Disable auto vectorization of xxhash64, when AVX512 is present.

[TocarIP]

AVX512 adds support for VPMULLQ instructions, which makes is possible to vectorize XXH64_update. Here is a minimal reproducer where clang already vectrozies it with avx512 enabled: https://godbolt.org/z/cnM3986dx . Unfortunately some architectures (e.g skylake) have a very significant latency (uops.info report 15+ cycles), so this makes xxhash ~2x slower. Trivial benchmark shows BM_hash/16M 1.29ms ± 3% vs BM_hash/16M 2.74ms ± 4%. We already disable vectorization for xxhash32 if sse4 is present, so we should probably do the same for xxhash64 and avx512.

[Cyan4973]

We already disable vectorization for xxhash32 if sse4 is present, so we should probably do the same for xxhash64 and avx512.

Indeed,
this seems like the same issue,
and we should probably use a similar solution.

[TocarIP]

Someone asked me about AMD offline so might as well post here. On genoa (Zen4) AVX512 is actually ~30% faster ( 1135871 ns/op vs 885935 ns/op).

[Cyan4973]

OK, so it makes the situation a bit less clear,
since avx512 is now sometimes beneficial, sometimes detrimental.
It probably opens the door towards offering a user choice on this topic.

However, it doesn't remove the question of "what's a good default",
and if I read the situation correctly, it seems that disabling avx512 for xxh64 remains a reasonable default at this point in time.

[TocarIP]

Regression on Intel is bigger than gain on AMD, so disabled by default makes sense. Using the same XXH_ENABLE_AUTOVECTORIZE to allow opt-in is probably the best.

[Cyan4973]

Cyan4973/xxHash#924