Stabilize the avx512 target features by sayantn · Pull Request #138940 · rust-lang/rust (original) (raw)

I discovered a related mistake in LLVM. Basically

• avx512f - GCC and Intel implies only avx2, LLVM also does fma and f16c
• avx512fp16 - GCC and Intel implies only avx512bw, LLVM also does avx512dq and avx512vl

I have filed llvm/llvm-project#136209

For avx512f, this requirement is probably justified (see this comment). We just need to ensure that fma and f16c are required for avx512f in std_detect. I have sent a PR (rust-lang/stdarch#1781).

But we need to settle the avx512fp16 business.

This has the potential to generate invalid code (even in safe Rust, e.g https://godbolt.org/z/d3M674nY6 - here only avx512fp16 is enabled, but LLVM is generating vaddph.128, which is not available without avx512vl). Even if #140389 passes though, this invalid codegen will persist, as it is an LLVM issue. So essentially we have 2 options

• Don't adhere to Intel's spec. We say that avx512fp16 implies avx512vl in rustc_target/target_feature.rs, and only enable it in std_detect when avx512vl is available. This is safe, with the caveat that if Intel releases a CPU with avx512fp16, but not avx512vl (currently none exist) Rust will just not detect avx512fp16 there.
• Delay the stabilization of avx512fp16 for now. We wait for the upstream LLVM issue to be fixed (which I believe will take a long time).

I don't think we have to delay the stabilization of any other avx512 target feature.

cc @rust-lang/lang
cc @rust-lang/wg-llvm