[llvm] [AMDGPU] Constant fold FMAD_FTZ (PR #69443)

[Jay Foad via llvm-commits]

================
@@ -5047,11 +5047,24 @@ SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
     SDValue N2 = N->getOperand(2);
     EVT VT = N->getValueType(0);
 
-    // FMAD_FTZ is a FMAD, but flushing to zero is allowed (not required).
-    // if all operands are constant we can just constant-fold it like a FMAD.
-    if (isa<ConstantFPSDNode>(N0) && isa<ConstantFPSDNode>(N1) &&
-        isa<ConstantFPSDNode>(N2))
-      return DAG.getNode(ISD::FMAD, DL, VT, {N0, N1, N2});
+    // FMAD_FTZ is a FMAD + flush denormals to zero.
+    // We flush the inputs, the intermediate step, and the output.
+    ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
+    ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
+    ConstantFPSDNode *N2CFP = dyn_cast<ConstantFPSDNode>(N2);
+    if (N0CFP && N1CFP && N2CFP) {
+      const auto FTZ = [](const APFloat &V) {
+        return V.isDenormal() ? APFloat(V.getSemantics(), 0) : V;
----------------
jayfoad wrote:

I think mad/mac flushing preserves sign, doesn't it? So -denormal should flush to -0.

https://github.com/llvm/llvm-project/pull/69443