[libc] Optimize mempcy size thresholds (#70049)

Adjust boundary conditions for sizes = 16/32/64.
See the added comment for explanations.

Results on a machine with AVX2, so sizes 64/128 affected:
```
                │   baseline   │               adjusted               │
                │    sec/op    │   sec/op     vs base                 │
memcpy/Google_A   5.701n ±  0%   5.551n ± 1%   -2.63% (n=100)
memcpy/Google_B   3.817n ±  0%   3.776n ± 0%   -1.07% (p=0.000 n=100)
memcpy/Google_D   11.35n ±  1%   11.32n ± 0%        ~ (p=0.066 n=100)
memcpy/Google_U   3.874n ± 1%    3.821n ± 1%   -1.37% (p=0.001 n=100)
memcpy/64         3.843n ±  0%   3.105n ± 3%  -19.22% (n=50)
memcpy/128        4.842n ±  0%   3.818n ± 0%  -21.15% (p=0.000 n=50)
```

Author: dvyukov

libc/src/string/memory_utils/x86_64/inline_memcpy.h

@@ -55,7 +55,7 @@ LIBC_INLINE_VAR constexpr size_t kRepMovsbThreshold =
 [[maybe_unused]] LIBC_INLINE void
 inline_memcpy_x86_sse2_ge64(Ptr __restrict dst, CPtr __restrict src,
                             size_t count) {
-  if (count < 128)
+  if (count <= 128)
     return builtin::Memcpy<64>::head_tail(dst, src, count);
   builtin::Memcpy<32>::block(dst, src);
   align_to_next_boundary<32, Arg::Dst>(dst, src, count);
@@ -65,7 +65,7 @@ inline_memcpy_x86_sse2_ge64(Ptr __restrict dst, CPtr __restrict src,
 [[maybe_unused]] LIBC_INLINE void
 inline_memcpy_x86_avx_ge64(Ptr __restrict dst, CPtr __restrict src,
                            size_t count) {
-  if (count < 128)
+  if (count <= 128)
     return builtin::Memcpy<64>::head_tail(dst, src, count);
   if (count < 256)
     return builtin::Memcpy<128>::head_tail(dst, src, count);
@@ -79,7 +79,7 @@ inline_memcpy_x86_sse2_ge64_sw_prefetching(Ptr __restrict dst,
                                            CPtr __restrict src, size_t count) {
   using namespace LIBC_NAMESPACE::x86;
   prefetch_to_local_cache(src + kOneCacheline);
-  if (count < 128)
+  if (count <= 128)
     return builtin::Memcpy<64>::head_tail(dst, src, count);
   prefetch_to_local_cache(src + kTwoCachelines);
   // Aligning 'dst' on a 32B boundary.
@@ -120,7 +120,7 @@ inline_memcpy_x86_avx_ge64_sw_prefetching(Ptr __restrict dst,
                                           CPtr __restrict src, size_t count) {
   using namespace LIBC_NAMESPACE::x86;
   prefetch_to_local_cache(src + kOneCacheline);
-  if (count < 128)
+  if (count <= 128)
     return builtin::Memcpy<64>::head_tail(dst, src, count);
   prefetch_to_local_cache(src + kTwoCachelines);
   prefetch_to_local_cache(src + kThreeCachelines);
@@ -149,6 +149,15 @@ inline_memcpy_x86_avx_ge64_sw_prefetching(Ptr __restrict dst,
 
 [[maybe_unused]] LIBC_INLINE void
 inline_memcpy_x86(Ptr __restrict dst, CPtr __restrict src, size_t count) {
+#if defined(__AVX512F__)
+  constexpr size_t vector_size = 64;
+#elif defined(__AVX__)
+  constexpr size_t vector_size = 32;
+#elif defined(__SSE2__)
+  constexpr size_t vector_size = 16;
+#else
+  constexpr size_t vector_size = 8;
+#endif
   if (count == 0)
     return;
   if (count == 1)
@@ -161,11 +170,20 @@ inline_memcpy_x86(Ptr __restrict dst, CPtr __restrict src, size_t count) {
     return builtin::Memcpy<4>::block(dst, src);
   if (count < 8)
     return builtin::Memcpy<4>::head_tail(dst, src, count);
-  if (count < 16)
+  // If count is equal to a power of 2, we can handle it as head-tail
+  // of both smaller size and larger size (head-tail are either
+  // non-overlapping for smaller size, or completely collapsed
+  // for larger size). It seems to be more profitable to do the copy
+  // with the larger size, if it's natively supported (e.g. doing
+  // 2 collapsed 32-byte moves for count=64 if AVX2 is supported).
+  // But it's not profitable to use larger size if it's not natively
+  // supported: we will both use more instructions and handle fewer
+  // sizes in earlier branches.
+  if (vector_size >= 16 ? count < 16 : count <= 16)
     return builtin::Memcpy<8>::head_tail(dst, src, count);
-  if (count < 32)
+  if (vector_size >= 32 ? count < 32 : count <= 32)
     return builtin::Memcpy<16>::head_tail(dst, src, count);
-  if (count < 64)
+  if (vector_size >= 64 ? count < 64 : count <= 64)
     return builtin::Memcpy<32>::head_tail(dst, src, count);
   if constexpr (x86::kAvx) {
     if constexpr (x86::kUseSoftwarePrefetching) {