, 0xa6, 0x23],
+    [0xec, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff], // kSmallOrderPoints #9
+    [0xEC, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0x7F], // kSmallOrderPoints #1
+]
+
+// https://eprint.iacr.org/2020/1244.pdf
+// signature = (R, S); public key A, h = SHA512(R||A||M )
+// 8(SB) = 8R + 8(hA) => (1)
+// SB = R + hA => (2)
+var kSmallOrderTestCases = {
+    "Ed25519": [
+        {
+            id: "0", // S = 0 | A's order = small | R's order = small | (1) = pass | (2) = pass
+            message : Uint8Array.from([0x8c, 0x93, 0x25, 0x5d, 0x71, 0xdc, 0xab, 0x10, 0xe8, 0xf3, 0x79, 0xc2, 0x62, 0x00, 0xf3, 0xc7, 0xbd, 0x5f, 0x09, 0xd9, 0xbc, 0x30, 0x68, 0xd3, 0xef, 0x4e, 0xde, 0xb4, 0x85, 0x30, 0x22, 0xb6]),
+            keyData : Uint8Array.from(pubKeys[0]),
+            signature : Uint8Array.from([0xc7, 0x17, 0x6a, 0x70, 0x3d, 0x4d, 0xd8, 0x4f, 0xba, 0x3c, 0x0b, 0x76, 0x0d, 0x10, 0x67, 0x0f, 0x2a, 0x20, 0x53, 0xfa, 0x2c, 0x39, 0xcc, 0xc6, 0x4e, 0xc7, 0xfd, 0x77, 0x92, 0xac, 0x03, 0x7a, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00]),
+            verified: false,  // small-order signature's R fail in the verification.
+        },
+        {
+            id: "1", // 0 < S < L | A's order = small | R's order = mixed | (1) = pass | (2) = pass
+            message : Uint8Array.from([0x9b, 0xd9, 0xf4, 0x4f, 0x4d, 0xcc, 0x75, 0xbd, 0x53, 0x1b, 0x56, 0xb2, 0xcd, 0x28, 0x0b, 0x0b, 0xb3, 0x8f, 0xc1, 0xcd, 0x6d, 0x12, 0x30, 0xe1, 0x48, 0x61, 0xd8, 0x61, 0xde, 0x09, 0x2e, 0x79]),
+            keyData : Uint8Array.from(pubKeys[0]),
+            signature : Uint8Array.from([0xf7, 0xba, 0xde, 0xc5, 0xb8, 0xab, 0xea, 0xf6, 0x99, 0x58, 0x39, 0x92, 0x21, 0x9b, 0x7b, 0x22, 0x3f, 0x1d, 0xf3, 0xfb, 0xbe, 0xa9, 0x19, 0x84, 0x4e, 0x3f, 0x7c, 0x55, 0x4a, 0x43, 0xdd, 0x43, 0xa5, 0xbb, 0x70, 0x47, 0x86, 0xbe, 0x79, 0xfc, 0x47, 0x6f, 0x91, 0xd3, 0xf3, 0xf8, 0x9b, 0x03, 0x98, 0x4d, 0x80, 0x68, 0xdc, 0xf1, 0xbb, 0x7d, 0xfc, 0x66, 0x37, 0xb4, 0x54, 0x50, 0xac, 0x04]),
+            verified: false, // small-order key's data fail in the verification.
+        },
+        {
+            id: "2", // 0 < S < L | A's order = mixed | R's order = small | (1) = pass | (2) = pass
+            message : Uint8Array.from([0xae, 0xbf, 0x3f, 0x26, 0x01, 0xa0, 0xc8, 0xc5, 0xd3, 0x9c, 0xc7, 0xd8, 0x91, 0x16, 0x42, 0xf7, 0x40, 0xb7, 0x81, 0x68, 0x21, 0x8d, 0xa8, 0x47, 0x17, 0x72, 0xb3, 0x5f, 0x9d, 0x35, 0xb9, 0xab]),
+            keyData : Uint8Array.from(pubKeys[1]),
+            signature : Uint8Array.from([0xc7, 0x17, 0x6a, 0x70, 0x3d, 0x4d, 0xd8, 0x4f, 0xba, 0x3c, 0x0b, 0x76, 0x0d, 0x10, 0x67, 0x0f, 0x2a, 0x20, 0x53, 0xfa, 0x2c, 0x39, 0xcc, 0xc6, 0x4e, 0xc7, 0xfd, 0x77, 0x92, 0xac, 0x03, 0xfa, 0x8c, 0x4b, 0xd4, 0x5a, 0xec, 0xac, 0xa5, 0xb2, 0x4f, 0xb9, 0x7b, 0xc1, 0x0a, 0xc2, 0x7a, 0xc8, 0x75, 0x1a, 0x7d, 0xfe, 0x1b, 0xaf, 0xf8, 0xb9, 0x53, 0xec, 0x9f, 0x58, 0x33, 0xca, 0x26, 0x0e]),
+            verified: false,  // small-order signature's R fail in the verification.
+        },
+        {
+            id: "3", // 0 < S < L | A's order = mixed | R's order = mixed | (1) = pass | (2) = pass
+            message : Uint8Array.from([0x9b, 0xd9, 0xf4, 0x4f, 0x4d, 0xcc, 0x75, 0xbd, 0x53, 0x1b, 0x56, 0xb2, 0xcd, 0x28, 0x0b, 0x0b, 0xb3, 0x8f, 0xc1, 0xcd, 0x6d, 0x12, 0x30, 0xe1, 0x48, 0x61, 0xd8, 0x61, 0xde, 0x09, 0x2e, 0x79]),
+            keyData : Uint8Array.from(pubKeys[2]),
+            signature : Uint8Array.from([0x90, 0x46, 0xa6, 0x47, 0x50, 0x44, 0x49, 0x38, 0xde, 0x19, 0xf2, 0x27, 0xbb, 0x80, 0x48, 0x5e, 0x92, 0xb8, 0x3f, 0xdb, 0x4b, 0x65, 0x06, 0xc1, 0x60, 0x48, 0x4c, 0x01, 0x6c, 0xc1, 0x85, 0x2f, 0x87, 0x90, 0x9e, 0x14, 0x42, 0x8a, 0x7a, 0x1d, 0x62, 0xe9, 0xf2, 0x2f, 0x3d, 0x3a, 0xd7, 0x80, 0x2d, 0xb0, 0x2e, 0xb2, 0xe6, 0x88, 0xb6, 0xc5, 0x2f, 0xcd, 0x66, 0x48, 0xa9, 0x8b, 0xd0, 0x09]),
+            verified: true, // mixed-order points are not checked.
+        },
+        {
+            id: "4", // 0 < S < L | A's order = mixed | R's order = mixed | (1) = pass | (2) = fail
+            message : Uint8Array.from([0xe4, 0x7d, 0x62, 0xc6, 0x3f, 0x83, 0x0d, 0xc7, 0xa6, 0x85, 0x1a, 0x0b, 0x1f, 0x33, 0xae, 0x4b, 0xb2, 0xf5, 0x07, 0xfb, 0x6c, 0xff, 0xec, 0x40, 0x11, 0xea, 0xcc, 0xd5, 0x5b, 0x53, 0xf5, 0x6c]),
+            keyData : Uint8Array.from(pubKeys[2]),
+            signature : Uint8Array.from([0x16, 0x0a, 0x1c, 0xb0, 0xdc, 0x9c, 0x02, 0x58, 0xcd, 0x0a, 0x7d, 0x23, 0xe9, 0x4d, 0x8f, 0xa8, 0x78, 0xbc, 0xb1, 0x92, 0x5f, 0x2c, 0x64, 0x24, 0x6b, 0x2d, 0xee, 0x17, 0x96, 0xbe, 0xd5, 0x12, 0x5e, 0xc6, 0xbc, 0x98, 0x2a, 0x26, 0x9b, 0x72, 0x3e, 0x06, 0x68, 0xe5, 0x40, 0x91, 0x1a, 0x9a, 0x6a, 0x58, 0x92, 0x1d, 0x69, 0x25, 0xe4, 0x34, 0xab, 0x10, 0xaa, 0x79, 0x40, 0x55, 0x1a, 0x09]),
+            verified: false, // expect a cofactorless verification algorithm.
+        },
+        {
+            id: "5", // 0 < S < L | A's order = mixed | R's order = L | (1) = pass | (2) = fail
+            message : Uint8Array.from([0xe4, 0x7d, 0x62, 0xc6, 0x3f, 0x83, 0x0d, 0xc7, 0xa6, 0x85, 0x1a, 0x0b, 0x1f, 0x33, 0xae, 0x4b, 0xb2, 0xf5, 0x07, 0xfb, 0x6c, 0xff, 0xec, 0x40, 0x11, 0xea, 0xcc, 0xd5, 0x5b, 0x53, 0xf5, 0x6c]),
+            keyData : Uint8Array.from(pubKeys[2]),
+            signature : Uint8Array.from([0x21, 0x12, 0x2a, 0x84, 0xe0, 0xb5, 0xfc, 0xa4, 0x05, 0x2f, 0x5b, 0x12, 0x35, 0xc8, 0x0a, 0x53, 0x78, 0x78, 0xb3, 0x8f, 0x31, 0x42, 0x35, 0x6b, 0x2c, 0x23, 0x84, 0xeb, 0xad, 0x46, 0x68, 0xb7, 0xe4, 0x0b, 0xc8, 0x36, 0xda, 0xc0, 0xf7, 0x10, 0x76, 0xf9, 0xab, 0xe3, 0xa5, 0x3f, 0x9c, 0x03, 0xc1, 0xce, 0xee, 0xdd, 0xb6, 0x58, 0xd0, 0x03, 0x04, 0x94, 0xac, 0xe5, 0x86, 0x68, 0x74, 0x05]),
+            verified: false, // expect a cofactorless verification algorithm.
+        },
+        {
+            id: "6", // S > L | A's order = L | R's order = L | (1) = pass | (2) = pass
+            message : Uint8Array.from([0x85, 0xe2, 0x41, 0xa0, 0x7d, 0x14, 0x8b, 0x41, 0xe4, 0x7d, 0x62, 0xc6, 0x3f, 0x83, 0x0d, 0xc7, 0xa6, 0x85, 0x1a, 0x0b, 0x1f, 0x33, 0xae, 0x4b, 0xb2, 0xf5, 0x07, 0xfb, 0x6c, 0xff, 0xec, 0x40]),
+            keyData : Uint8Array.from(pubKeys[3]),
+            signature : Uint8Array.from([0xe9, 0x6f, 0x66, 0xbe, 0x97, 0x6d, 0x82, 0xe6, 0x01, 0x50, 0xba, 0xec, 0xff, 0x99, 0x06, 0x68, 0x4a, 0xeb, 0xb1, 0xef, 0x18, 0x1f, 0x67, 0xa7, 0x18, 0x9a, 0xc7, 0x8e, 0xa2, 0x3b, 0x6c, 0x0e, 0x54, 0x7f, 0x76, 0x90, 0xa0, 0xe2, 0xdd, 0xcd, 0x04, 0xd8, 0x7d, 0xbc, 0x34, 0x90, 0xdc, 0x19, 0xb3, 0xb3, 0x05, 0x2f, 0x7f, 0xf0, 0x53, 0x8c, 0xb6, 0x8a, 0xfb, 0x36, 0x9b, 0xa3, 0xa5, 0x14]),
+            verified: false, // S out of bounds
+        },
+        {
+            id: "7", // S >> L | A's order = L | R's order = L | (1) = pass | (2) = pass
+            message : Uint8Array.from([0x85, 0xe2, 0x41, 0xa0, 0x7d, 0x14, 0x8b, 0x41, 0xe4, 0x7d, 0x62, 0xc6, 0x3f, 0x83, 0x0d, 0xc7, 0xa6, 0x85, 0x1a, 0x0b, 0x1f, 0x33, 0xae, 0x4b, 0xb2, 0xf5, 0x07, 0xfb, 0x6c, 0xff, 0xec, 0x40]),
+            keyData : Uint8Array.from(pubKeys[3]),
+            signature : Uint8Array.from([0x8c, 0xe5, 0xb9, 0x6c, 0x8f, 0x26, 0xd0, 0xab, 0x6c, 0x47, 0x95, 0x8c, 0x9e, 0x68, 0xb9, 0x37, 0x10, 0x4c, 0xd3, 0x6e, 0x13, 0xc3, 0x35, 0x66, 0xac, 0xd2, 0xfe, 0x8d, 0x38, 0xaa, 0x19, 0x42, 0x7e, 0x71, 0xf9, 0x8a, 0x47, 0x34, 0xe7, 0x4f, 0x2f, 0x13, 0xf0, 0x6f, 0x97, 0xc2, 0x0d, 0x58, 0xcc, 0x3f, 0x54, 0xb8, 0xbd, 0x0d, 0x27, 0x2f, 0x42, 0xb6, 0x95, 0xdd, 0x7e, 0x89, 0xa8, 0xc2, 0x02]),
+            verified: false, // S out of bounds
+        },
+        {
+            id: "8", // 0 < S < L | A's order = mixed | R's order = small (non-canonical) | (1) = ? | (2) = ? Implementations that reduce A before hashing will accept #8 and accept #9, and viceversa
+            message : Uint8Array.from([0x9b, 0xed, 0xc2, 0x67, 0x42, 0x37, 0x25, 0xd4, 0x73, 0x88, 0x86, 0x31, 0xeb, 0xf4, 0x59, 0x88, 0xba, 0xd3, 0xdb, 0x83, 0x85, 0x1e, 0xe8, 0x5c, 0x85, 0xe2, 0x41, 0xa0, 0x7d, 0x14, 0x8b, 0x41]),
+            keyData : Uint8Array.from(pubKeys[1]),
+            signature : Uint8Array.from([0xec, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x03, 0xbe, 0x96, 0x78, 0xac, 0x10, 0x2e, 0xdc, 0xd9, 0x2b, 0x02, 0x10, 0xbb, 0x34, 0xd7, 0x42, 0x8d, 0x12, 0xff, 0xc5, 0xdf, 0x5f, 0x37, 0xe3, 0x59, 0x94, 0x12, 0x66, 0xa4, 0xe3, 0x5f, 0x0f]),
+            verified: false, // non-canonical point should fail in the verificaton (RFC8032)
+        },
+        {
+            id: "9", // 0 < S < L | A's order = mixed | R's order = small (non-canonical) | (1) = ? | (2) = ?
+            message : Uint8Array.from([0x9b, 0xed, 0xc2, 0x67, 0x42, 0x37, 0x25, 0xd4, 0x73, 0x88, 0x86, 0x31, 0xeb, 0xf4, 0x59, 0x88, 0xba, 0xd3, 0xdb, 0x83, 0x85, 0x1e, 0xe8, 0x5c, 0x85, 0xe2, 0x41, 0xa0, 0x7d, 0x14, 0x8b, 0x41]),
+            keyData : Uint8Array.from(pubKeys[1]),
+            signature : Uint8Array.from([0xec, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xca, 0x8c, 0x5b, 0x64, 0xcd, 0x20, 0x89, 0x82, 0xaa, 0x38, 0xd4, 0x93, 0x66, 0x21, 0xa4, 0x77, 0x5a, 0xa2, 0x33, 0xaa, 0x05, 0x05, 0x71, 0x1d, 0x8f, 0xdc, 0xfd, 0xaa, 0x94, 0x3d, 0x49, 0x08]),
+            verified: false, // non-canonical point should fail in the verificaton (RFC8032)
+        },
+        {
+            id: "10", // 0 < S < L | A's order = small (non-canonical) | R's order = mixed | (1) = ? | (2) = ? Implementations that reduce A before hashing will accept #10 and accept #11, and viceversa
+            message : Uint8Array.from([0xe9, 0x6b, 0x70, 0x21, 0xeb, 0x39, 0xc1, 0xa1, 0x63, 0xb6, 0xda, 0x4e, 0x30, 0x93, 0xdc, 0xd3, 0xf2, 0x13, 0x87, 0xda, 0x4c, 0xc4, 0x57, 0x2b, 0xe5, 0x88, 0xfa, 0xfa, 0xe2, 0x3c, 0x15, 0x5b]),
+            keyData : Uint8Array.from(pubKeys[4]),
+            signature : Uint8Array.from([0xa9, 0xd5, 0x52, 0x60, 0xf7, 0x65, 0x26, 0x1e, 0xb9, 0xb8, 0x4e, 0x10, 0x6f, 0x66, 0x5e, 0x00, 0xb8, 0x67, 0x28, 0x7a, 0x76, 0x19, 0x90, 0xd7, 0x13, 0x59, 0x63, 0xee, 0x0a, 0x7d, 0x59, 0xdc, 0xa5, 0xbb, 0x70, 0x47, 0x86, 0xbe, 0x79, 0xfc, 0x47, 0x6f, 0x91, 0xd3, 0xf3, 0xf8, 0x9b, 0x03, 0x98, 0x4d, 0x80, 0x68, 0xdc, 0xf1, 0xbb, 0x7d, 0xfc, 0x66, 0x37, 0xb4, 0x54, 0x50, 0xac, 0x04]),
+            verified: false, // non-canonical point should fail in the verificaton (RFC8032)
+        },
+        {
+            id: "11", // 0 < S < L | A's order = small (non-canonical) | R's order = mixed | (1) = ? | (2) = ? Implementations that reduce A before hashing will accept #10 and accept #11, and viceversa
+            message : Uint8Array.from([0x39, 0xa5, 0x91, 0xf5, 0x32, 0x1b, 0xbe, 0x07, 0xfd, 0x5a, 0x23, 0xdc, 0x2f, 0x39, 0xd0, 0x25, 0xd7, 0x45, 0x26, 0x61, 0x57, 0x46, 0x72, 0x7c, 0xee, 0xfd, 0x6e, 0x82, 0xae, 0x65, 0xc0, 0x6f]),
+            keyData : Uint8Array.from(pubKeys[4]),
+            signature : Uint8Array.from([0xa9, 0xd5, 0x52, 0x60, 0xf7, 0x65, 0x26, 0x1e, 0xb9, 0xb8, 0x4e, 0x10, 0x6f, 0x66, 0x5e, 0x00, 0xb8, 0x67, 0x28, 0x7a, 0x76, 0x19, 0x90, 0xd7, 0x13, 0x59, 0x63, 0xee, 0x0a, 0x7d, 0x59, 0xdc, 0xa5, 0xbb, 0x70, 0x47, 0x86, 0xbe, 0x79, 0xfc, 0x47, 0x6f, 0x91, 0xd3, 0xf3, 0xf8, 0x9b, 0x03, 0x98, 0x4d, 0x80, 0x68, 0xdc, 0xf1, 0xbb, 0x7d, 0xfc, 0x66, 0x37, 0xb4, 0x54, 0x50, 0xac, 0x04]),
+            verified: false, // non-canonical point should fail in the verificaton (RFC8032)
+        },
+        // https://eprint.iacr.org/2020/1244.pdf#section.A.2
+        // cases breaking non-repudiation
+        {
+            id: "12", // 0 < S < L | A's order = small | R's order = mixed | (1) = ? | (2) = ?
+            message : Uint8Array.from([0x53, 0x65, 0x6e, 0x64, 0x20, 0x31, 0x30, 0x30, 0x20, 0x55, 0x53, 0x44, 0x20, 0x74, 0x6f, 0x20, 0x41, 0x6c, 0x69, 0x63, 0x65]),
+            keyData : Uint8Array.from(pubKeys[5]),
+            signature : Uint8Array.from([0xa9, 0xd5, 0x52, 0x60, 0xf7, 0x65, 0x26, 0x1e, 0xb9, 0xb8, 0x4e, 0x10, 0x6f, 0x66, 0x5e, 0x00, 0xb8, 0x67, 0x28, 0x7a, 0x76, 0x19, 0x90, 0xd7, 0x13, 0x59, 0x63, 0xee, 0x0a, 0x7d, 0x59, 0xdc, 0xa5, 0xbb, 0x70, 0x47, 0x86, 0xbe, 0x79, 0xfc, 0x47, 0x6f, 0x91, 0xd3, 0xf3, 0xf8, 0x9b, 0x03, 0x98, 0x4d, 0x80, 0x68, 0xdc, 0xf1, 0xbb, 0x7d, 0xfc, 0x66, 0x37, 0xb4, 0x54, 0x50, 0xac, 0x04]),
+            verified: false,
+        },
+        {
+            id: "13", // 0 < S < L | A's order = small | R's order = mixed | (1) = ? | (2) = ?
+            message : Uint8Array.from([0x53, 0x65, 0x6e, 0x64, 0x20, 0x31, 0x30, 0x30, 0x30, 0x30, 0x30, 0x20, 0x55, 0x53, 0x44, 0x20, 0x74, 0x6f, 0x20, 0x41, 0x6c, 0x69, 0x63, 0x65]),
+            keyData : Uint8Array.from(pubKeys[5]),
+            signature : Uint8Array.from([0xa9, 0xd5, 0x52, 0x60, 0xf7, 0x65, 0x26, 0x1e, 0xb9, 0xb8, 0x4e, 0x10, 0x6f, 0x66, 0x5e, 0x00, 0xb8, 0x67, 0x28, 0x7a, 0x76, 0x19, 0x90, 0xd7, 0x13, 0x59, 0x63, 0xee, 0x0a, 0x7d, 0x59, 0xdc, 0xa5, 0xbb, 0x70, 0x47, 0x86, 0xbe, 0x79, 0xfc, 0x47, 0x6f, 0x91, 0xd3, 0xf3, 0xf8, 0x9b, 0x03, 0x98, 0x4d, 0x80, 0x68, 0xdc, 0xf1, 0xbb, 0x7d, 0xfc, 0x66, 0x37, 0xb4, 0x54, 0x50, 0xac, 0x04]),
+            verified: false,
+        }
+    ]
+};
diff --git a/test/fixtures/wpt/WebCryptoAPI/util/helpers.js b/test/fixtures/wpt/WebCryptoAPI/util/helpers.js
index bda97003263ff4..c60371dc6adac9 100644
--- a/test/fixtures/wpt/WebCryptoAPI/util/helpers.js
+++ b/test/fixtures/wpt/WebCryptoAPI/util/helpers.js
@@ -259,3 +259,41 @@ function allNameVariants(name, slowTest) {
     if (slowTest) return [mixedCaseName];
     return unique([upCaseName, lowCaseName, mixedCaseName]);
 }
+
+// Builds a hex string representation for an array-like input.
+// "bytes" can be an Array of bytes, an ArrayBuffer, or any TypedArray.
+// The output looks like this:
+//    ab034c99
+function bytesToHexString(bytes)
+{
+    if (!bytes)
+        return null;
+
+    bytes = new Uint8Array(bytes);
+    var hexBytes = [];
+
+    for (var i = 0; i < bytes.length; ++i) {
+        var byteString = bytes[i].toString(16);
+        if (byteString.length < 2)
+            byteString = "0" + byteString;
+        hexBytes.push(byteString);
+    }
+
+    return hexBytes.join("");
+}
+
+function hexStringToUint8Array(hexString)
+{
+    if (hexString.length % 2 != 0)
+        throw "Invalid hexString";
+    var arrayBuffer = new Uint8Array(hexString.length / 2);
+
+    for (var i = 0; i < hexString.length; i += 2) {
+        var byteValue = parseInt(hexString.substr(i, 2), 16);
+        if (byteValue == NaN)
+            throw "Invalid hexString";
+        arrayBuffer[i/2] = byteValue;
+    }
+
+    return arrayBuffer;
+}
diff --git a/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey.https.any.js b/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey.https.any.js
index edb67d9e30fdba..9dd837b3bf60a9 100644
--- a/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey.https.any.js
+++ b/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey.https.any.js
@@ -1,238 +1,256 @@
 // META: title=WebCryptoAPI: wrapKey() and unwrapKey()
 // META: timeout=long
 // META: script=../util/helpers.js
+// META: script=wrapKey_unwrapKey_vectors.js
 
 // Tests for wrapKey and unwrapKey round tripping
 
     var subtle = self.crypto.subtle;
 
-    var wrappers = [];  // Things we wrap (and upwrap) keys with
-    var keys = [];      // Things to wrap and unwrap
-
-    // Generate all the keys needed, then iterate over all combinations
+    var wrappers = {};  // Things we wrap (and upwrap) keys with
+    var keys = {};      // Things to wrap and unwrap
+
+    // There are five algorithms that can be used for wrapKey/unwrapKey.
+    // Generate one key with typical parameters for each kind.
+    //
+    // Note: we don't need cryptographically strong parameters for things
+    // like IV - just any legal value will do.
+    var wrappingKeysParameters = [
+        {
+            name: "RSA-OAEP",
+            importParameters: {name: "RSA-OAEP", hash: "SHA-256"},
+            wrapParameters: {name: "RSA-OAEP", label: new Uint8Array(8)}
+        },
+        {
+            name: "AES-CTR",
+            importParameters: {name: "AES-CTR", length: 128},
+            wrapParameters: {name: "AES-CTR", counter: new Uint8Array(16), length: 64}
+        },
+        {
+            name: "AES-CBC",
+            importParameters: {name: "AES-CBC", length: 128},
+            wrapParameters: {name: "AES-CBC", iv: new Uint8Array(16)}
+        },
+        {
+            name: "AES-GCM",
+            importParameters: {name: "AES-GCM", length: 128},
+            wrapParameters: {name: "AES-GCM", iv: new Uint8Array(16), additionalData: new Uint8Array(16), tagLength: 128}
+        },
+        {
+            name: "AES-KW",
+            importParameters: {name: "AES-KW", length: 128},
+            wrapParameters: {name: "AES-KW"}
+        }
+    ];
+
+    var keysToWrapParameters = [
+        {algorithm: {name: "RSASSA-PKCS1-v1_5", hash: "SHA-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
+        {algorithm: {name: "RSA-PSS", hash: "SHA-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
+        {algorithm: {name: "RSA-OAEP", hash: "SHA-256"}, privateUsages: ["decrypt"], publicUsages: ["encrypt"]},
+        {algorithm: {name: "ECDSA", namedCurve: "P-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
+        {algorithm: {name: "ECDH", namedCurve: "P-256"}, privateUsages: ["deriveBits"], publicUsages: []},
+        {algorithm: {name: "Ed25519" }, privateUsages: ["sign"], publicUsages: ["verify"]},
+        {algorithm: {name: "Ed448" }, privateUsages: ["sign"], publicUsages: ["verify"]},
+        {algorithm: {name: "X25519" }, privateUsages: ["deriveBits"], publicUsages: []},
+        {algorithm: {name: "X448" }, privateUsages: ["deriveBits"], publicUsages: []},
+        {algorithm: {name: "AES-CTR", length: 128}, usages: ["encrypt", "decrypt"]},
+        {algorithm: {name: "AES-CBC", length: 128}, usages: ["encrypt", "decrypt"]},
+        {algorithm: {name: "AES-GCM", length: 128}, usages: ["encrypt", "decrypt"]},
+        {algorithm: {name: "AES-KW", length: 128}, usages: ["wrapKey", "unwrapKey"]},
+        {algorithm: {name: "HMAC", length: 128, hash: "SHA-256"}, usages: ["sign", "verify"]}
+    ];
+
+    // Import all the keys needed, then iterate over all combinations
     // to test wrapping and unwrapping.
     promise_test(function() {
-    return Promise.all([generateWrappingKeys(), generateKeysToWrap()])
+    return Promise.all([importWrappingKeys(), importKeysToWrap()])
     .then(function(results) {
-        var promises = [];
-        wrappers.forEach(function(wrapper) {
-            keys.forEach(function(key) {
-                promises.push(testWrapping(wrapper, key));
-            })
+        wrappingKeysParameters.filter((param) => Object.keys(wrappers).includes(param.name)).forEach(function(wrapperParam) {
+            var wrapper = wrappers[wrapperParam.name];
+            keysToWrapParameters.filter((param) => Object.keys(keys).includes(param.algorithm.name)).forEach(function(toWrapParam) {
+                var keyData = keys[toWrapParam.algorithm.name];
+                ["raw", "spki", "pkcs8"].filter((fmt) => Object.keys(keyData).includes(fmt)).forEach(function(keyDataFormat) {
+                    var toWrap = keyData[keyDataFormat];
+                    [keyDataFormat, "jwk"].forEach(function(format) {
+                        if (wrappingIsPossible(toWrap.originalExport[format], wrapper.parameters.name)) {
+                            testWrapping(wrapper, toWrap, format);
+                            if (canCompareNonExtractableKeys(toWrap.key)) {
+                                testWrappingNonExtractable(wrapper, toWrap, format);
+                                if (format === "jwk") {
+                                    testWrappingNonExtractableAsExtractable(wrapper, toWrap);
+                                }
+                            }
+                        }
+                    });
+                });
+            });
         });
-        return Promise.allSettled(promises);
+        return Promise.resolve("setup done");
+    }, function(err) {
+        return Promise.reject("setup failed: " + err.name + ': "' + err.message + '"');
     });
     }, "setup");
 
-    function generateWrappingKeys() {
-        // There are five algorithms that can be used for wrapKey/unwrapKey.
-        // Generate one key with typical parameters for each kind.
-        //
-        // Note: we don't need cryptographically strong parameters for things
-        // like IV - just any legal value will do.
-        var parameters = [
-            {
-                name: "RSA-OAEP",
-                generateParameters: {name: "RSA-OAEP", modulusLength: 4096, publicExponent: new Uint8Array([1,0,1]), hash: "SHA-256"},
-                wrapParameters: {name: "RSA-OAEP", label: new Uint8Array(8)}
-            },
-            {
-                name: "AES-CTR",
-                generateParameters: {name: "AES-CTR", length: 128},
-                wrapParameters: {name: "AES-CTR", counter: new Uint8Array(16), length: 64}
-            },
-            {
-                name: "AES-CBC",
-                generateParameters: {name: "AES-CBC", length: 128},
-                wrapParameters: {name: "AES-CBC", iv: new Uint8Array(16)}
-            },
-            {
-                name: "AES-GCM",
-                generateParameters: {name: "AES-GCM", length: 128},
-                wrapParameters: {name: "AES-GCM", iv: new Uint8Array(16), additionalData: new Uint8Array(16), tagLength: 128}
-            },
-            {
-                name: "AES-KW",
-                generateParameters: {name: "AES-KW", length: 128},
-                wrapParameters: {name: "AES-KW"}
+    function importWrappingKeys() {
+        // Using allSettled to skip unsupported test cases.
+        var promises = [];
+        wrappingKeysParameters.forEach(function(params) {
+            if (params.name === "RSA-OAEP") { // we have a key pair, not just a key
+                var algorithm = {name: "RSA-OAEP", hash: "SHA-256"};
+                wrappers[params.name] = {wrappingKey: undefined, unwrappingKey: undefined, parameters: params};
+                promises.push(subtle.importKey("spki", wrappingKeyData["RSA"].spki, algorithm, true, ["wrapKey"])
+                              .then(function(key) {
+                                  wrappers["RSA-OAEP"].wrappingKey = key;
+                              }));
+                promises.push(subtle.importKey("pkcs8", wrappingKeyData["RSA"].pkcs8, algorithm, true, ["unwrapKey"])
+                              .then(function(key) {
+                                  wrappers["RSA-OAEP"].unwrappingKey = key;
+                              }));
+            } else {
+                var algorithm = {name: params.name};
+                promises.push(subtle.importKey("raw", wrappingKeyData["SYMMETRIC"].raw, algorithm, true, ["wrapKey", "unwrapKey"])
+                              .then(function(key) {
+                                  wrappers[params.name] = {wrappingKey: key, unwrappingKey: key, parameters: params};
+                              }));
             }
-        ];
-
+        });
         // Using allSettled to skip unsupported test cases.
-        return Promise.allSettled(parameters.map(function(params) {
-            return subtle.generateKey(params.generateParameters, true, ["wrapKey", "unwrapKey"])
-            .then(function(key) {
-                var wrapper;
-                if (params.name === "RSA-OAEP") { // we have a key pair, not just a key
-                    wrapper = {wrappingKey: key.publicKey, unwrappingKey: key.privateKey, parameters: params};
-                } else {
-                    wrapper = {wrappingKey: key, unwrappingKey: key, parameters: params};
-                }
-                wrappers.push(wrapper);
-                return true;
-            })
-        }));
+        return Promise.allSettled(promises);
     }
 
+    async function importAndExport(format, keyData, algorithm, keyUsages, keyType) {
+        var importedKey;
+        try {
+            importedKey = await subtle.importKey(format, keyData, algorithm, true, keyUsages);
+            keys[algorithm.name][format] = { name: algorithm.name + " " + keyType, algorithm: algorithm, usages: keyUsages, key: importedKey, originalExport: {} };
+        } catch (err) {
+            delete keys[algorithm.name][format];
+            throw("Error importing " + algorithm.name + " " + keyType + " key in '" + format + "' -" + err.name + ': "' + err.message + '"');
+        };
+        try {
+            var exportedKey = await subtle.exportKey(format, importedKey);
+            keys[algorithm.name][format].originalExport[format] = exportedKey;
+        } catch (err) {
+            delete keys[algorithm.name][format];
+            throw("Error exporting " + algorithm.name + " '" + format + "' key -" + err.name + ': "' + err.message + '"');
+        };
+        try {
+            var jwkExportedKey = await subtle.exportKey("jwk", importedKey);
+            keys[algorithm.name][format].originalExport["jwk"] = jwkExportedKey;
+        } catch (err) {
+            delete keys[algorithm.name][format];
+            throw("Error exporting " + algorithm.name + " '" + format + "' key to 'jwk' -" + err.name + ': "' + err.message + '"');
+        };
+    }
 
-    function generateKeysToWrap() {
-        var parameters = [
-            {algorithm: {name: "RSASSA-PKCS1-v1_5", modulusLength: 1024, publicExponent: new Uint8Array([1,0,1]), hash: "SHA-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
-            {algorithm: {name: "RSA-PSS", modulusLength: 1024, publicExponent: new Uint8Array([1,0,1]), hash: "SHA-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
-            {algorithm: {name: "RSA-OAEP", modulusLength: 1024, publicExponent: new Uint8Array([1,0,1]), hash: "SHA-256"}, privateUsages: ["decrypt"], publicUsages: ["encrypt"]},
-            {algorithm: {name: "ECDSA", namedCurve: "P-256"}, privateUsages: ["sign"], publicUsages: ["verify"]},
-            {algorithm: {name: "ECDH", namedCurve: "P-256"}, privateUsages: ["deriveBits"], publicUsages: []},
-            {algorithm: {name: "Ed25519" }, privateUsages: ["sign"], publicUsages: ["verify"]},
-            {algorithm: {name: "Ed448" }, privateUsages: ["sign"], publicUsages: ["verify"]},
-            {algorithm: {name: "X25519" }, privateUsages: ["deriveBits"], publicUsages: []},
-            {algorithm: {name: "X448" }, privateUsages: ["deriveBits"], publicUsages: []},
-            {algorithm: {name: "AES-CTR", length: 128}, usages: ["encrypt", "decrypt"]},
-            {algorithm: {name: "AES-CBC", length: 128}, usages: ["encrypt", "decrypt"]},
-            {algorithm: {name: "AES-GCM", length: 128}, usages: ["encrypt", "decrypt"]},
-            {algorithm: {name: "AES-KW", length: 128}, usages: ["wrapKey", "unwrapKey"]},
-            {algorithm: {name: "HMAC", length: 128, hash: "SHA-256"}, usages: ["sign", "verify"]}
-        ];
-
-        // Using allSettled to skip unsupported test cases.
-        return Promise.allSettled(parameters.map(function(params) {
-            var usages;
-            if ("usages" in params) {
-                usages = params.usages;
+    function importKeysToWrap() {
+        var promises = [];
+        keysToWrapParameters.forEach(function(params) {
+            if ("publicUsages" in params) {
+                keys[params.algorithm.name] = {};
+                var keyData = toWrapKeyDataFromAlg(params.algorithm.name);
+                promises.push(importAndExport("spki", keyData.spki, params.algorithm, params.publicUsages, "public key "));
+                promises.push(importAndExport("pkcs8", keyData.pkcs8, params.algorithm, params.privateUsages, "private key "));
             } else {
-                usages = params.publicUsages.concat(params.privateUsages);
+                keys[params.algorithm.name] = {};
+                promises.push(importAndExport("raw", toWrapKeyData["SYMMETRIC"].raw, params.algorithm, params.usages, ""));
             }
-
-            return subtle.generateKey(params.algorithm, true, usages)
-            .then(function(result) {
-                if (result.constructor === CryptoKey) {
-                    keys.push({name: params.algorithm.name, algorithm: params.algorithm, usages: params.usages, key: result});
-                } else {
-                    keys.push({name: params.algorithm.name + " public key", algorithm: params.algorithm, usages: params.publicUsages, key: result.publicKey});
-                    keys.push({name: params.algorithm.name + " private key", algorithm: params.algorithm, usages: params.privateUsages, key: result.privateKey});
-                }
-                return true;
-            });
-        }));
+        });
+        // Using allSettled to skip unsupported test cases.
+        return Promise.allSettled(promises);
     }
 
     // Can we successfully "round-trip" (wrap, then unwrap, a key)?
-    function testWrapping(wrapper, toWrap) {
-        var formats;
+    function testWrapping(wrapper, toWrap, fmt) {
+        promise_test(async() => {
+            try {
+                var wrappedResult = await subtle.wrapKey(fmt, toWrap.key, wrapper.wrappingKey, wrapper.parameters.wrapParameters);
+                var unwrappedResult = await subtle.unwrapKey(fmt, wrappedResult, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, true, toWrap.usages);
+                assert_goodCryptoKey(unwrappedResult, toWrap.algorithm, true, toWrap.usages, toWrap.key.type);
+                var roundTripExport = await subtle.exportKey(fmt, unwrappedResult);
+                assert_true(equalExport(toWrap.originalExport[fmt], roundTripExport), "Post-wrap export matches original export");
+            } catch (err) {
+                if (err instanceof AssertionError) {
+                    throw err;
+                }
+                assert_unreached("Round trip for extractable key threw an error - " + err.name + ': "' + err.message + '"');
+            }
+        }, "Can wrap and unwrap " + toWrap.name + "keys using " + fmt + " and " + wrapper.parameters.name);
+    }
 
-        if (toWrap.name.includes("private")) {
-            formats = ["pkcs8", "jwk"];
-        } else if (toWrap.name.includes("public")) {
-            formats = ["spki", "jwk"]
-        } else {
-            formats = ["raw", "jwk"]
-        }
+    function testWrappingNonExtractable(wrapper, toWrap, fmt) {
+        promise_test(async() => {
+            try {
+                var wrappedResult = await subtle.wrapKey(fmt, toWrap.key, wrapper.wrappingKey, wrapper.parameters.wrapParameters);
+                var unwrappedResult = await subtle.unwrapKey(fmt, wrappedResult, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, false, toWrap.usages);
+                assert_goodCryptoKey(unwrappedResult, toWrap.algorithm, false, toWrap.usages, toWrap.key.type);
+                var result = await equalKeys(toWrap.key, unwrappedResult);
+                assert_true(result, "Unwrapped key matches original");
+            } catch (err) {
+                if (err instanceof AssertionError) {
+                    throw err;
+                }
+                assert_unreached("Round trip for key unwrapped non-extractable threw an error - " + err.name + ': "' + err.message + '"');
+            };
+        }, "Can wrap and unwrap " + toWrap.name + "keys as non-extractable using " + fmt + " and " + wrapper.parameters.name);
+    }
 
-        return Promise.all(formats.map(function(fmt) {
-            var originalExport;
-            return subtle.exportKey(fmt, toWrap.key).then(function(exportedKey) {
-                originalExport = exportedKey;
-                const isPossible = wrappingIsPossible(originalExport, wrapper.parameters.name);
-                promise_test(function(test) {
-                    if (!isPossible) {
-                        return Promise.resolve().then(() => {
-                            assert_false(false, "Wrapping is not possible");
-                        })
-                    }
-                    return subtle.wrapKey(fmt, toWrap.key, wrapper.wrappingKey, wrapper.parameters.wrapParameters)
-                    .then(function(wrappedResult) {
-                        return subtle.unwrapKey(fmt, wrappedResult, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, true, toWrap.usages);
-                    }).then(function(unwrappedResult) {
-                        assert_goodCryptoKey(unwrappedResult, toWrap.algorithm, true, toWrap.usages, toWrap.key.type);
-                        return subtle.exportKey(fmt, unwrappedResult)
-                    }).then(function(roundTripExport) {
-                        assert_true(equalExport(originalExport, roundTripExport), "Post-wrap export matches original export");
-                    }, function(err) {
-                        assert_unreached("Round trip for extractable key threw an error - " + err.name + ': "' + err.message + '"');
-                    });
-                }, "Can wrap and unwrap " + toWrap.name + " keys using " + fmt + " and " + wrapper.parameters.name);
-
-                if (canCompareNonExtractableKeys(toWrap.key)) {
-                    promise_test(function(test){
-                        if (!isPossible) {
-                            return Promise.resolve().then(() => {
-                                assert_false(false, "Wrapping is not possible");
-                            })
-                        }
-                        return subtle.wrapKey(fmt, toWrap.key, wrapper.wrappingKey, wrapper.parameters.wrapParameters)
-                        .then(function(wrappedResult) {
-                            return subtle.unwrapKey(fmt, wrappedResult, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, false, toWrap.usages);
-                        }).then(function(unwrappedResult){
-                            assert_goodCryptoKey(unwrappedResult, toWrap.algorithm, false, toWrap.usages, toWrap.key.type);
-                            return equalKeys(toWrap.key, unwrappedResult);
-                        }).then(function(result){
-                            assert_true(result, "Unwrapped key matches original");
-                        }).catch(function(err){
-                            assert_unreached("Round trip for key unwrapped non-extractable threw an error - " + err.name + ': "' + err.message + '"');
-                        });
-                    }, "Can wrap and unwrap " + toWrap.name + " keys as non-extractable using " + fmt + " and " + wrapper.parameters.name);
-
-                    if (fmt === "jwk") {
-                        promise_test(function(test){
-                            if (!isPossible) {
-                                return Promise.resolve().then(() => {
-                                    assert_false(false, "Wrapping is not possible");
-                                })
-                            }
-                            var wrappedKey;
-                            return wrapAsNonExtractableJwk(toWrap.key,wrapper).then(function(wrappedResult){
-                                wrappedKey = wrappedResult;
-                                return subtle.unwrapKey("jwk", wrappedKey, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, false, toWrap.usages);
-                            }).then(function(unwrappedResult){
-                                assert_false(unwrappedResult.extractable, "Unwrapped key is non-extractable");
-                                return equalKeys(toWrap.key,unwrappedResult);
-                            }).then(function(result){
-                                assert_true(result, "Unwrapped key matches original");
-                            }).catch(function(err){
-                                assert_unreached("Round trip for non-extractable key threw an error - " + err.name + ': "' + err.message + '"');
-                            }).then(function(){
-                                return subtle.unwrapKey("jwk", wrappedKey, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, true, toWrap.usages);
-                            }).then(function(unwrappedResult){
-                                assert_unreached("Unwrapping a non-extractable JWK as extractable should fail");
-                            }).catch(function(err){
-                                assert_equals(err.name, "DataError", "Unwrapping a non-extractable JWK as extractable fails with DataError");
-                            });
-                        }, "Can unwrap " + toWrap.name + " non-extractable keys using jwk and " + wrapper.parameters.name);
-                    }
+    function testWrappingNonExtractableAsExtractable(wrapper, toWrap) {
+        promise_test(async() => {
+            var wrappedKey;
+            try {
+                var wrappedResult = await wrapAsNonExtractableJwk(toWrap.key,wrapper);
+                wrappedKey = wrappedResult;
+                var unwrappedResult = await subtle.unwrapKey("jwk", wrappedKey, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, false, toWrap.usages);
+                assert_false(unwrappedResult.extractable, "Unwrapped key is non-extractable");
+                var result = await equalKeys(toWrap.key,unwrappedResult);
+                assert_true(result, "Unwrapped key matches original");
+            } catch (err) {
+                if (err instanceof AssertionError) {
+                    throw err;
                 }
-            });
-        }));
+                assert_unreached("Round trip for non-extractable key threw an error - " + err.name + ': "' + err.message + '"');
+            };
+            try {
+                var unwrappedResult = await subtle.unwrapKey("jwk", wrappedKey, wrapper.unwrappingKey, wrapper.parameters.wrapParameters, toWrap.algorithm, true, toWrap.usages);
+                assert_unreached("Unwrapping a non-extractable JWK as extractable should fail");
+            } catch (err) {
+                if (err instanceof AssertionError) {
+                    throw err;
+                }
+                assert_equals(err.name, "DataError", "Unwrapping a non-extractable JWK as extractable fails with DataError");
+            }
+        }, "Can unwrap " + toWrap.name + "non-extractable keys using jwk and " + wrapper.parameters.name);
     }
 
     // Implement key wrapping by hand to wrap a key as non-extractable JWK
-    function wrapAsNonExtractableJwk(key, wrapper){
+    async function wrapAsNonExtractableJwk(key, wrapper) {
         var wrappingKey = wrapper.wrappingKey,
             encryptKey;
 
-        return subtle.exportKey("jwk",wrappingKey)
-        .then(function(jwkWrappingKey){
-            // Update the key generation parameters to work as key import parameters
-            var params = Object.create(wrapper.parameters.generateParameters);
-            if(params.name === "AES-KW") {
-                params.name = "AES-CBC";
-                jwkWrappingKey.alg = "A"+params.length+"CBC";
-            } else if (params.name === "RSA-OAEP") {
-                params.modulusLength = undefined;
-                params.publicExponent = undefined;
-            }
-            jwkWrappingKey.key_ops = ["encrypt"];
-            return subtle.importKey("jwk", jwkWrappingKey, params, true, ["encrypt"]);
-        }).then(function(importedWrappingKey){
-            encryptKey = importedWrappingKey;
-            return subtle.exportKey("jwk",key);
-        }).then(function(exportedKey){
-            exportedKey.ext = false;
-            var jwk = JSON.stringify(exportedKey)
-            if (wrappingKey.algorithm.name === "AES-KW") {
-                return aeskw(encryptKey, str2ab(jwk.slice(0,-1) + " ".repeat(jwk.length%8 ? 8-jwk.length%8 : 0) + "}"));
-            } else {
-                return subtle.encrypt(wrapper.parameters.wrapParameters,encryptKey,str2ab(jwk));
-            }
-        });
+        var jwkWrappingKey = await subtle.exportKey("jwk",wrappingKey);
+        // Update the key generation parameters to work as key import parameters
+        var params = Object.create(wrapper.parameters.importParameters);
+        if(params.name === "AES-KW") {
+            params.name = "AES-CBC";
+            jwkWrappingKey.alg = "A"+params.length+"CBC";
+        } else if (params.name === "RSA-OAEP") {
+            params.modulusLength = undefined;
+            params.publicExponent = undefined;
+        }
+        jwkWrappingKey.key_ops = ["encrypt"];
+        var importedWrappingKey = await subtle.importKey("jwk", jwkWrappingKey, params, true, ["encrypt"]);
+        encryptKey = importedWrappingKey;
+        var exportedKey = await subtle.exportKey("jwk",key);
+        exportedKey.ext = false;
+        var jwk = JSON.stringify(exportedKey)
+        var result;
+        if (wrappingKey.algorithm.name === "AES-KW") {
+            result = await aeskw(encryptKey, str2ab(jwk.slice(0,-1) + " ".repeat(jwk.length%8 ? 8-jwk.length%8 : 0) + "}"));
+        } else {
+            result = await subtle.encrypt(wrapper.parameters.wrapParameters,encryptKey,str2ab(jwk));
+        }
+        return result;
     }
 
 
@@ -365,9 +383,9 @@
     }
 
     // Compare two keys by using them (works for non-extractable keys)
-    function equalKeys(expected, got){
+    async function equalKeys(expected, got){
         if ( expected.algorithm.name !== got.algorithm.name ) {
-            return Promise.resolve(false);
+            return false;
         }
 
         var cryptParams, signParams, wrapParams, deriveParams;
@@ -419,75 +437,60 @@
         }
 
         if (cryptParams) {
-            return subtle.exportKey("jwk",expected)
-            .then(function(jwkExpectedKey){
-                if (expected.algorithm.name === "RSA-OAEP") {
-                    ["d","p","q","dp","dq","qi","oth"].forEach(function(field){ delete jwkExpectedKey[field]; });
-                }
-                jwkExpectedKey.key_ops = ["encrypt"];
-                return subtle.importKey("jwk", jwkExpectedKey, expected.algorithm, true, ["encrypt"]);
-            }).then(function(expectedEncryptKey){
-                return subtle.encrypt(cryptParams, expectedEncryptKey, new Uint8Array(32));
-            }).then(function(encryptedData){
-                return subtle.decrypt(cryptParams, got, encryptedData);
-            }).then(function(decryptedData){
-                var result = new Uint8Array(decryptedData);
-                return !result.some(x => x);
-            });
+            var jwkExpectedKey = await subtle.exportKey("jwk", expected);
+            if (expected.algorithm.name === "RSA-OAEP") {
+                ["d","p","q","dp","dq","qi","oth"].forEach(function(field){ delete jwkExpectedKey[field]; });
+            }
+            jwkExpectedKey.key_ops = ["encrypt"];
+            var expectedEncryptKey = await subtle.importKey("jwk", jwkExpectedKey, expected.algorithm, true, ["encrypt"]);
+            var encryptedData = await subtle.encrypt(cryptParams, expectedEncryptKey, new Uint8Array(32));
+            var decryptedData = await subtle.decrypt(cryptParams, got, encryptedData);
+            var result = new Uint8Array(decryptedData);
+            return !result.some(x => x);
         } else if (signParams) {
             var verifyKey;
-            return subtle.exportKey("jwk",expected)
-            .then(function(jwkExpectedKey){
-                if (expected.algorithm.name === "RSA-PSS" || expected.algorithm.name === "RSASSA-PKCS1-v1_5") {
-                    ["d","p","q","dp","dq","qi","oth"].forEach(function(field){ delete jwkExpectedKey[field]; });
-                }
-                if (expected.algorithm.name === "ECDSA" || expected.algorithm.name.startsWith("Ed")) {
-                    delete jwkExpectedKey["d"];
-                }
-                jwkExpectedKey.key_ops = ["verify"];
-                return subtle.importKey("jwk", jwkExpectedKey, expected.algorithm, true, ["verify"]);
-            }).then(function(expectedVerifyKey){
-                verifyKey = expectedVerifyKey;
-                return subtle.sign(signParams, got, new Uint8Array(32));
-            }).then(function(signature){
-                return subtle.verify(signParams, verifyKey, signature, new Uint8Array(32));
-            });
+            var jwkExpectedKey = await subtle.exportKey("jwk",expected);
+            if (expected.algorithm.name === "RSA-PSS" || expected.algorithm.name === "RSASSA-PKCS1-v1_5") {
+                ["d","p","q","dp","dq","qi","oth"].forEach(function(field){ delete jwkExpectedKey[field]; });
+            }
+            if (expected.algorithm.name === "ECDSA" || expected.algorithm.name.startsWith("Ed")) {
+                delete jwkExpectedKey["d"];
+            }
+            jwkExpectedKey.key_ops = ["verify"];
+            var expectedVerifyKey = await subtle.importKey("jwk", jwkExpectedKey, expected.algorithm, true, ["verify"]);
+            verifyKey = expectedVerifyKey;
+            var signature = await subtle.sign(signParams, got, new Uint8Array(32));
+            var result = await subtle.verify(signParams, verifyKey, signature, new Uint8Array(32));
+            return result;
         } else if (wrapParams) {
             var aKeyToWrap, wrappedWithExpected;
-            return subtle.importKey("raw", new Uint8Array(16), "AES-CBC", true, ["encrypt"])
-            .then(function(key){
-                aKeyToWrap = key;
-                return subtle.wrapKey("raw", aKeyToWrap, expected, wrapParams);
-            }).then(function(wrapResult){
-                wrappedWithExpected = Array.from((new Uint8Array(wrapResult)).values());
-                return subtle.wrapKey("raw", aKeyToWrap, got, wrapParams);
-            }).then(function(wrapResult){
-                var wrappedWithGot = Array.from((new Uint8Array(wrapResult)).values());
-                return wrappedWithGot.every((x,i) => x === wrappedWithExpected[i]);
-            });
+            var key = await subtle.importKey("raw",new Uint8Array(16), "AES-CBC", true, ["encrypt"])
+            aKeyToWrap = key;
+            var wrapResult = await subtle.wrapKey("raw", aKeyToWrap, expected, wrapParams);
+            wrappedWithExpected = Array.from((new Uint8Array(wrapResult)).values());
+            wrapResult = await subtle.wrapKey("raw", aKeyToWrap, got, wrapParams);
+            var wrappedWithGot = Array.from((new Uint8Array(wrapResult)).values());
+            return wrappedWithGot.every((x,i) => x === wrappedWithExpected[i]);
         } else if (deriveParams) {
             var expectedDerivedBits;
-            return subtle.generateKey(expected.algorithm, true, ['deriveBits']).then(({ publicKey }) => {
-                deriveParams.public = publicKey;
-                return subtle.deriveBits(deriveParams, expected, 128)
-            })
-            .then(function(result){
-                expectedDerivedBits = Array.from((new Uint8Array(result)).values());
-                return subtle.deriveBits(deriveParams, got, 128);
-            }).then(function(result){
-                var gotDerivedBits = Array.from((new Uint8Array(result)).values());
-                return gotDerivedBits.every((x,i) => x === expectedDerivedBits[i]);
-            });
+            var key = await subtle.generateKey(expected.algorithm, true, ['deriveBits']);
+            deriveParams.public = key.publicKey;
+            var result = await subtle.deriveBits(deriveParams, expected, 128);
+            expectedDerivedBits = Array.from((new Uint8Array(result)).values());
+            result = await subtle.deriveBits(deriveParams, got, 128);
+            var gotDerivedBits = Array.from((new Uint8Array(result)).values());
+            return gotDerivedBits.every((x,i) => x === expectedDerivedBits[i]);
         }
     }
 
     // Raw AES encryption
-    function aes( k, p ) {
-        return subtle.encrypt({name: "AES-CBC", iv: new Uint8Array(16) }, k, p).then(function(ciphertext){return ciphertext.slice(0,16);});
+    async function aes(k, p) {
+        const ciphertext = await subtle.encrypt({ name: "AES-CBC", iv: new Uint8Array(16) }, k, p);
+        return ciphertext.slice(0, 16);
     }
 
     // AES Key Wrap
-    function aeskw(key, data) {
+    async function aeskw(key, data) {
         if (data.byteLength % 8 !== 0) {
             throw new Error("AES Key Wrap data must be a multiple of 8 bytes in length");
         }
@@ -501,7 +504,7 @@
             R.push(new Uint8Array(data.slice(i,i+8)));
         }
 
-        function aeskw_step(j, i, final, B) {
+        async function aeskw_step(j, i, final, B) {
             A.set(new Uint8Array(B.slice(0,8)));
             Av.setUint32(4,Av.getUint32(4) ^ (n*j+i+1));
             R[i] = new Uint8Array(B.slice(8,16));
@@ -516,18 +519,16 @@
             }
         }
 
-        var p = new Promise(function(resolve){
-            A.set(R[0],8);
-            resolve(aes(key,A));
-        });
+        A.set(R[0], 8);
+        let B = await aes(key, A);
 
         for(var j=0;j<6;++j) {
             for(var i=0;i<n;++i) {
-                p = p.then(aeskw_step.bind(undefined, j, i,j===5 && i===(n-1)));
+                B = await aeskw_step(j, i, j === 5 && i === (n - 1), B);
             }
         }
 
-        return p;
+        return B;
     }
 
     function str2ab(str)        { return Uint8Array.from( str.split(''), function(s){return s.charCodeAt(0)} ); }
diff --git a/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey_vectors.js b/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey_vectors.js
new file mode 100644
index 00000000000000..0c29a7df88b15e
--- /dev/null
+++ b/test/fixtures/wpt/WebCryptoAPI/wrapKey_unwrapKey/wrapKey_unwrapKey_vectors.js
@@ -0,0 +1,114 @@
+let wrappingKeyData = {
+    "RSA": { // modulusLength: 4096, hash: "SHA-256"
+        spki: new Uint8Array([48, 130, 2, 34, 48, 13, 6, 9, 42, 134, 72, 134, 247, 13, 1, 1, 1, 5, 0, 3, 130, 2, 15, 0, 48, 130, 2, 10, 2, 130, 2, 1, 0, 218, 170, 246, 76, 189, 156, 216, 153, 155, 176, 221, 14, 44, 132, 103, 104, 0, 127, 100, 166, 245, 248, 104, 125, 31, 74, 155, 226, 90, 193, 184, 54, 170, 145, 111, 222, 20, 252, 19, 248, 146, 44, 190, 115, 73, 188, 52, 251, 4, 178, 121, 238, 212, 204, 34, 62, 122, 100, 203, 111, 233, 231, 210, 73, 53, 146, 147, 211, 14, 161, 109, 137, 212, 175, 226, 18, 183, 173, 103, 103, 30, 128, 31, 218, 69, 126, 234, 65, 88, 231, 160, 91, 51, 245, 77, 54, 4, 167, 192, 33, 68, 244, 163, 242, 187, 111, 209, 180, 241, 221, 107, 172, 5, 40, 134, 47, 210, 85, 8, 112, 57, 186, 29, 131, 176, 93, 116, 198, 202, 82, 108, 251, 209, 3, 72, 75, 143, 59, 44, 222, 56, 89, 69, 103, 159, 211, 160, 19, 214, 173, 77, 133, 0, 68, 219, 164, 79, 64, 238, 65, 189, 201, 248, 173, 180, 146, 196, 238, 86, 232, 215, 109, 39, 165, 162, 16, 230, 46, 134, 234, 148, 106, 34, 230, 198, 63, 231, 143, 16, 179, 208, 109, 22, 100, 54, 156, 107, 132, 28, 208, 118, 205, 217, 89, 228, 75, 196, 169, 181, 5, 85, 157, 144, 110, 129, 186, 141, 119, 104, 162, 206, 170, 115, 7, 96, 82, 240, 33, 143, 81, 243, 215, 67, 96, 137, 207, 209, 22, 162, 251, 108, 208, 232, 32, 236, 205, 167, 174, 161, 116, 13, 249, 187, 22, 240, 185, 172, 160, 103, 94, 162, 147, 26, 15, 143, 183, 147, 98, 231, 117, 134, 185, 50, 64, 40, 30, 27, 13, 152, 132, 40, 138, 32, 78, 158, 162, 207, 212, 229, 210, 251, 88, 116, 67, 229, 164, 164, 147, 59, 32, 94, 217, 197, 242, 149, 102, 74, 219, 46, 127, 68, 28, 116, 10, 2, 249, 231, 130, 123, 29, 45, 73, 56, 17, 195, 208, 45, 25, 60, 252, 98, 189, 109, 25, 0, 253, 151, 254, 124, 211, 48, 23, 156, 78, 163, 154, 188, 17, 69, 14, 188, 16, 64, 59, 190, 136, 70, 162, 253, 237, 156, 111, 41, 27, 40, 63, 205, 204, 94, 0, 50, 237, 62, 87, 211, 115, 91, 68, 194, 104, 119, 72, 106, 226, 160, 48, 165, 138, 134, 2, 138, 153, 181, 38, 249, 48, 120, 72, 15, 245, 227, 15, 164, 64, 188, 74, 4, 84, 213, 83, 67, 73, 87, 181, 72, 94, 46, 54, 193, 252, 188, 14, 207, 28, 82, 159, 131, 168, 238, 168, 145, 28, 230, 27, 126, 151, 93, 5, 96, 68, 126, 66, 174, 155, 101, 123, 20, 218, 131, 92, 124, 78, 82, 44, 55, 139, 77, 105, 177, 136, 121, 177, 43, 77, 12, 240, 0, 76, 20, 133, 121, 129, 73, 15, 160, 200, 150, 114, 95, 59, 59, 165, 240, 204, 13, 156, 134, 194, 4, 70, 158, 213, 111, 229, 103, 216, 239, 132, 16, 184, 151, 206, 254, 229, 62, 23, 58, 125, 49, 144, 208, 215, 2, 3, 1, 0, 1]),
+        pkcs8: new Uint8Array([48, 130, 9, 68, 2, 1, 0, 48, 13, 6, 9, 42, 134, 72, 134, 247, 13, 1, 1, 1, 5, 0, 4, 130, 9, 46, 48, 130, 9, 42, 2, 1, 0, 2, 130, 2, 1, 0, 218, 170, 246, 76, 189, 156, 216, 153, 155, 176, 221, 14, 44, 132, 103, 104, 0, 127, 100, 166, 245, 248, 104, 125, 31, 74, 155, 226, 90, 193, 184, 54, 170, 145, 111, 222, 20, 252, 19, 248, 146, 44, 190, 115, 73, 188, 52, 251, 4, 178, 121, 238, 212, 204, 34, 62, 122, 100, 203, 111, 233, 231, 210, 73, 53, 146, 147, 211, 14, 161, 109, 137, 212, 175, 226, 18, 183, 173, 103, 103, 30, 128, 31, 218, 69, 126, 234, 65, 88, 231, 160, 91, 51, 245, 77, 54, 4, 167, 192, 33, 68, 244, 163, 242, 187, 111, 209, 180, 241, 221, 107, 172, 5, 40, 134, 47, 210, 85, 8, 112, 57, 186, 29, 131, 176, 93, 116, 198, 202, 82, 108, 251, 209, 3, 72, 75, 143, 59, 44, 222, 56, 89, 69, 103, 159, 211, 160, 19, 214, 173, 77, 133, 0, 68, 219, 164, 79, 64, 238, 65, 189, 201, 248, 173, 180, 146, 196, 238, 86, 232, 215, 109, 39, 165, 162, 16, 230, 46, 134, 234, 148, 106, 34, 230, 198, 63, 231, 143, 16, 179, 208, 109, 22, 100, 54, 156, 107, 132, 28, 208, 118, 205, 217, 89, 228, 75, 196, 169, 181, 5, 85, 157, 144, 110, 129, 186, 141, 119, 104, 162, 206, 170, 115, 7, 96, 82, 240, 33, 143, 81, 243, 215, 67, 96, 137, 207, 209, 22, 162, 251, 108, 208, 232, 32, 236, 205, 167, 174, 161, 116, 13, 249, 187, 22, 240, 185, 172, 160, 103, 94, 162, 147, 26, 15, 143, 183, 147, 98, 231, 117, 134, 185, 50, 64, 40, 30, 27, 13, 152, 132, 40, 138, 32, 78, 158, 162, 207, 212, 229, 210, 251, 88, 116, 67, 229, 164, 164, 147, 59, 32, 94, 217, 197, 242, 149, 102, 74, 219, 46, 127, 68, 28, 116, 10, 2, 249, 231, 130, 123, 29, 45, 73, 56, 17, 195, 208, 45, 25, 60, 252, 98, 189, 109, 25, 0, 253, 151, 254, 124, 211, 48, 23, 156, 78, 163, 154, 188, 17, 69, 14, 188, 16, 64, 59, 190, 136, 70, 162, 253, 237, 156, 111, 41, 27, 40, 63, 205, 204, 94, 0, 50, 237, 62, 87, 211, 115, 91, 68, 194, 104, 119, 72, 106, 226, 160, 48, 165, 138, 134, 2, 138, 153, 181, 38, 249, 48, 120, 72, 15, 245, 227, 15, 164, 64, 188, 74, 4, 84, 213, 83, 67, 73, 87, 181, 72, 94, 46, 54, 193, 252, 188, 14, 207, 28, 82, 159, 131, 168, 238, 168, 145, 28, 230, 27, 126, 151, 93, 5, 96, 68, 126, 66, 174, 155, 101, 123, 20, 218, 131, 92, 124, 78, 82, 44, 55, 139, 77, 105, 177, 136, 121, 177, 43, 77, 12, 240, 0, 76, 20, 133, 121, 129, 73, 15, 160, 200, 150, 114, 95, 59, 59, 165, 240, 204, 13, 156, 134, 194, 4, 70, 158, 213, 111, 229, 103, 216, 239, 132, 16, 184, 151, 206, 254, 229, 62, 23, 58, 125, 49, 144, 208, 215, 2, 3, 1, 0, 1, 2, 130, 2, 1, 0, 185, 115, 209, 92, 24, 92, 19, 159, 131, 89, 166, 193, 68, 164, 46, 135, 24, 20, 243, 42, 94, 230, 4, 200, 73, 103, 159, 121, 131, 251, 83, 222, 153, 30, 171, 191, 176, 16, 114, 103, 152, 161, 118, 12, 148, 246, 152, 0, 100, 101, 113, 224, 74, 125, 174, 117, 74, 156, 125, 165, 54, 189, 179, 172, 255, 80, 135, 42, 178, 247, 217, 204, 209, 163, 49, 155, 42, 72, 88, 176, 46, 63, 255, 195, 192, 184, 248, 183, 223, 76, 226, 197, 54, 245, 206, 60, 8, 10, 181, 122, 1, 223, 113, 196, 133, 143, 58, 77, 185, 235, 78, 76, 32, 59, 212, 66, 110, 162, 75, 123, 210, 153, 180, 58, 97, 179, 129, 60, 175, 142, 228, 123, 85, 50, 241, 119, 147, 204, 94, 43, 65, 163, 4, 167, 243, 247, 41, 134, 105, 197, 165, 63, 45, 145, 56, 174, 203, 192, 135, 209, 29, 195, 83, 179, 14, 184, 131, 104, 152, 48, 245, 179, 207, 178, 60, 23, 21, 1, 84, 207, 82, 124, 9, 137, 171, 141, 187, 55, 172, 180, 180, 10, 48, 185, 97, 79, 156, 39, 249, 192, 27, 98, 77, 250, 93, 18, 157, 130, 72, 210, 115, 96, 36, 132, 116, 101, 225, 96, 234, 79, 89, 243, 89, 135, 97, 252, 53, 72, 97, 34, 226, 41, 41, 45, 144, 243, 189, 162, 243, 43, 69, 136, 143, 182, 140, 223, 134, 93, 38, 245, 36, 125, 46, 93, 48, 94, 215, 39, 156, 57, 86, 93, 207, 204, 72, 106, 112, 215, 203, 230, 80, 20, 137, 224, 242, 33, 146, 33, 108, 188, 185, 254, 117, 189, 240, 82, 64, 60, 186, 247, 190, 138, 170, 159, 147, 75, 49, 148, 101, 174, 130, 21, 177, 211, 121, 6, 153, 144, 230, 166, 181, 155, 94, 232, 2, 4, 119, 236, 35, 133, 253, 223, 14, 30, 199, 57, 215, 31, 251, 90, 167, 19, 231, 154, 54, 225, 85, 68, 17, 234, 158, 53, 50, 243, 182, 149, 193, 214, 60, 188, 6, 38, 2, 200, 161, 232, 193, 30, 153, 231, 221, 57, 140, 55, 69, 35, 21, 153, 34, 238, 175, 65, 253, 210, 119, 125, 120, 116, 153, 127, 67, 204, 9, 66, 210, 200, 165, 212, 216, 2, 62, 19, 15, 171, 77, 183, 247, 127, 224, 138, 41, 208, 170, 227, 36, 158, 176, 111, 128, 172, 70, 73, 241, 148, 172, 50, 174, 126, 80, 177, 235, 93, 89, 102, 84, 76, 221, 30, 216, 49, 125, 142, 35, 45, 96, 224, 60, 161, 63, 48, 85, 143, 20, 76, 182, 111, 15, 156, 139, 55, 155, 113, 226, 248, 239, 130, 252, 241, 197, 247, 124, 61, 39, 197, 170, 119, 76, 136, 195, 180, 169, 106, 240, 234, 101, 114, 207, 11, 160, 170, 139, 194, 187, 48, 22, 114, 84, 64, 151, 30, 212, 99, 213, 176, 106, 79, 232, 127, 197, 153, 133, 8, 56, 210, 83, 67, 106, 124, 231, 96, 2, 145, 2, 130, 1, 1, 0, 244, 218, 215, 194, 174, 36, 99, 217, 1, 4, 236, 11, 160, 86, 85, 65, 206, 36, 36, 143, 205, 108, 166, 191, 91, 209, 75, 117, 7, 81, 33, 179, 44, 101, 145, 215, 39, 117, 195, 81, 31, 111, 36, 7, 26, 105, 30, 249, 91, 2, 2, 237, 126, 141, 231, 153, 213, 181, 100, 234, 219, 192, 114, 179, 215, 229, 39, 212, 107, 9, 55, 220, 136, 233, 237, 28, 74, 97, 6, 22, 26, 47, 150, 83, 82, 95, 186, 146, 22, 38, 176, 231, 255, 166, 199, 223, 217, 86, 142, 56, 43, 199, 25, 247, 249, 122, 59, 142, 152, 20, 49, 147, 13, 132, 249, 203, 251, 146, 116, 96, 88, 81, 232, 45, 106, 100, 187, 99, 73, 32, 203, 134, 30, 223, 100, 179, 179, 128, 81, 242, 25, 85, 137, 125, 96, 153, 240, 224, 86, 20, 206, 24, 26, 197, 233, 164, 158, 50, 222, 103, 197, 211, 144, 101, 182, 205, 201, 51, 23, 231, 125, 229, 130, 61, 139, 204, 195, 243, 69, 38, 185, 187, 48, 249, 140, 107, 137, 39, 234, 21, 13, 43, 24, 112, 108, 109, 15, 25, 57, 55, 127, 40, 152, 238, 227, 96, 86, 157, 114, 35, 52, 54, 38, 140, 85, 42, 119, 53, 99, 35, 133, 208, 240, 65, 171, 8, 71, 255, 243, 248, 176, 166, 17, 178, 92, 62, 203, 56, 158, 31, 169, 223, 123, 7, 118, 216, 166, 132, 83, 62, 112, 160, 99, 244, 132, 29, 2, 130, 1, 1, 0, 228, 158, 249, 243, 243, 94, 42, 189, 87, 61, 152, 139, 197, 122, 33, 97, 4, 39, 135, 66, 219, 225, 11, 70, 103, 92, 115, 10, 8, 225, 5, 2, 220, 32, 23, 147, 56, 111, 237, 98, 48, 174, 122, 207, 109, 152, 187, 125, 220, 186, 73, 127, 42, 82, 39, 228, 163, 12, 188, 36, 71, 107, 52, 235, 223, 200, 7, 38, 6, 167, 28, 158, 26, 213, 126, 186, 90, 152, 133, 44, 53, 156, 61, 130, 92, 163, 3, 27, 35, 185, 141, 112, 236, 246, 210, 107, 75, 245, 33, 126, 134, 215, 41, 1, 244, 220, 36, 93, 22, 232, 50, 62, 68, 141, 153, 118, 62, 1, 167, 197, 202, 113, 187, 196, 186, 251, 161, 128, 66, 211, 145, 103, 133, 69, 207, 155, 117, 65, 76, 251, 125, 43, 224, 105, 171, 6, 29, 254, 31, 111, 144, 5, 158, 166, 180, 143, 163, 205, 212, 151, 7, 11, 50, 234, 82, 37, 143, 75, 104, 124, 97, 69, 220, 246, 202, 45, 25, 40, 220, 23, 92, 116, 112, 114, 204, 198, 140, 48, 111, 191, 53, 28, 9, 134, 234, 90, 168, 243, 108, 75, 197, 99, 162, 173, 31, 194, 97, 224, 184, 76, 227, 170, 199, 106, 129, 14, 77, 234, 231, 38, 192, 197, 233, 174, 150, 240, 55, 252, 241, 27, 97, 169, 49, 49, 115, 9, 218, 65, 253, 14, 253, 217, 91, 141, 44, 68, 32, 247, 219, 199, 31, 45, 212, 68, 46, 131, 2, 130, 1, 1, 0, 225, 142, 199, 187, 155, 88, 2, 114, 225, 49, 123, 144, 170, 63, 93, 130, 165, 55, 62, 71, 10, 97, 208, 169, 239, 23, 58, 127, 176, 33, 216, 253, 137, 36, 119, 216, 207, 140, 248, 68, 62, 196, 207, 87, 139, 200, 210, 179, 186, 86, 124, 3, 243, 213, 29, 72, 229, 73, 152, 145, 145, 166, 19, 4, 1, 26, 36, 58, 213, 239, 67, 250, 112, 85, 174, 11, 165, 169, 3, 70, 81, 17, 13, 85, 236, 72, 43, 66, 112, 13, 108, 98, 11, 107, 196, 44, 61, 182, 50, 133, 36, 46, 225, 137, 65, 212, 140, 16, 171, 159, 206, 155, 60, 149, 6, 216, 22, 3, 176, 25, 32, 195, 51, 50, 195, 19, 208, 91, 129, 254, 39, 254, 129, 106, 33, 6, 57, 145, 55, 235, 225, 210, 158, 57, 85, 71, 250, 81, 110, 122, 243, 239, 216, 154, 0, 197, 152, 198, 27, 131, 85, 5, 179, 187, 63, 79, 10, 205, 122, 115, 209, 210, 30, 204, 59, 128, 129, 242, 19, 253, 188, 146, 232, 102, 186, 40, 69, 204, 243, 34, 57, 99, 61, 188, 50, 229, 180, 70, 244, 34, 95, 141, 50, 116, 190, 24, 253, 49, 68, 247, 145, 29, 97, 29, 93, 71, 37, 81, 148, 230, 32, 91, 125, 55, 193, 42, 123, 201, 25, 34, 58, 248, 128, 204, 225, 149, 38, 248, 29, 17, 230, 22, 236, 234, 207, 92, 124, 232, 225, 22, 96, 2, 32, 146, 27, 49, 2, 130, 1, 1, 0, 129, 62, 34, 61, 183, 242, 31, 37, 68, 193, 108, 144, 111, 133, 248, 130, 184, 239, 131, 182, 215, 72, 164, 176, 27, 84, 151, 48, 48, 14, 205, 95, 109, 131, 178, 240, 38, 50, 152, 55, 47, 32, 36, 11, 73, 128, 211, 85, 118, 199, 213, 46, 207, 132, 252, 74, 115, 166, 138, 97, 212, 2, 22, 59, 214, 25, 101, 121, 40, 191, 166, 28, 247, 60, 132, 84, 227, 76, 95, 212, 187, 69, 229, 59, 226, 20, 193, 119, 193, 61, 111, 105, 76, 124, 200, 61, 162, 6, 36, 246, 59, 82, 61, 59, 126, 234, 72, 160, 91, 135, 206, 135, 135, 7, 169, 158, 191, 180, 253, 220, 129, 242, 195, 220, 150, 124, 20, 51, 199, 19, 133, 154, 201, 43, 203, 14, 174, 61, 201, 64, 78, 229, 212, 10, 200, 133, 63, 197, 94, 142, 26, 20, 35, 57, 72, 207, 255, 33, 40, 50, 108, 231, 246, 211, 162, 182, 219, 8, 29, 60, 91, 93, 60, 106, 67, 167, 53, 22, 245, 61, 59, 166, 19, 191, 194, 101, 231, 240, 165, 235, 169, 33, 125, 125, 72, 213, 17, 183, 243, 27, 238, 173, 193, 212, 47, 37, 27, 98, 7, 174, 103, 242, 46, 163, 213, 235, 121, 62, 247, 135, 223, 232, 194, 143, 81, 130, 225, 147, 219, 213, 199, 226, 247, 13, 102, 100, 70, 127, 145, 136, 189, 22, 248, 123, 153, 111, 182, 87, 136, 102, 76, 9, 3, 123, 187, 243, 2, 130, 1, 0, 36, 121, 149, 41, 189, 115, 193, 110, 98, 69, 30, 145, 9, 231, 177, 98, 120, 118, 126, 102, 62, 220, 58, 207, 73, 211, 60, 15, 24, 107, 208, 95, 29, 107, 40, 190, 182, 84, 106, 17, 217, 198, 210, 27, 233, 227, 153, 252, 128, 181, 44, 145, 101, 156, 7, 209, 23, 149, 66, 78, 109, 145, 138, 13, 241, 174, 198, 3, 26, 222, 15, 241, 120, 176, 54, 190, 97, 80, 215, 99, 49, 62, 204, 135, 226, 32, 141, 102, 251, 32, 152, 108, 113, 237, 59, 142, 30, 185, 195, 135, 145, 1, 86, 115, 56, 253, 215, 186, 221, 202, 196, 36, 227, 118, 177, 130, 60, 59, 56, 190, 198, 157, 142, 18, 96, 43, 218, 199, 150, 42, 174, 44, 198, 65, 103, 139, 167, 177, 46, 26, 155, 248, 209, 56, 155, 209, 204, 42, 89, 224, 212, 75, 80, 135, 106, 203, 4, 81, 181, 85, 128, 247, 73, 134, 41, 48, 183, 57, 127, 28, 234, 26, 244, 177, 159, 113, 90, 249, 120, 32, 248, 134, 79, 99, 123, 155, 173, 201, 185, 216, 166, 32, 152, 181, 6, 154, 118, 18, 181, 245, 106, 25, 37, 146, 118, 16, 215, 30, 83, 96, 35, 154, 93, 0, 13, 5, 206, 156, 129, 147, 118, 87, 248, 155, 49, 135, 7, 39, 157, 226, 171, 96, 16, 112, 122, 173, 58, 145, 19, 6, 90, 11, 221, 109, 208, 16, 251, 188, 18, 120, 106, 170, 143, 149, 79, 192]),
+        jwk: {
+            kty: "RSA",
+            n: "2qr2TL2c2JmbsN0OLIRnaAB_ZKb1-Gh9H0qb4lrBuDaqkW_eFPwT-JIsvnNJvDT7BLJ57tTMIj56ZMtv6efSSTWSk9MOoW2J1K_iEretZ2cegB_aRX7qQVjnoFsz9U02BKfAIUT0o_K7b9G08d1rrAUohi_SVQhwObodg7BddMbKUmz70QNIS487LN44WUVnn9OgE9atTYUARNukT0DuQb3J-K20ksTuVujXbSelohDmLobqlGoi5sY_548Qs9BtFmQ2nGuEHNB2zdlZ5EvEqbUFVZ2QboG6jXdoos6qcwdgUvAhj1Hz10Ngic_RFqL7bNDoIOzNp66hdA35uxbwuaygZ16ikxoPj7eTYud1hrkyQCgeGw2YhCiKIE6eos_U5dL7WHRD5aSkkzsgXtnF8pVmStsuf0QcdAoC-eeCex0tSTgRw9AtGTz8Yr1tGQD9l_580zAXnE6jmrwRRQ68EEA7vohGov3tnG8pGyg_zcxeADLtPlfTc1tEwmh3SGrioDClioYCipm1JvkweEgP9eMPpEC8SgRU1VNDSVe1SF4uNsH8vA7PHFKfg6juqJEc5ht-l10FYER-Qq6bZXsU2oNcfE5SLDeLTWmxiHmxK00M8ABMFIV5gUkPoMiWcl87O6XwzA2chsIERp7Vb-Vn2O-EELiXzv7lPhc6fTGQ0Nc",
+            e: "AQAB",
+            d: "uXPRXBhcE5-DWabBRKQuv,hxgU8ype5gTISWefeYP7U96ZHqu_sBByZ5ihdgyU9pgAZGVx4Ep9rnVKnH2lNr2zrP9Qhyqy99nM0aMxmypIWLAuP__DwLj4t99M4sU29c48CAq1egHfccSFjzpNuetOTCA71EJuokt70pm0OmGzgTyvjuR7VTLxd5PMXitBowSn8_cphmnFpT8tkTiuy8CH0R3DU7MOuINomDD1s8-yPBcVAVTPUnwJiauNuzestLQKMLlhT5wn-cAbYk36XRKdgkjSc2AkhHRl4WDqT1nzWYdh_DVIYSLiKSktkPO9ovMrRYiPtozfhl0m9SR9Ll0wXtcnnDlWXc_MSGpw18vmUBSJ4PIhkiFsvLn-db3wUkA8uve-iqqfk0sxlGWughWx03kGmZDmprWbXugCBHfsI4X93w4exznXH_tapxPnmjbhVUQR6p41MvO2lcHWPLwGJgLIoejBHpnn3TmMN0UjFZki7q9B_dJ3fXh0mX9DzAlC0sil1NgCPhMPq02393_giinQquMknrBvgKxGSfGUrDKuflCx611ZZlRM3R7YMX2OIy1g4DyhPzBVjxRMtm8PnIs3m3Hi-O-C_PHF93w9J8Wqd0yIw7SpavDqZXLPC6Cqi8K7MBZyVECXHtRj1bBqT-h_xZmFCDjSU0NqfOdgApE",
+            p: "9NrXwq4kY9kBBOwLoFZVQc4kJI_NbKa_W9FLdQdRIbMsZZHXJ3XDUR9vJAcaaR75WwIC7X6N55nVtWTq28Bys9flJ9RrCTfciOntHEphBhYaL5ZTUl-6khYmsOf_psff2VaOOCvHGff5ejuOmBQxkw2E-cv7knRgWFHoLWpku2NJIMuGHt9ks7OAUfIZVYl9YJnw4FYUzhgaxemknjLeZ8XTkGW2zckzF-d95YI9i8zD80Umubsw-YxriSfqFQ0rGHBsbQ8ZOTd_KJju42BWnXIjNDYmjFUqdzVjI4XQ8EGrCEf_8_iwphGyXD7LOJ4fqd97B3bYpoRTPnCgY_SEHQ",
+            q: "5J758_NeKr1XPZiLxXohYQQnh0Lb4QtGZ1xzCgjhBQLcIBeTOG_tYjCues9tmLt93LpJfypSJ-SjDLwkR2s069_IByYGpxyeGtV-ulqYhSw1nD2CXKMDGyO5jXDs9tJrS_UhfobXKQH03CRdFugyPkSNmXY-AafFynG7xLr7oYBC05FnhUXPm3VBTPt9K-BpqwYd_h9vkAWeprSPo83UlwcLMupSJY9LaHxhRdz2yi0ZKNwXXHRwcszGjDBvvzUcCYbqWqjzbEvFY6KtH8Jh4LhM46rHaoEOTernJsDF6a6W8Df88RthqTExcwnaQf0O_dlbjSxEIPfbxx8t1EQugw",
+            dp: "4Y7Hu5tYAnLhMXuQqj9dgqU3PkcKYdCp7xc6f7Ah2P2JJHfYz4z4RD7Ez1eLyNKzulZ8A_PVHUjlSZiRkaYTBAEaJDrV70P6cFWuC6WpA0ZREQ1V7EgrQnANbGILa8QsPbYyhSQu4YlB1IwQq5_OmzyVBtgWA7AZIMMzMsMT0FuB_if-gWohBjmRN-vh0p45VUf6UW568-_YmgDFmMYbg1UFs7s_TwrNenPR0h7MO4CB8hP9vJLoZrooRczzIjljPbwy5bRG9CJfjTJ0vhj9MUT3kR1hHV1HJVGU5iBbfTfBKnvJGSI6-IDM4ZUm-B0R5hbs6s9cfOjhFmACIJIbMQ",
+            dq: "gT4iPbfyHyVEwWyQb4X4grjvg7bXSKSwG1SXMDAOzV9tg7LwJjKYNy8gJAtJgNNVdsfVLs-E_Epzpoph1AIWO9YZZXkov6Yc9zyEVONMX9S7ReU74hTBd8E9b2lMfMg9ogYk9jtSPTt-6kigW4fOh4cHqZ6_tP3cgfLD3JZ8FDPHE4WaySvLDq49yUBO5dQKyIU_xV6OGhQjOUjP_yEoMmzn9tOittsIHTxbXTxqQ6c1FvU9O6YTv8Jl5_Cl66khfX1I1RG38xvurcHULyUbYgeuZ_Iuo9XreT73h9_owo9RguGT29XH4vcNZmRGf5GIvRb4e5lvtleIZkwJA3u78w",
+            qi: "JHmVKb1zwW5iRR6RCeexYnh2fmY-3DrPSdM8Dxhr0F8dayi-tlRqEdnG0hvp45n8gLUskWWcB9EXlUJObZGKDfGuxgMa3g_xeLA2vmFQ12MxPsyH4iCNZvsgmGxx7TuOHrnDh5EBVnM4_de63crEJON2sYI8Ozi-xp2OEmAr2seWKq4sxkFni6exLhqb-NE4m9HMKlng1EtQh2rLBFG1VYD3SYYpMLc5fxzqGvSxn3Fa-Xgg-IZPY3ubrcm52KYgmLUGmnYStfVqGSWSdhDXHlNgI5pdAA0FzpyBk3ZX-JsxhwcnneKrYBBweq06kRMGWgvdbdAQ-7wSeGqqj5VPwA"
+        },
+    },
+    "SYMMETRIC": {
+        raw: new Uint8Array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]),
+    }
+}
+
+let toWrapKeyData = {
+    "RSA": { //modulusLength: 1024, hash: "SHA-256"
+        spki: new Uint8Array([48, 129, 159, 48, 13, 6, 9, 42, 134, 72, 134, 247, 13, 1, 1, 1, 5, 0, 3, 129, 141, 0, 48, 129, 137, 2, 129, 129, 0, 205, 153, 248, 177, 17, 159, 141, 10, 44, 231, 172, 139, 253, 12, 181, 71, 211, 72, 249, 49, 204, 156, 92, 167, 159, 222, 32, 229, 28, 64, 235, 1, 171, 38, 30, 1, 37, 61, 241, 232, 143, 113, 208, 134, 233, 75, 122, 190, 119, 131, 145, 3, 164, 118, 190, 224, 204, 135, 199, 67, 21, 26, 253, 68, 49, 250, 93, 143, 160, 81, 39, 28, 245, 78, 73, 207, 117, 0, 216, 169, 149, 126, 192, 155, 157, 67, 239, 112, 9, 140, 87, 241, 13, 3, 191, 211, 23, 72, 175, 86, 59, 136, 22, 135, 114, 13, 60, 123, 16, 161, 205, 85, 58, 199, 29, 41, 107, 110, 222, 236, 165, 185, 156, 138, 251, 54, 221, 151, 2, 3, 1, 0, 1]),
+        pkcs8: new Uint8Array([48, 130, 2, 120, 2, 1, 0, 48, 13, 6, 9, 42, 134, 72, 134, 247, 13, 1, 1, 1, 5, 0, 4, 130, 2, 98, 48, 130, 2, 94, 2, 1, 0, 2, 129, 129, 0, 205, 153, 248, 177, 17, 159, 141, 10, 44, 231, 172, 139, 253, 12, 181, 71, 211, 72, 249, 49, 204, 156, 92, 167, 159, 222, 32, 229, 28, 64, 235, 1, 171, 38, 30, 1, 37, 61, 241, 232, 143, 113, 208, 134, 233, 75, 122, 190, 119, 131, 145, 3, 164, 118, 190, 224, 204, 135, 199, 67, 21, 26, 253, 68, 49, 250, 93, 143, 160, 81, 39, 28, 245, 78, 73, 207, 117, 0, 216, 169, 149, 126, 192, 155, 157, 67, 239, 112, 9, 140, 87, 241, 13, 3, 191, 211, 23, 72, 175, 86, 59, 136, 22, 135, 114, 13, 60, 123, 16, 161, 205, 85, 58, 199, 29, 41, 107, 110, 222, 236, 165, 185, 156, 138, 251, 54, 221, 151, 2, 3, 1, 0, 1, 2, 129, 128, 98, 162, 10, 252, 103, 71, 243, 145, 126, 25, 102, 93, 129, 248, 38, 191, 94, 77, 19, 191, 32, 57, 162, 249, 135, 104, 56, 191, 176, 222, 51, 223, 137, 11, 176, 57, 60, 116, 139, 40, 214, 39, 243, 177, 197, 25, 192, 184, 190, 253, 15, 4, 128, 81, 183, 32, 128, 254, 98, 73, 124, 70, 134, 88, 228, 85, 8, 229, 210, 6, 149, 141, 122, 147, 24, 166, 42, 57, 218, 125, 240, 230, 232, 249, 81, 145, 44, 6, 118, 237, 101, 205, 4, 181, 104, 85, 23, 96, 46, 169, 174, 213, 110, 34, 171, 89, 196, 20, 18, 1, 8, 241, 93, 32, 19, 144, 248, 183, 32, 96, 240, 101, 239, 247, 222, 249, 117, 1, 2, 65, 0, 244, 26, 192, 131, 146, 245, 205, 250, 134, 62, 229, 137, 14, 224, 194, 5, 127, 147, 154, 214, 93, 172, 226, 55, 98, 206, 25, 104, 223, 178, 48, 249, 83, 143, 5, 146, 16, 243, 180, 170, 119, 227, 17, 151, 48, 217, 88, 23, 30, 2, 73, 153, 181, 92, 163, 164, 241, 114, 66, 66, 152, 70, 42, 121, 2, 65, 0, 215, 158, 227, 12, 157, 88, 107, 153, 230, 66, 244, 207, 110, 18, 128, 60, 7, 140, 90, 136, 49, 11, 38, 144, 78, 64, 107, 167, 125, 41, 16, 167, 122, 152, 100, 129, 223, 206, 97, 170, 190, 1, 34, 79, 44, 221, 254, 204, 117, 122, 76, 249, 68, 169, 105, 152, 20, 161, 62, 40, 255, 101, 68, 143, 2, 65, 0, 169, 215, 127, 65, 76, 220, 104, 31, 186, 142, 66, 168, 213, 72, 62, 215, 18, 136, 2, 0, 203, 22, 194, 35, 37, 69, 31, 90, 223, 226, 28, 191, 45, 139, 98, 165, 217, 211, 167, 77, 192, 178, 166, 7, 155, 62, 110, 83, 79, 86, 234, 28, 223, 154, 128, 102, 0, 116, 174, 115, 165, 125, 148, 137, 2, 65, 0, 132, 212, 95, 192, 228, 169, 148, 215, 225, 46, 252, 75, 80, 222, 218, 218, 160, 55, 201, 137, 190, 212, 196, 179, 255, 80, 214, 64, 254, 236, 174, 82, 206, 70, 85, 28, 96, 248, 109, 216, 86, 102, 178, 113, 30, 13, 192, 42, 202, 112, 70, 61, 5, 28, 108, 109, 128, 191, 248, 96, 31, 61, 142, 103, 2, 65, 0, 205, 186, 73, 64, 8, 98, 158, 188, 82, 109, 82, 177, 5, 13, 132, 100, 97, 84, 15, 103, 183, 88, 37, 219, 0, 148, 88, 166, 79, 7, 85, 14, 64, 3, 157, 142, 132, 164, 226, 112, 236, 158, 218, 17, 7, 158, 184, 41, 20, 172, 194, 242, 44, 231, 78, 192, 134, 220, 83, 36, 191, 7, 35, 225]),
+        jwk: {
+            kty: "RSA",
+            n: "zZn4sRGfjQos56yL_Qy1R9NI-THMnFynn94g5RxA6wGrJh4BJT3x6I9x0IbpS3q-d4ORA6R2vuDMh8dDFRr9RDH6XY-gUScc9U5Jz3UA2KmVfsCbnUPvcAmMV_ENA7_TF0ivVjuIFodyDTx7EKHNVTrHHSlrbt7spbmcivs23Zc",
+            e: "AQAB",
+            d: "YqIK_GdH85F-GWZdgfgmv15NE78gOaL5h2g4v7DeM9-JC7A5PHSLKNYn87HFGcC4vv0PBIBRtyCA_mJJfEaGWORVCOXSBpWNepMYpio52n3w5uj5UZEsBnbtZc0EtWhVF2Auqa7VbiKrWcQUEgEI8V0gE5D4tyBg8GXv9975dQE",
+            p: "9BrAg5L1zfqGPuWJDuDCBX-TmtZdrOI3Ys4ZaN-yMPlTjwWSEPO0qnfjEZcw2VgXHgJJmbVco6TxckJCmEYqeQ",
+            q: "157jDJ1Ya5nmQvTPbhKAPAeMWogxCyaQTkBrp30pEKd6mGSB385hqr4BIk8s3f7MdXpM-USpaZgUoT4o_2VEjw",
+            dp: "qdd_QUzcaB-6jkKo1Ug-1xKIAgDLFsIjJUUfWt_iHL8ti2Kl2dOnTcCypgebPm5TT1bqHN-agGYAdK5zpX2UiQ",
+            dq: "hNRfwOSplNfhLvxLUN7a2qA3yYm-1MSz_1DWQP7srlLORlUcYPht2FZmsnEeDcAqynBGPQUcbG2Av_hgHz2OZw",
+            qi: "zbpJQAhinrxSbVKxBQ2EZGFUD2e3WCXbAJRYpk8HVQ5AA52OhKTicOye2hEHnrgpFKzC8iznTsCG3FMkvwcj4Q"
+        }
+    },
+    "EC": {
+        spki: new Uint8Array([48, 89, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97, 232]),
+        pkcs8: new Uint8Array([48, 129, 135, 2, 1, 0, 48, 19, 6, 7, 42, 134, 72, 206, 61, 2, 1, 6, 8, 42, 134, 72, 206, 61, 3, 1, 7, 4, 109, 48, 107, 2, 1, 1, 4, 32, 19, 211, 58, 45, 90, 191, 156, 249, 235, 178, 31, 248, 96, 212, 174, 254, 110, 86, 231, 119, 144, 244, 222, 233, 180, 8, 132, 235, 211, 53, 68, 234, 161, 68, 3, 66, 0, 4, 210, 16, 176, 166, 249, 217, 240, 18, 134, 128, 88, 180, 63, 164, 244, 113, 1, 133, 67, 187, 160, 12, 146, 80, 223, 146, 87, 194, 172, 174, 93, 209, 206, 3, 117, 82, 212, 129, 69, 12, 227, 155, 77, 16, 149, 112, 27, 23, 91, 250, 179, 75, 142, 108, 9, 158, 24, 241, 193, 152, 53, 131, 97, 232]),
+        jwk: {
+            kty: "EC",
+            crv: "P-256",
+            x: "0hCwpvnZ8BKGgFi0P6T0cQGFQ7ugDJJQ35JXwqyuXdE",
+            y: "zgN1UtSBRQzjm00QlXAbF1v6s0uObAmeGPHBmDWDYeg",
+            d: "E9M6LVq_nPnrsh_4YNSu_m5W53eQ9N7ptAiE69M1ROo"
+        }
+    },
+    "ED25519": {
+        spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 112, 3, 33, 0, 216, 225, 137, 99, 216, 9, 212, 135, 217, 84, 154, 204, 174, 198, 116, 46, 126, 235, 162, 77, 138, 13, 59, 20, 183, 227, 202, 234, 6, 137, 61, 204]),
+        pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 112, 4, 34, 4, 32, 243, 200, 244, 196, 141, 248, 120, 20, 110, 140, 211, 191, 109, 244, 229, 14, 56, 155, 167, 7, 78, 21, 194, 53, 45, 205, 93, 48, 141, 76, 168, 31]),
+        jwk: {
+            crv: "Ed25519",
+            d: "88j0xI34eBRujNO_bfTlDjibpwdOFcI1Lc1dMI1MqB8",
+            x: "2OGJY9gJ1IfZVJrMrsZ0Ln7rok2KDTsUt-PK6gaJPcw",
+            kty: "OKP"
+        }
+    },
+    "ED448": {
+        spki: new Uint8Array([48, 67, 48, 5, 6, 3, 43, 101, 113, 3, 58, 0, 171, 75, 184, 133, 253, 125, 44, 90, 242, 78, 131, 113, 12, 255, 160, 199, 74, 87, 226, 116, 128, 29, 178, 5, 123, 11, 220, 94, 160, 50, 182, 254, 107, 199, 139, 128, 69, 54, 90, 235, 38, 232, 110, 31, 20, 253, 52, 157, 7, 196, 132, 149, 245, 164, 106, 90, 128]),
+        pkcs8: new Uint8Array([48, 71, 2, 1, 0, 48, 5, 6, 3, 43, 101, 113, 4, 59, 4, 57, 14, 255, 3, 69, 140, 40, 224, 23, 156, 82, 29, 227, 18, 201, 105, 183, 131, 67, 72, 236, 171, 153, 26, 96, 227, 178, 233, 167, 158, 76, 217, 228, 128, 239, 41, 23, 18, 210, 200, 61, 4, 114, 114, 213, 201, 244, 40, 102, 79, 105, 109, 38, 112, 69, 143, 29, 46]),
+        jwk: {
+            crv: "Ed448",
+            d: "Dv8DRYwo4BecUh3jEslpt4NDSOyrmRpg47Lpp55M2eSA7ykXEtLIPQRyctXJ9ChmT2ltJnBFjx0u",
+            x: "q0u4hf19LFryToNxDP-gx0pX4nSAHbIFewvcXqAytv5rx4uARTZa6ybobh8U_TSdB8SElfWkalqA",
+            kty: "OKP"
+        }
+    },
+    "X25519": {
+        spki: new Uint8Array([48, 42, 48, 5, 6, 3, 43, 101, 110, 3, 33, 0, 28, 242, 177, 230, 2, 46, 197, 55, 55, 30, 215, 245, 62, 84, 250, 17, 84, 216, 62, 152, 235, 100, 234, 81, 250, 229, 179, 48, 124, 254, 151, 6]),
+        pkcs8: new Uint8Array([48, 46, 2, 1, 0, 48, 5, 6, 3, 43, 101, 110, 4, 34, 4, 32, 200, 131, 142, 118, 208, 87, 223, 183, 216, 201, 90, 105, 225, 56, 22, 10, 221, 99, 115, 253, 113, 164, 210, 118, 187, 86, 227, 168, 27, 100, 255, 97]),
+        jwk: {
+            crv: "X25519",
+            d: "yIOOdtBX37fYyVpp4TgWCt1jc_1xpNJ2u1bjqBtk_2E",
+            x: "HPKx5gIuxTc3Htf1PlT6EVTYPpjrZOpR-uWzMHz-lwY",
+            kty: "OKP"
+        }
+    },
+    "X448": {
+        spki: new Uint8Array([48, 66, 48, 5, 6, 3, 43, 101, 111, 3, 57, 0, 182, 4, 161, 209, 165, 205, 29, 148, 38, 213, 97, 239, 99, 10, 158, 177, 108, 190, 105, 213, 185, 202, 97, 94, 220, 83, 99, 62, 251, 82, 234, 49, 230, 230, 160, 161, 219, 172, 198, 231, 108, 188, 230, 72, 45, 126, 75, 163, 213, 93, 158, 128, 39, 101, 206, 111]),
+        pkcs8: new Uint8Array([48, 70, 2, 1, 0, 48, 5, 6, 3, 43, 101, 111, 4, 58, 4, 56, 88, 199, 210, 154, 62, 181, 25, 178, 157, 0, 207, 177, 145, 187, 100, 252, 109, 138, 66, 216, 241, 113, 118, 39, 43, 137, 242, 39, 45, 24, 25, 41, 92, 101, 37, 192, 130, 150, 113, 176, 82, 239, 7, 39, 83, 15, 24, 142, 49, 208, 204, 83, 191, 38, 146, 158]),
+        jwk: {
+            crv: "X448",
+            d: "WMfSmj61GbKdAM-xkbtk_G2KQtjxcXYnK4nyJy0YGSlcZSXAgpZxsFLvBydTDxiOMdDMU78mkp4",
+            x: "tgSh0aXNHZQm1WHvYwqesWy-adW5ymFe3FNjPvtS6jHm5qCh26zG52y85kgtfkuj1V2egCdlzm8",
+            kty: "OKP"
+        }
+    },
+    "SYMMETRIC": {
+        raw: new Uint8Array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16]),
+    }
+}
+
+function toWrapKeyDataFromAlg(alg) {
+    switch(alg) {
+    case "RSASSA-PKCS1-v1_5":
+    case "RSA-PSS":
+    case "RSA-OAEP":
+        return toWrapKeyData["RSA"];
+    case "ECDSA":
+    case "ECDH":
+        return toWrapKeyData["EC"];
+    case "Ed25519":
+        return toWrapKeyData["ED25519"];
+    case "X25519":
+        return toWrapKeyData["X25519"];
+    case "Ed448":
+        return toWrapKeyData["ED448"];
+    case "X448":
+        return toWrapKeyData["X448"];
+    default:
+        return undefined;
+    }
+}
diff --git a/test/fixtures/wpt/versions.json b/test/fixtures/wpt/versions.json
index b2aea8fdbff061..0011df74768f28 100644
--- a/test/fixtures/wpt/versions.json
+++ b/test/fixtures/wpt/versions.json
@@ -84,7 +84,7 @@
     "path": "wasm/webapi"
   },
   "WebCryptoAPI": {
-    "commit": "5e042cbc4ecab7b2279a5fd411c6daa24ca886c6",
+    "commit": "edd42c005cf8192fbae41ec061c14342e7bcac15",
     "path": "WebCryptoAPI"
   },
   "webidl/ecmascript-binding/es-exceptions": {
@@ -95,4 +95,4 @@
     "commit": "e97fac4791931fb7455ba3fad759d362c7108b09",
     "path": "webmessaging/broadcastchannel"
   }
-}
\ No newline at end of file
+}
diff --git a/test/parallel/parallel.status b/test/parallel/parallel.status
index a7860449225092..8caeb3ccef959f 100644
--- a/test/parallel/parallel.status
+++ b/test/parallel/parallel.status
@@ -28,6 +28,9 @@ test-runner-watch-mode-complex: PASS, FLAKY
 # https://github.com/nodejs/node/issues/54808
 test-async-context-frame: PASS, FLAKY
 
+# https://github.com/nodejs/node/issues/56751
+test-without-async-context-frame: PASS, FLAKY
+
 # Windows on x86
 [$system==win32 && $arch==ia32]
 test-worker-nearheaplimit-deadlock: PASS, FLAKY
diff --git a/test/parallel/test-buffer-tostring-rangeerror.js b/test/parallel/test-buffer-tostring-rangeerror.js
index 0ebea759b5c42b..59b5b8315024e6 100644
--- a/test/parallel/test-buffer-tostring-rangeerror.js
+++ b/test/parallel/test-buffer-tostring-rangeerror.js
@@ -18,8 +18,22 @@ const message = {
   code: 'ERR_STRING_TOO_LONG',
   name: 'Error',
 };
-assert.throws(() => Buffer(len).toString('utf8'), message);
-assert.throws(() => SlowBuffer(len).toString('utf8'), message);
-assert.throws(() => Buffer.alloc(len).toString('utf8'), message);
-assert.throws(() => Buffer.allocUnsafe(len).toString('utf8'), message);
-assert.throws(() => Buffer.allocUnsafeSlow(len).toString('utf8'), message);
+
+function test(getBuffer) {
+  let buf;
+  try {
+    buf = getBuffer();
+  } catch (e) {
+    // If the buffer allocation fails, we skip the test.
+    if (e.code === 'ERR_MEMORY_ALLOCATION_FAILED' || /Array buffer allocation failed/.test(e.message)) {
+      return;
+    }
+  }
+  assert.throws(() => { buf.toString('utf8'); }, message);
+}
+
+test(() => Buffer(len));
+test(() => SlowBuffer(len));
+test(() => Buffer.alloc(len));
+test(() => Buffer.allocUnsafe(len));
+test(() => Buffer.allocUnsafeSlow(len));
diff --git a/test/parallel/test-child-process-kill.js b/test/parallel/test-child-process-kill.js
index 26bdc029c04fe8..4d045feba434fb 100644
--- a/test/parallel/test-child-process-kill.js
+++ b/test/parallel/test-child-process-kill.js
@@ -58,3 +58,23 @@ if (common.isWindows) {
   });
   process.kill('SIGHUP');
 }
+
+// Test that the process is not killed when sending a 0 signal.
+// This is a no-op signal that is used to check if the process is alive.
+const code = `const interval = setInterval(() => {}, 1000);
+process.stdin.on('data', () => { clearInterval(interval); });
+process.stdout.write('x');`;
+
+const checkProcess = spawn(process.execPath, ['-e', code]);
+
+checkProcess.on('exit', (code, signal) => {
+  assert.strictEqual(code, 0);
+  assert.strictEqual(signal, null);
+});
+
+checkProcess.stdout.on('data', common.mustCall((chunk) => {
+  assert.strictEqual(chunk.toString(), 'x');
+  checkProcess.kill(0);
+  checkProcess.stdin.write('x');
+  checkProcess.stdin.end();
+}));
diff --git a/test/parallel/test-file-write-stream4.js b/test/parallel/test-file-write-stream4.js
index 6b3862fa714d7c..e741cdd79036b4 100644
--- a/test/parallel/test-file-write-stream4.js
+++ b/test/parallel/test-file-write-stream4.js
@@ -1,3 +1,4 @@
+// Flags: --expose-gc
 'use strict';
 
 // Test that 'close' emits once and not twice when `emitClose: true` is set.
@@ -17,4 +18,8 @@ const fileWriteStream = fs.createWriteStream(filepath, {
 });
 
 fileReadStream.pipe(fileWriteStream);
-fileWriteStream.on('close', common.mustCall());
+fileWriteStream.on('close', common.mustCall(() => {
+  // TODO(lpinca): Remove the forced GC when
+  // https://github.com/nodejs/node/issues/54918 is fixed.
+  globalThis.gc({ type: 'major' });
+}));
diff --git a/test/parallel/test-http-server-method.query.js b/test/parallel/test-http-server-method.query.js
new file mode 100644
index 00000000000000..8159fb75050263
--- /dev/null
+++ b/test/parallel/test-http-server-method.query.js
@@ -0,0 +1,27 @@
+'use strict';
+
+const common = require('../common');
+const { strictEqual } = require('assert');
+const { createServer, request } = require('http');
+
+const server = createServer(common.mustCall((req, res) => {
+  strictEqual(req.method, 'QUERY');
+  res.end('OK');
+}));
+
+server.listen(0, common.mustCall(() => {
+  const req = request({ port: server.address().port, method: 'QUERY' }, common.mustCall((res) => {
+    strictEqual(res.statusCode, 200);
+
+    let buffer = '';
+    res.setEncoding('utf-8');
+
+    res.on('data', (c) => buffer += c);
+    res.on('end', common.mustCall(() => {
+      strictEqual(buffer, 'OK');
+      server.close();
+    }));
+  }));
+
+  req.end();
+}));
diff --git a/test/parallel/test-http2-client-rststream-before-connect.js b/test/parallel/test-http2-client-rststream-before-connect.js
index bc0cb5ff619dc0..788253d29ae22f 100644
--- a/test/parallel/test-http2-client-rststream-before-connect.js
+++ b/test/parallel/test-http2-client-rststream-before-connect.js
@@ -5,16 +5,23 @@ if (!common.hasCrypto)
   common.skip('missing crypto');
 const assert = require('assert');
 const h2 = require('http2');
+let client;
 
 const server = h2.createServer();
 server.on('stream', (stream) => {
-  stream.on('close', common.mustCall());
-  stream.respond();
-  stream.end('ok');
+  stream.on('close', common.mustCall(() => {
+    client.close();
+    server.close();
+  }));
+  stream.on('error', common.expectsError({
+    code: 'ERR_HTTP2_STREAM_ERROR',
+    name: 'Error',
+    message: 'Stream closed with error code NGHTTP2_PROTOCOL_ERROR'
+  }));
 });
 
 server.listen(0, common.mustCall(() => {
-  const client = h2.connect(`http://localhost:${server.address().port}`);
+  client = h2.connect(`http://localhost:${server.address().port}`);
   const req = client.request();
   const closeCode = 1;
 
@@ -52,8 +59,6 @@ server.listen(0, common.mustCall(() => {
   req.on('close', common.mustCall(() => {
     assert.strictEqual(req.destroyed, true);
     assert.strictEqual(req.rstCode, closeCode);
-    server.close();
-    client.close();
   }));
 
   req.on('error', common.expectsError({
diff --git a/test/parallel/test-http2-compat-write-head-after-close.js b/test/parallel/test-http2-compat-write-head-after-close.js
new file mode 100644
index 00000000000000..b082400537b7e2
--- /dev/null
+++ b/test/parallel/test-http2-compat-write-head-after-close.js
@@ -0,0 +1,23 @@
+'use strict';
+
+const common = require('../common');
+if (!common.hasCrypto) { common.skip('missing crypto'); }
+const h2 = require('http2');
+
+const server = h2.createServer((req, res) => {
+  const stream = req.stream;
+  stream.close();
+  res.writeHead(200, { 'content-type': 'text/plain' });
+});
+
+server.listen(0, common.mustCall(() => {
+  const port = server.address().port;
+  const client = h2.connect(`http://localhost:${port}`);
+  const req = client.request({ ':path': '/' });
+  req.on('response', common.mustNotCall('head after close should not be sent'));
+  req.on('end', common.mustCall(() => {
+    client.close();
+    server.close();
+  }));
+  req.end();
+}));
diff --git a/test/parallel/test-http2-options-max-headers-block-length.js b/test/parallel/test-http2-options-max-headers-block-length.js
index 15b142ac89b811..0a2645c4ccc921 100644
--- a/test/parallel/test-http2-options-max-headers-block-length.js
+++ b/test/parallel/test-http2-options-max-headers-block-length.js
@@ -22,6 +22,8 @@ server.listen(0, common.mustCall(() => {
   const client = h2.connect(`http://localhost:${server.address().port}`,
                             options);
 
+  client.on('error', () => {});
+
   const req = client.request();
   req.on('response', common.mustNotCall());
 
diff --git a/test/parallel/test-http2-session-graceful-close.js b/test/parallel/test-http2-session-graceful-close.js
new file mode 100644
index 00000000000000..174eb037dce5b4
--- /dev/null
+++ b/test/parallel/test-http2-session-graceful-close.js
@@ -0,0 +1,48 @@
+'use strict';
+
+const common = require('../common');
+if (!common.hasCrypto)
+  common.skip('missing crypto');
+const assert = require('assert');
+const h2 = require('http2');
+
+const server = h2.createServer();
+let session;
+
+server.on('session', common.mustCall(function(s) {
+  session = s;
+  session.on('close', common.mustCall(function() {
+    server.close();
+  }));
+}));
+
+server.listen(0, common.mustCall(function() {
+  const port = server.address().port;
+
+  const url = `http://localhost:${port}`;
+  const client = h2.connect(url, common.mustCall(function() {
+    const headers = {
+      ':path': '/',
+      ':method': 'GET',
+      ':scheme': 'http',
+      ':authority': `localhost:${port}`
+    };
+    const request = client.request(headers);
+    request.on('response', common.mustCall(function(headers) {
+      assert.strictEqual(headers[':status'], 200);
+    }, 1));
+    request.on('end', common.mustCall(function() {
+      client.close();
+    }));
+    request.end();
+    request.resume();
+  }));
+  client.on('goaway', common.mustCallAtLeast(1));
+}));
+
+server.once('request', common.mustCall(function(request, response) {
+  response.on('finish', common.mustCall(function() {
+    session.close();
+  }));
+  response.end();
+}));
diff --git a/test/parallel/test-internal-util-getCIDR.js b/test/parallel/test-internal-util-getCIDR.js
new file mode 100644
index 00000000000000..f7dd3c1194b6a2
--- /dev/null
+++ b/test/parallel/test-internal-util-getCIDR.js
@@ -0,0 +1,23 @@
+// Flags: --expose-internals
+'use strict';
+require('../common');
+
+// These are tests that verify that the subnetmask is used
+// to create the correct CIDR address.
+// Tests that it returns null if the subnetmask is not in the correct format.
+// (ref: https://www.rfc-editor.org/rfc/rfc1878)
+
+const assert = require('node:assert');
+const { getCIDR } = require('internal/util');
+
+assert.strictEqual(getCIDR('127.0.0.1', '255.0.0.0', 'IPv4'), '127.0.0.1/8');
+assert.strictEqual(getCIDR('127.0.0.1', '255.255.0.0', 'IPv4'), '127.0.0.1/16');
+
+// 242 = 11110010(2)
+assert.strictEqual(getCIDR('127.0.0.1', '242.0.0.0', 'IPv4'), null);
+
+assert.strictEqual(getCIDR('::1', 'ffff:ffff:ffff:ffff::', 'IPv6'), '::1/64');
+assert.strictEqual(getCIDR('::1', 'ffff:ffff:ffff:ffff:ffff:ffff:ffff:ffff', 'IPv6'), '::1/128');
+
+// ff00:ffff = 11111111 00000000 : 11111111 11111111(2)
+assert.strictEqual(getCIDR('::1', 'ffff:ff00:ffff::', 'IPv6'), null);
diff --git a/test/parallel/test-module-loading-error.js b/test/parallel/test-module-loading-error.js
index 8cbe8b2c675663..6aa91f3675e8ed 100644
--- a/test/parallel/test-module-loading-error.js
+++ b/test/parallel/test-module-loading-error.js
@@ -28,7 +28,7 @@ const errorMessagesByPlatform = {
   win32: ['is not a valid Win32 application'],
   linux: ['file too short', 'Exec format error'],
   sunos: ['unknown file type', 'not an ELF file'],
-  darwin: ['file too short', 'not a mach-o file'],
+  darwin: ['file too short', 'not a mach-o file', 'not valid mach-o file'],
   aix: ['Cannot load module',
         'Cannot run a file that does not have a valid format.',
         'Exec format error'],
diff --git a/test/parallel/test-runner-output.mjs b/test/parallel/test-runner-output.mjs
index 9fc6a6bfb44e9a..49627d3ba6d347 100644
--- a/test/parallel/test-runner-output.mjs
+++ b/test/parallel/test-runner-output.mjs
@@ -32,7 +32,7 @@ function replaceJunitDuration(str) {
   return str
     .replaceAll(/time="[0-9.]+"/g, 'time="*"')
     .replaceAll(/duration_ms [0-9.]+/g, 'duration_ms *')
-    .replaceAll(hostname(), 'HOSTNAME')
+    .replaceAll(`hostname="${hostname()}"`, 'hostname="HOSTNAME"')
     .replace(stackTraceBasePath, '$3');
 }
 
diff --git a/test/parallel/test-webcrypto-derivebits-cfrg.js b/test/parallel/test-webcrypto-derivebits-cfrg.js
index 3e28774d05a9ce..f5c602b6deb630 100644
--- a/test/parallel/test-webcrypto-derivebits-cfrg.js
+++ b/test/parallel/test-webcrypto-derivebits-cfrg.js
@@ -140,9 +140,11 @@ async function prepareKeys() {
           public: publicKey
         }, privateKey, 8 * size - 11);
 
-        assert.strictEqual(
-          Buffer.from(bits).toString('hex'),
-          result.slice(0, -2));
+        const expected = Buffer.from(result.slice(0, -2), 'hex');
+        expected[size - 2] = expected[size - 2] & 0b11111000;
+        assert.deepStrictEqual(
+          Buffer.from(bits),
+          expected);
       }
     }));
 
diff --git a/test/parallel/test-webcrypto-derivebits-ecdh.js b/test/parallel/test-webcrypto-derivebits-ecdh.js
index 4dba34d84a7907..6c99946752b4b6 100644
--- a/test/parallel/test-webcrypto-derivebits-ecdh.js
+++ b/test/parallel/test-webcrypto-derivebits-ecdh.js
@@ -161,9 +161,11 @@ async function prepareKeys() {
           public: publicKey
         }, privateKey, 8 * size - 11);
 
-        assert.strictEqual(
-          Buffer.from(bits).toString('hex'),
-          result.slice(0, -2));
+        const expected = Buffer.from(result.slice(0, -2), 'hex');
+        expected[size - 2] = expected[size - 2] & 0b11111000;
+        assert.deepStrictEqual(
+          Buffer.from(bits),
+          expected);
       }
     }));
 
diff --git a/test/parallel/test-webcrypto-derivebits-hkdf.js b/test/parallel/test-webcrypto-derivebits-hkdf.js
index bef6abdc19d0d6..0e3b2992e1c137 100644
--- a/test/parallel/test-webcrypto-derivebits-hkdf.js
+++ b/test/parallel/test-webcrypto-derivebits-hkdf.js
@@ -261,11 +261,6 @@ async function testDeriveBitsBadLengths(
       subtle.deriveBits(algorithm, baseKeys[size], undefined), {
         name: 'OperationError',
       }),
-    assert.rejects(
-      subtle.deriveBits(algorithm, baseKeys[size], 0), {
-        message: /length cannot be zero/,
-        name: 'OperationError',
-      }),
     assert.rejects(
       subtle.deriveBits(algorithm, baseKeys[size], null), {
         message: 'length cannot be null',
@@ -562,3 +557,18 @@ async function testWrongKeyType(
   await Promise.all(variations);
 
 })().then(common.mustCall());
+
+// https://github.com/w3c/webcrypto/pull/380
+{
+  crypto.subtle.importKey('raw', new Uint8Array(0), 'HKDF', false, ['deriveBits']).then((key) => {
+    return crypto.subtle.deriveBits({
+      name: 'HKDF',
+      hash: { name: 'SHA-256' },
+      info: new Uint8Array(0),
+      salt: new Uint8Array(0),
+    }, key, 0);
+  }).then((bits) => {
+    assert.deepStrictEqual(bits, new ArrayBuffer(0));
+  })
+  .then(common.mustCall());
+}
diff --git a/test/parallel/test-webcrypto-export-import-cfrg.js b/test/parallel/test-webcrypto-export-import-cfrg.js
index 84d969eb1c5850..ea1dd8176ae413 100644
--- a/test/parallel/test-webcrypto-export-import-cfrg.js
+++ b/test/parallel/test-webcrypto-export-import-cfrg.js
@@ -12,7 +12,7 @@ const { subtle } = globalThis.crypto;
 
 const keyData = {
   'Ed25519': {
-    jwsAlg: 'EdDSA',
+    jwsAlg: 'Ed25519',
     spki: Buffer.from(
       '302a300506032b6570032100a054b618c12b26c8d43595a5c38dd2b0140b944a' +
       '151f75003278c2b6c58ec08f', 'hex'),
@@ -27,7 +27,7 @@ const keyData = {
     }
   },
   'Ed448': {
-    jwsAlg: 'EdDSA',
+    jwsAlg: 'Ed448',
     spki: Buffer.from(
       '3043300506032b6571033a0008cc38160c85bca5656ac4924af7ea97a9161b20' +
       '2528273dcb84afd2eeb99ac912a401b34ef15ef4d9486406a6eecc31e5909219' +
@@ -183,10 +183,7 @@ async function testImportJwk({ name, publicUsages, privateUsages }, extractable)
 
   const jwk = keyData[name].jwk;
 
-  const [
-    publicKey,
-    privateKey,
-  ] = await Promise.all([
+  const tests = [
     subtle.importKey(
       'jwk',
       {
@@ -221,7 +218,37 @@ async function testImportJwk({ name, publicUsages, privateUsages }, extractable)
       { name },
       extractable,
       privateUsages),
-  ]);
+  ];
+
+  // Test the deprecated "alg" value
+  if (keyData[name].jwsAlg?.startsWith('Ed')) {
+    tests.push(
+      subtle.importKey(
+        'jwk',
+        {
+          alg: 'EdDSA',
+          kty: jwk.kty,
+          crv: jwk.crv,
+          x: jwk.x,
+        },
+        { name },
+        extractable, publicUsages),
+      subtle.importKey(
+        'jwk',
+        {
+          ...jwk,
+          alg: 'EdDSA',
+        },
+        { name },
+        extractable,
+        privateUsages),
+    );
+  }
+
+  const [
+    publicKey,
+    privateKey,
+  ] = await Promise.all(tests);
 
   assert.strictEqual(publicKey.type, 'public');
   assert.strictEqual(privateKey.type, 'private');
@@ -259,8 +286,13 @@ async function testImportJwk({ name, publicUsages, privateUsages }, extractable)
     assert.strictEqual(pvtJwk.crv, jwk.crv);
     assert.strictEqual(pvtJwk.d, jwk.d);
 
-    assert.strictEqual(pubJwk.alg, undefined);
-    assert.strictEqual(pvtJwk.alg, undefined);
+    if (jwk.crv.startsWith('Ed')) {
+      assert.strictEqual(pubJwk.alg, jwk.crv);
+      assert.strictEqual(pvtJwk.alg, jwk.crv);
+    } else {
+      assert.strictEqual(pubJwk.alg, undefined);
+      assert.strictEqual(pvtJwk.alg, undefined);
+    }
   } else {
     await assert.rejects(
       subtle.exportKey('jwk', publicKey), {
@@ -284,22 +316,24 @@ async function testImportJwk({ name, publicUsages, privateUsages }, extractable)
       { message: 'Invalid JWK "use" Parameter' });
   }
 
-  // The JWK alg member is ignored
-  // https://github.com/WICG/webcrypto-secure-curves/pull/24
   if (name.startsWith('Ed')) {
-    await subtle.importKey(
-      'jwk',
-      { kty: jwk.kty, x: jwk.x, crv: jwk.crv, alg: 'foo' },
-      { name },
-      extractable,
-      publicUsages);
+    await assert.rejects(
+      subtle.importKey(
+        'jwk',
+        { kty: jwk.kty, x: jwk.x, crv: jwk.crv, alg: 'foo' },
+        { name },
+        extractable,
+        publicUsages),
+      { message: 'JWK "alg" does not match the requested algorithm' });
 
-    await subtle.importKey(
-      'jwk',
-      { ...jwk, alg: 'foo' },
-      { name },
-      extractable,
-      privateUsages);
+    await assert.rejects(
+      subtle.importKey(
+        'jwk',
+        { ...jwk, alg: 'foo' },
+        { name },
+        extractable,
+        privateUsages),
+      { message: 'JWK "alg" does not match the requested algorithm' });
   }
 
   for (const crv of [undefined, name === 'Ed25519' ? 'Ed448' : 'Ed25519']) {
diff --git a/test/parallel/test-worker-stdio-flush-inflight.js b/test/parallel/test-worker-stdio-flush-inflight.js
new file mode 100644
index 00000000000000..34b81152811e7b
--- /dev/null
+++ b/test/parallel/test-worker-stdio-flush-inflight.js
@@ -0,0 +1,24 @@
+'use strict';
+const common = require('../common');
+const assert = require('assert');
+const { Worker, isMainThread } = require('worker_threads');
+
+if (isMainThread) {
+  const w = new Worker(__filename, { stdout: true });
+  const expected = 'hello world';
+
+  let data = '';
+  w.stdout.setEncoding('utf8');
+  w.stdout.on('data', (chunk) => {
+    data += chunk;
+  });
+
+  w.on('exit', common.mustCall(() => {
+    assert.strictEqual(data, expected);
+  }));
+} else {
+  process.stdout.write('hello');
+  process.stdout.write(' ');
+  process.stdout.write('world');
+  process.exit(0);
+}
diff --git a/test/parallel/test-worker-stdio-flush.js b/test/parallel/test-worker-stdio-flush.js
new file mode 100644
index 00000000000000..e52e721fc69483
--- /dev/null
+++ b/test/parallel/test-worker-stdio-flush.js
@@ -0,0 +1,25 @@
+'use strict';
+const common = require('../common');
+const assert = require('assert');
+const { Worker, isMainThread } = require('worker_threads');
+
+if (isMainThread) {
+  const w = new Worker(__filename, { stdout: true });
+  const expected = 'hello world';
+
+  let data = '';
+  w.stdout.setEncoding('utf8');
+  w.stdout.on('data', (chunk) => {
+    data += chunk;
+  });
+
+  w.on('exit', common.mustCall(() => {
+    assert.strictEqual(data, expected);
+  }));
+} else {
+  process.on('exit', () => {
+    process.stdout.write(' ');
+    process.stdout.write('world');
+  });
+  process.stdout.write('hello');
+}
diff --git a/test/pummel/test-blob-slice-with-large-size.js b/test/pummel/test-blob-slice-with-large-size.js
index 639c2217db79e3..af38d6d5760740 100644
--- a/test/pummel/test-blob-slice-with-large-size.js
+++ b/test/pummel/test-blob-slice-with-large-size.js
@@ -1,4 +1,7 @@
 'use strict';
+
+// This tests that Blob.prototype.slice() works correctly when the size of the
+// Blob is outside the range of 32-bit signed integers.
 const common = require('../common');
 
 // Buffer with size > INT32_MAX
@@ -14,8 +17,11 @@ try {
   const slicedBlob = blob.slice(size - 1, size);
   assert.strictEqual(slicedBlob.size, 1);
 } catch (e) {
-  if (e.code !== 'ERR_MEMORY_ALLOCATION_FAILED') {
-    throw e;
+  if (e.code === 'ERR_MEMORY_ALLOCATION_FAILED') {
+    common.skip('insufficient space for Buffer.allocUnsafe');
+  }
+  if (/Array buffer allocation failed/.test(e.message)) {
+    common.skip('insufficient space for Blob.prototype.slice()');
   }
-  common.skip('insufficient space for Buffer.allocUnsafe');
+  throw e;
 }
diff --git a/test/pummel/test-webcrypto-derivebits-pbkdf2.js b/test/pummel/test-webcrypto-derivebits-pbkdf2.js
index 382dadf1b35e45..242bb080d82368 100644
--- a/test/pummel/test-webcrypto-derivebits-pbkdf2.js
+++ b/test/pummel/test-webcrypto-derivebits-pbkdf2.js
@@ -449,11 +449,6 @@ async function testDeriveBitsBadLengths(
       subtle.deriveBits(algorithm, baseKeys[size], undefined), {
         name: 'OperationError',
       }),
-    assert.rejects(
-      subtle.deriveBits(algorithm, baseKeys[size], 0), {
-        message: /length cannot be zero/,
-        name: 'OperationError',
-      }),
     assert.rejects(
       subtle.deriveBits(algorithm, baseKeys[size], null), {
         message: 'length cannot be null',
@@ -693,3 +688,19 @@ async function testWrongKeyType(
 
   await Promise.all(variations);
 })().then(common.mustCall());
+
+
+// https://github.com/w3c/webcrypto/pull/380
+{
+  crypto.subtle.importKey('raw', new Uint8Array(0), 'PBKDF2', false, ['deriveBits']).then((key) => {
+    return crypto.subtle.deriveBits({
+      name: 'PBKDF2',
+      hash: { name: 'SHA-256' },
+      iterations: 10,
+      salt: new Uint8Array(0),
+    }, key, 0);
+  }).then((bits) => {
+    assert.deepStrictEqual(bits, new ArrayBuffer(0));
+  })
+  .then(common.mustCall());
+}
diff --git a/test/sequential/sequential.status b/test/sequential/sequential.status
index a3199b385dd99d..6893abfe6e1c03 100644
--- a/test/sequential/sequential.status
+++ b/test/sequential/sequential.status
@@ -52,3 +52,15 @@ test-tls-securepair-client: PASS, FLAKY
 [$arch==arm]
 # https://github.com/nodejs/node/issues/49933
 test-watch-mode-inspect: SKIP
+
+[$system==linux && $arch==ppc64]
+# https://github.com/nodejs/node/pull/58588
+test-single-executable-application: SKIP
+test-single-executable-application-assets: SKIP
+test-single-executable-application-assets-raw: SKIP
+test-single-executable-application-disable-experimental-sea-warning: SKIP
+test-single-executable-application-empty: SKIP
+test-single-executable-application-snapshot: SKIP
+test-single-executable-application-snapshot-and-code-cache: SKIP
+test-single-executable-application-snapshot-worker: SKIP
+test-single-executable-application-use-code-cache: SKIP
diff --git a/test/wpt/README.md b/test/wpt/README.md
index a378e3244a1f0b..3ca3c674b31105 100644
--- a/test/wpt/README.md
+++ b/test/wpt/README.md
@@ -27,7 +27,7 @@ it's not yet clear how compliant the implementation is,
 the requirements and expected failures can be figured out in a later step
 when the tests are run for the first time.
 
-See [Format of a status JSON file](#status-format) for details.
+See [Format of a status file](#status-format) for details.
 
 ### 2. Pull the WPT files
 
@@ -98,7 +98,7 @@ add this to `test/wpt/status/url.json`:
   }
 ```
 
-See [Format of a status JSON file](#status-format) for details.
+See [Format of a status file](#status-format) for details.
 
 ### 5. Commit the changes and submit a Pull Request
 
@@ -147,7 +147,7 @@ expected failures.
 
 <a id="status-format"></a>
 
-## Format of a status JSON file
+## Format of a status file
 
 ```json
 {
@@ -177,6 +177,10 @@ A test may have to be skipped because it depends on another irrelevant
 Web API, or certain harness has not been ported in our test runner yet.
 In that case it needs to be marked with `skip` instead of `fail`.
 
+The status file may optionally also be a CJS module that exports the object.
+This allows for more complex logic to be used to determine the expected status
+of a test.
+
 [Web Platform Tests]: https://github.com/web-platform-tests/wpt
 [`test/fixtures/wpt/README.md`]: ../fixtures/wpt/README.md
 [git node wpt]: https://github.com/nodejs/node-core-utils/blob/HEAD/docs/git-node.md#git-node-wpt
diff --git a/test/wpt/status/WebCryptoAPI.cjs b/test/wpt/status/WebCryptoAPI.cjs
new file mode 100644
index 00000000000000..96fd989ba41539
--- /dev/null
+++ b/test/wpt/status/WebCryptoAPI.cjs
@@ -0,0 +1,45 @@
+'use strict';
+
+const os = require('node:os');
+
+const s390x = os.arch() === 's390x';
+
+module.exports = {
+  'algorithm-discards-context.https.window.js': {
+    'skip': 'Not relevant in Node.js context',
+  },
+  'historical.any.js': {
+    'skip': 'Not relevant in Node.js context',
+  },
+  'getRandomValues.any.js': {
+    'fail': {
+      'note': 'Node.js does not support Float16Array',
+      'expected': [
+        'Float16 arrays',
+      ],
+    },
+  },
+  'sign_verify/eddsa_small_order_points.https.any.js': {
+    'fail': {
+      'note': 'see https://github.com/nodejs/node/issues/54572',
+      'expected': [
+        'Ed25519 Verification checks with small-order key of order - Test 1',
+        'Ed25519 Verification checks with small-order key of order - Test 2',
+        'Ed25519 Verification checks with small-order key of order - Test 12',
+        'Ed25519 Verification checks with small-order key of order - Test 13',
+        ...(s390x ? [] : [
+          'Ed25519 Verification checks with small-order key of order - Test 0',
+          'Ed25519 Verification checks with small-order key of order - Test 11',
+        ]),
+      ],
+    },
+  },
+  'idlharness.https.any.js': {
+    'fail': {
+      'note': 'WPT not updated for https://github.com/w3c/webcrypto/pull/345 yet',
+      'expected': [
+        'SubtleCrypto interface: operation deriveBits(AlgorithmIdentifier, CryptoKey, unsigned long)'
+      ]
+    }
+  }
+};
diff --git a/test/wpt/status/WebCryptoAPI.json b/test/wpt/status/WebCryptoAPI.json
deleted file mode 100644
index 69f86168f5a9df..00000000000000
--- a/test/wpt/status/WebCryptoAPI.json
+++ /dev/null
@@ -1,16 +0,0 @@
-{
-  "algorithm-discards-context.https.window.js": {
-    "skip": "Not relevant in Node.js context"
-  },
-  "historical.any.js": {
-    "skip": "Not relevant in Node.js context"
-  },
-  "idlharness.https.any.js": {
-    "fail": {
-      "note": "WPT not updated for https://github.com/w3c/webcrypto/pull/345 yet",
-      "expected": [
-        "SubtleCrypto interface: operation deriveBits(AlgorithmIdentifier, CryptoKey, unsigned long)"
-      ]
-    }
-  }
-}
diff --git a/tools/actions/create-release-proposal.sh b/tools/actions/create-release-proposal.sh
index cc5e8b5b7834fd..a17752ece95958 100755
--- a/tools/actions/create-release-proposal.sh
+++ b/tools/actions/create-release-proposal.sh
@@ -31,8 +31,9 @@ HEAD_SHA="$(git rev-parse HEAD^)"
 
 TITLE="$(git log -1 --format=%s)"
 
-# Use a temporary file for the PR body
-TEMP_BODY="$(awk "/## ${RELEASE_DATE}/,/^<a id=/{ if (!/^<a id=/) print }" "doc/changelogs/CHANGELOG_V${RELEASE_LINE}.md")"
+TEMP_BODY="$(awk -v MAX_BODY_LENGTH="65536" \
+    "/^## ${RELEASE_DATE}/,/^<a id=/{ if (/^<a id=/) {exit;} if ((output_length += length(\$0)) > MAX_BODY_LENGTH) {exit 1;} print }" \
+    "doc/changelogs/CHANGELOG_V${RELEASE_LINE}.md" || echo "…")"
 
 # Create the proposal branch
 gh api \
diff --git a/tools/certdata.txt b/tools/certdata.txt
index e0f60abcd6cf62..1ed5a248b57a62 100644
--- a/tools/certdata.txt
+++ b/tools/certdata.txt
@@ -2347,174 +2347,6 @@ CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
 CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
 
-#
-# Certificate "SwissSign Silver CA - G2"
-#
-# Issuer: CN=SwissSign Silver CA - G2,O=SwissSign AG,C=CH
-# Serial Number:4f:1b:d4:2f:54:bb:2f:4b
-# Subject: CN=SwissSign Silver CA - G2,O=SwissSign AG,C=CH
-# Not Valid Before: Wed Oct 25 08:32:46 2006
-# Not Valid After : Sat Oct 25 08:32:46 2036
-# Fingerprint (SHA-256): BE:6C:4D:A2:BB:B9:BA:59:B6:F3:93:97:68:37:42:46:C3:C0:05:99:3F:A9:8F:02:0D:1D:ED:BE:D4:8A:81:D5
-# Fingerprint (SHA1): 9B:AA:E5:9F:56:EE:21:CB:43:5A:BE:25:93:DF:A7:F0:40:D1:1D:CB
-CKA_CLASS CK_OBJECT_CLASS CKO_CERTIFICATE
-CKA_TOKEN CK_BBOOL CK_TRUE
-CKA_PRIVATE CK_BBOOL CK_FALSE
-CKA_MODIFIABLE CK_BBOOL CK_FALSE
-CKA_LABEL UTF8 "SwissSign Silver CA - G2"
-CKA_CERTIFICATE_TYPE CK_CERTIFICATE_TYPE CKC_X_509
-CKA_SUBJECT MULTILINE_OCTAL
-\060\107\061\013\060\011\006\003\125\004\006\023\002\103\110\061
-\025\060\023\006\003\125\004\012\023\014\123\167\151\163\163\123
-\151\147\156\040\101\107\061\041\060\037\006\003\125\004\003\023
-\030\123\167\151\163\163\123\151\147\156\040\123\151\154\166\145
-\162\040\103\101\040\055\040\107\062
-END
-CKA_ID UTF8 "0"
-CKA_ISSUER MULTILINE_OCTAL
-\060\107\061\013\060\011\006\003\125\004\006\023\002\103\110\061
-\025\060\023\006\003\125\004\012\023\014\123\167\151\163\163\123
-\151\147\156\040\101\107\061\041\060\037\006\003\125\004\003\023
-\030\123\167\151\163\163\123\151\147\156\040\123\151\154\166\145
-\162\040\103\101\040\055\040\107\062
-END
-CKA_SERIAL_NUMBER MULTILINE_OCTAL
-\002\010\117\033\324\057\124\273\057\113
-END
-CKA_VALUE MULTILINE_OCTAL
-\060\202\005\275\060\202\003\245\240\003\002\001\002\002\010\117
-\033\324\057\124\273\057\113\060\015\006\011\052\206\110\206\367
-\015\001\001\005\005\000\060\107\061\013\060\011\006\003\125\004
-\006\023\002\103\110\061\025\060\023\006\003\125\004\012\023\014
-\123\167\151\163\163\123\151\147\156\040\101\107\061\041\060\037
-\006\003\125\004\003\023\030\123\167\151\163\163\123\151\147\156
-\040\123\151\154\166\145\162\040\103\101\040\055\040\107\062\060
-\036\027\015\060\066\061\060\062\065\060\070\063\062\064\066\132
-\027\015\063\066\061\060\062\065\060\070\063\062\064\066\132\060
-\107\061\013\060\011\006\003\125\004\006\023\002\103\110\061\025
-\060\023\006\003\125\004\012\023\014\123\167\151\163\163\123\151
-\147\156\040\101\107\061\041\060\037\006\003\125\004\003\023\030
-\123\167\151\163\163\123\151\147\156\040\123\151\154\166\145\162
-\040\103\101\040\055\040\107\062\060\202\002\042\060\015\006\011
-\052\206\110\206\367\015\001\001\001\005\000\003\202\002\017\000
-\060\202\002\012\002\202\002\001\000\304\361\207\177\323\170\061
-\367\070\311\370\303\231\103\274\307\367\274\067\347\116\161\272
-\113\217\245\163\035\134\156\230\256\003\127\256\070\067\103\057
-\027\075\037\310\316\150\020\301\170\256\031\003\053\020\372\054
-\171\203\366\350\271\150\271\125\362\004\104\247\071\371\374\004
-\213\036\361\242\115\047\371\141\173\272\267\345\242\023\266\353
-\141\076\320\154\321\346\373\372\136\355\035\264\236\240\065\133
-\241\222\313\360\111\222\376\205\012\005\076\346\331\013\342\117
-\273\334\225\067\374\221\351\062\065\042\321\037\072\116\047\205
-\235\260\025\224\062\332\141\015\107\115\140\102\256\222\107\350
-\203\132\120\130\351\212\213\271\135\241\334\335\231\112\037\066
-\147\273\110\344\203\266\067\353\110\072\257\017\147\217\027\007
-\350\004\312\357\152\061\207\324\300\266\371\224\161\173\147\144
-\270\266\221\112\102\173\145\056\060\152\014\365\220\356\225\346
-\362\315\202\354\331\241\112\354\366\262\113\345\105\205\346\155
-\170\223\004\056\234\202\155\066\251\304\061\144\037\206\203\013
-\052\364\065\012\170\311\125\317\101\260\107\351\060\237\231\276
-\141\250\006\204\271\050\172\137\070\331\033\251\070\260\203\177
-\163\301\303\073\110\052\202\017\041\233\270\314\250\065\303\204
-\033\203\263\076\276\244\225\151\001\072\211\000\170\004\331\311
-\364\231\031\253\126\176\133\213\206\071\025\221\244\020\054\011
-\062\200\140\263\223\300\052\266\030\013\235\176\215\111\362\020
-\112\177\371\325\106\057\031\222\243\231\247\046\254\273\214\074
-\346\016\274\107\007\334\163\121\361\160\144\057\010\371\264\107
-\035\060\154\104\352\051\067\205\222\150\146\274\203\070\376\173
-\071\056\323\120\360\037\373\136\140\266\251\246\372\047\101\361
-\233\030\162\362\365\204\164\112\311\147\304\124\256\110\144\337
-\214\321\156\260\035\341\007\217\010\036\231\234\161\351\114\330
-\245\367\107\022\037\164\321\121\236\206\363\302\242\043\100\013
-\163\333\113\246\347\163\006\214\301\240\351\301\131\254\106\372
-\346\057\370\317\161\234\106\155\271\304\025\215\070\171\003\105
-\110\357\304\135\327\010\356\207\071\042\206\262\015\017\130\103
-\367\161\251\110\056\375\352\326\037\002\003\001\000\001\243\201
-\254\060\201\251\060\016\006\003\125\035\017\001\001\377\004\004
-\003\002\001\006\060\017\006\003\125\035\023\001\001\377\004\005
-\060\003\001\001\377\060\035\006\003\125\035\016\004\026\004\024
-\027\240\315\301\344\101\266\072\133\073\313\105\235\275\034\302
-\230\372\206\130\060\037\006\003\125\035\043\004\030\060\026\200
-\024\027\240\315\301\344\101\266\072\133\073\313\105\235\275\034
-\302\230\372\206\130\060\106\006\003\125\035\040\004\077\060\075
-\060\073\006\011\140\205\164\001\131\001\003\001\001\060\056\060
-\054\006\010\053\006\001\005\005\007\002\001\026\040\150\164\164
-\160\072\057\057\162\145\160\157\163\151\164\157\162\171\056\163
-\167\151\163\163\163\151\147\156\056\143\157\155\057\060\015\006
-\011\052\206\110\206\367\015\001\001\005\005\000\003\202\002\001
-\000\163\306\201\340\047\322\055\017\340\225\060\342\232\101\177
-\120\054\137\137\142\141\251\206\152\151\030\014\164\111\326\135
-\204\352\101\122\030\157\130\255\120\126\040\152\306\275\050\151
-\130\221\334\221\021\065\251\072\035\274\032\245\140\236\330\037
-\177\105\221\151\331\176\273\170\162\301\006\017\052\316\217\205
-\160\141\254\240\315\013\270\071\051\126\204\062\116\206\273\075
-\304\052\331\327\037\162\356\376\121\241\042\101\261\161\002\143
-\032\202\260\142\253\136\127\022\037\337\313\335\165\240\300\135
-\171\220\214\033\340\120\346\336\061\376\230\173\160\137\245\220
-\330\255\370\002\266\157\323\140\335\100\113\042\305\075\255\072
-\172\237\032\032\107\221\171\063\272\202\334\062\151\003\226\156
-\037\113\360\161\376\343\147\162\240\261\277\134\213\344\372\231
-\042\307\204\271\033\215\043\227\077\355\045\340\317\145\273\365
-\141\004\357\335\036\262\132\101\042\132\241\237\135\054\350\133
-\311\155\251\014\014\170\252\140\306\126\217\001\132\014\150\274
-\151\031\171\304\037\176\227\005\277\305\351\044\121\136\324\325
-\113\123\355\331\043\132\066\003\145\243\301\003\255\101\060\363
-\106\033\205\220\257\145\265\325\261\344\026\133\170\165\035\227
-\172\155\131\251\052\217\173\336\303\207\211\020\231\111\163\170
-\310\075\275\121\065\164\052\325\361\176\151\033\052\273\073\275
-\045\270\232\132\075\162\141\220\146\207\356\014\326\115\324\021
-\164\013\152\376\013\003\374\243\125\127\211\376\112\313\256\133
-\027\005\310\362\215\043\061\123\070\322\055\152\077\202\271\215
-\010\152\367\136\101\164\156\303\021\176\007\254\051\140\221\077
-\070\312\127\020\015\275\060\057\307\245\346\101\240\332\256\005
-\207\232\240\244\145\154\114\011\014\211\272\270\323\271\300\223
-\212\060\372\215\345\232\153\025\001\116\147\252\332\142\126\076
-\204\010\146\322\304\066\175\247\076\020\374\210\340\324\200\345
-\000\275\252\363\116\006\243\172\152\371\142\162\343\011\117\353
-\233\016\001\043\361\237\273\174\334\334\154\021\227\045\262\362
-\264\143\024\322\006\052\147\214\203\365\316\352\007\330\232\152
-\036\354\344\012\273\052\114\353\011\140\071\316\312\142\330\056
-\156
-END
-CKA_NSS_MOZILLA_CA_POLICY CK_BBOOL CK_TRUE
-CKA_NSS_SERVER_DISTRUST_AFTER CK_BBOOL CK_FALSE
-CKA_NSS_EMAIL_DISTRUST_AFTER CK_BBOOL CK_FALSE
-
-# Trust for "SwissSign Silver CA - G2"
-# Issuer: CN=SwissSign Silver CA - G2,O=SwissSign AG,C=CH
-# Serial Number:4f:1b:d4:2f:54:bb:2f:4b
-# Subject: CN=SwissSign Silver CA - G2,O=SwissSign AG,C=CH
-# Not Valid Before: Wed Oct 25 08:32:46 2006
-# Not Valid After : Sat Oct 25 08:32:46 2036
-# Fingerprint (SHA-256): BE:6C:4D:A2:BB:B9:BA:59:B6:F3:93:97:68:37:42:46:C3:C0:05:99:3F:A9:8F:02:0D:1D:ED:BE:D4:8A:81:D5
-# Fingerprint (SHA1): 9B:AA:E5:9F:56:EE:21:CB:43:5A:BE:25:93:DF:A7:F0:40:D1:1D:CB
-CKA_CLASS CK_OBJECT_CLASS CKO_NSS_TRUST
-CKA_TOKEN CK_BBOOL CK_TRUE
-CKA_PRIVATE CK_BBOOL CK_FALSE
-CKA_MODIFIABLE CK_BBOOL CK_FALSE
-CKA_LABEL UTF8 "SwissSign Silver CA - G2"
-CKA_CERT_SHA1_HASH MULTILINE_OCTAL
-\233\252\345\237\126\356\041\313\103\132\276\045\223\337\247\360
-\100\321\035\313
-END
-CKA_CERT_MD5_HASH MULTILINE_OCTAL
-\340\006\241\311\175\317\311\374\015\300\126\165\226\330\142\023
-END
-CKA_ISSUER MULTILINE_OCTAL
-\060\107\061\013\060\011\006\003\125\004\006\023\002\103\110\061
-\025\060\023\006\003\125\004\012\023\014\123\167\151\163\163\123
-\151\147\156\040\101\107\061\041\060\037\006\003\125\004\003\023
-\030\123\167\151\163\163\123\151\147\156\040\123\151\154\166\145
-\162\040\103\101\040\055\040\107\062
-END
-CKA_SERIAL_NUMBER MULTILINE_OCTAL
-\002\010\117\033\324\057\124\273\057\113
-END
-CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
-CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
-CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
-CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
-
 #
 # Certificate "SecureTrust CA"
 #
@@ -21231,7 +21063,7 @@ CKA_SERIAL_NUMBER MULTILINE_OCTAL
 \002\011\000\326\135\233\263\170\201\056\353
 END
 CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
-CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
+CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
 
@@ -21399,7 +21231,7 @@ CKA_SERIAL_NUMBER MULTILINE_OCTAL
 \154\040
 END
 CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
-CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
+CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
 
@@ -21514,7 +21346,7 @@ CKA_SERIAL_NUMBER MULTILINE_OCTAL
 \112\353
 END
 CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
-CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
+CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
 
@@ -25970,3 +25802,339 @@ CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
 CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
 CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
+
+#
+# Certificate "D-TRUST BR Root CA 2 2023"
+#
+# Issuer: CN=D-TRUST BR Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Serial Number:73:3b:30:04:48:5b:d9:4d:78:2e:73:4b:c9:a1:dc:66
+# Subject: CN=D-TRUST BR Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Not Valid Before: Tue May 09 08:56:31 2023
+# Not Valid After : Sun May 09 08:56:30 2038
+# Fingerprint (SHA-256): 05:52:E6:F8:3F:DF:65:E8:FA:96:70:E6:66:DF:28:A4:E2:13:40:B5:10:CB:E5:25:66:F9:7C:4F:B9:4B:2B:D1
+# Fingerprint (SHA1): 2D:B0:70:EE:71:94:AF:69:68:17:DB:79:CE:58:9F:A0:6B:96:F7:87
+CKA_CLASS CK_OBJECT_CLASS CKO_CERTIFICATE
+CKA_TOKEN CK_BBOOL CK_TRUE
+CKA_PRIVATE CK_BBOOL CK_FALSE
+CKA_MODIFIABLE CK_BBOOL CK_FALSE
+CKA_LABEL UTF8 "D-TRUST BR Root CA 2 2023"
+CKA_CERTIFICATE_TYPE CK_CERTIFICATE_TYPE CKC_X_509
+CKA_SUBJECT MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\102\122\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_ID UTF8 "0"
+CKA_ISSUER MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\102\122\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_SERIAL_NUMBER MULTILINE_OCTAL
+\002\020\163\073\060\004\110\133\331\115\170\056\163\113\311\241
+\334\146
+END
+CKA_VALUE MULTILINE_OCTAL
+\060\202\005\251\060\202\003\221\240\003\002\001\002\002\020\163
+\073\060\004\110\133\331\115\170\056\163\113\311\241\334\146\060
+\015\006\011\052\206\110\206\367\015\001\001\015\005\000\060\110
+\061\013\060\011\006\003\125\004\006\023\002\104\105\061\025\060
+\023\006\003\125\004\012\023\014\104\055\124\162\165\163\164\040
+\107\155\142\110\061\042\060\040\006\003\125\004\003\023\031\104
+\055\124\122\125\123\124\040\102\122\040\122\157\157\164\040\103
+\101\040\062\040\062\060\062\063\060\036\027\015\062\063\060\065
+\060\071\060\070\065\066\063\061\132\027\015\063\070\060\065\060
+\071\060\070\065\066\063\060\132\060\110\061\013\060\011\006\003
+\125\004\006\023\002\104\105\061\025\060\023\006\003\125\004\012
+\023\014\104\055\124\162\165\163\164\040\107\155\142\110\061\042
+\060\040\006\003\125\004\003\023\031\104\055\124\122\125\123\124
+\040\102\122\040\122\157\157\164\040\103\101\040\062\040\062\060
+\062\063\060\202\002\042\060\015\006\011\052\206\110\206\367\015
+\001\001\001\005\000\003\202\002\017\000\060\202\002\012\002\202
+\002\001\000\256\377\011\131\221\200\012\112\150\346\044\077\270
+\247\344\310\072\012\072\026\315\311\043\141\240\223\161\362\253
+\213\163\217\240\147\145\140\322\124\153\143\121\157\111\063\340
+\162\007\023\175\070\315\006\222\007\051\122\153\116\167\154\004
+\323\225\372\335\114\214\331\135\301\141\175\113\347\050\263\104
+\201\173\121\257\335\063\261\150\174\326\116\114\376\053\150\271
+\312\146\151\304\354\136\127\177\367\015\307\234\066\066\345\007
+\140\254\300\114\352\010\154\357\006\174\117\133\050\172\010\374
+\223\135\233\366\234\264\213\206\272\041\271\364\360\350\131\132
+\050\241\064\204\032\045\221\266\265\217\357\262\371\200\372\371
+\075\074\021\162\330\343\057\206\166\305\171\054\301\251\220\223
+\106\230\147\313\203\152\240\120\043\247\073\366\201\071\340\355
+\360\271\277\145\361\330\313\172\373\357\163\003\316\000\364\175
+\327\340\135\073\146\270\334\216\272\203\313\207\166\003\374\045
+\331\347\043\157\006\375\147\363\340\377\204\274\107\277\265\026
+\030\106\151\024\314\005\367\333\323\111\254\153\314\253\344\265
+\013\103\044\136\113\153\115\147\337\326\265\076\117\170\037\224
+\161\044\352\336\160\374\361\223\376\236\223\132\344\224\132\227
+\124\014\065\173\137\154\356\000\037\044\354\003\272\002\365\166
+\364\237\324\232\355\205\054\070\042\057\307\330\057\166\021\117
+\375\154\134\350\365\216\047\207\177\031\112\041\107\220\035\171
+\215\034\133\370\317\112\205\344\355\263\133\215\276\304\144\050
+\135\101\304\156\254\070\132\117\043\164\164\251\022\303\366\322
+\271\021\025\063\007\221\330\073\067\072\143\060\006\321\305\042
+\066\050\142\043\020\340\106\314\227\254\326\053\135\144\044\325
+\356\034\016\336\373\010\132\165\052\366\143\155\316\013\102\276
+\321\272\160\034\234\041\345\017\061\151\027\327\374\012\264\336
+\355\200\234\313\222\264\213\365\336\131\242\130\011\245\143\107
+\013\341\101\062\064\101\331\232\261\331\250\260\033\132\336\015
+\015\364\342\262\135\065\200\271\201\324\204\151\221\002\313\165
+\320\215\305\265\075\011\221\011\217\024\241\024\164\171\076\326
+\311\025\035\244\131\131\042\334\366\212\105\075\074\022\326\076
+\135\062\057\002\003\001\000\001\243\201\216\060\201\213\060\017
+\006\003\125\035\023\001\001\377\004\005\060\003\001\001\377\060
+\035\006\003\125\035\016\004\026\004\024\147\220\360\326\336\265
+\030\325\106\051\176\134\253\370\236\010\274\144\225\020\060\016
+\006\003\125\035\017\001\001\377\004\004\003\002\001\006\060\111
+\006\003\125\035\037\004\102\060\100\060\076\240\074\240\072\206
+\070\150\164\164\160\072\057\057\143\162\154\056\144\055\164\162
+\165\163\164\056\156\145\164\057\143\162\154\057\144\055\164\162
+\165\163\164\137\142\162\137\162\157\157\164\137\143\141\137\062
+\137\062\060\062\063\056\143\162\154\060\015\006\011\052\206\110
+\206\367\015\001\001\015\005\000\003\202\002\001\000\064\367\263
+\167\123\333\060\026\271\055\245\041\361\100\041\165\353\353\110
+\026\201\075\163\340\236\047\052\353\167\251\023\244\152\012\132
+\132\024\063\075\150\037\201\256\151\375\214\237\145\154\064\102
+\331\055\320\177\170\026\261\072\254\043\061\255\136\177\256\347
+\256\053\372\272\374\074\227\225\100\223\137\303\055\003\243\355
+\244\157\123\327\372\100\016\060\365\000\040\054\000\114\214\073
+\264\243\037\266\277\221\062\253\257\222\230\323\026\346\324\321
+\124\134\103\133\056\256\357\127\052\250\264\157\244\357\015\126
+\024\332\041\253\040\166\236\003\374\046\270\236\077\076\003\046
+\346\114\333\235\137\102\204\075\105\003\003\034\131\210\312\334
+\056\141\044\132\244\352\047\013\163\022\276\122\263\012\317\062
+\027\342\036\207\032\026\225\110\155\132\340\320\317\011\222\046
+\146\221\330\243\141\016\252\201\201\177\350\122\202\321\102\347
+\340\035\030\372\244\205\066\347\206\340\015\353\274\324\311\326
+\074\103\361\135\111\156\176\201\233\151\265\211\142\217\210\122
+\330\327\376\047\301\043\305\313\053\002\273\261\137\376\373\103
+\205\003\106\276\135\306\312\041\046\377\327\002\236\164\112\334
+\370\023\025\261\201\127\066\313\145\134\321\035\061\167\351\045
+\303\303\262\062\067\325\361\230\011\344\155\143\200\010\253\006
+\222\201\324\351\160\217\247\077\262\355\206\214\202\152\065\310
+\102\132\202\321\122\032\105\017\025\245\000\360\224\173\145\047
+\127\071\103\317\174\177\346\275\065\263\173\361\031\114\336\072
+\226\317\351\166\356\003\347\302\103\122\074\152\201\350\301\132
+\200\275\021\135\223\153\373\307\346\144\077\273\151\034\351\335
+\045\213\257\164\311\124\100\312\313\223\023\012\355\373\146\222
+\021\312\365\300\372\330\203\125\003\174\323\305\042\106\165\160
+\153\171\110\006\052\202\232\277\346\353\026\016\042\105\001\274
+\335\066\224\064\251\065\046\212\327\227\271\356\010\162\277\064
+\222\160\203\200\253\070\252\131\150\335\100\244\030\220\262\363
+\325\003\312\046\312\357\325\307\340\217\123\216\360\000\343\250
+\355\237\371\255\167\340\053\143\117\236\303\356\067\273\170\011
+\204\236\271\156\373\051\231\220\350\200\323\237\044
+END
+CKA_NSS_MOZILLA_CA_POLICY CK_BBOOL CK_TRUE
+CKA_NSS_SERVER_DISTRUST_AFTER CK_BBOOL CK_FALSE
+CKA_NSS_EMAIL_DISTRUST_AFTER CK_BBOOL CK_FALSE
+
+# Trust for "D-TRUST BR Root CA 2 2023"
+# Issuer: CN=D-TRUST BR Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Serial Number:73:3b:30:04:48:5b:d9:4d:78:2e:73:4b:c9:a1:dc:66
+# Subject: CN=D-TRUST BR Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Not Valid Before: Tue May 09 08:56:31 2023
+# Not Valid After : Sun May 09 08:56:30 2038
+# Fingerprint (SHA-256): 05:52:E6:F8:3F:DF:65:E8:FA:96:70:E6:66:DF:28:A4:E2:13:40:B5:10:CB:E5:25:66:F9:7C:4F:B9:4B:2B:D1
+# Fingerprint (SHA1): 2D:B0:70:EE:71:94:AF:69:68:17:DB:79:CE:58:9F:A0:6B:96:F7:87
+CKA_CLASS CK_OBJECT_CLASS CKO_NSS_TRUST
+CKA_TOKEN CK_BBOOL CK_TRUE
+CKA_PRIVATE CK_BBOOL CK_FALSE
+CKA_MODIFIABLE CK_BBOOL CK_FALSE
+CKA_LABEL UTF8 "D-TRUST BR Root CA 2 2023"
+CKA_CERT_SHA1_HASH MULTILINE_OCTAL
+\055\260\160\356\161\224\257\151\150\027\333\171\316\130\237\240
+\153\226\367\207
+END
+CKA_CERT_MD5_HASH MULTILINE_OCTAL
+\341\011\355\323\140\324\126\033\107\037\267\014\137\033\137\205
+END
+CKA_ISSUER MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\102\122\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_SERIAL_NUMBER MULTILINE_OCTAL
+\002\020\163\073\060\004\110\133\331\115\170\056\163\113\311\241
+\334\146
+END
+CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
+CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
+CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
+CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
+
+#
+# Certificate "D-TRUST EV Root CA 2 2023"
+#
+# Issuer: CN=D-TRUST EV Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Serial Number:69:26:09:7e:80:4b:4c:a0:a7:8c:78:62:53:5f:5a:6f
+# Subject: CN=D-TRUST EV Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Not Valid Before: Tue May 09 09:10:33 2023
+# Not Valid After : Sun May 09 09:10:32 2038
+# Fingerprint (SHA-256): 8E:82:21:B2:E7:D4:00:78:36:A1:67:2F:0D:CC:29:9C:33:BC:07:D3:16:F1:32:FA:1A:20:6D:58:71:50:F1:CE
+# Fingerprint (SHA1): A5:5B:D8:47:6C:8F:19:F7:4C:F4:6D:6B:B6:C2:79:82:22:DF:54:8B
+CKA_CLASS CK_OBJECT_CLASS CKO_CERTIFICATE
+CKA_TOKEN CK_BBOOL CK_TRUE
+CKA_PRIVATE CK_BBOOL CK_FALSE
+CKA_MODIFIABLE CK_BBOOL CK_FALSE
+CKA_LABEL UTF8 "D-TRUST EV Root CA 2 2023"
+CKA_CERTIFICATE_TYPE CK_CERTIFICATE_TYPE CKC_X_509
+CKA_SUBJECT MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\105\126\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_ID UTF8 "0"
+CKA_ISSUER MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\105\126\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_SERIAL_NUMBER MULTILINE_OCTAL
+\002\020\151\046\011\176\200\113\114\240\247\214\170\142\123\137
+\132\157
+END
+CKA_VALUE MULTILINE_OCTAL
+\060\202\005\251\060\202\003\221\240\003\002\001\002\002\020\151
+\046\011\176\200\113\114\240\247\214\170\142\123\137\132\157\060
+\015\006\011\052\206\110\206\367\015\001\001\015\005\000\060\110
+\061\013\060\011\006\003\125\004\006\023\002\104\105\061\025\060
+\023\006\003\125\004\012\023\014\104\055\124\162\165\163\164\040
+\107\155\142\110\061\042\060\040\006\003\125\004\003\023\031\104
+\055\124\122\125\123\124\040\105\126\040\122\157\157\164\040\103
+\101\040\062\040\062\060\062\063\060\036\027\015\062\063\060\065
+\060\071\060\071\061\060\063\063\132\027\015\063\070\060\065\060
+\071\060\071\061\060\063\062\132\060\110\061\013\060\011\006\003
+\125\004\006\023\002\104\105\061\025\060\023\006\003\125\004\012
+\023\014\104\055\124\162\165\163\164\040\107\155\142\110\061\042
+\060\040\006\003\125\004\003\023\031\104\055\124\122\125\123\124
+\040\105\126\040\122\157\157\164\040\103\101\040\062\040\062\060
+\062\063\060\202\002\042\060\015\006\011\052\206\110\206\367\015
+\001\001\001\005\000\003\202\002\017\000\060\202\002\012\002\202
+\002\001\000\330\216\243\211\200\013\262\127\122\334\251\123\114
+\067\271\177\143\027\023\357\247\133\043\133\151\165\260\231\012
+\027\301\213\304\333\250\340\314\061\272\302\362\315\135\351\267
+\370\035\257\152\304\225\207\327\107\311\225\330\202\004\120\075
+\201\010\377\344\075\263\261\326\305\262\375\210\011\333\234\204
+\354\045\027\024\207\177\060\170\233\152\130\311\266\163\050\074
+\064\367\231\367\177\323\246\370\034\105\174\255\054\214\224\077
+\330\147\020\123\176\042\315\116\045\121\360\045\044\065\021\136
+\020\306\354\207\146\211\201\150\272\314\053\235\107\163\037\275
+\315\221\244\162\152\234\242\033\030\240\157\354\120\364\175\100
+\302\250\060\317\275\163\310\023\053\020\023\036\213\232\250\072
+\224\163\323\030\151\012\112\377\301\001\003\377\171\177\265\110
+\177\173\356\350\051\157\066\114\225\141\206\330\371\242\163\212
+\356\256\057\226\356\150\315\075\115\050\102\371\105\053\062\033
+\106\125\026\152\246\113\051\371\273\225\126\277\106\035\354\035
+\223\035\300\145\262\037\241\103\256\126\236\240\261\217\153\022
+\267\140\155\170\013\312\212\134\355\036\226\016\203\246\110\225
+\215\073\243\041\304\256\130\306\000\262\204\264\043\244\226\206
+\065\270\330\236\330\254\064\111\230\143\225\305\313\155\110\107
+\342\362\056\030\036\320\061\253\335\164\354\371\334\214\270\034
+\216\150\043\272\320\363\120\334\317\145\217\163\072\062\307\174
+\376\312\202\042\117\276\216\142\107\146\345\315\207\342\350\325
+\017\030\237\345\004\162\113\106\074\020\362\104\302\144\126\161
+\116\165\350\234\311\046\164\305\175\131\321\012\133\017\155\376
+\236\165\034\030\306\032\072\174\330\015\004\314\315\267\105\145
+\172\261\217\270\256\204\110\076\263\172\115\250\003\342\342\176
+\001\026\131\150\030\103\063\260\322\334\260\032\103\065\356\245
+\332\251\106\134\256\206\201\101\001\112\164\046\354\237\006\277
+\302\005\067\144\165\170\051\150\375\305\365\353\376\107\371\344
+\205\260\341\173\061\235\246\177\162\243\271\304\054\056\314\231
+\127\016\041\014\105\001\224\145\353\145\011\306\143\042\013\063
+\111\222\110\074\374\315\316\260\076\216\236\213\370\376\111\305
+\065\162\107\002\003\001\000\001\243\201\216\060\201\213\060\017
+\006\003\125\035\023\001\001\377\004\005\060\003\001\001\377\060
+\035\006\003\125\035\016\004\026\004\024\252\374\221\020\033\207
+\221\137\026\271\277\117\113\221\136\000\034\261\062\200\060\016
+\006\003\125\035\017\001\001\377\004\004\003\002\001\006\060\111
+\006\003\125\035\037\004\102\060\100\060\076\240\074\240\072\206
+\070\150\164\164\160\072\057\057\143\162\154\056\144\055\164\162
+\165\163\164\056\156\145\164\057\143\162\154\057\144\055\164\162
+\165\163\164\137\145\166\137\162\157\157\164\137\143\141\137\062
+\137\062\060\062\063\056\143\162\154\060\015\006\011\052\206\110
+\206\367\015\001\001\015\005\000\003\202\002\001\000\223\313\245
+\037\231\021\354\232\015\137\054\025\223\306\077\276\020\215\170
+\102\360\156\220\107\107\216\243\222\062\215\160\217\366\133\215
+\276\211\316\107\001\152\033\040\040\211\133\310\202\020\154\340
+\347\231\252\153\306\052\240\143\065\221\152\205\045\255\027\070
+\245\233\176\120\362\166\352\205\005\052\047\101\053\261\201\321
+\242\366\100\165\251\016\313\361\125\110\330\354\321\354\263\350
+\316\024\241\065\354\302\136\065\032\253\246\026\001\006\216\352
+\334\057\243\212\312\054\221\353\122\216\137\014\233\027\317\313
+\163\007\031\304\152\302\163\124\357\174\103\122\143\301\021\312
+\302\105\261\364\073\123\365\151\256\074\343\245\336\254\350\124
+\267\262\221\375\254\251\037\362\207\344\027\306\111\250\174\330
+\012\101\364\362\076\347\167\064\004\122\335\350\201\362\115\057
+\124\105\235\025\341\117\314\345\336\064\127\020\311\043\162\027
+\160\215\120\160\037\126\154\314\271\377\072\132\117\143\172\303
+\156\145\007\035\204\241\377\251\014\143\211\155\262\100\210\071
+\327\037\167\150\265\374\234\325\326\147\151\133\250\164\333\374
+\211\366\033\062\367\244\044\246\166\267\107\123\357\215\111\217
+\251\266\203\132\245\226\220\105\141\365\336\003\117\046\017\250
+\213\360\003\226\260\254\025\320\161\132\152\173\224\346\160\223
+\332\361\151\340\262\142\115\236\217\377\211\235\233\135\315\105
+\351\224\002\042\215\340\065\177\350\361\004\171\161\154\124\203
+\370\063\271\005\062\033\130\125\021\117\320\345\047\107\161\354
+\355\332\147\326\142\246\113\115\017\151\242\311\274\354\042\113
+\224\307\150\224\027\176\342\216\050\076\266\306\352\365\064\154
+\237\067\210\007\070\333\206\161\372\315\225\110\103\156\243\117
+\202\207\327\064\230\156\113\223\171\140\165\151\017\360\032\325
+\123\372\041\014\302\077\351\077\037\030\214\222\135\170\247\166
+\147\031\273\262\352\177\351\160\011\126\126\243\260\014\013\055
+\066\136\305\351\304\325\203\313\206\027\227\054\154\023\157\207
+\132\257\111\246\035\333\315\070\004\056\137\342\112\065\016\055
+\113\370\242\044\004\215\330\341\143\136\002\222\064\332\230\141
+\134\034\157\130\166\144\263\374\002\270\365\235\012
+END
+CKA_NSS_MOZILLA_CA_POLICY CK_BBOOL CK_TRUE
+CKA_NSS_SERVER_DISTRUST_AFTER CK_BBOOL CK_FALSE
+CKA_NSS_EMAIL_DISTRUST_AFTER CK_BBOOL CK_FALSE
+
+# Trust for "D-TRUST EV Root CA 2 2023"
+# Issuer: CN=D-TRUST EV Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Serial Number:69:26:09:7e:80:4b:4c:a0:a7:8c:78:62:53:5f:5a:6f
+# Subject: CN=D-TRUST EV Root CA 2 2023,O=D-Trust GmbH,C=DE
+# Not Valid Before: Tue May 09 09:10:33 2023
+# Not Valid After : Sun May 09 09:10:32 2038
+# Fingerprint (SHA-256): 8E:82:21:B2:E7:D4:00:78:36:A1:67:2F:0D:CC:29:9C:33:BC:07:D3:16:F1:32:FA:1A:20:6D:58:71:50:F1:CE
+# Fingerprint (SHA1): A5:5B:D8:47:6C:8F:19:F7:4C:F4:6D:6B:B6:C2:79:82:22:DF:54:8B
+CKA_CLASS CK_OBJECT_CLASS CKO_NSS_TRUST
+CKA_TOKEN CK_BBOOL CK_TRUE
+CKA_PRIVATE CK_BBOOL CK_FALSE
+CKA_MODIFIABLE CK_BBOOL CK_FALSE
+CKA_LABEL UTF8 "D-TRUST EV Root CA 2 2023"
+CKA_CERT_SHA1_HASH MULTILINE_OCTAL
+\245\133\330\107\154\217\031\367\114\364\155\153\266\302\171\202
+\042\337\124\213
+END
+CKA_CERT_MD5_HASH MULTILINE_OCTAL
+\226\264\170\011\360\011\313\167\353\273\033\115\157\066\274\266
+END
+CKA_ISSUER MULTILINE_OCTAL
+\060\110\061\013\060\011\006\003\125\004\006\023\002\104\105\061
+\025\060\023\006\003\125\004\012\023\014\104\055\124\162\165\163
+\164\040\107\155\142\110\061\042\060\040\006\003\125\004\003\023
+\031\104\055\124\122\125\123\124\040\105\126\040\122\157\157\164
+\040\103\101\040\062\040\062\060\062\063
+END
+CKA_SERIAL_NUMBER MULTILINE_OCTAL
+\002\020\151\046\011\176\200\113\114\240\247\214\170\142\123\137
+\132\157
+END
+CKA_TRUST_SERVER_AUTH CK_TRUST CKT_NSS_TRUSTED_DELEGATOR
+CKA_TRUST_EMAIL_PROTECTION CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
+CKA_TRUST_CODE_SIGNING CK_TRUST CKT_NSS_MUST_VERIFY_TRUST
+CKA_TRUST_STEP_UP_APPROVED CK_BBOOL CK_FALSE
diff --git a/tools/icu/current_ver.dep b/tools/icu/current_ver.dep
index a2c73ee25ae6c4..da52397ed40e5d 100644
--- a/tools/icu/current_ver.dep
+++ b/tools/icu/current_ver.dep
@@ -1,6 +1,6 @@
 [
   {
-    "url": "https://github.com/unicode-org/icu/releases/download/release-76-1/icu4c-76_1-src.tgz",
-    "md5": "857fdafff8127139cc175a3ec9b43bd6"
+    "url": "https://github.com/unicode-org/icu/releases/download/release-77-1/icu4c-77_1-src.tgz",
+    "md5": "bc0132b4c43db8455d2446c3bae58898"
   }
 ]
diff --git a/tools/osx-pkg-postinstall.sh b/tools/osx-pkg-postinstall.sh
deleted file mode 100644
index ab0dfa8d234b28..00000000000000
--- a/tools/osx-pkg-postinstall.sh
+++ /dev/null
@@ -1,9 +0,0 @@
-#!/bin/sh
-# TODO Can this be done inside the .pmdoc?
-# TODO Can we extract $PREFIX from the installer?
-cd /usr/local/bin || exit
-
-ln -sf ../lib/node_modules/npm/bin/npm-cli.js npm
-ln -sf ../lib/node_modules/npm/bin/npx-cli.js npx
-
-ln -sf ../lib/node_modules/corepack/dist/corepack.js corepack