[llvm] [InstCombine][X86] Fold blendv(x, y, shuffle(bitcast(sext(m)))) -> select(shuffle(m),x,y) (PR #96882)

[Simon Pilgrim via llvm-commits]

================
@@ -2809,11 +2826,26 @@ X86TTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
         BoolVec->getType()->getScalarSizeInBits() == 1) {
       auto *MaskTy = cast<FixedVectorType>(Mask->getType());
       auto *OpTy = cast<FixedVectorType>(II.getType());
+      unsigned NumMaskElts = MaskTy->getNumElements();
+      unsigned NumOperandElts = OpTy->getNumElements();
+
+      // If we peeked through a shuffle, reapply the shuffle to the bool vector.
+      if (MaskSrc) {
+        unsigned NumMaskSrcElts =
+            cast<FixedVectorType>(MaskSrc->getType())->getNumElements();
+        NumMaskElts = (ShuffleMask.size() * NumMaskElts) / NumMaskSrcElts;
+        // Multiple mask bits maps to the same operand element - bail out.
+        if (NumMaskElts > NumOperandElts)
----------------
RKSimon wrote:

we don't handle NumMaskElts > NumOperandElts (IIRC due to endianess issues) - this is an early out so we don't bother creating an unused shuffle instruction

https://github.com/llvm/llvm-project/pull/96882