[llvm] [X86] Fold AND(Y, XOR(X, SUB(0, X))) to ANDN(Y, BLSMSK(X)) (PR #128348)

[Simon Pilgrim via llvm-commits]

================
@@ -51045,6 +51045,41 @@ static SDValue combineBMILogicOp(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+/// Fold AND(Y, XOR(X, NEG(X))) -> ANDN(Y, BLSMSK(X)) if BMI is available.
+static SDValue combineAndXorSubWithBMI(SDValue Op, SDValue OtherOp, SDLoc DL,
+                                       SelectionDAG &DAG,
+                                       const X86Subtarget &Subtarget) {
+  EVT VT = Op.getValueType();
+  // Make sure this node is a candidate for BMI instructions.
+  if (!Subtarget.hasBMI() || !VT.isScalarInteger() ||
+      (VT != MVT::i32 && VT != MVT::i64))
+    return SDValue();
+
+  if (Op.getOpcode() != ISD::XOR || !Op.hasOneUse())
----------------
RKSimon wrote:

You can probably remove all the commutation/swapping by using SDPatternMatch matches

https://github.com/llvm/llvm-project/pull/128348