[FPEnv][InstSimplify] Enable more folds for constrained fsub

[kpn]

Summary

Currently there are optimizations for the fsub instruction that do not fire for a constrained fsub. Add some of these optimizations.

Optimizations that require teaching the IR matchers about the constrained intrinsics are for a future patch.

[sepavloff]

llvm/lib/Analysis/InstructionSimplify.cpp

4963

This transformation is valid only if FMF.noNaNs() is set, because it is incorrect if X==SNaN. Similar for other cases below.

[kpn]

Update for review comments. Instructions that are "maytrap" where a SNaN may be seen must not be eliminated.

I've also added tests for transforms that won't fire due to a lack of IR matcher support. We're going through codepaths we weren't before so we're testing what we weren't before.

[spatel]

Please add the test file as a preliminary commit. Then update here to show the code diffs. Also as part of this patch, apply comments on the tests that are intentionally not changing and/or are expected to change later (TODO).

llvm/lib/Analysis/InstructionSimplify.cpp

4963–4965

This clause is going to be repeated in many places. Make it or its inverse a helper function and give it a name (canIgnoreSNAN or mustPreserveSNAN?).

[kpn]

Update for review comments.

[lebedev.ri]

This review may be stuck/dead, consider abandoning if no longer relevant.
Removing myself as reviewer in attempt to clean dashboard.

llvm/lib/Analysis/InstructionSimplify.cpp

SimplifyFSubInst(Value *Op0, Value *Op1, FastMathFlags FMF,
                  RoundingMode Rounding = RoundingMode::NearestTiesToEven) {
   if (isDefaultFPEnvironment(ExBehavior, Rounding))
     if (Constant *C = foldOrCommuteConstant(Instruction::FSub, Op0, Op1, Q))
       return C;
 
   if (Constant *C = simplifyFPOp({Op0, Op1}, FMF, Q, ExBehavior, Rounding))
     return C;
 
-  if (!isDefaultFPEnvironment(ExBehavior, Rounding))
-    return nullptr;
-
   // fsub X, +0 ==> X
+  if (ExBehavior != fp::ebStrict || FMF.noNaNs())
     if (match(Op1, m_PosZeroFP()))
       return Op0;
 
   // fsub X, -0 ==> X, when we know X is not -0
+  if (ExBehavior != fp::ebStrict || FMF.noNaNs())
     if (match(Op1, m_NegZeroFP()) &&
         (FMF.noSignedZeros() || CannotBeNegativeZero(Op0, Q.TLI)))
     return Op0;
 
   // fsub -0.0, (fsub -0.0, X) ==> X
   // fsub -0.0, (fneg X) ==> X
   Value *X;
+  if (ExBehavior != fp::ebStrict || FMF.noNaNs())
     if (match(Op0, m_NegZeroFP()) &&
         match(Op1, m_FNeg(m_Value(X))))
     return X;
 
+  if (!isDefaultFPEnvironment(ExBehavior, Rounding))
+    return nullptr;
+
   // fsub 0.0, (fsub 0.0, X) ==> X if signed zeros are ignored.
   // fsub 0.0, (fneg X) ==> X if signed zeros are ignored.
   if (FMF.noSignedZeros() && match(Op0, m_AnyZeroFP()) &&
       (match(Op1, m_FSub(m_AnyZeroFP(), m_Value(X))) ||
        match(Op1, m_FNeg(m_Value(X)))))
     return X;
 
   // fsub nnan x, x ==> 0.0

llvm/test/Transforms/InstSimplify/strictfp-fsub.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -instsimplify -S | FileCheck %s
+
+;
+; constrained fsub
+;
+
+;
+; fsub X, +0 ==> X
+;
+
+define float @fsub_x_p0_defaultenv(float %a) #0 {
+; CHECK-LABEL: @fsub_x_p0_defaultenv(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float %a, float 0.0, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_vec_x_p0_defaultenv(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_vec_x_p0_defaultenv(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float><float 0.0, float 0.0>, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_x_p0_ebmaytrap(float %a) #0 {
+; CHECK-LABEL: @fsub_x_p0_ebmaytrap(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float %a, float 0.0, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_vec_x_p0_ebmaytrap(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_vec_x_p0_ebmaytrap(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float><float 0.0, float 0.0>, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_x_p0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float 0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0:[0-9]+]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float %a, float 0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_vec_x_p0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_vec_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[A:%.*]], <2 x float> zeroinitializer, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float><float 0.0, float 0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_nnan_x_p0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_nnan_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call nnan float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float 0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[A]]
+;
+  %ret = call nnan float @llvm.experimental.constrained.fsub.f32(float %a, float 0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_vec_nnan_x_p0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_vec_nnan_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[A:%.*]], <2 x float> zeroinitializer, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[A]]
+;
+  %ret = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float><float 0.0, float 0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+; Test with a fast math flag set but that flag is not "nnan".
+define float @fsub_ninf_x_p0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_ninf_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call ninf float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float 0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %ret = call ninf float @llvm.experimental.constrained.fsub.f32(float %a, float 0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+; Test with a fast math flag set but that flag is not "nnan".
+define <2 x float> @fsub_vec_ninf_x_p0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_vec_ninf_x_p0_ebstrict(
+; CHECK-NEXT:    [[RET:%.*]] = call ninf <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[A:%.*]], <2 x float> zeroinitializer, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %ret = call ninf <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float><float 0.0, float 0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+;
+; fsub X, -0 ==> X, when we know X is not -0
+; (fast math flag: nsz)
+;
+
+define float @fold_fsub_nsz_x_n0_defaultenv(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_nsz_x_n0_defaultenv(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %sub = call nsz float @llvm.experimental.constrained.fsub.f32(float %a, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_nsz_x_n0_defaultenv(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_nsz_x_n0_defaultenv(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %sub = call nsz <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_nsz_x_n0_ebmaytrap(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_nsz_x_n0_ebmaytrap(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %sub = call nsz float @llvm.experimental.constrained.fsub.f32(float %a, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_nsz_x_n0_ebmaytrap(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_nsz_x_n0_ebmaytrap(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %sub = call nsz <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_nsz_x_n0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_nsz_x_n0_ebstrict(
+; CHECK-NEXT:    [[SUB:%.*]] = call nsz float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float -0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[SUB]]
+;
+  %sub = call nsz float @llvm.experimental.constrained.fsub.f32(float %a, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_nsz_x_n0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_nsz_x_n0_ebstrict(
+; CHECK-NEXT:    [[SUB:%.*]] = call nsz <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[A:%.*]], <2 x float> <float -0.000000e+00, float -0.000000e+00>, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[SUB]]
+;
+  %sub = call nsz <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_nsz_nnan_x_n0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_nsz_nnan_x_n0_ebstrict(
+; CHECK-NEXT:    [[SUB:%.*]] = call nnan nsz float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float -0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[A]]
+;
+  %sub = call nsz nnan float @llvm.experimental.constrained.fsub.f32(float %a, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_nsz_nnan_x_n0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_nsz_nnan_x_n0_ebstrict(
+; CHECK-NEXT:    [[SUB:%.*]] = call nnan nsz <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[A:%.*]], <2 x float> <float -0.000000e+00, float -0.000000e+00>, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[A]]
+;
+  %sub = call nsz nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %a, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %sub
+}
+
+;
+; fsub X, -0 ==> X, when we know X is not -0
+; (No "nsz" flags)
+;
+
+define float @fold_fsub_fabs_x_n0_defaultenv(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_fabs_x_n0_defaultenv(
+; CHECK-NEXT:    [[ABSA:%.*]] = call float @llvm.fabs.f32(float [[A:%.*]])
+; CHECK-NEXT:    ret float [[ABSA]]
+;
+  %absa = call float @llvm.fabs.f32(float %a)
+  %sub = call float @llvm.experimental.constrained.fsub.f32(float %absa, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_fabs_x_n0_defaultenv(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_fabs_x_n0_defaultenv(
+; CHECK-NEXT:    [[ABSA:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[A:%.*]])
+; CHECK-NEXT:    ret <2 x float> [[ABSA]]
+;
+  %absa = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
+  %sub = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %absa, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_fabs_x_n0_ebmaytrap(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_fabs_x_n0_ebmaytrap(
+; CHECK-NEXT:    [[ABSA:%.*]] = call float @llvm.fabs.f32(float [[A:%.*]])
+; CHECK-NEXT:    ret float [[ABSA]]
+;
+  %absa = call float @llvm.fabs.f32(float %a)
+  %sub = call float @llvm.experimental.constrained.fsub.f32(float %absa, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_fabs_x_n0_ebmaytrap(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_fabs_x_n0_ebmaytrap(
+; CHECK-NEXT:    [[ABSA:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[A:%.*]])
+; CHECK-NEXT:    ret <2 x float> [[ABSA]]
+;
+  %absa = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
+  %sub = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %absa, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_fabs_x_n0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_fabs_x_n0_ebstrict(
+; CHECK-NEXT:    [[SUB:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float [[A:%.*]], float -0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[SUB]]
+;
+  %absa = call float @llvm.fabs.f32(float %a)
+  %sub = call float @llvm.experimental.constrained.fsub.f32(float %a, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_fabs_x_n0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_fabs_x_n0_ebstrict(
+; CHECK-NEXT:    [[ABSA:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[A:%.*]])
+; CHECK-NEXT:    [[SUB:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[ABSA]], <2 x float> <float -0.000000e+00, float -0.000000e+00>, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[SUB]]
+;
+  %absa = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
+  %sub = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %absa, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %sub
+}
+
+define float @fold_fsub_fabs_nnan_x_n0_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fold_fsub_fabs_nnan_x_n0_ebstrict(
+; CHECK-NEXT:    [[ABSA:%.*]] = call float @llvm.fabs.f32(float [[A:%.*]])
+; CHECK-NEXT:    [[SUB:%.*]] = call nnan float @llvm.experimental.constrained.fsub.f32(float [[ABSA]], float -0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[ABSA]]
+;
+  %absa = call float @llvm.fabs.f32(float %a)
+  %sub = call nnan float @llvm.experimental.constrained.fsub.f32(float %absa, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %sub
+}
+
+define <2 x float> @fold_fsub_vec_fabs_nnan_x_n0_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_fabs_nnan_x_n0_ebstrict(
+; CHECK-NEXT:    [[ABSA:%.*]] = call <2 x float> @llvm.fabs.v2f32(<2 x float> [[A:%.*]])
+; CHECK-NEXT:    [[SUB:%.*]] = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[ABSA]], <2 x float> <float -0.000000e+00, float -0.000000e+00>, metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[ABSA]]
+;
+  %absa = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
+  %sub = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %absa, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %sub
+}
+
+; TODO: This can't fire currently due to missing constrained support:
+define float @fold_fsub_sitofp_x_n0_defaultenv(i32 %a) #0 {
+; CHECK-LABEL: @fold_fsub_sitofp_x_n0_defaultenv(
+; CHECK-NEXT:    [[FPA:%.*]] = call float @llvm.experimental.constrained.sitofp.f32.i32(i32 [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    [[SUB:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float [[FPA]], float -0.000000e+00, metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[SUB]]
+;
+  %fpa = call float @llvm.experimental.constrained.sitofp.f32.i32(i32 %a, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  %sub = call float @llvm.experimental.constrained.fsub.f32(float %fpa, float -0.0, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %sub
+}
+
+; TODO: This can't fire currently due to missing constrained support:
+define <2 x float> @fold_fsub_vec_sitofp_x_n0_defaultenv(<2 x i32> %a) #0 {
+; CHECK-LABEL: @fold_fsub_vec_sitofp_x_n0_defaultenv(
+; CHECK-NEXT:    [[FPA:%.*]] = call <2 x float> @llvm.experimental.constrained.sitofp.v2f32.v2i32(<2 x i32> [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    [[SUB:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> [[FPA]], <2 x float> <float -0.000000e+00, float -0.000000e+00>, metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[SUB]]
+;
+  %fpa = call <2 x float> @llvm.experimental.constrained.sitofp.v2f32.v2i32(<2 x i32> %a, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  %sub = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> %fpa, <2 x float> <float -0.0, float -0.0>, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %sub
+}
+
+;
+; fsub -0.0, (fneg X) ==> X
+;
+
+define float @fsub_fneg_n0_fnX_defaultenv(float %a) #0 {
+; CHECK-LABEL: @fsub_fneg_n0_fnX_defaultenv(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %nega = fneg float %a
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fneg_vec_n0_fnX_defaultenv(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fneg_vec_n0_fnX_defaultenv(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %nega = fneg <2 x float> %a
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fneg_n0_fnX_ebmaytrap(float %a) #0 {
+; CHECK-LABEL: @fsub_fneg_n0_fnX_ebmaytrap(
+; CHECK-NEXT:    ret float [[A:%.*]]
+;
+  %nega = fneg float %a
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fneg_vec_n0_fnX_ebmaytrap(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fneg_vec_n0_fnX_ebmaytrap(
+; CHECK-NEXT:    ret <2 x float> [[A:%.*]]
+;
+  %nega = fneg <2 x float> %a
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fneg_n0_fnX_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_fneg_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = fneg float [[A:%.*]]
+; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %nega = fneg float %a
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fneg_vec_n0_fnX_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fneg_vec_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = fneg <2 x float> [[A:%.*]]
+; CHECK-NEXT:    [[RET:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %nega = fneg <2 x float> %a
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fneg_nnan_n0_fnX_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_fneg_nnan_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = fneg float [[A:%.*]]
+; CHECK-NEXT:    [[RET:%.*]] = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[A]]
+;
+  %nega = fneg float %a
+  %ret = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fneg_vec_nnan_n0_fnX_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fneg_vec_nnan_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = fneg <2 x float> [[A:%.*]]
+; CHECK-NEXT:    [[RET:%.*]] = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[A]]
+;
+  %nega = fneg <2 x float> %a
+  %ret = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+;
+; fsub -0.0, (fsub -0.0, X) ==> X
+; TODO: These generally don't fire because m_FNeg doesn't know about
+; TODO: the constrained intrinsics.
+;
+
+define float @fsub_fsub_n0_fnX_defaultenv(float %a) #0 {
+; CHECK-LABEL: @fsub_fsub_n0_fnX_defaultenv(
+; CHECK-NEXT:    [[NEGA:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %nega = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %a, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fsub_vec_n0_fnX_defaultenv(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fsub_vec_n0_fnX_defaultenv(
+; CHECK-NEXT:    [[NEGA:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %nega = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %a, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.ignore") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fsub_n0_fnX_ebmaytrap(float %a) #0 {
+; CHECK-LABEL: @fsub_fsub_n0_fnX_ebmaytrap(
+; CHECK-NEXT:    [[NEGA:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %nega = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %a, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fsub_vec_n0_fnX_ebmaytrap(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fsub_vec_n0_fnX_ebmaytrap(
+; CHECK-NEXT:    [[NEGA:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.maytrap") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %nega = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %a, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.maytrap") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fsub_n0_fnX_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_fsub_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %nega = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %a, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  %ret = call float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fsub_vec_n0_fnX_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fsub_vec_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %nega = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %a, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  %ret = call <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+define float @fsub_fsub_nnan_n0_fnX_ebstrict(float %a) #0 {
+; CHECK-LABEL: @fsub_fsub_nnan_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.000000e+00, float [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret float [[RET]]
+;
+  %nega = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.0, float %a, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  %ret = call nnan float @llvm.experimental.constrained.fsub.f32(float -0.0, float %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret float %ret
+}
+
+define <2 x float> @fsub_fsub_vec_nnan_n0_fnX_ebstrict(<2 x float> %a) #0 {
+; CHECK-LABEL: @fsub_fsub_vec_nnan_n0_fnX_ebstrict(
+; CHECK-NEXT:    [[NEGA:%.*]] = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[A:%.*]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    [[RET:%.*]] = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.000000e+00, float -0.000000e+00>, <2 x float> [[NEGA]], metadata !"round.tonearest", metadata !"fpexcept.strict") #[[ATTR0]]
+; CHECK-NEXT:    ret <2 x float> [[RET]]
+;
+  %nega = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %a, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  %ret = call nnan <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float> <float -0.0, float -0.0>, <2 x float> %nega, metadata !"round.tonearest", metadata !"fpexcept.strict") #0
+  ret <2 x float> %ret
+}
+
+declare float @llvm.fabs.f32(float)
+declare <2 x float> @llvm.fabs.v2f32(<2 x float>)
+
+declare float @llvm.experimental.constrained.fsub.f32(float, float, metadata, metadata) #0
+declare <2 x float> @llvm.experimental.constrained.fsub.v2f32(<2 x float>, <2 x float>, metadata, metadata) #0
+
+declare float @llvm.experimental.constrained.sitofp.f32.i32(i32, metadata, metadata) #0
+declare <2 x float> @llvm.experimental.constrained.sitofp.v2f32.v2i32(<2 x i32>, metadata, metadata) #0
+
+attributes #0 = { strictfp }