[AArch64] Emit vector FP cmp when LE is used with fast-math

Reviewed By: paulwalker-arm

Differential Revision: https://reviews.llvm.org/D130093

(cherry-picked from bf268a0)

Author: c-rhodes

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -11697,6 +11697,9 @@ SDValue DAGCombiner::foldSextSetcc(SDNode *N) {
   EVT N00VT = N00.getValueType();
   SDLoc DL(N);
 
+  // Propagate fast-math-flags.
+  SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags());
+
   // On some architectures (such as SSE/NEON/etc) the SETCC result type is
   // the same size as the compared operands. Try to optimize sext(setcc())
   // if this is the case.
@@ -12230,6 +12233,9 @@ SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
     return V;
 
   if (N0.getOpcode() == ISD::SETCC) {
+    // Propagate fast-math-flags.
+    SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags());
+
     // Only do this before legalize for now.
     if (!LegalOperations && VT.isVector() &&
         N0.getValueType().getVectorElementType() == MVT::i1) {
@@ -12417,6 +12423,9 @@ SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
   }
 
   if (N0.getOpcode() == ISD::SETCC) {
+    // Propagate fast-math-flags.
+    SelectionDAG::FlagInserter FlagsInserter(DAG, N0->getFlags());
+
     // For vectors:
     // aext(setcc) -> vsetcc
     // aext(setcc) -> truncate(vsetcc)

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -12087,6 +12087,11 @@ static SDValue EmitVectorComparison(SDValue LHS, SDValue RHS,
       if (IsZero)
         return DAG.getNode(AArch64ISD::FCMGTz, dl, VT, LHS);
       return DAG.getNode(AArch64ISD::FCMGT, dl, VT, LHS, RHS);
+    case AArch64CC::LE:
+      if (!NoNans)
+        return SDValue();
+      // If we ignore NaNs then we can use to the LS implementation.
+      LLVM_FALLTHROUGH;
     case AArch64CC::LS:
       if (IsZero)
         return DAG.getNode(AArch64ISD::FCMLEz, dl, VT, LHS);
@@ -12192,7 +12197,7 @@ SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
   bool ShouldInvert;
   changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
 
-  bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath;
+  bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath || Op->getFlags().hasNoNaNs();
   SDValue Cmp =
       EmitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, dl, DAG);
   if (!Cmp.getNode())