Faster PoT verification for CPUs that support AVX512F+VAES

[nazar-pc]

Backport of nazar-pc/abundance#260, doubles PoT verification performance on my Zen 4 machine on top of #3551, which itself is ~10% faster than main.

Would be nice for someone to invest time into optimizing verification for Aarch64, there are both generic and low-level optimized examples now.

To be merged after #3551.

Code contributor checklist:

• I have read, understood and followed contributing guide

[nazar-pc]

Looks like Rust toolchain upgrade is needed here, will wait for #3535

[teor2345]

Looks fine to me once compilation is fixed.

[teor2345]

Looks like Rust toolchain upgrade is needed here, will wait for #3535

From the tracking issue it looks like there's a CPU feature missing, and it's still unstable in nightly:
rust-lang/rust#44839

[nazar-pc]

There are more problems with code than this. And this will not be needed once Rust toolchain is upgraded: rust-lang/rust#138940

In fact soon stdarch_x86_avx512 will not be needed either: rust-lang/rust#111137 (comment)

[nazar-pc]

Last commit is neither necessary nor sufficient, just wait for toolchain to be upgraded and it will all light up green and compile successfully.

[teor2345]

Last commit is neither necessary nor sufficient, just wait for toolchain to be upgraded and it will all light up green and compile successfully.

We'll need a later nightly, the current one in PR #3535 does not include stabilisation of avx512
Edit: #3535 (comment)

[nazar-pc]

True, but it is fairly trivial to upgrade there from where #3535 stands

[teor2345]

True, but it is fairly trivial to upgrade there from where #3535 stands

I've asked for the later nightly, we might as well do it all at once.

[nazar-pc]

Pushed another commit that reduces usage of unsafe to the very minimum required by the fact that platform-specific features are used in certain functions. VAES intrinsics are unsafe for unknown reason, so I filed an upstream issue for now.

This should make life easier for auditors.

[nazar-pc]

Backported nazar-pc/abundance#269 too in the last commit, it implements VAES support without AVX512F and even regular AES-NI support for older CPUs, basically covering everything on x86-64 side.

[nazar-pc]

And I ended up implementing optimized version of both proving and verification for aarch64 too (no SVE AES because no intrinsics in standard library yet): nazar-pc/abundance#270

Will submit as a separate PR once this is merged.

[teor2345]

VAES intrinsics are unsafe for unknown reason, so I filed an upstream issue for now.

These have now been marked as safe:
rust-lang/stdarch#1810

[nazar-pc]

Yes, but stdarch git module has not yet been updated in the compiler repository: https://github.com/rust-lang/rust/tree/master/library
Last update was 2 weeks ago.

[teor2345]

Yes, but stdarch git module has not yet been updated in the compiler repository: https://github.com/rust-lang/rust/tree/master/library Last update was 2 weeks ago.

I understand it will take some time to upstream, this was more about any reviews/audits of the "unsafe" code.

[nazar-pc]

That unsafe isn't really critical, I just find it quite pleasant that with experimental SIMD and recent compiler version it is possible to write mostly safe code even when using low-level intrinsics, it is pretty cool.

[teor2345]

This is now ready for a merge with the 2025-05-31 nightly update.

[teor2345]

Thanks!

[nazar-pc]

AVX512 intrinsics were just stabilized in nightly-2025-06-08 and unsafe is no longer needed for them as well.