[PATCH] D137036: [X86] Enable EVEX GFNI instructions without avx512bw.

Mon Oct 31 00:56:52 PDT 2022

craig.topper added inline comments.

================
Comment at: llvm/test/CodeGen/X86/avx512-gfni-intrinsics.ll:29
+;
+; X86NOBW-LABEL: test_vgf2p8affineinvqb_128:
+; X86NOBW:       # %bb.0:
----------------
pengfei wrote:
> Since we are able to lower the mask version intrinsics, we have 3 choices for the FE support:
> 1. Still disallow Clang intrinsics for AVX512F. This matches with GCC too;
> 2. Allow the mask intrinsics as well. This matches with Intrinsic Guide and SDM somehow;
> 3. And mixed for AVX512F and AVX512BW as in this revision;
> I slightly prefer to 1). I think the intention of the EVEX design is to use the masked instructions directly, so it should always imply AVX512BW in reality. But I'm fine with any way.
I think gcc allows the 128-bit mask with avx512vl only. The 256 and 512 require avx512bw due to needing kmovd and kmovq to have a large enough mask register.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D137036/new/

https://reviews.llvm.org/D137036