[PATCH] D103180: [InstSimplify] Add constant fold for extractelement + splat for scalable vectors

[Sander de Smalen via Phabricator via llvm-commits]

sdesmalen added inline comments.


================
Comment at: llvm/lib/IR/ConstantFold.cpp:916-917
         return CAZ->getElementValue(CIdx->getZExtValue());
+      if (Constant *SplatVal = Val->getSplatValue())
+        return SplatVal;
     }
----------------
CarolineConcatto wrote:
> CarolineConcatto wrote:
> > foad wrote:
> > > You should be able to remove the two lines above handling the CAZ case, because your new code will handle it.
> > a
> Hi @foad 
> I believe we should keep these 2 tests:
> !CIdx->uge(ValSVTy->getMinNumElements())
> because it limits the index to be lower than the minimum width of the scalable vector.
> 
> and it should be only for scalable vector because of this 
>         return CAZ->getElementValue(CIdx->getZExtValue());
> transformation. 
> If we change to be vector type  many other  tests fail. For instance this one:
>  CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
> Transforms/LoopVectorize/X86/gather_scatter.ll
Hi @CarolineConcatto, I think @foad meant removing the lines:

  if (auto *CAZ = dyn_cast<ConstantAggregateZero>(Val))
    return CAZ->getElementValue(CIdx->getZExtValue());

If `Val` is a constant aggregate zero (<=> all zeroes), then this is a splat value that's recognized by `getSplatValue`, so your new lines added in this patch will handle the same case making the above lines redundant.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D103180/new/

https://reviews.llvm.org/D103180