[PATCH] D81327: [X86] Teach combineBitcastvxi1 to prefer movmsk on avx512 in more cases

[Craig Topper via Phabricator via llvm-commits]

craig.topper marked an inline comment as done.
craig.topper added inline comments.


================
Comment at: llvm/test/CodeGen/X86/avx512bwvl-intrinsics-upgrade.ll:6410
 ; CHECK-NEXT:    ret{{[l|q]}} # encoding: [0xc3]
     %res = call i16 @llvm.x86.avx512.cvtb2mask.128(<16 x i8> %x0)
     ret i16 %res
----------------
RKSimon wrote:
> I don't recognise what (c/c++) intrinsic this would have come from - is this likely to be a problem? Same for the other *-intrinsics-upgrade.ll cases
I believe this is _mm256_movepi8_mask which is the intrinsic for vpmovb2m. The intrinsic itself is defined to return unsigned short which introduced the bitcast. If it got used with another intrinsic that new how to bitcast back those bitcasts would cancel.

But I guess we have a bunch of intrinsics that take scalar integer as arguments for mask and introduce a bitcast to vXi1 in intrinsic lowering. Which is too late to cancel the bitcast that would form MOVMSK. Maybe we need a combine to reverse MOVMSK? I guess I need to do some experiments.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D81327/new/

https://reviews.llvm.org/D81327