[SVE] Fix bad FixedVectorType cast in simplifyDivRem

[ctetreau]

Summary

simplifyDivRem attempts to walk a VectorType elementwise. Ensure that it
only does so for FixedVectorType

[spatel]

Do you have a regression test with a scalable vector for this pattern?

[ctetreau]

rebase

[ctetreau]

This codepath is tested for scalable vectors by LLVM.Transforms/InstCombine::udiv-pow2-vscale.ll

[ctetreau]

@spatel Are you satisfied with the existing test? May I merge this?

[spatel]

In D81856#2095970, @ctetreau wrote:

@spatel Are you satisfied with the existing test? May I merge this?

I'm confused - if the test is already there, but it doesn't change with this patch, then what does the failure/bug look like?

[ctetreau]

I'm confused - if the test is already there, but it doesn't change with this patch, then what does the failure/bug look like?

The issue that this patch is solving is that this code calls getNumElements through a base VectorType pointer, which is going to be removed soon. getAggregateElement() will only return a value for a scalable vector constant if it is zeroinitializer or undef. Both of these cases are handled before the lines changed by this patch. The test I listed above hits this code path. What currently happens is getAggregateElement returns nullptr for each of 0 .. VTy->getElementCount().Min, and then the block exits. Essentially, this is NFC.

[spatel]

In D81856#2096429, @ctetreau wrote:

I'm confused - if the test is already there, but it doesn't change with this patch, then what does the failure/bug look like?

The issue that this patch is solving is that this code calls getNumElements through a base VectorType pointer, which is going to be removed soon. getAggregateElement() will only return a value for a scalable vector constant if it is zeroinitializer or undef. Both of these cases are handled before the lines changed by this patch. The test I listed above hits this code path. What currently happens is getAggregateElement returns nullptr for each of 0 .. VTy->getElementCount().Min, and then the block exits. Essentially, this is NFC.

OK, thanks for explaining. LGTM.

llvm/lib/Analysis/InstructionSimplify.cpp

if (match(Op1, m_Undef()))
     return Op1;
 
   // X / 0 -> undef
   // X % 0 -> undef
   // We don't need to preserve faults!
   if (match(Op1, m_Zero()))
     return UndefValue::get(Ty);
 
-  // If any element of a constant divisor vector is zero or undef, the whole op
-  // is undef.
+  // If any element of a constant divisor fixed width vector is zero or undef,
+  // the whole op is undef.
   auto *Op1C = dyn_cast<Constant>(Op1);
-  auto *VTy = dyn_cast<VectorType>(Ty);
+  auto *VTy = dyn_cast<FixedVectorType>(Ty);
   if (Op1C && VTy) {
     unsigned NumElts = VTy->getNumElements();
     for (unsigned i = 0; i != NumElts; ++i) {
       Constant *Elt = Op1C->getAggregateElement(i);
       if (Elt && (Elt->isNullValue() || isa<UndefValue>(Elt)))
         return UndefValue::get(Ty);
     }
   }