[llvm] [AMDGPU] Optimize rotate instruction selection patterns (PR #143551)

[Matt Arsenault via llvm-commits]

================
@@ -3610,6 +3610,26 @@ bool AMDGPUDAGToDAGISel::SelectSWMMACIndex16(SDValue In, SDValue &Src,
   return true;
 }
 
+bool AMDGPUDAGToDAGISel::SelectImmSub(SDValue In, SDValue &Src,
+                                      SDValue &InvSrc) const {
+  Src = In;
+
+  // Handle constant operands
+  ConstantSDNode *ImmVal = dyn_cast<ConstantSDNode>(In);
+  if (ImmVal)
+    InvSrc = CurDAG->getTargetConstant(32 - ImmVal->getZExtValue(), SDLoc(In),
+                                       MVT::i32);
+  else {
+    // Fallback: generate SUB instruction for non-constant, non-negation cases
----------------
arsenm wrote:

You can let the ordinary pattern do its job, you can fail and let the ordinary select pattern hit 

https://github.com/llvm/llvm-project/pull/143551