Skip to content

Commit

Permalink
[APFloat] convertToDouble/Float can work on shorter types
Browse files Browse the repository at this point in the history 
Previously APFloat::convertToDouble may be called only for APFloats that
were built using double semantics. Other semantics like single precision
were not allowed although corresponding numbers could be converted to
double without loss of precision. The similar restriction applied to
APFloat::convertToFloat.

With this change any APFloat that can be precisely represented by double
can be handled with convertToDouble. Behavior of convertToFloat was
updated similarly. It make the conversion operations more convenient and
adds support for formats like half and bfloat.

Differential Revision: https://reviews.llvm.org/D102671
  • Loading branch information
@spavloff
spavloff committed May 21, 2021
1 parent 4902885 commit c162f08
Show file tree
Hide file tree
Showing 7 changed files with 313 additions and 18 deletions.
16 changes: 14 additions & 2 deletions llvm/include/llvm/ADT/APFloat.h
View file
Expand Up @@ -1132,8 +1132,20 @@ class APFloat : public APFloatBase {
APInt bitcastToAPInt() const {
APFLOAT_DISPATCH_ON_SEMANTICS(bitcastToAPInt());
}
double convertToDouble() const { return getIEEE().convertToDouble(); }
float convertToFloat() const { return getIEEE().convertToFloat(); }

/// Converts this APFloat to host double value.
///
/// \pre The APFloat must be built using semantics, that can be represented by
/// the host double type without loss of precision. It can be IEEEdouble and
/// shorter semantics, like IEEEsingle and others.
double convertToDouble() const;

/// Converts this APFloat to host float value.
///
/// \pre The APFloat must be built using semantics, that can be represented by
/// the host float type without loss of precision. It can be IEEEsingle and
/// shorter semantics, like IEEEhalf.
float convertToFloat() const;

bool operator==(const APFloat &RHS) const { return compare(RHS) == cmpEqual; }

Expand Down
5 changes: 1 addition & 4 deletions llvm/lib/Analysis/ConstantFolding.cpp
View file
Expand Up @@ -1794,10 +1794,7 @@ Constant *ConstantFoldSSEConvertToInt(const APFloat &Val, bool roundTowardZero,
double getValueAsDouble(ConstantFP *Op) {
Type *Ty = Op->getType();

if (Ty->isFloatTy())
return Op->getValueAPF().convertToFloat();

if (Ty->isDoubleTy())
if (Ty->isBFloatTy() || Ty->isHalfTy() || Ty->isFloatTy() || Ty->isDoubleTy())
return Op->getValueAPF().convertToDouble();

bool unused;
Expand Down
6 changes: 3 additions & 3 deletions llvm/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
View file
Expand Up @@ -953,9 +953,9 @@ static bool emitDebugValueComment(const MachineInstr *MI, AsmPrinter &AP) {
switch (Op.getType()) {
case MachineOperand::MO_FPImmediate: {
APFloat APF = APFloat(Op.getFPImm()->getValueAPF());
if (Op.getFPImm()->getType()->isFloatTy()) {
OS << (double)APF.convertToFloat();
} else if (Op.getFPImm()->getType()->isDoubleTy()) {
Type *ImmTy = Op.getFPImm()->getType();
if (ImmTy->isBFloatTy() || ImmTy->isHalfTy() || ImmTy->isFloatTy() ||
ImmTy->isDoubleTy()) {
OS << APF.convertToDouble();
} else {
// There is no good way to print long double. Convert a copy to
Expand Down
2 changes: 1 addition & 1 deletion llvm/lib/IR/AsmWriter.cpp
View file
Expand Up @@ -1368,7 +1368,7 @@ static void WriteConstantInternal(raw_ostream &Out, const Constant *CV,
bool isInf = APF.isInfinity();
bool isNaN = APF.isNaN();
if (!isInf && !isNaN) {
double Val = isDouble ? APF.convertToDouble() : APF.convertToFloat();
double Val = APF.convertToDouble();
SmallString<128> StrVal;
APF.toString(StrVal, 6, 0, false);
// Check to make sure that the stringized number is not some string like
Expand Down
8 changes: 2 additions & 6 deletions llvm/lib/IR/Core.cpp
View file
Expand Up @@ -1399,12 +1399,8 @@ double LLVMConstRealGetDouble(LLVMValueRef ConstantVal, LLVMBool *LosesInfo) {
ConstantFP *cFP = unwrap<ConstantFP>(ConstantVal) ;
Type *Ty = cFP->getType();

if (Ty->isFloatTy()) {
*LosesInfo = false;
return cFP->getValueAPF().convertToFloat();
}

if (Ty->isDoubleTy()) {
if (Ty->isHalfTy() || Ty->isBFloatTy() || Ty->isFloatTy() ||
Ty->isDoubleTy()) {
*LosesInfo = false;
return cFP->getValueAPF().convertToDouble();
}
Expand Down
33 changes: 33 additions & 0 deletions llvm/lib/Support/APFloat.cpp
View file
Expand Up @@ -66,6 +66,13 @@ namespace llvm {

/* Number of bits actually used in the semantics. */
unsigned int sizeInBits;

// Returns true if any number described by this semantics can be precisely
// represented by the specified semantics.
bool isRepresentableBy(const fltSemantics &S) const {
return maxExponent <= S.maxExponent && minExponent >= S.minExponent &&
precision <= S.precision;
}
};

static const fltSemantics semIEEEhalf = {15, -14, 11, 16};
Expand Down Expand Up @@ -4875,6 +4882,32 @@ APFloat::opStatus APFloat::convertToInteger(APSInt &result,
return status;
}

double APFloat::convertToDouble() const {
if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEdouble)
return getIEEE().convertToDouble();
assert(getSemantics().isRepresentableBy(semIEEEdouble) &&
"Float semantics is not representable by IEEEdouble");
APFloat Temp = *this;
bool LosesInfo;
opStatus St = Temp.convert(semIEEEdouble, rmNearestTiesToEven, &LosesInfo);
assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
(void)St;
return Temp.getIEEE().convertToDouble();
}

float APFloat::convertToFloat() const {
if (&getSemantics() == (const llvm::fltSemantics *)&semIEEEsingle)
return getIEEE().convertToFloat();
assert(getSemantics().isRepresentableBy(semIEEEsingle) &&
"Float semantics is not representable by IEEEsingle");
APFloat Temp = *this;
bool LosesInfo;
opStatus St = Temp.convert(semIEEEsingle, rmNearestTiesToEven, &LosesInfo);
assert(!(St & opInexact) && !LosesInfo && "Unexpected imprecision");
(void)St;
return Temp.getIEEE().convertToFloat();
}

} // namespace llvm

#undef APFLOAT_DISPATCH_ON_SEMANTICS
261 changes: 259 additions & 2 deletions llvm/unittests/ADT/APFloatTest.cpp
View file
Expand Up @@ -1271,8 +1271,10 @@ TEST(APFloatTest, makeNaN) {
#ifdef GTEST_HAS_DEATH_TEST
#ifndef NDEBUG
TEST(APFloatTest, SemanticsDeath) {
EXPECT_DEATH(APFloat(APFloat::IEEEsingle(), 0).convertToDouble(), "Float semantics are not IEEEdouble");
EXPECT_DEATH(APFloat(APFloat::IEEEdouble(), 0).convertToFloat(), "Float semantics are not IEEEsingle");
EXPECT_DEATH(APFloat(APFloat::IEEEquad(), 0).convertToDouble(),
"Float semantics is not representable by IEEEdouble");
EXPECT_DEATH(APFloat(APFloat::IEEEdouble(), 0).convertToFloat(),
"Float semantics is not representable by IEEEsingle");
}
#endif
#endif
Expand Down Expand Up @@ -4709,4 +4711,259 @@ TEST(APFloatTest, x87Next) {
F.next(false);
EXPECT_TRUE(ilogb(F) == -1);
}

TEST(APFloatTest, ToDouble) {
APFloat DPosZero(0.0);
APFloat DPosZeroToDouble(DPosZero.convertToDouble());
EXPECT_TRUE(DPosZeroToDouble.isPosZero());
APFloat DNegZero(-0.0);
APFloat DNegZeroToDouble(DNegZero.convertToDouble());
EXPECT_TRUE(DNegZeroToDouble.isNegZero());

APFloat DOne(1.0);
EXPECT_EQ(1.0, DOne.convertToDouble());
APFloat DPosLargest = APFloat::getLargest(APFloat::IEEEdouble(), false);
EXPECT_EQ(std::numeric_limits<double>::max(), DPosLargest.convertToDouble());
APFloat DNegLargest = APFloat::getLargest(APFloat::IEEEdouble(), true);
EXPECT_EQ(-std::numeric_limits<double>::max(), DNegLargest.convertToDouble());
APFloat DPosSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEdouble(), false);
EXPECT_EQ(std::numeric_limits<double>::min(), DPosSmallest.convertToDouble());
APFloat DNegSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEdouble(), true);
EXPECT_EQ(-std::numeric_limits<double>::min(),
DNegSmallest.convertToDouble());

APFloat DSmallestDenorm = APFloat::getSmallest(APFloat::IEEEdouble(), false);
EXPECT_EQ(std::numeric_limits<double>::denorm_min(),
DSmallestDenorm.convertToDouble());
APFloat DLargestDenorm(APFloat::IEEEdouble(), "0x0.FFFFFFFFFFFFFp-1022");
EXPECT_EQ(/*0x0.FFFFFFFFFFFFFp-1022*/ 2.225073858507201e-308,
DLargestDenorm.convertToDouble());

APFloat DPosInf = APFloat::getInf(APFloat::IEEEdouble());
EXPECT_EQ(std::numeric_limits<double>::infinity(), DPosInf.convertToDouble());
APFloat DNegInf = APFloat::getInf(APFloat::IEEEdouble(), true);
EXPECT_EQ(-std::numeric_limits<double>::infinity(),
DNegInf.convertToDouble());
APFloat DQNaN = APFloat::getQNaN(APFloat::IEEEdouble());
EXPECT_TRUE(std::isnan(DQNaN.convertToDouble()));

APFloat FPosZero(0.0F);
APFloat FPosZeroToDouble(FPosZero.convertToDouble());
EXPECT_TRUE(FPosZeroToDouble.isPosZero());
APFloat FNegZero(-0.0F);
APFloat FNegZeroToDouble(FNegZero.convertToDouble());
EXPECT_TRUE(FNegZeroToDouble.isNegZero());

APFloat FOne(1.0F);
EXPECT_EQ(1.0, FOne.convertToDouble());
APFloat FPosLargest = APFloat::getLargest(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::max(), FPosLargest.convertToDouble());
APFloat FNegLargest = APFloat::getLargest(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<float>::max(), FNegLargest.convertToDouble());
APFloat FPosSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::min(), FPosSmallest.convertToDouble());
APFloat FNegSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<float>::min(), FNegSmallest.convertToDouble());

APFloat FSmallestDenorm = APFloat::getSmallest(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::denorm_min(),
FSmallestDenorm.convertToDouble());
APFloat FLargestDenorm(APFloat::IEEEdouble(), "0x0.FFFFFEp-126");
EXPECT_EQ(/*0x0.FFFFFEp-126*/ 1.1754942106924411e-38,
FLargestDenorm.convertToDouble());

APFloat FPosInf = APFloat::getInf(APFloat::IEEEsingle());
EXPECT_EQ(std::numeric_limits<double>::infinity(), FPosInf.convertToDouble());
APFloat FNegInf = APFloat::getInf(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<double>::infinity(),
FNegInf.convertToDouble());
APFloat FQNaN = APFloat::getQNaN(APFloat::IEEEsingle());
EXPECT_TRUE(std::isnan(FQNaN.convertToDouble()));

APFloat HPosZero = APFloat::getZero(APFloat::IEEEhalf());
APFloat HPosZeroToDouble(HPosZero.convertToDouble());
EXPECT_TRUE(HPosZeroToDouble.isPosZero());
APFloat HNegZero = APFloat::getZero(APFloat::IEEEhalf(), true);
APFloat HNegZeroToDouble(HNegZero.convertToDouble());
EXPECT_TRUE(HNegZeroToDouble.isNegZero());

APFloat HOne(APFloat::IEEEhalf(), "1.0");
EXPECT_EQ(1.0, HOne.convertToDouble());
APFloat HPosLargest = APFloat::getLargest(APFloat::IEEEhalf(), false);
EXPECT_EQ(65504.0, HPosLargest.convertToDouble());
APFloat HNegLargest = APFloat::getLargest(APFloat::IEEEhalf(), true);
EXPECT_EQ(-65504.0, HNegLargest.convertToDouble());
APFloat HPosSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEhalf(), false);
EXPECT_EQ(/*0x1.p-14*/ 6.103515625e-05, HPosSmallest.convertToDouble());
APFloat HNegSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEhalf(), true);
EXPECT_EQ(/*-0x1.p-14*/ -6.103515625e-05, HNegSmallest.convertToDouble());

APFloat HSmallestDenorm = APFloat::getSmallest(APFloat::IEEEhalf(), false);
EXPECT_EQ(/*0x1.p-24*/ 5.960464477539063e-08,
HSmallestDenorm.convertToDouble());
APFloat HLargestDenorm(APFloat::IEEEhalf(), "0x1.FFCp-14");
EXPECT_EQ(/*0x1.FFCp-14*/ 0.00012201070785522461,
HLargestDenorm.convertToDouble());

APFloat HPosInf = APFloat::getInf(APFloat::IEEEhalf());
EXPECT_EQ(std::numeric_limits<double>::infinity(), HPosInf.convertToDouble());
APFloat HNegInf = APFloat::getInf(APFloat::IEEEhalf(), true);
EXPECT_EQ(-std::numeric_limits<double>::infinity(),
HNegInf.convertToDouble());
APFloat HQNaN = APFloat::getQNaN(APFloat::IEEEhalf());
EXPECT_TRUE(std::isnan(HQNaN.convertToDouble()));

APFloat BPosZero = APFloat::getZero(APFloat::IEEEhalf());
APFloat BPosZeroToDouble(BPosZero.convertToDouble());
EXPECT_TRUE(BPosZeroToDouble.isPosZero());
APFloat BNegZero = APFloat::getZero(APFloat::IEEEhalf(), true);
APFloat BNegZeroToDouble(BNegZero.convertToDouble());
EXPECT_TRUE(BNegZeroToDouble.isNegZero());

APFloat BOne(APFloat::BFloat(), "1.0");
EXPECT_EQ(1.0, BOne.convertToDouble());
APFloat BPosLargest = APFloat::getLargest(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.FEp127*/ 3.3895313892515355e+38,
BPosLargest.convertToDouble());
APFloat BNegLargest = APFloat::getLargest(APFloat::BFloat(), true);
EXPECT_EQ(/*-0x1.FEp127*/ -3.3895313892515355e+38,
BNegLargest.convertToDouble());
APFloat BPosSmallest =
APFloat::getSmallestNormalized(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.p-126*/ 1.1754943508222875e-38,
BPosSmallest.convertToDouble());
APFloat BNegSmallest =
APFloat::getSmallestNormalized(APFloat::BFloat(), true);
EXPECT_EQ(/*-0x1.p-126*/ -1.1754943508222875e-38,
BNegSmallest.convertToDouble());

APFloat BSmallestDenorm = APFloat::getSmallest(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.p-133*/ 9.183549615799121e-41,
BSmallestDenorm.convertToDouble());
APFloat BLargestDenorm(APFloat::BFloat(), "0x1.FCp-127");
EXPECT_EQ(/*0x1.FCp-127*/ 1.1663108012064884e-38,
BLargestDenorm.convertToDouble());

APFloat BPosInf = APFloat::getInf(APFloat::BFloat());
EXPECT_EQ(std::numeric_limits<double>::infinity(), BPosInf.convertToDouble());
APFloat BNegInf = APFloat::getInf(APFloat::BFloat(), true);
EXPECT_EQ(-std::numeric_limits<double>::infinity(),
BNegInf.convertToDouble());
APFloat BQNaN = APFloat::getQNaN(APFloat::BFloat());
EXPECT_TRUE(std::isnan(BQNaN.convertToDouble()));
}

TEST(APFloatTest, ToFloat) {
APFloat FPosZero(0.0F);
APFloat FPosZeroToFloat(FPosZero.convertToFloat());
EXPECT_TRUE(FPosZeroToFloat.isPosZero());
APFloat FNegZero(-0.0F);
APFloat FNegZeroToFloat(FNegZero.convertToFloat());
EXPECT_TRUE(FNegZeroToFloat.isNegZero());

APFloat FOne(1.0F);
EXPECT_EQ(1.0F, FOne.convertToFloat());
APFloat FPosLargest = APFloat::getLargest(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::max(), FPosLargest.convertToFloat());
APFloat FNegLargest = APFloat::getLargest(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<float>::max(), FNegLargest.convertToFloat());
APFloat FPosSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::min(), FPosSmallest.convertToFloat());
APFloat FNegSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<float>::min(), FNegSmallest.convertToFloat());

APFloat FSmallestDenorm = APFloat::getSmallest(APFloat::IEEEsingle(), false);
EXPECT_EQ(std::numeric_limits<float>::denorm_min(),
FSmallestDenorm.convertToFloat());
APFloat FLargestDenorm(APFloat::IEEEsingle(), "0x1.FFFFFEp-126");
EXPECT_EQ(/*0x1.FFFFFEp-126*/ 2.3509885615147286e-38F,
FLargestDenorm.convertToFloat());

APFloat FPosInf = APFloat::getInf(APFloat::IEEEsingle());
EXPECT_EQ(std::numeric_limits<float>::infinity(), FPosInf.convertToFloat());
APFloat FNegInf = APFloat::getInf(APFloat::IEEEsingle(), true);
EXPECT_EQ(-std::numeric_limits<float>::infinity(), FNegInf.convertToFloat());
APFloat FQNaN = APFloat::getQNaN(APFloat::IEEEsingle());
EXPECT_TRUE(std::isnan(FQNaN.convertToFloat()));

APFloat HPosZero = APFloat::getZero(APFloat::IEEEhalf());
APFloat HPosZeroToFloat(HPosZero.convertToFloat());
EXPECT_TRUE(HPosZeroToFloat.isPosZero());
APFloat HNegZero = APFloat::getZero(APFloat::IEEEhalf(), true);
APFloat HNegZeroToFloat(HNegZero.convertToFloat());
EXPECT_TRUE(HNegZeroToFloat.isNegZero());

APFloat HOne(APFloat::IEEEhalf(), "1.0");
EXPECT_EQ(1.0F, HOne.convertToFloat());
APFloat HPosLargest = APFloat::getLargest(APFloat::IEEEhalf(), false);
EXPECT_EQ(/*0x1.FFCp15*/ 65504.0F, HPosLargest.convertToFloat());
APFloat HNegLargest = APFloat::getLargest(APFloat::IEEEhalf(), true);
EXPECT_EQ(/*-0x1.FFCp15*/ -65504.0F, HNegLargest.convertToFloat());
APFloat HPosSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEhalf(), false);
EXPECT_EQ(/*0x1.p-14*/ 6.103515625e-05F, HPosSmallest.convertToFloat());
APFloat HNegSmallest =
APFloat::getSmallestNormalized(APFloat::IEEEhalf(), true);
EXPECT_EQ(/*-0x1.p-14*/ -6.103515625e-05F, HNegSmallest.convertToFloat());

APFloat HSmallestDenorm = APFloat::getSmallest(APFloat::IEEEhalf(), false);
EXPECT_EQ(/*0x1.p-24*/ 5.960464477539063e-08F,
HSmallestDenorm.convertToFloat());
APFloat HLargestDenorm(APFloat::IEEEhalf(), "0x1.FFCp-14");
EXPECT_EQ(/*0x1.FFCp-14*/ 0.00012201070785522461F,
HLargestDenorm.convertToFloat());

APFloat HPosInf = APFloat::getInf(APFloat::IEEEhalf());
EXPECT_EQ(std::numeric_limits<float>::infinity(), HPosInf.convertToFloat());
APFloat HNegInf = APFloat::getInf(APFloat::IEEEhalf(), true);
EXPECT_EQ(-std::numeric_limits<float>::infinity(), HNegInf.convertToFloat());
APFloat HQNaN = APFloat::getQNaN(APFloat::IEEEhalf());
EXPECT_TRUE(std::isnan(HQNaN.convertToFloat()));

APFloat BPosZero = APFloat::getZero(APFloat::BFloat());
APFloat BPosZeroToDouble(BPosZero.convertToFloat());
EXPECT_TRUE(BPosZeroToDouble.isPosZero());
APFloat BNegZero = APFloat::getZero(APFloat::BFloat(), true);
APFloat BNegZeroToDouble(BNegZero.convertToFloat());
EXPECT_TRUE(BNegZeroToDouble.isNegZero());

APFloat BOne(APFloat::BFloat(), "1.0");
EXPECT_EQ(1.0F, BOne.convertToFloat());
APFloat BPosLargest = APFloat::getLargest(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.FEp127*/ 3.3895313892515355e+38F,
BPosLargest.convertToFloat());
APFloat BNegLargest = APFloat::getLargest(APFloat::BFloat(), true);
EXPECT_EQ(/*-0x1.FEp127*/ -3.3895313892515355e+38F,
BNegLargest.convertToFloat());
APFloat BPosSmallest =
APFloat::getSmallestNormalized(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.p-126*/ 1.1754943508222875e-38F,
BPosSmallest.convertToFloat());
APFloat BNegSmallest =
APFloat::getSmallestNormalized(APFloat::BFloat(), true);
EXPECT_EQ(/*-0x1.p-126*/ -1.1754943508222875e-38F,
BNegSmallest.convertToFloat());

APFloat BSmallestDenorm = APFloat::getSmallest(APFloat::BFloat(), false);
EXPECT_EQ(/*0x1.p-133*/ 9.183549615799121e-41F,
BSmallestDenorm.convertToFloat());
APFloat BLargestDenorm(APFloat::BFloat(), "0x1.FCp-127");
EXPECT_EQ(/*0x1.FCp-127*/ 1.1663108012064884e-38F,
BLargestDenorm.convertToFloat());

APFloat BPosInf = APFloat::getInf(APFloat::BFloat());
EXPECT_EQ(std::numeric_limits<float>::infinity(), BPosInf.convertToFloat());
APFloat BNegInf = APFloat::getInf(APFloat::BFloat(), true);
EXPECT_EQ(-std::numeric_limits<float>::infinity(), BNegInf.convertToFloat());
APFloat BQNaN = APFloat::getQNaN(APFloat::BFloat());
EXPECT_TRUE(std::isnan(BQNaN.convertToFloat()));
}
}

0 comments on commit c162f08

Please sign in to comment.