[llvm] [DAGCombiner] Fold pattern for srl-shl-zext (PR #138290)

[Alexander Peskov via llvm-commits]

================
@@ -10979,6 +10979,39 @@ SDValue DAGCombiner::visitSRL(SDNode *N) {
       return DAG.getNode(ISD::SRL, DL, VT, N0, NewOp1);
   }
 
+  // fold (srl (or x, (shl (zext y), c1)), c1) -> (or (srl x, c1), (zext y))
+  // c1 <= leadingzeros(zext(y))
+  if (N1C && (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND ||
+              N0.getOpcode() == ISD::XOR)) {
+    SDValue lhs = N0.getOperand(0);
+    SDValue rhs = N0.getOperand(1);
+    SDValue shl;
+    SDValue other;
+    if (lhs.getOpcode() == ISD::SHL) {
+      shl = lhs;
+      other = rhs;
+    } else if (rhs.getOpcode() == ISD::SHL) {
+      shl = rhs;
+      other = lhs;
+    }
+    if (shl.getNode()) {
+      if (shl.getOperand(1).getNode() == N1C) {
----------------
apeskov wrote:

Agree. Switched to comparison with N1. The corresponding test is also provided.

https://github.com/llvm/llvm-project/pull/138290