[PATCH] D67874: Prefer AVX512 memcpy when applicable

Sun Sep 22 12:33:57 PDT 2019

RKSimon added a reviewer: RKSimon.
RKSimon added inline comments.

================
Comment at: llvm/test/CodeGen/X86/memset-nonzero.ll:512
+; AVX512F-zmm-NEXT:  vpbroadcastd    %eax, %zmm0
+; AVX512-ymm-NEXT:   vmovdqu %ymm0, 96(%rdi)
+; AVX512-ymm-NEXT:   vmovdqu %ymm0, 64(%rdi)
----------------
craig.topper wrote:
> davezarzycki wrote:
> > craig.topper wrote:
> > > This doesn't look like it was generated by the update_llc_test_checks script.
> > How so? I tried running update_llc_test_checks against the test and nothing changed.
> I've never seen the script mix prefixes like this. As far as I know it always creates a block with the same prefix separated by a blank line from the other prefixes.
Something's definitely gone wrong here, it might be that you've not set the  --llc-binary args correctly? I'd recommend manually deleting the all the 'AVX512' checks from all these cases and then regenerating to see what happens.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D67874/new/

https://reviews.llvm.org/D67874