[PATCH] D103820: [X86] Prefer vpmovq2m over vpternlogd + vpcmpgtq

[Craig Topper via Phabricator via llvm-commits]

craig.topper added inline comments.


================
Comment at: llvm/lib/Target/X86/X86ISelLowering.cpp:22802
+  if (SetCCOpcode == ISD::SETGT && ISD::isBuildVectorAllOnes(Op1.getNode())) {
+    SetCCOpcode = ISD::SETGE;
+    Op1 = DAG.getConstant(0, dl, Op1.getValueType());
----------------
davezarzycki wrote:
> craig.topper wrote:
> > I'm a little confused. the vpmovq2m patter is (setgt imm_allzeros, X). Does this SETGE get canonicalized again?
> Oh, sorry about that. I originally made a logic error and converted > -1 to < 0 instead of >= 0; but at that time, I did verify that VPMOVQ2M was generated. You are right that the desired instruction is not generated with SETGE now that the logic error is fixed. In any case, I'd like to update the tests ahead of any code change. Is there a preferred file to put the current suboptimal code gen?
avx512-vec-cmp.ll seems like a decent file to put it in. 


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103820/new/

https://reviews.llvm.org/D103820