[llvm] [NVPTX] Don't propagate `ninf` and `nnan` in `lowerFREM` (PR #147125)

[Matt Arsenault via llvm-commits]

================
@@ -2793,14 +2793,18 @@ static SDValue lowerFREM(SDValue Op, SelectionDAG &DAG,
   EVT Ty = Op.getValueType();
   SDNodeFlags Flags = Op->getFlags();
 
-  SDValue Div = DAG.getNode(ISD::FDIV, DL, Ty, X, Y, Flags);
-  SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, Ty, Div, Flags);
+  // fdiv can still generate inf and nan when nnan and ninf are set.
+  SDNodeFlags NewFlags = Flags;
+  NewFlags.setNoNaNs(false);
+  NewFlags.setNoInfs(false);
+  SDValue Div = DAG.getNode(ISD::FDIV, DL, Ty, X, Y, NewFlags);
+  SDValue Trunc = DAG.getNode(ISD::FTRUNC, DL, Ty, Div, NewFlags);
   SDValue Mul = DAG.getNode(ISD::FMUL, DL, Ty, Trunc, Y,
-                            Flags | SDNodeFlags::AllowContract);
+                            NewFlags | SDNodeFlags::AllowContract);
   SDValue Sub = DAG.getNode(ISD::FSUB, DL, Ty, X, Mul,
-                            Flags | SDNodeFlags::AllowContract);
+                            NewFlags | SDNodeFlags::AllowContract);
 
-  if (AllowUnsafeFPMath || Flags.hasNoInfs())
+  if (AllowUnsafeFPMath || (Flags.hasNoInfs() && Flags.hasApproximateFuncs()))
----------------
arsenm wrote:

You shouldn't touch the afn thing here. I also did not mean this tail section of the function is wrong, I meant the entire implementation of this function is wrong 

https://github.com/llvm/llvm-project/pull/147125