[PATCH] D159250: [X86][RFC] Add new option `-m[no-]evex512` to disable ZMM and 64-bit mask instructions for AVX512 features

[Phoebe Wang via Phabricator via cfe-commits]

pengfei added a comment.

In D159250#4634774 <https://reviews.llvm.org/D159250#4634774>, @jyknight wrote:

> In D159250#4633530 <https://reviews.llvm.org/D159250#4633530>, @pengfei wrote:
>
>> In D159250#4631786 <https://reviews.llvm.org/D159250#4631786>, @RKSimon wrote:
>>
>>> Would it be possible to add function multiversioning tests to ensure the evex512 attribute would work with it?
>>
>> Function multiversioning is orthogonal to evex512 feature.
>>
>> When user uses `-mno-evex512` in command line, all the code generation, including function multiversioning are limited to 256-bit vector and 32-bit mask.
>>
>> User is not suggested to use `avx512xxx,evex512` in function attributes for function multiversioning, because EVEX512 is a software concept and the dispatcher cannot distinguish between `avx512xxx` and `avx512xxx,evex512`.
>
> If the dispatcher is updated to take into account AVX10.1 CPUID, it could distinguish the different hardware support.
>
> That is:
>
> - to check for AVX512xxx with evex512 //enabled//, the dispatcher need only check for the AVX512xxx CPUID bit, since according to the doc, a CPU which implements AVX10.1 with 512-bit register size will also set the corresponding AVX512 CPUID bits. No change there.
> - to check for AVX512xxx with evex512 //disabled//, the dispatcher function should check that either CPUID reports the AVX512xxx bit OR that the CPUID reports AVX10.1 with support for at least 256-bit register size. (But only for the 'AVX512xxx' features which are actually included in AVX10.1, of course).

Let's not to mix evex512 with AVX10.1 512-bit register size enumeration bit. EVEX512 is intended for AVX512xxx only. It's not supposed to use for AVX10. And it conflicts with the functionality of the AVX10.1 bit in same way.
For example, to maintain backward compatibility, EVEX512 is designed to be a default by on feature. That says, if users don't disable EVEX512 in the command line explicitly, and use `avx512xxx` only in function attributes, compiler will attach a `evex512` implicitly. If we map `evex512` to the AVX10.1 bit, the function will never be dispatched on prior-AVX10 targets.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D159250/new/

https://reviews.llvm.org/D159250