[X86] Add isSimple check to the load combine in combineExtractVectorElt.

[craig.topper]

Summary

I think we need to be sure the load isn't volatile before we
duplicate and shrink it.

Noticed while looking at this code for other reasons. So I don't have
a test case. If we agree this is a bug I can try to add one.

[craig.topper]

Rebase. I used the wrong repo so it might have been older than I meant it to be.

[spatel]

Test added with:
d5ebba2aa683

Please rebase and LGTM - this is conservatively correct, and the fact that there were no tests indicates that the missing check was an oversight.

llvm/lib/Target/X86/X86ISelLowering.cpp

   bool LikelyUsedAsVector = any_of(N->uses(), [](SDNode *Use) {
     return Use->getOpcode() == ISD::STORE ||
            Use->getOpcode() == ISD::INSERT_VECTOR_ELT ||
            Use->getOpcode() == ISD::SCALAR_TO_VECTOR;
   });
   auto *LoadVec = dyn_cast<LoadSDNode>(InputVector);
   if (LoadVec && CIdx && ISD::isNormalLoad(LoadVec) && VT.isInteger() &&
       SrcVT.getVectorElementType() == VT && DCI.isAfterLegalizeDAG() &&
-      !LikelyUsedAsVector) {
+      !LikelyUsedAsVector && LoadVec->isSimple()) {
     const TargetLowering &TLI = DAG.getTargetLoweringInfo();
     SDValue NewPtr =
         TLI.getVectorElementPointer(DAG, LoadVec->getBasePtr(), SrcVT, EltIdx);
     unsigned PtrOff = VT.getSizeInBits() * CIdx->getZExtValue() / 8;
     MachinePointerInfo MPI = LoadVec->getPointerInfo().getWithOffset(PtrOff);
     Align Alignment = commonAlignment(LoadVec->getAlign(), PtrOff);
     SDValue Load =
         DAG.getLoad(VT, dl, LoadVec->getChain(), NewPtr, MPI, Alignment,