[llvm] [X86] Vectorize non-power-of-two integer division (PR #215076)

[via llvm-commits]

================
@@ -50660,7 +50660,9 @@ static SDValue combineIntDivRem(SDNode *N, SelectionDAG &DAG,
       continue;
     ExtractedElts.setBit(IdxVal.getZExtValue());
   }
-  if (OnlyExtracts && !ExtractedElts.isAllOnes())
+  if (OnlyExtracts &&
+      (isPowerOf2_32(VT.getVectorNumElements()) ||
----------------
Andarwinux wrote:

Why?

What I mean is the sample is supposed to keep vectorization rather than scalarization, which is a positive side effect of this PR, so should add coverage.

https://github.com/llvm/llvm-project/pull/215076