[llvm] [AMDGPU][SDAG] Legalise v2i32 or/xor/and instructions to make use of 64-bit wide instructions (PR #140694)

[Janek van Oirschot via llvm-commits]

================
@@ -12872,6 +12887,46 @@ SDValue SITargetLowering::performOrCombine(SDNode *N,
     }
   }
 
+  // Detect identity v2i32 OR and replace with identity source node.
+  // Specifically an Or that has operands constructed from the same source node
+  // via extract_vector_elt and build_vector.
+  if (VT == MVT::v2i32) {
+    if (LHS->getOpcode() == ISD::BUILD_VECTOR &&
+        RHS->getOpcode() == ISD::BUILD_VECTOR) {
+      LLVM_DEBUG(dbgs() << "### Performing v2i32 SIISelLowering "
+                           "DAGCombine::CombineOR\n";);
+
+      auto *LC = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
+      auto *RC = dyn_cast<ConstantSDNode>(RHS->getOperand(0));
+
+      if (LC && RC) {
+
+        // Test for and normalise build vectors.
+        if (LHS->getOpcode() == ISD::BUILD_VECTOR &&
+            RHS->getOpcode() == ISD::BUILD_VECTOR &&
+            // Check cast to constantnode here
+            LHS->getConstantOperandVal(1) == 0 &&
+            RHS->getConstantOperandVal(0) == 0) {
----------------
JanekvO wrote:

Can use LC/RC directly instead of converting to ConstantSDNode again within getConstantOperandVal

https://github.com/llvm/llvm-project/pull/140694