[llvm] [AMDGPU] Fold fneg into select source modifiers (PR #201426)

[Matt Arsenault via llvm-commits]

================
@@ -5070,13 +5070,19 @@ AMDGPUTargetLowering::foldFreeOpFromSelect(TargetLowering::DAGCombinerInfo &DCI,
   SDValue RHS = N.getOperand(2);
 
   EVT VT = N.getValueType();
-  if ((LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) ||
-      (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG)) {
+  if (LHS.getOpcode() == ISD::FNEG && RHS.getOpcode() == ISD::FNEG) {
+    if (shouldFoldFNegIntoSrc(N.getNode(), SDValue()) &&
+        shouldFoldFNegIntoSelect(N, LHS.getOperand(0), RHS.getOperand(0)))
+      return SDValue();
+
+    return distributeOpThroughSelect(DCI, ISD::FNEG, SDLoc(N), Cond, LHS, RHS);
+  }
+
+  if (LHS.getOpcode() == ISD::FABS && RHS.getOpcode() == ISD::FABS) {
----------------
arsenm wrote:

The same probably applies to fabs, but best to look into that separately 

https://github.com/llvm/llvm-project/pull/201426