wasm: browser compatibility policy

[neelance]

I would like to propose a discussion on how we want to handle browser compatibility. There are several features being worked on: https://github.com/WebAssembly/proposals When do we want to use them? Do we want to offer backwards-compatibility?

Here are my own thoughts, open for debate:

(1) Go's support for WebAssembly is experimental and it should stay experimental at least until WebAssembly itself is fully standardized (right now the main spec is still in "Phase 2 - Proposed Spec Text Available").

(2) The WebAssembly project itself has no clear answer yet on how feature tests and backwards compatibility should work, see https://webassembly.org/docs/feature-test/

(3) Modern browsers (except Safari) have auto-update mechanisms, so staying with the latest version shouldn't be an issue for most users.

Due to (1), (2) and (3) I would opt for no backwards compatibility as long as Go's support for WebAssembly is experimental.

New WebAssembly features can be adopted as soon as they are marked as "standardized" and are supported by the latest stable Chrome, Firefox, Edge and Safari (on desktop).

[neelance]

/cc @bradfitz @myitcv @dennwc @theclapp

[bradfitz]

This bug is a little abstract, so this answer is somewhat vague, but:

I'm fine with treating experimental things as experimental, but we should start trying to minimize how often we break users for doing basic things using just the base "MVP WebAssembly" bits. It sounds like everybody's cool with the syscall/js.NewCallback change, but that's the sort of change we shouldn't be making every release.

For experimental new features of WebAssembly, I'd prefer them to be opt-in somehow with associated "This is super experimental!" documentation on the opt-in switch, at least where possible. At least document the state of the compatibility policy in, say, the syscall/js package comment and/or release notes.

[neelance]

Yes, I agree that we should aim for keeping the API stable. This issue however was mainly about browser compatibility.

The features in question wouldn't add new API, but improve performance or reduce output size. Some examples:

• a builtin memcpy operation
• other operations that currently require a slow workaround
• advanced memory operations so we don't have to request 1 GB of memory at startup
• support for threads

Here are some options we have:

• Not use the new features, stay with the MVP features. (This would mean suboptimal performance.)
• Add compiler flags to choose which features to use. (Go usually avoids compiler flags.)
• Output multiple wasm binaries using different feature sets.
• Build a special decoder that can polyfill unsupported features at load time. (This will probably be the best solution at some point, but is also complicated and will likely be tackled by the WebAssembly project itself.)
• No backwards compatibility for now.

Given these options I would prefer to not have backwards compatibility right now and as soon as feature testing is possible I would add some warning mechanism that tells the user to upgrade the browser.

[myitcv]

I'm guessing we don't want to use different values of GOOS here? GOOS=chrome etc...

[neelance]

I'm guessing we don't want to use different values of GOOS here? GOOS=chrome etc...

More GOOS or GOARCH values are cumbersome. Having just one for chrome wouldn't be enough, you would need to target browser versions or generations, e.g. wasm-10-2018.

[bradfitz]

Let's use memcpy as an example:

Even without an official feature detection API, can we try to use it in runtime init & catch a fault & set a bool about whether to use it? We do that elsewhere in the runtime for other platforms.

Likewise with other stuff.

If we really can't feature detect at all, I'd say we shoudn't use stuff by default until it's in all the major browser's stable releases.

Worst case we could add a GOWASM environment variable like GOARM, GO386, GOMIPS, etc.

[theclapp]

Add compiler flags to choose which features to use. (Go usually avoids compiler flags.)

Do you mean actual compiler flags like -a or -n, or build tags?

Of course, I imagine build tags would suffer from the same issue as GOOS mentioned previously: "Having just one for chrome wouldn't be enough, you would need to target browser versions or generations". But even so, it seems like build tags might allow more flexibility than compiler flags.

[agnivade]

can we try to use it in runtime init & catch a fault & set a bool about whether to use it? We do that elsewhere in the runtime for other platforms.

I think the issue is that the entire wasm code is validated AOT. As explained here (https://webassembly.org/docs/feature-test/) -

Since some WebAssembly features add operators and all WebAssembly code in a module is validated ahead-of-time, the usual JavaScript feature detection pattern:

if (foo)
    foo();
else
    alternativeToFoo();
won’t work in WebAssembly (if foo isn’t supported, foo() will fail to validate).

I would recommend against more flags or environment variables. Given that wasm is a fast moving platform and it is much easier to upgrade browsers than operating systems.

Not use the new features, stay with the MVP features. (This would mean suboptimal performance.)

Perhaps we can delay new features by one release, so that we give enough time for everybody to catch up to the latest browser version ?

[neelance]

can we try to use it in runtime init & catch a fault & set a bool about whether to use it? We do that elsewhere in the runtime for other platforms.

I think the issue is that the entire wasm code is validated AOT.

Right. The WebAssembly loader will reject the binary if there are any unsupported operations. The runtime init will not be executed. Currently the only way to do feature detection is to try to load a wasm binary and if it fails then load a different binary.

If we really can't feature detect at all, I'd say we shoudn't use stuff by default until it's in all the major browser's stable releases.

In my first post I suggested "New WebAssembly features can be adopted as soon as they are marked as "standardized" and are supported by the latest stable Chrome, Firefox, Edge and Safari (on desktop)." @bradfitz This is exactly what you mean, isn't it?

Perhaps we can delay new features by one release, so that we give enough time for everybody to catch up to the latest browser version ?

Sounds like a good idea. Maybe 2 or 3 months after the feature has been released?

[bradfitz]

In my first post I suggested "New WebAssembly features can be adopted as soon as they are marked as "standardized" and are supported by the latest stable Chrome, Firefox, Edge and Safari (on desktop)." @bradfitz This is exactly what you mean, isn't it?

Yeah.

But if you want to do better earlier, I think it'd be fine to use a GOWASM environment variable to opt-in to newer features that aren't as widely available.

The default should be to produce a binary that works everywhere, but if you want to use something more experimental that only works on one browser, GOWASM=threads,foo,bar seems fine to me at least.

[djhworld]

New WebAssembly features can be adopted as soon as they are marked as "standardized"

I think this should be the gate to decide whether to implement, rather than relying on the release train of the browsers.

I agree with @bradfitz though to maybe put the bleeding edge stuff behind some flags.

On the flipside that's getting back to the bad old days of "best ran on X browser!" badges, but that's not a Go problem 😄