[FPEnv][ARM] Don't call mutateStrictFPToFP when lowering

mutateStrictFPToFP can delete the node and replace it with another with the same
value which can later cause problems, and returning the result of
mutateStrictFPToFP doesn't work because SelectionDAGLegalize expects that the
returned value has the same number of results as the original. Instead handle
things by doing the mutation manually.

Differential Revision: https://reviews.llvm.org/D74726

(cherry picked from commit 594a89f)

Author: john-brawn-arm

llvm/lib/Target/ARM/ARMISelLowering.cpp

@@ -5421,7 +5421,12 @@ SDValue ARMTargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) const {
 
   // FIXME: Remove this when we have strict fp instruction selection patterns
   if (IsStrict) {
-    DAG.mutateStrictFPToFP(Op.getNode());
+    SDLoc Loc(Op);
+    SDValue Result =
+        DAG.getNode(Op.getOpcode() == ISD::STRICT_FP_TO_SINT ? ISD::FP_TO_SINT
+                                                             : ISD::FP_TO_UINT,
+                    Loc, Op.getValueType(), SrcVal);
+    return DAG.getMergeValues({Result, Op.getOperand(0)}, Loc);
   }
 
   return Op;
@@ -16384,7 +16389,10 @@ SDValue ARMTargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
   if (SrcSz == 32 && DstSz == 64 && Subtarget->hasFP64()) {
     // FIXME: Remove this when we have strict fp instruction selection patterns
     if (IsStrict) {
-      DAG.mutateStrictFPToFP(Op.getNode());
+      SDLoc Loc(Op);
+      SDValue Result = DAG.getNode(ISD::FP_EXTEND,
+                                   Loc, Op.getValueType(), SrcVal);
+      return DAG.getMergeValues({Result, Op.getOperand(0)}, Loc);
     }
     return Op;
   }

llvm/test/CodeGen/ARM/fp-intrinsics.ll

@@ -72,6 +72,21 @@ define i32 @fptosi_f32(float %x) #0 {
   ret i32 %val
 }
 
+; CHECK-LABEL: fptosi_f32_twice:
+; CHECK-NOSP: bl __aeabi_f2iz
+; CHECK-NOSP: bl __aeabi_f2iz
+; CHECK-SP: vcvt.s32.f32
+; FIXME-CHECK-SP: vcvt.s32.f32
+define void @fptosi_f32_twice(float %arg, i32* %ptr) #0 {
+entry:
+  %conv = call i32 @llvm.experimental.constrained.fptosi.f32(float %arg, metadata !"fpexcept.strict") #0
+  store i32 %conv, i32* %ptr, align 4
+  %conv1 = call i32 @llvm.experimental.constrained.fptosi.f32(float %arg, metadata !"fpexcept.strict") #0
+  %idx = getelementptr inbounds i32, i32* %ptr, i32 1
+  store i32 %conv1, i32* %idx, align 4
+  ret void
+}
+
 ; CHECK-LABEL: fptoui_f32:
 ; CHECK-NOSP: bl __aeabi_f2uiz
 ; FIXME-CHECK-SP: vcvt.u32.f32
@@ -80,6 +95,21 @@ define i32 @fptoui_f32(float %x) #0 {
   ret i32 %val
 }
 
+; CHECK-LABEL: fptoui_f32_twice:
+; CHECK-NOSP: bl __aeabi_f2uiz
+; CHECK-NOSP: bl __aeabi_f2uiz
+; FIXME-CHECK-SP: vcvt.u32.f32
+; FIXME-CHECK-SP: vcvt.u32.f32
+define void @fptoui_f32_twice(float %arg, i32* %ptr) #0 {
+entry:
+  %conv = call i32 @llvm.experimental.constrained.fptoui.f32(float %arg, metadata !"fpexcept.strict") #0
+  store i32 %conv, i32* %ptr, align 4
+  %conv1 = call i32 @llvm.experimental.constrained.fptoui.f32(float %arg, metadata !"fpexcept.strict") #0
+  %idx = getelementptr inbounds i32, i32* %ptr, i32 1
+  store i32 %conv1, i32* %idx, align 4
+  ret void
+}
+
 ; CHECK-LABEL: sqrt_f32:
 ; CHECK-NOSP: bl sqrtf
 ; CHECK-SP: vsqrt.f32
@@ -947,6 +977,21 @@ define double @fpext_f32(float %x) #0 {
   ret double %val
 }
 
+; CHECK-LABEL: fpext_f32_twice:
+; CHECK-NODP: bl __aeabi_f2d
+; CHECK-NODP: bl __aeabi_f2d
+; CHECK-DP: vcvt.f64.f32
+; FIXME-CHECK-DP: vcvt.f64.f32
+define void @fpext_f32_twice(float %arg, double* %ptr) #0 {
+entry:
+  %conv1 = call double @llvm.experimental.constrained.fpext.f64.f32(float %arg, metadata !"fpexcept.strict") #0
+  store double %conv1, double* %ptr, align 8
+  %conv2 = call double @llvm.experimental.constrained.fpext.f64.f32(float %arg, metadata !"fpexcept.strict") #0
+  %idx = getelementptr inbounds double, double* %ptr, i32 1
+  store double %conv2, double* %idx, align 8
+  ret void
+}
+
 ; CHECK-LABEL: sitofp_f32_i32:
 ; CHECK-NOSP: bl __aeabi_i2f
 ; FIXME-CHECK-SP: vcvt.f32.s32