Index: llvm/include/llvm/Analysis/ValueTracking.h =================================================================== --- llvm/include/llvm/Analysis/ValueTracking.h +++ llvm/include/llvm/Analysis/ValueTracking.h @@ -293,6 +293,9 @@ /// floating-point mode for the function interprets denormals as zero. bool isKnownNeverLogicalZero(const Function &F, Type *Ty) const; + /// Return true if it's know this can never be interpreted as a negative zero. + bool isKnownNeverLogicalNegZero(const Function &F, Type *Ty) const; + static constexpr FPClassTest OrderedLessThanZeroMask = fcNegSubnormal | fcNegNormal | fcNegInf; static constexpr FPClassTest OrderedGreaterThanZeroMask = Index: llvm/lib/Analysis/ValueTracking.cpp =================================================================== --- llvm/lib/Analysis/ValueTracking.cpp +++ llvm/lib/Analysis/ValueTracking.cpp @@ -4296,11 +4296,24 @@ return F.getDenormalMode(Ty->getFltSemantics()).Input == DenormalMode::IEEE; } +static bool inputDenormalIsIEEEOrPosZero(const Function &F, const Type *Ty) { + Ty = Ty->getScalarType(); + DenormalMode Mode = F.getDenormalMode(Ty->getFltSemantics()); + return Mode.Input == DenormalMode::IEEE || + Mode.Input == DenormalMode::PositiveZero; +} + bool KnownFPClass::isKnownNeverLogicalZero(const Function &F, Type *Ty) const { return isKnownNeverZero() && (isKnownNeverSubnormal() || inputDenormalIsIEEE(F, Ty)); } +bool KnownFPClass::isKnownNeverLogicalNegZero(const Function &F, + Type *Ty) const { + return isKnownNeverNegZero() && + (isKnownNeverNegSubnormal() || inputDenormalIsIEEEOrPosZero(F, Ty)); +} + /// Returns a pair of values, which if passed to llvm.is.fpclass, returns the /// same result as an fcmp with the given operands. std::pair llvm::fcmpToClassTest(FCmpInst::Predicate Pred, @@ -4714,6 +4727,39 @@ Known.knownNot(fcNegative); break; } + case Intrinsic::sqrt: { + KnownFPClass KnownSrc; + FPClassTest InterestedSrcs = InterestedClasses; + if (InterestedClasses & fcNan) + InterestedSrcs |= OrderedLessThanZeroMask; + + computeKnownFPClass(II->getArgOperand(0), DemandedElts, + InterestedSrcs, KnownSrc, Depth + 1, Q, TLI); + + if (KnownSrc.isKnownNeverPosInfinity()) + Known.knownNot(fcPosInf); + if (KnownSrc.isKnownNever(fcSNan)) + Known.knownNot(fcSNan); + + // Any negative value besides -0 returns a nan. + if (KnownSrc.isKnownNeverNaN() && + KnownSrc.cannotBeOrderedLessThanZero()) + Known.knownNot(fcNan); + + // The only negative value that can be returned is -0 for -0 inputs. + Known.knownNot(fcNegInf | fcNegSubnormal | fcNegNormal); + + // If the input denormal mode could be PreserveSign, a negative + // subnormal input could produce a negative zero output. + if (KnownSrc.isKnownNeverLogicalNegZero(*II->getFunction(), + II->getType())) { + Known.knownNot(fcNegZero); + if (KnownSrc.isKnownNeverNaN()) + Known.SignBit = false; + } + + break; + } case Intrinsic::sin: case Intrinsic::cos: { // Return NaN on infinite inputs. Index: llvm/test/Transforms/Attributor/nofpclass-sqrt.ll =================================================================== --- llvm/test/Transforms/Attributor/nofpclass-sqrt.ll +++ llvm/test/Transforms/Attributor/nofpclass-sqrt.ll @@ -6,9 +6,9 @@ declare float @llvm.experimental.constrained.sqrt.f32(float, metadata, metadata) define float @ret_sqrt(float %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt +; CHECK-LABEL: define nofpclass(ninf nsub nnorm) float @ret_sqrt ; CHECK-SAME: (float [[ARG0:%.*]]) #[[ATTR2:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10:[0-9]+]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(ninf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10:[0-9]+]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -16,9 +16,9 @@ } define float @ret_sqrt_noinf(float nofpclass(inf) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_noinf +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_noinf ; CHECK-SAME: (float nofpclass(inf) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -26,9 +26,9 @@ } define float @ret_sqrt_nopinf(float nofpclass(pinf) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_nopinf +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_nopinf ; CHECK-SAME: (float nofpclass(pinf) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -36,9 +36,9 @@ } define float @ret_sqrt_noninf(float nofpclass(ninf) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_noninf +; CHECK-LABEL: define nofpclass(ninf nsub nnorm) float @ret_sqrt_noninf ; CHECK-SAME: (float nofpclass(ninf) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(ninf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -46,9 +46,9 @@ } define float @ret_sqrt_nonan(float nofpclass(nan) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_nonan +; CHECK-LABEL: define nofpclass(snan ninf nsub nnorm) float @ret_sqrt_nonan ; CHECK-SAME: (float nofpclass(nan) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan ninf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -56,9 +56,9 @@ } define float @ret_sqrt_nonan_noinf(float nofpclass(nan inf) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_nonan_noinf +; CHECK-LABEL: define nofpclass(snan inf nsub nnorm) float @ret_sqrt_nonan_noinf ; CHECK-SAME: (float nofpclass(nan inf) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -66,9 +66,9 @@ } define float @ret_sqrt_nonan_noinf_nozero(float nofpclass(nan inf zero) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_nonan_noinf_nozero +; CHECK-LABEL: define nofpclass(snan inf nzero nsub nnorm) float @ret_sqrt_nonan_noinf_nozero ; CHECK-SAME: (float nofpclass(nan inf zero) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -76,9 +76,9 @@ } define float @ret_sqrt_noinf_nozero(float nofpclass(inf zero) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_noinf_nozero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_noinf_nozero ; CHECK-SAME: (float nofpclass(inf zero) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -86,9 +86,9 @@ } define float @ret_sqrt_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #0 { -; CHECK-LABEL: define float @ret_sqrt_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR2]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -96,10 +96,10 @@ } define float @ret_sqrt_positive_source(i32 %arg) #0 { -; CHECK-LABEL: define float @ret_sqrt_positive_source +; CHECK-LABEL: define nofpclass(nan inf nzero nsub nnorm) float @ret_sqrt_positive_source ; CHECK-SAME: (i32 [[ARG:%.*]]) #[[ATTR2]] { ; CHECK-NEXT: [[UITOFP:%.*]] = uitofp i32 [[ARG]] to float -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[UITOFP]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(nan inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[UITOFP]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %uitofp = uitofp i32 %arg to float @@ -109,10 +109,10 @@ ; Could produce a nan because we don't know if the multiply is negative. define float @ret_sqrt_unknown_sign(float nofpclass(nan) %arg0, float nofpclass(nan) %arg1) #0 { -; CHECK-LABEL: define float @ret_sqrt_unknown_sign +; CHECK-LABEL: define nofpclass(snan ninf nsub nnorm) float @ret_sqrt_unknown_sign ; CHECK-SAME: (float nofpclass(nan) [[ARG0:%.*]], float nofpclass(nan) [[ARG1:%.*]]) #[[ATTR2]] { ; CHECK-NEXT: [[UNKNOWN_SIGN_NOT_NAN:%.*]] = fmul nnan float [[ARG0]], [[ARG1]] -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[UNKNOWN_SIGN_NOT_NAN]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(snan ninf nsub nnorm) float @llvm.sqrt.f32(float [[UNKNOWN_SIGN_NOT_NAN]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %unknown.sign.not.nan = fmul nnan float %arg0, %arg1 @@ -121,9 +121,9 @@ } define float @ret_sqrt_daz_noinf_nozero(float nofpclass(inf zero) %arg0) #1 { -; CHECK-LABEL: define float @ret_sqrt_daz_noinf_nozero +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_daz_noinf_nozero ; CHECK-SAME: (float nofpclass(inf zero) [[ARG0:%.*]]) #[[ATTR3:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -131,9 +131,9 @@ } define <2 x float> @ret_sqrt_daz_noinf_nozero_v2f32(<2 x float> nofpclass(inf zero) %arg0) #1 { -; CHECK-LABEL: define <2 x float> @ret_sqrt_daz_noinf_nozero_v2f32 +; CHECK-LABEL: define nofpclass(inf nsub nnorm) <2 x float> @ret_sqrt_daz_noinf_nozero_v2f32 ; CHECK-SAME: (<2 x float> nofpclass(inf zero) [[ARG0:%.*]]) #[[ATTR3]] { -; CHECK-NEXT: [[CALL:%.*]] = call <2 x float> @llvm.sqrt.v2f32(<2 x float> [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) <2 x float> @llvm.sqrt.v2f32(<2 x float> [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret <2 x float> [[CALL]] ; %call = call <2 x float> @llvm.sqrt.v2f32(<2 x float> %arg0) @@ -141,9 +141,9 @@ } define float @ret_sqrt_daz_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #1 { -; CHECK-LABEL: define float @ret_sqrt_daz_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_daz_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR3]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -151,9 +151,9 @@ } define float @ret_sqrt_dapz_noinf_nozero(float nofpclass(inf zero) %arg0) #2 { -; CHECK-LABEL: define float @ret_sqrt_dapz_noinf_nozero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_dapz_noinf_nozero ; CHECK-SAME: (float nofpclass(inf zero) [[ARG0:%.*]]) #[[ATTR4:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -161,9 +161,9 @@ } define float @ret_sqrt_dapz_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #2 { -; CHECK-LABEL: define float @ret_sqrt_dapz_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_dapz_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR4]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -171,9 +171,9 @@ } define float @ret_sqrt_dynamic_noinf_nozero(float nofpclass(inf zero) %arg0) #3 { -; CHECK-LABEL: define float @ret_sqrt_dynamic_noinf_nozero +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_dynamic_noinf_nozero ; CHECK-SAME: (float nofpclass(inf zero) [[ARG0:%.*]]) #[[ATTR5:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -181,9 +181,9 @@ } define float @ret_sqrt_dynamic_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #3 { -; CHECK-LABEL: define float @ret_sqrt_dynamic_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nsub nnorm) float @ret_sqrt_dynamic_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR5]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -191,9 +191,9 @@ } define float @ret_sqrt_ftz_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #4 { -; CHECK-LABEL: define float @ret_sqrt_ftz_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_ftz_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR6:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -201,9 +201,9 @@ } define float @ret_sqrt_ftpz_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #5 { -; CHECK-LABEL: define float @ret_sqrt_ftpz_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_ftpz_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR7:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) @@ -211,9 +211,9 @@ } define float @ret_sqrt_ftz_dynamic_noinf_nonegzero(float nofpclass(inf nzero) %arg0) #6 { -; CHECK-LABEL: define float @ret_sqrt_ftz_dynamic_noinf_nonegzero +; CHECK-LABEL: define nofpclass(inf nzero nsub nnorm) float @ret_sqrt_ftz_dynamic_noinf_nonegzero ; CHECK-SAME: (float nofpclass(inf nzero) [[ARG0:%.*]]) #[[ATTR8:[0-9]+]] { -; CHECK-NEXT: [[CALL:%.*]] = call float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] +; CHECK-NEXT: [[CALL:%.*]] = call nofpclass(inf nzero nsub nnorm) float @llvm.sqrt.f32(float [[ARG0]]) #[[ATTR10]] ; CHECK-NEXT: ret float [[CALL]] ; %call = call float @llvm.sqrt.f32(float %arg0) Index: llvm/unittests/Analysis/ValueTrackingTest.cpp =================================================================== --- llvm/unittests/Analysis/ValueTrackingTest.cpp +++ llvm/unittests/Analysis/ValueTrackingTest.cpp @@ -1512,10 +1512,10 @@ TEST_F(ComputeKnownFPClassTest, CopySignNInfSrc0_NegSign) { parseAssembly( - "declare float @llvm.sqrt.f32(float)\n" + "declare float @llvm.log.f32(float)\n" "declare float @llvm.copysign.f32(float, float)\n" "define float @test(float %arg0, float %arg1) {\n" - " %ninf = call ninf float @llvm.sqrt.f32(float %arg0)" + " %ninf = call ninf float @llvm.log.f32(float %arg0)" " %A = call float @llvm.copysign.f32(float %ninf, float -1.0)" " ret float %A\n" "}\n");