[libc-commits] [libc] [libc][SVE] add sve handling for memcpy with count less than 32b (PR #167446)

[Guillaume Chatelet via libc-commits]

gchatelet wrote:

We finally have good evidence that the patch is positive in production.
As said earlier we need to treat the `0` size apart and have a special path for size `<=16`.

Something along these lines
```c++
  if (count == 0) return;
#ifdef LIBC_TARGET_CPU_HAS_SVE
  auto src_ptr = reinterpret_cast<const uint8_t*>(src);
  auto dst_ptr = reinterpret_cast<uint8_t*>(dst);
  if (count <= 16) {
    const svbool_t mask = svwhilelt_b8_u64(0, count);
    svst1_u8(mask, dst_ptr, svld1_u8(mask, src_ptr));
    return;
  }
  if (count <= 32) {
    const size_t vlen = svcntb();
    svbool_t m0 = svwhilelt_b8_u64(0, count);
    svbool_t m1 = svwhilelt_b8_u64(vlen, count);
    svst1_u8(m0, dst_ptr, svld1_u8(m0, src_ptr));
    svst1_u8(m1, dst_ptr + vlen, svld1_u8(m1, src_ptr + vlen));
    return;
  }
#else
  if (count == 1) return builtin::Memcpy<1>::block(dst, src);
  ...
```

I'm still a bit bothered by how this patch will age when `vlen` bumps to 32 or higher but for now this seems to be a win.

https://github.com/llvm/llvm-project/pull/167446