[llvm] [AMDGPU] Correctly restore FP mode in FDIV32 lowering (PR #66346)

[Matt Arsenault via llvm-commits]

================
@@ -4723,10 +4722,23 @@ bool AMDGPULegalizerInfo::legalizeFDIV32(MachineInstr &MI,
                        .setMIFlags(Flags);
   auto NegDivScale0 = B.buildFNeg(S32, DenominatorScaled, Flags);
 
+  const bool PreservesDenormals = Mode.FP32Denormals == DenormalMode::getIEEE();
+  const bool HasDynamicDenormals =
+      (Mode.FP32Denormals.Input == DenormalMode::Dynamic) ||
+      (Mode.FP32Denormals.Output == DenormalMode::Dynamic);
+
   // FIXME: Doesn't correctly model the FP mode switch, and the FP operations
----------------
arsenm wrote:

Still has the fixme?

https://github.com/llvm/llvm-project/pull/66346