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

[Chris Jackson via llvm-commits]

================
@@ -4654,8 +4654,27 @@ AMDGPUTargetLowering::foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
     if (!AMDGPUTargetLowering::allUsesHaveSourceMods(N.getNode()))
       return SDValue();
 
-    return distributeOpThroughSelect(DCI, LHS.getOpcode(),
-                                     SDLoc(N), Cond, LHS, RHS);
+    // select c, (fneg (f32 bitcast i32 x)), (fneg (f32 bitcast i32 y)) can be
+    // lowered directly to a V_CNDMASK_. So prevent the fneg from being pulled
+    // out in this case. For now I've made the logic as specific to the case as
+    // possible, hopefully this can be relaxed in future.
+    if (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG) {
+      SDValue LHSB = LHS.getOperand(0);
+      SDValue RHSB = RHS.getOperand(0);
+      if (LHSB.getOpcode() == ISD::BITCAST &&
+          RHSB->getOpcode() == ISD::BITCAST) {
+        EVT LHSBOpTy = LHSB->getOperand(0).getValueType();
+        EVT RHSBOpTy = RHSB->getOperand(0).getValueType();
+        if (LHSB.getValueType() == MVT::f32 &&
+            RHSB.getValueType() == MVT::f32 && LHSBOpTy == MVT::i32 &&
+            RHSBOpTy == MVT::i32) {
+          return SDValue();
+        }
+      }
+    }
+
+    return distributeOpThroughSelect(DCI, LHS.getOpcode(), SDLoc(N), Cond, LHS,
+                                     RHS);
----------------
chrisjbris wrote:

Just double-checked. Yes, the work of clang-format.

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