Skip to content

[ConstantFPRange] Add support for mul/div #163063

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Oct 13, 2025
+262 −0
Merged

[ConstantFPRange] Add support for mul/div #163063

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions llvm/include/llvm/IR/ConstantFPRange.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -231,6 +231,15 @@ class [[nodiscard]] ConstantFPRange {
/// from a subtraction of a value in this range and a value in \p Other.
LLVM_ABI ConstantFPRange sub(const ConstantFPRange &Other) const;

/// Return a new range representing the possible values resulting
/// from a multiplication of a value in this range and a value in \p Other.
LLVM_ABI ConstantFPRange mul(const ConstantFPRange &Other) const;

/// Return a new range representing the possible values resulting
/// from a division of a value in this range and a value in
/// \p Other.
LLVM_ABI ConstantFPRange div(const ConstantFPRange &Other) const;

/// Flush denormal values to zero according to the specified mode.
/// For dynamic mode, we return the union of all possible results.
LLVM_ABI void flushDenormals(DenormalMode::DenormalModeKind Mode);
Expand Down
144 changes: 144 additions & 0 deletions llvm/lib/IR/ConstantFPRange.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -528,3 +528,147 @@ void ConstantFPRange::flushDenormals(DenormalMode::DenormalModeKind Mode) {
Lower = minnum(Lower, APFloat::getZero(Sem, ZeroLowerNegative));
Upper = maxnum(Upper, APFloat::getZero(Sem, ZeroUpperNegative));
}

/// Represent a contiguous range of values sharing the same sign.
struct SameSignRange {
bool HasZero;
bool HasNonZero;
bool HasInf;
// The lower and upper bounds of the range (inclusive).
// The sign is dropped and infinities are excluded.
std::optional<std::pair<APFloat, APFloat>> FinitePart;

explicit SameSignRange(const APFloat &Lower, const APFloat &Upper)
: HasZero(Lower.isZero()), HasNonZero(!Upper.isZero()),
HasInf(Upper.isInfinity()) {
assert(!Lower.isNegative() && !Upper.isNegative() &&
"The sign should be dropped.");
assert(strictCompare(Lower, Upper) != APFloat::cmpGreaterThan &&
"Empty set.");
if (!Lower.isInfinity())
FinitePart = {Lower,
HasInf ? APFloat::getLargest(Lower.getSemantics()) : Upper};
}
};

/// Split the range into positive and negative components.
static void splitPosNeg(const APFloat &Lower, const APFloat &Upper,
std::optional<SameSignRange> &NegPart,
std::optional<SameSignRange> &PosPart) {
assert(strictCompare(Lower, Upper) != APFloat::cmpGreaterThan &&
"Non-NaN part is empty.");
if (Lower.isNegative() == Upper.isNegative()) {
if (Lower.isNegative())
NegPart = SameSignRange{abs(Upper), abs(Lower)};
else
PosPart = SameSignRange{Lower, Upper};
return;
}
auto &Sem = Lower.getSemantics();
NegPart = SameSignRange{APFloat::getZero(Sem), abs(Lower)};
PosPart = SameSignRange{APFloat::getZero(Sem), Upper};
}

ConstantFPRange ConstantFPRange::mul(const ConstantFPRange &Other) const {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At some point should probably thread the rounding mode arguments through

auto &Sem = getSemantics();
bool ResMayBeQNaN = ((MayBeQNaN || MayBeSNaN) && !Other.isEmptySet()) ||
((Other.MayBeQNaN || Other.MayBeSNaN) && !isEmptySet());
if (isNaNOnly() || Other.isNaNOnly())
return getNaNOnly(Sem, /*MayBeQNaN=*/ResMayBeQNaN,
/*MayBeSNaN=*/false);
std::optional<SameSignRange> LHSNeg, LHSPos, RHSNeg, RHSPos;
splitPosNeg(Lower, Upper, LHSNeg, LHSPos);
splitPosNeg(Other.Lower, Other.Upper, RHSNeg, RHSPos);
APFloat ResLower = APFloat::getInf(Sem, /*Negative=*/false);
APFloat ResUpper = APFloat::getInf(Sem, /*Negative=*/true);
auto Update = [&](std::optional<SameSignRange> &LHS,
std::optional<SameSignRange> &RHS, bool Negative) {
if (!LHS || !RHS)
return;
// 0 * inf = QNaN
ResMayBeQNaN |= LHS->HasZero && RHS->HasInf;
ResMayBeQNaN |= RHS->HasZero && LHS->HasInf;
// NonZero * inf = inf
if ((LHS->HasInf && RHS->HasNonZero) || (RHS->HasInf && LHS->HasNonZero))
(Negative ? ResLower : ResUpper) = APFloat::getInf(Sem, Negative);
// Finite * Finite
if (LHS->FinitePart && RHS->FinitePart) {
APFloat NewLower = LHS->FinitePart->first * RHS->FinitePart->first;
APFloat NewUpper = LHS->FinitePart->second * RHS->FinitePart->second;
if (Negative) {
ResLower = minnum(ResLower, -NewUpper);
ResUpper = maxnum(ResUpper, -NewLower);
} else {
ResLower = minnum(ResLower, NewLower);
ResUpper = maxnum(ResUpper, NewUpper);
}
}
};
Update(LHSNeg, RHSNeg, /*Negative=*/false);
Update(LHSNeg, RHSPos, /*Negative=*/true);
Update(LHSPos, RHSNeg, /*Negative=*/true);
Update(LHSPos, RHSPos, /*Negative=*/false);
return ConstantFPRange(ResLower, ResUpper, ResMayBeQNaN, /*MayBeSNaN=*/false);
}

ConstantFPRange ConstantFPRange::div(const ConstantFPRange &Other) const {
auto &Sem = getSemantics();
bool ResMayBeQNaN = ((MayBeQNaN || MayBeSNaN) && !Other.isEmptySet()) ||
((Other.MayBeQNaN || Other.MayBeSNaN) && !isEmptySet());
if (isNaNOnly() || Other.isNaNOnly())
return getNaNOnly(Sem, /*MayBeQNaN=*/ResMayBeQNaN,
/*MayBeSNaN=*/false);
std::optional<SameSignRange> LHSNeg, LHSPos, RHSNeg, RHSPos;
splitPosNeg(Lower, Upper, LHSNeg, LHSPos);
splitPosNeg(Other.Lower, Other.Upper, RHSNeg, RHSPos);
APFloat ResLower = APFloat::getInf(Sem, /*Negative=*/false);
APFloat ResUpper = APFloat::getInf(Sem, /*Negative=*/true);
auto Update = [&](std::optional<SameSignRange> &LHS,
std::optional<SameSignRange> &RHS, bool Negative) {
if (!LHS || !RHS)
return;
// inf / inf = QNaN 0 / 0 = QNaN
ResMayBeQNaN |= LHS->HasInf && RHS->HasInf;
ResMayBeQNaN |= LHS->HasZero && RHS->HasZero;
// It is not straightforward to infer HasNonZeroFinite = HasFinite &&
// HasNonZero. By definitions we have:
// HasFinite = HasNonZeroFinite || HasZero
// HasNonZero = HasNonZeroFinite || HasInf
// Since the range is contiguous, if both HasFinite and HasNonZero are true,
// HasNonZeroFinite must be true.
bool LHSHasNonZeroFinite = LHS->FinitePart && LHS->HasNonZero;
bool RHSHasNonZeroFinite = RHS->FinitePart && RHS->HasNonZero;
// inf / Finite = inf FiniteNonZero / 0 = inf
if ((LHS->HasInf && RHS->FinitePart) ||
(LHSHasNonZeroFinite && RHS->HasZero))
(Negative ? ResLower : ResUpper) = APFloat::getInf(Sem, Negative);
// Finite / inf = 0
if (LHS->FinitePart && RHS->HasInf) {
APFloat Zero = APFloat::getZero(Sem, /*Negative=*/Negative);
ResLower = minnum(ResLower, Zero);
ResUpper = maxnum(ResUpper, Zero);
}
// Finite / FiniteNonZero
if (LHS->FinitePart && RHSHasNonZeroFinite) {
assert(!RHS->FinitePart->second.isZero() &&
"Divisor should be non-zero.");
APFloat NewLower = LHS->FinitePart->first / RHS->FinitePart->second;
APFloat NewUpper = LHS->FinitePart->second /
(RHS->FinitePart->first.isZero()
? APFloat::getSmallest(Sem, /*Negative=*/false)
: RHS->FinitePart->first);
if (Negative) {
ResLower = minnum(ResLower, -NewUpper);
ResUpper = maxnum(ResUpper, -NewLower);
} else {
ResLower = minnum(ResLower, NewLower);
ResUpper = maxnum(ResUpper, NewUpper);
}
}
};
Update(LHSNeg, RHSNeg, /*Negative=*/false);
Update(LHSNeg, RHSPos, /*Negative=*/true);
Update(LHSPos, RHSNeg, /*Negative=*/true);
Update(LHSPos, RHSPos, /*Negative=*/false);
return ConstantFPRange(ResLower, ResUpper, ResMayBeQNaN, /*MayBeSNaN=*/false);
}
109 changes: 109 additions & 0 deletions llvm/unittests/IR/ConstantFPRangeTest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1066,6 +1066,115 @@ TEST_F(ConstantFPRangeTest, sub) {
#endif
}

TEST_F(ConstantFPRangeTest, mul) {
EXPECT_EQ(Full.mul(Full), NonNaN.unionWith(QNaN));
EXPECT_EQ(Full.mul(Empty), Empty);
EXPECT_EQ(Empty.mul(Full), Empty);
EXPECT_EQ(Empty.mul(Empty), Empty);
EXPECT_EQ(One.mul(One), ConstantFPRange(APFloat(1.0)));
EXPECT_EQ(Some.mul(Some),
ConstantFPRange::getNonNaN(APFloat(-9.0), APFloat(9.0)));
EXPECT_EQ(SomePos.mul(SomeNeg),
ConstantFPRange::getNonNaN(APFloat(-9.0), APFloat(-0.0)));
EXPECT_EQ(PosInf.mul(PosInf), PosInf);
EXPECT_EQ(NegInf.mul(NegInf), PosInf);
EXPECT_EQ(PosInf.mul(Finite), NonNaN.unionWith(QNaN));
EXPECT_EQ(NegInf.mul(Finite), NonNaN.unionWith(QNaN));
EXPECT_EQ(PosInf.mul(NegInf), NegInf);
EXPECT_EQ(NegInf.mul(PosInf), NegInf);
EXPECT_EQ(PosZero.mul(NegZero), NegZero);
EXPECT_EQ(PosZero.mul(Zero), Zero);
EXPECT_EQ(NegZero.mul(NegZero), PosZero);
EXPECT_EQ(NegZero.mul(Zero), Zero);
EXPECT_EQ(NaN.mul(NaN), QNaN);
EXPECT_EQ(NaN.mul(Finite), QNaN);

#if defined(EXPENSIVE_CHECKS)
EnumerateTwoInterestingConstantFPRanges(
[](const ConstantFPRange &LHS, const ConstantFPRange &RHS) {
ConstantFPRange Res = LHS.mul(RHS);
ConstantFPRange Expected =
ConstantFPRange::getEmpty(LHS.getSemantics());
EnumerateValuesInConstantFPRange(
LHS,
[&](const APFloat &LHSC) {
EnumerateValuesInConstantFPRange(
RHS,
[&](const APFloat &RHSC) {
APFloat Prod = LHSC * RHSC;
EXPECT_TRUE(Res.contains(Prod))
<< "Wrong result for " << LHS << " * " << RHS
<< ". The result " << Res << " should contain " << Prod;
if (!Expected.contains(Prod))
Expected = Expected.unionWith(ConstantFPRange(Prod));
},
/*IgnoreNaNPayload=*/true);
},
/*IgnoreNaNPayload=*/true);
EXPECT_EQ(Res, Expected)
<< "Suboptimal result for " << LHS << " * " << RHS << ". Expected "
<< Expected << ", but got " << Res;
},
SparseLevel::SpecialValuesOnly);
#endif
}

TEST_F(ConstantFPRangeTest, div) {
EXPECT_EQ(Full.div(Full), NonNaN.unionWith(QNaN));
EXPECT_EQ(Full.div(Empty), Empty);
EXPECT_EQ(Empty.div(Full), Empty);
EXPECT_EQ(Empty.div(Empty), Empty);
EXPECT_EQ(One.div(One), ConstantFPRange(APFloat(1.0)));
EXPECT_EQ(Some.div(Some), NonNaN.unionWith(QNaN));
EXPECT_EQ(SomePos.div(SomeNeg),
ConstantFPRange(APFloat::getInf(Sem, /*Negative=*/true),
APFloat::getZero(Sem, /*Negative=*/true),
/*MayBeQNaN=*/true, /*MayBeSNaN=*/false));
EXPECT_EQ(PosInf.div(PosInf), QNaN);
EXPECT_EQ(NegInf.div(NegInf), QNaN);
EXPECT_EQ(PosInf.div(Finite), NonNaN);
EXPECT_EQ(NegInf.div(Finite), NonNaN);
EXPECT_EQ(PosInf.div(NegInf), QNaN);
EXPECT_EQ(NegInf.div(PosInf), QNaN);
EXPECT_EQ(Zero.div(Zero), QNaN);
EXPECT_EQ(SomePos.div(PosInf), PosZero);
EXPECT_EQ(SomeNeg.div(PosInf), NegZero);
EXPECT_EQ(PosInf.div(SomePos), PosInf);
EXPECT_EQ(NegInf.div(SomeNeg), PosInf);
EXPECT_EQ(NegInf.div(Some), NonNaN);
EXPECT_EQ(NaN.div(NaN), QNaN);
EXPECT_EQ(NaN.div(Finite), QNaN);

#if defined(EXPENSIVE_CHECKS)
EnumerateTwoInterestingConstantFPRanges(
[](const ConstantFPRange &LHS, const ConstantFPRange &RHS) {
ConstantFPRange Res = LHS.div(RHS);
ConstantFPRange Expected =
ConstantFPRange::getEmpty(LHS.getSemantics());
EnumerateValuesInConstantFPRange(
LHS,
[&](const APFloat &LHSC) {
EnumerateValuesInConstantFPRange(
RHS,
[&](const APFloat &RHSC) {
APFloat Val = LHSC / RHSC;
EXPECT_TRUE(Res.contains(Val))
<< "Wrong result for " << LHS << " / " << RHS
<< ". The result " << Res << " should contain " << Val;
if (!Expected.contains(Val))
Expected = Expected.unionWith(ConstantFPRange(Val));
},
/*IgnoreNaNPayload=*/true);
},
/*IgnoreNaNPayload=*/true);
EXPECT_EQ(Res, Expected)
<< "Suboptimal result for " << LHS << " / " << RHS << ". Expected "
<< Expected << ", but got " << Res;
},
SparseLevel::SpecialValuesOnly);
#endif
}

TEST_F(ConstantFPRangeTest, flushDenormals) {
const fltSemantics &FP8Sem = APFloat::Float8E4M3();
APFloat NormalVal = APFloat::getSmallestNormalized(FP8Sem);
Expand Down
Loading