[llvm] [InstCombine] [X86] pblendvb intrinsics must be replaced by select when possible (PR #137322)

[Simon Pilgrim via llvm-commits]

================
@@ -801,6 +802,58 @@ bool VectorCombine::foldInsExtBinop(Instruction &I) {
   return true;
 }
 
+bool VectorCombine::foldBitOpOfBitcasts(Instruction &I) {
+  // Match: bitop(bitcast(x), bitcast(y)) -> bitcast(bitop(x, y))
+  auto *BinOp = dyn_cast<BinaryOperator>(&I);
+  if (!BinOp || !BinOp->isBitwiseLogicOp())
+    return false;
+
+  Value *LHS = BinOp->getOperand(0);
+  Value *RHS = BinOp->getOperand(1);
+
+  // Both operands must be bitcasts
+  auto *LHSCast = dyn_cast<BitCastInst>(LHS);
+  auto *RHSCast = dyn_cast<BitCastInst>(RHS);
+  if (!LHSCast || !RHSCast)
+    return false;
+
+  Value *LHSSrc = LHSCast->getOperand(0);
+  Value *RHSSrc = RHSCast->getOperand(0);
+
+  // Source types must match
+  if (LHSSrc->getType() != RHSSrc->getType())
+    return false;
+
+  // Only handle vector types
+  auto *SrcVecTy = dyn_cast<FixedVectorType>(LHSSrc->getType());
+  auto *DstVecTy = dyn_cast<FixedVectorType>(I.getType());
+  if (!SrcVecTy || !DstVecTy)
+    return false;
+
+  // Same total bit width
+  if (SrcVecTy->getPrimitiveSizeInBits() != DstVecTy->getPrimitiveSizeInBits())
+    return false;
+
+  // Cost check: prefer operations on narrower element types
----------------
RKSimon wrote:

Should probably replace this with a real cost check and not use elt bitwidths - OldCost = bitlogic + 2*bitcasts, NewCost = bitlogic + bitcast, and ensure NewCost<= OldCost - see other VectorCombine examples

https://github.com/llvm/llvm-project/pull/137322