SimplifyLibCalls: Replace fabs libcalls with intrinsics

[arsenm]

Summary

Add missing fabs(fpext) optimzation that worked with the call,
and also fixes it creating a second fpext when there were multiple
uses.

[efriedma]

Canonicalizing @fabs() -> @llvm.fabs() makes sense.

I'd like to see a regression test which specifically checks the fabs/fabsf/fabsl -> llvm.fabs transform. Please include testcases which transform fabsl -> llvm.fabs.f64, fabsl -> llvm.fabs.f80, and fabsl -> llvm.fabs.f128, to make sure they all work as expected.

Please make sure to kill off all the existing optimizations besides this one which check for LibFunc::fabs or the string "fabsf" in a followup.

test/Transforms/InstCombine/double-float-shrink-2.ll

27

"replacable" is a little unclear... I think the point is that the intrinsic does the right thing on all platforms?

test/Transforms/InstCombine/fabs.ll

19–20

This is probably going to fail on Windows... maybe change the test to call llvm.fabs rather than fabs?

[arsenm]

Test adjustments

[arsenm]

Preserve fast math flags

[efriedma]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1642

Do we need to preserve fast-math flags here?

lib/Transforms/Utils/SimplifyLibCalls.cpp

1221

Do we need to preserve fast-math flags here?

test/Transforms/InstCombine/double-float-shrink-2.ll

91

We allow putting fast-math flags on arbitrary calls...? Please update LangRef with an explanation of what this means.

[arsenm]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1642

I don't think fpext supports fast math flags

[arsenm]

test/Transforms/InstCombine/double-float-shrink-2.ll

91

FMF on calls was added in r255555

[arsenm]

test/Transforms/InstCombine/double-float-shrink-2.ll

91

This seems to already be noted in the langref for call

[arsenm]

Preserve fast math flags

[efriedma]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1642

Sorry, I meant fast-math flags on the call.

test/Transforms/InstCombine/double-float-shrink-2.ll

91

Oh, I was looking in the wrong place; it looks like there's a missing note in the section on fast-math flags.

[arsenm]

Minor test improvement

[arsenm]

Anything else?

[efriedma]

I'd still like to see a test which checks the transform fabsl() -> llvm.fabs.f80().

lib/Transforms/InstCombine/InstCombineCalls.cpp

1640

Fast-math flags?

test/Transforms/InstCombine/fabs.ll

19–20

Still probably going to fail on Windows, where fabsf() doesn't exist.

[arsenm]

test/Transforms/InstCombine/fabs.ll

19–20

Do you mean the intrinsic won't work? I would expect codegen to promote it

[arsenm]

Test fp80

test/Transforms/InstCombine/fabs.ll

19–20

It looks like a fabs instruction is emitted for the intrinsic

[efriedma]

lib/Transforms/InstCombine/InstCombineCalls.cpp

1640

Fast-math flags?

test/Transforms/InstCombine/fabs.ll

19–20

No... I mean we call TLI.setUnavailable(LibFunc::fabsf); on Windows, so it isn't a library call, so we'll never call optimizeFabs.

[arsenm]

test/Transforms/InstCombine/fabs.ll

19–20

I thought the intent there was to not recognize fabsf as a special builtin function since it isn't available, i.e. fabsf could be a user function

[efriedma]

test/Transforms/InstCombine/fabs.ll

19–20

Yes, that's working as intended; the only problem is that you didn't specify a triple for this test, so it will inherit the triple from the host, therefore on a Windows host the CHECK line won't match.

[arsenm]

Preserve fast math flags in another place

[arsenm]

Add triple to test

[efriedma]

LGTM.

[arsenm]

r292172

include/llvm/IR/PatternMatch.h

 }
 
 /// \brief Matches SIToFP.
 template <typename OpTy>
 inline CastClass_match<OpTy, Instruction::SIToFP> m_SIToFP(const OpTy &Op) {
   return CastClass_match<OpTy, Instruction::SIToFP>(Op);
 }
 
+/// \brief Matches FPTrunc
+template <typename OpTy>
+inline CastClass_match<OpTy, Instruction::FPTrunc> m_FPTrunc(const OpTy &Op) {
+  return CastClass_match<OpTy, Instruction::FPTrunc>(Op);
+}
+
+/// \brief Matches FPExt
+template <typename OpTy>
+inline CastClass_match<OpTy, Instruction::FPExt> m_FPExt(const OpTy &Op) {
+  return CastClass_match<OpTy, Instruction::FPExt>(Op);
+}
+
 //===----------------------------------------------------------------------===//
 // Matchers for unary operators
 //
 
 template <typename LHS_t> struct not_match {
   LHS_t L;
 
   not_match(const LHS_t &LHS) : L(LHS) {}

lib/Transforms/InstCombine/InstCombineCalls.cpp

case Intrinsic::fabs: {
     Constant *LHS, *RHS;
     if (match(II->getArgOperand(0),
               m_Select(m_Value(Cond), m_Constant(LHS), m_Constant(RHS)))) {
       CallInst *Call0 = Builder->CreateCall(II->getCalledFunction(), {LHS});
       CallInst *Call1 = Builder->CreateCall(II->getCalledFunction(), {RHS});
       return SelectInst::Create(Cond, Call0, Call1);
     }
 
+    Value *ExtSrc;
+    if (match(II->getArgOperand(0), m_FPExt(m_Value(ExtSrc))) &&
+        II->getArgOperand(0)->hasOneUse()) {
+      // fabs (fpext x) -> fpext (fabs x)
+      Value *F = Intrinsic::getDeclaration(II->getModule(), Intrinsic::fabs,
+                                           { ExtSrc->getType() });
+      CallInst *NewFabs = Builder->CreateCall(F, ExtSrc);
+      NewFabs->copyFastMathFlags(II);
+      NewFabs->takeName(II);
+      return new FPExtInst(NewFabs, II->getType());
+    }
+
     break;
   }
   case Intrinsic::cos:
   case Intrinsic::amdgcn_cos: {
     Value *SrcSrc;
     Value *Src = II->getArgOperand(0);
     if (match(Src, m_FNeg(m_Value(SrcSrc))) ||
         match(Src, m_Intrinsic<Intrinsic::fabs>(m_Value(SrcSrc)))) {

lib/Transforms/InstCombine/InstCombineCasts.cpp

if (SI &&
     Value *RHSTrunc = Builder->CreateFPTrunc(SI->getOperand(2),
                                              CI.getType());
     return SelectInst::Create(SI->getOperand(0), LHSTrunc, RHSTrunc);
   }
 
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI.getOperand(0));
   if (II) {
     switch (II->getIntrinsicID()) {
-      default: break;
-      case Intrinsic::fabs: {
-        // (fptrunc (fabs x)) -> (fabs (fptrunc x))
-        Value *InnerTrunc = Builder->CreateFPTrunc(II->getArgOperand(0),
-                                                   CI.getType());
-        Type *IntrinsicType[] = { CI.getType() };
-        Function *Overload = Intrinsic::getDeclaration(
-            CI.getModule(), II->getIntrinsicID(), IntrinsicType);
 
-        SmallVector<OperandBundleDef, 1> OpBundles;
-        II->getOperandBundlesAsDefs(OpBundles);
 
-        Value *Args[] = { InnerTrunc };
-        return CallInst::Create(Overload, Args, OpBundles, II->getName());
-      }
+      CallInst *NewCI =  CallInst::Create(Overload, Args,
+                                          OpBundles, II->getName());
+      NewCI->copyFastMathFlags(II);
+      return NewCI;
     }
   }
 
   return nullptr;
 }
 
 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
   return commonCastTransforms(CI);

lib/Transforms/Utils/SimplifyLibCalls.cpp

if (LdExpArg) {
       return CI;
     }
   }
   return Ret;
 }
 
 Value *LibCallSimplifier::optimizeFabs(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
-  StringRef Name = Callee->getName();
-  if (Name == "fabs" && hasFloatVersion(Name))
-    return optimizeUnaryDoubleFP(CI, B, false);
+  IRBuilder<>::FastMathFlagGuard Guard(B);
+  B.setFastMathFlags(CI->getFastMathFlags());
 
-  return nullptr;
+  // fabs/fabsf -> llvm.fabs.*
+  Value *F = Intrinsic::getDeclaration(Callee->getParent(), Intrinsic::fabs,
+                                       CI->getType());
+  Value *NewCall = B.CreateCall(F, { CI->getArgOperand(0) });
+  NewCall->takeName(CI);
+  return NewCall;
 }
 
 Value *LibCallSimplifier::optimizeFMinFMax(CallInst *CI, IRBuilder<> &B) {
   Function *Callee = CI->getCalledFunction();
   // If we can shrink the call to a float function rather than a double
   // function, do that first.
   StringRef Name = Callee->getName();
   if ((Name == "fmin" || Name == "fmax") && hasFloatVersion(Name))
Value *LibCallSimplifier::optimizeCall(CallInst *CI) {
   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
     if (!isCallingConvC)
       return nullptr;
     switch (II->getIntrinsicID()) {
     case Intrinsic::pow:
       return optimizePow(CI, Builder);
     case Intrinsic::exp2:
       return optimizeExp2(CI, Builder);
-    case Intrinsic::fabs:
-      return optimizeFabs(CI, Builder);
     case Intrinsic::log:
       return optimizeLog(CI, Builder);
     case Intrinsic::sqrt:
       return optimizeSqrt(CI, Builder);
     // TODO: Use foldMallocMemset() with memset intrinsic.
     default:
       return nullptr;
     }

test/Transforms/InstCombine/double-float-shrink-2.ll

 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-linux" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-win32" | FileCheck -check-prefix=DONT-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-win32" | FileCheck -check-prefix=C89-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "i386-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "x86_64-pc-mingw32" | FileCheck -check-prefix=DO-SIMPLIFY %s
 ; RUN: opt < %s -instcombine -S -mtriple "sparc-sun-solaris" | FileCheck -check-prefix=DO-SIMPLIFY %s
 
 ; DO-SIMPLIFY: call float @floorf(
 ; DO-SIMPLIFY: call float @ceilf(
 ; DO-SIMPLIFY: call float @roundf(
 ; DO-SIMPLIFY: call float @nearbyintf(
 ; DO-SIMPLIFY: call float @truncf(
-; DO-SIMPLIFY: call float @fabsf(
+; DO-SIMPLIFY: call float @llvm.fabs.f32(
+; DO-SIMPLIFY: call float fast @llvm.fabs.f32(
 
 ; C89-SIMPLIFY: call float @floorf(
 ; C89-SIMPLIFY: call float @ceilf(
 ; C89-SIMPLIFY: call double @round(
 ; C89-SIMPLIFY: call double @nearbyint(
 
 ; DONT-SIMPLIFY: call double @floor(
 ; DONT-SIMPLIFY: call double @ceil(
 ; DONT-SIMPLIFY: call double @round(
 ; DONT-SIMPLIFY: call double @nearbyint(
 ; DONT-SIMPLIFY: call double @trunc(
-; DONT-SIMPLIFY: call double @fabs(
+
+; This is replaced with the intrinsic, which does the right thing on
+; all platforms.
+; DONT-SIMPLIFY: call float @llvm.fabs.f32(
 
 declare double @floor(double)
 declare double @ceil(double)
 declare double @round(double)
 declare double @nearbyint(double)
 declare double @trunc(double)
 declare double @fabs(double)
+declare double @llvm.fabs.f64(double)
 
 define float @test_floor(float %C) {
   %D = fpext float %C to double
   ; --> floorf
   %E = call double @floor(double %D)
   %F = fptrunc double %E to float
   ret float %F
 }

 
 define float @test_fabs(float %C) {
   %D = fpext float %C to double
   ; --> fabsf
   %E = call double @fabs(double %D)
   %F = fptrunc double %E to float
   ret float %F
 }
+
+; Make sure fast math flags are preserved
+define float @test_fabs_fast(float %C) {
+  %D = fpext float %C to double
+  ; --> fabsf
+  %E = call fast double @fabs(double %D)
+  %F = fptrunc double %E to float
+  ret float %F
+}

test/Transforms/InstCombine/fabs-libcall.ll

+; RUN: opt -S -mtriple=i686-apple-macosx -instcombine %s | FileCheck %s
+
+declare x86_fp80 @fabsl(x86_fp80)
+
+; CHECK-LABEL: @replace_fabs_call_f80(
+; CHECK-NEXT: %fabsl = call x86_fp80 @llvm.fabs.f80(x86_fp80 %x)
+; CHECK-NEXT: ret x86_fp80 %fabsl
+define x86_fp80 @replace_fabs_call_f80(x86_fp80 %x) {
+  %fabsl = tail call x86_fp80 @fabsl(x86_fp80 %x)
+  ret x86_fp80 %fabsl
+
+}
+
+; CHECK-LABEL: @fmf_replace_fabs_call_f80(
+; CHECK-NEXT: %fabsl = call nnan x86_fp80 @llvm.fabs.f80(x86_fp80 %x)
+; CHECK-NEXT: ret x86_fp80 %fabsl
+define x86_fp80 @fmf_replace_fabs_call_f80(x86_fp80 %x) {
+  %fabsl = tail call nnan x86_fp80 @fabsl(x86_fp80 %x)
+  ret x86_fp80 %fabsl
+}
+

test/Transforms/InstCombine/fabs.ll

 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
-; Make sure all library calls are eliminated when the input is known positive.
+; Make sure libcalls are replaced with intrinsic calls.
+
+declare float @llvm.fabs.f32(float)
+declare double @llvm.fabs.f64(double)
+declare fp128 @llvm.fabs.f128(fp128)
 
 declare float @fabsf(float)
 declare double @fabs(double)
 declare fp128 @fabsl(fp128)
 declare float @llvm.fma.f32(float, float, float)
 declare float @llvm.fmuladd.f32(float, float, float)
 
-define float @square_fabs_call_f32(float %x) {
-  %mul = fmul float %x, %x
-  %fabsf = tail call float @fabsf(float %mul)
+define float @replace_fabs_call_f32(float %x) {
+  %fabsf = tail call float @fabsf(float %x)
   ret float %fabsf
 
-; CHECK-LABEL: square_fabs_call_f32(
-; CHECK-NEXT: %mul = fmul float %x, %x
-; CHECK-NEXT: %fabsf = tail call float @fabsf(float %mul)
+; CHECK-LABEL: @replace_fabs_call_f32(
+; CHECK-NEXT: %fabsf = call float @llvm.fabs.f32(float %x)
 ; CHECK-NEXT: ret float %fabsf
 }
 
-define double @square_fabs_call_f64(double %x) {
-  %mul = fmul double %x, %x
-  %fabs = tail call double @fabs(double %mul)
+define double @replace_fabs_call_f64(double %x) {
+  %fabs = tail call double @fabs(double %x)
   ret double %fabs
 
-; CHECK-LABEL: square_fabs_call_f64(
-; CHECK-NEXT: %mul = fmul double %x, %x
-; CHECK-NEXT: %fabs = tail call double @fabs(double %mul)
+; CHECK-LABEL: @replace_fabs_call_f64(
+; CHECK-NEXT: %fabs = call double @llvm.fabs.f64(double %x)
 ; CHECK-NEXT: ret double %fabs
 }
 
-define fp128 @square_fabs_call_f128(fp128 %x) {
-  %mul = fmul fp128 %x, %x
-  %fabsl = tail call fp128 @fabsl(fp128 %mul)
+define fp128 @replace_fabs_call_f128(fp128 %x) {
+  %fabsl = tail call fp128 @fabsl(fp128 %x)
   ret fp128 %fabsl
 
-; CHECK-LABEL: square_fabs_call_f128(
-; CHECK-NEXT: %mul = fmul fp128 %x, %x
-; CHECK-NEXT: %fabsl = tail call fp128 @fabsl(fp128 %mul)
+; CHECK-LABEL: replace_fabs_call_f128(
+; CHECK-NEXT: %fabsl = call fp128 @llvm.fabs.f128(fp128 %x)
 ; CHECK-NEXT: ret fp128 %fabsl
 }
 
+; Make sure fast math flags are preserved when replacing the libcall.
+define float @fmf_replace_fabs_call_f32(float %x) {
+  %fabsf = tail call nnan float @fabsf(float %x)
+  ret float %fabsf
+
+; CHECK-LABEL: @fmf_replace_fabs_call_f32(
+; CHECK-NEXT: %fabsf = call nnan float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: ret float %fabsf
+}
+
 ; Make sure all intrinsic calls are eliminated when the input is known
 ; positive.
 
-declare float @llvm.fabs.f32(float)
-declare double @llvm.fabs.f64(double)
-declare fp128 @llvm.fabs.f128(fp128)
-
 ; The fabs cannot be eliminated because %x may be a NaN
 define float @square_fabs_intrinsic_f32(float %x) {
   %mul = fmul float %x, %x
   %fabsf = tail call float @llvm.fabs.f32(float %mul)
   ret float %fabsf
 
 ; CHECK-LABEL: square_fabs_intrinsic_f32(
 ; CHECK-NEXT: %mul = fmul float %x, %x

 define float @square_fabs_shrink_call1(float %x) {
   %ext = fpext float %x to double
   %sq = fmul double %ext, %ext
   %fabs = call double @fabs(double %sq)
   %trunc = fptrunc double %fabs to float
   ret float %trunc
 
 ; CHECK-LABEL: square_fabs_shrink_call1(
-; CHECK-NEXT: %ext = fpext float %x to double
-; CHECK-NEXT: %sq = fmul double %ext, %ext
-; CHECK-NEXT: call double @fabs(double %sq)
-; CHECK-NEXT: %trunc = fptrunc double %fabs to float
+; CHECK-NEXT: fmul float %x, %x
+; CHECK-NEXT: %trunc = call float @llvm.fabs.f32(float
 ; CHECK-NEXT: ret float %trunc
 }
 
 define float @square_fabs_shrink_call2(float %x) {
   %sq = fmul float %x, %x
   %ext = fpext float %sq to double
   %fabs = call double @fabs(double %ext)
   %trunc = fptrunc double %fabs to float
   ret float %trunc
 
 ; CHECK-LABEL: square_fabs_shrink_call2(
 ; CHECK-NEXT: %sq = fmul float %x, %x
-; CHECK-NEXT: %fabsf = call float @fabsf(float %sq)
-; CHECK-NEXT: ret float %fabsf
+; CHECK-NEXT: %trunc = call float @llvm.fabs.f32(float %sq)
+; CHECK-NEXT: ret float %trunc
 }
 
 ; CHECK-LABEL: @fabs_select_constant_negative_positive(
 ; CHECK: %fabs = select i1 %cmp, float 1.000000e+00, float 2.000000e+00
 ; CHECK-NEXT: ret float %fabs
 define float @fabs_select_constant_negative_positive(i32 %c) {
   %cmp = icmp eq i32 %c, 0
   %select = select i1 %cmp, float -1.0, float 2.0
define float @square_nnan_fmuladd_fabs_intrinsic_f32(float %x) {
   %fmuladd = call nnan float @llvm.fmuladd.f32(float %x, float %x, float 1.0)
   %fabsf = call float @llvm.fabs.f32(float %fmuladd)
   ret float %fabsf
 
 ; CHECK-LABEL: @square_nnan_fmuladd_fabs_intrinsic_f32(
 ; CHECK-NEXT: %fmuladd = call nnan float @llvm.fmuladd.f32(float %x, float %x, float 1.000000e+00)
 ; CHECK-NEXT: ret float %fmuladd
 }
+
+; Don't introduce a second fpext
+; CHECK-LABEL: @multi_use_fabs_fpext(
+; CHECK: %fpext = fpext float %x to double
+; CHECK-NEXT: %fabs = call double @llvm.fabs.f64(double %fpext)
+; CHECK-NEXT: store volatile double %fpext, double* undef, align 8
+; CHECK-NEXT: ret double %fabs
+define double @multi_use_fabs_fpext(float %x) {
+  %fpext = fpext float %x to double
+  %fabs = call double @llvm.fabs.f64(double %fpext)
+  store volatile double %fpext, double* undef
+  ret double %fabs
+}

test/Transforms/InstCombine/float-shrink-compare.ll

 define i32 @test2(float %x, float %y) nounwind uwtable {
   %1 = fpext float %x to double
   %2 = call double @fabs(double %1) nounwind readnone
   %3 = fpext float %y to double
   %4 = fcmp oeq double %2, %3
   %5 = zext i1 %4 to i32
   ret i32 %5
 ; CHECK-LABEL: @test2(
-; CHECK-NEXT: %fabsf = call float @fabsf(float %x)
-; CHECK-NEXT: fcmp oeq float %fabsf, %y
+; CHECK-NEXT: [[FABS:%[0-9]+]] = call float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float [[FABS]], %y
+}
+
+define i32 @fmf_test2(float %x, float %y) nounwind uwtable {
+  %1 = fpext float %x to double
+  %2 = call nnan double @fabs(double %1) nounwind readnone
+  %3 = fpext float %y to double
+  %4 = fcmp oeq double %2, %3
+  %5 = zext i1 %4 to i32
+  ret i32 %5
+; CHECK-LABEL: @fmf_test2(
+; CHECK-NEXT: [[FABS:%[0-9]+]] = call nnan float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float [[FABS]], %y
 }
 
 define i32 @test3(float %x, float %y) nounwind uwtable {
   %1 = fpext float %x to double
   %2 = call double @floor(double %1) nounwind readnone
   %3 = fpext float %y to double
   %4 = fcmp oeq double %2, %3
   %5 = zext i1 %4 to i32
define i32 @test8(float %x, float %y) nounwind uwtable {
   %5 = zext i1 %4 to i32
   ret i32 %5
 ; CHECK-LABEL: @test8(
 ; CHECK-NEXT: %ceilf = call float @ceilf(float %x)
 ; CHECK-NEXT: fcmp oeq float %ceilf, %y
 }
 
 define i32 @test9(float %x, float %y) nounwind uwtable {
-  %1 = fpext float %y to double
-  %2 = fpext float %x to double
-  %3 = call double @fabs(double %2) nounwind readnone
-  %4 = fcmp oeq double %1, %3
-  %5 = zext i1 %4 to i32
-  ret i32 %5
+  %x.ext = fpext float %x to double
+  %y.ext = fpext float %y to double
+  %fabs = call double @fabs(double %x.ext) nounwind readnone
+  %cmp = fcmp oeq double %y.ext, %fabs
+  %cmp.ext = zext i1 %cmp to i32
+  ret i32 %cmp.ext
 ; CHECK-LABEL: @test9(
-; CHECK-NEXT: %fabsf = call float @fabsf(float %x)
-; CHECK-NEXT: fcmp oeq float %fabsf, %y
+; CHECK-NEXT: %fabs = call float @llvm.fabs.f32(float %x)
+; CHECK-NEXT: fcmp oeq float %fabs, %y
 }
 
 define i32 @test10(float %x, float %y) nounwind uwtable {
   %1 = fpext float %y to double
   %2 = fpext float %x to double
   %3 = call double @floor(double %2) nounwind readnone
   %4 = fcmp oeq double %1, %3
   %5 = zext i1 %4 to i32

test/Transforms/InstCombine/pow-1.ll

 }
 
 ; Check pow(x, 0.5) -> fabs(sqrt(x)), where x != -infinity.
 
 define float @test_simplify7(float %x) {
 ; CHECK-LABEL: @test_simplify7(
   %retval = call float @powf(float %x, float 0.5)
 ; CHECK-NEXT: [[SQRTF:%[a-z0-9]+]] = call float @sqrtf(float %x) [[NUW_RO:#[0-9]+]]
-; CHECK-NEXT: [[FABSF:%[a-z0-9]+]] = call float @fabsf(float [[SQRTF]]) [[NUW_RO]]
+; CHECK-NEXT: [[FABSF:%[a-z0-9]+]] = call float @llvm.fabs.f32(float [[SQRTF]])
 ; CHECK-NEXT: [[FCMP:%[a-z0-9]+]] = fcmp oeq float %x, 0xFFF0000000000000
 ; CHECK-NEXT: [[SELECT:%[a-z0-9]+]] = select i1 [[FCMP]], float 0x7FF0000000000000, float [[FABSF]]
   ret float %retval
 ; CHECK-NEXT: ret float [[SELECT]]
 }
 
 define double @test_simplify8(double %x) {
 ; CHECK-LABEL: @test_simplify8(
   %retval = call double @pow(double %x, double 0.5)
 ; CHECK-NEXT: [[SQRT:%[a-z0-9]+]] = call double @sqrt(double %x) [[NUW_RO]]
-; CHECK-NEXT: [[FABS:%[a-z0-9]+]] = call double @fabs(double [[SQRT]]) [[NUW_RO]]
+; CHECK-NEXT: [[FABS:%[a-z0-9]+]] = call double @llvm.fabs.f64(double [[SQRT]])
 ; CHECK-NEXT: [[FCMP:%[a-z0-9]+]] = fcmp oeq double %x, 0xFFF0000000000000
 ; CHECK-NEXT: [[SELECT:%[a-z0-9]+]] = select i1 [[FCMP]], double 0x7FF0000000000000, double [[FABS]]
   ret double %retval
 ; CHECK-NEXT: ret double [[SELECT]]
 }
 
 ; Check pow(-infinity, 0.5) -> +infinity.
 

 ; CHECK-NEXT: ret double [[RECIPROCAL]]
 }
 
 declare double @llvm.pow.f64(double %Val, double %Power)
 define double @test_simplify17(double %x) {
 ; CHECK-LABEL: @test_simplify17(
   %retval = call double @llvm.pow.f64(double %x, double 0.5)
 ; CHECK-NEXT: [[SQRT:%[a-z0-9]+]] = call double @sqrt(double %x)
-; CHECK-NEXT: [[FABS:%[a-z0-9]+]] = call double @fabs(double [[SQRT]])
+; CHECK-NEXT: [[FABS:%[a-z0-9]+]] = call double @llvm.fabs.f64(double [[SQRT]])
 ; CHECK-NEXT: [[FCMP:%[a-z0-9]+]] = fcmp oeq double %x, 0xFFF0000000000000
 ; CHECK-NEXT: [[SELECT:%[a-z0-9]+]] = select i1 [[FCMP]], double 0x7FF0000000000000, double [[FABS]]
   ret double %retval
 ; CHECK-NEXT: ret double [[SELECT]]
 }
 
 ; Check pow(10.0, x) -> __exp10(x) on OS X 10.9+ and iOS 7.0+.
 

test/Transforms/InstCombine/win-math.ll

 ; WIN32-LABEL: @float_powsqrt(
 ; WIN32-NOT: float @sqrtf
 ; WIN32: float @powf
 ; WIN64-LABEL: @float_powsqrt(
 ; WIN64-NOT: float @sqrtf
 ; WIN64: float @powf
 ; MINGW32-LABEL: @float_powsqrt(
 ; MINGW32: float @sqrtf
-; MINGW32: float @fabsf
+; MINGW32: float @llvm.fabs.f32
 ; MINGW32-NOT: float @powf
 ; MINGW64-LABEL: @float_powsqrt(
 ; MINGW64: float @sqrtf
-; MINGW64: float @fabsf
+; MINGW64: float @llvm.fabs.f32(
 ; MINGW64-NOT: float @powf
     %1 = call float @powf(float %x, float 0.5)
     ret float %1
 }

test/Transforms/InstCombine/zero-point-zero-add.ll

 ;
   %Y = fadd double %X, 0.0          ;; Should be a single add x, 0.0
   %Z = fadd double %Y, 0.0
   ret double %Z
 }
 
 define double @test1(double %X) {
 ; CHECK-LABEL: @test1(
-; CHECK-NEXT:    [[Y:%.*]] = call double @fabs(double %X)
+; CHECK-NEXT:    [[Y:%.*]] = call double @llvm.fabs.f64(double %X)
 ; CHECK-NEXT:    ret double [[Y]]
 ;
   %Y = call double @fabs(double %X)
   %Z = fadd double %Y, 0.0
   ret double %Z
 }