d28adc2e7f
Original-commit: flang-compiler/f18@b0d1dd434b Reviewed-on: https://github.com/flang-compiler/f18/pull/119 Tree-same-pre-rewrite: false
431 lines
15 KiB
C++
431 lines
15 KiB
C++
// Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 (the "License");
|
|
// you may not use this file except in compliance with the License.
|
|
// You may obtain a copy of the License at
|
|
//
|
|
// http://www.apache.org/licenses/LICENSE-2.0
|
|
//
|
|
// Unless required by applicable law or agreed to in writing, software
|
|
// distributed under the License is distributed on an "AS IS" BASIS,
|
|
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
// See the License for the specific language governing permissions and
|
|
// limitations under the License.
|
|
|
|
#include "real.h"
|
|
#include "../common/idioms.h"
|
|
|
|
namespace Fortran::evaluate::value {
|
|
|
|
template<typename W, int P, bool IM>
|
|
Relation Real<W, P, IM>::Compare(const Real &y) const {
|
|
if (IsNotANumber() || y.IsNotANumber()) { // NaN vs x, x vs NaN
|
|
return Relation::Unordered;
|
|
} else if (IsInfinite()) {
|
|
if (y.IsInfinite()) {
|
|
if (IsNegative()) { // -Inf vs +/-Inf
|
|
return y.IsNegative() ? Relation::Equal : Relation::Less;
|
|
} else { // +Inf vs +/-Inf
|
|
return y.IsNegative() ? Relation::Greater : Relation::Equal;
|
|
}
|
|
} else { // +/-Inf vs finite
|
|
return IsNegative() ? Relation::Less : Relation::Greater;
|
|
}
|
|
} else if (y.IsInfinite()) { // finite vs +/-Inf
|
|
return y.IsNegative() ? Relation::Greater : Relation::Less;
|
|
} else { // two finite numbers
|
|
bool isNegative{IsNegative()};
|
|
if (isNegative != y.IsNegative()) {
|
|
if (word_.IOR(y.word_).IBCLR(bits - 1).IsZero()) {
|
|
return Relation::Equal; // +/-0.0 == -/+0.0
|
|
} else {
|
|
return isNegative ? Relation::Less : Relation::Greater;
|
|
}
|
|
} else {
|
|
// same sign
|
|
Ordering order{CompareUnsigned(Exponent(), y.Exponent())};
|
|
if (order == Ordering::Equal) {
|
|
order = GetSignificand().CompareUnsigned(y.GetSignificand());
|
|
}
|
|
if (isNegative) {
|
|
order = Reverse(order);
|
|
}
|
|
return RelationFromOrdering(order);
|
|
}
|
|
}
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
ValueWithRealFlags<Real<W, P, IM>> Real<W, P, IM>::Add(
|
|
const Real &y, Rounding rounding) const {
|
|
ValueWithRealFlags<Real> result;
|
|
if (IsNotANumber() || y.IsNotANumber()) {
|
|
result.value.word_ = NaNWord(); // NaN + x -> NaN
|
|
if (IsSignalingNaN() || y.IsSignalingNaN()) {
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
}
|
|
return result;
|
|
}
|
|
bool isNegative{IsNegative()};
|
|
bool yIsNegative{y.IsNegative()};
|
|
if (IsInfinite()) {
|
|
if (y.IsInfinite()) {
|
|
if (isNegative == yIsNegative) {
|
|
result.value = *this; // +/-Inf + +/-Inf -> +/-Inf
|
|
} else {
|
|
result.value.word_ = NaNWord(); // +/-Inf + -/+Inf -> NaN
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
}
|
|
} else {
|
|
result.value = *this; // +/-Inf + x -> +/-Inf
|
|
}
|
|
return result;
|
|
}
|
|
if (y.IsInfinite()) {
|
|
result.value = y; // x + +/-Inf -> +/-Inf
|
|
return result;
|
|
}
|
|
std::uint64_t exponent{Exponent()};
|
|
std::uint64_t yExponent{y.Exponent()};
|
|
if (exponent < yExponent) {
|
|
// y is larger in magnitude; simplify by reversing operands
|
|
return y.Add(*this, rounding);
|
|
}
|
|
if (exponent == yExponent && isNegative != yIsNegative) {
|
|
Ordering order{GetSignificand().CompareUnsigned(y.GetSignificand())};
|
|
if (order == Ordering::Less) {
|
|
// Same exponent, opposite signs, and y is larger in magnitude
|
|
return y.Add(*this, rounding);
|
|
}
|
|
if (order == Ordering::Equal) {
|
|
// x + (-x) -> +0.0 unless rounding is directed downwards
|
|
if (rounding == Rounding::Down) {
|
|
result.value.word_ = result.value.word_.IBSET(bits - 1); // -0.0
|
|
}
|
|
return result;
|
|
}
|
|
}
|
|
// Our exponent is greater than y's, or the exponents match and y is not
|
|
// of the opposite sign and greater magnitude. So (x+y) will have the
|
|
// same sign as x.
|
|
Fraction fraction{GetFraction()};
|
|
Fraction yFraction{y.GetFraction()};
|
|
int rshift = exponent - yExponent;
|
|
if (exponent > 0 && yExponent == 0) {
|
|
--rshift; // correct overshift when only y is denormal
|
|
}
|
|
RoundingBits roundingBits{yFraction, rshift};
|
|
yFraction = yFraction.SHIFTR(rshift);
|
|
bool carry{false};
|
|
if (isNegative != yIsNegative) {
|
|
// Opposite signs: subtract via addition of two's complement of y and
|
|
// the rounding bits.
|
|
yFraction = yFraction.NOT();
|
|
carry = roundingBits.Negate();
|
|
}
|
|
auto sum{fraction.AddUnsigned(yFraction, carry)};
|
|
fraction = sum.value;
|
|
if (isNegative == yIsNegative && sum.carry) {
|
|
roundingBits.ShiftRight(sum.value.BTEST(0));
|
|
fraction = fraction.SHIFTR(1).IBSET(fraction.bits - 1);
|
|
++exponent;
|
|
}
|
|
NormalizeAndRound(
|
|
result, isNegative, exponent, fraction, rounding, roundingBits);
|
|
return result;
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
ValueWithRealFlags<Real<W, P, IM>> Real<W, P, IM>::Multiply(
|
|
const Real &y, Rounding rounding) const {
|
|
ValueWithRealFlags<Real> result;
|
|
if (IsNotANumber() || y.IsNotANumber()) {
|
|
result.value.word_ = NaNWord(); // NaN * x -> NaN
|
|
if (IsSignalingNaN() || y.IsSignalingNaN()) {
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
}
|
|
} else {
|
|
bool isNegative{IsNegative() != y.IsNegative()};
|
|
if (IsInfinite() || y.IsInfinite()) {
|
|
if (IsZero() || y.IsZero()) {
|
|
result.value.word_ = NaNWord(); // 0 * Inf -> NaN
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
} else {
|
|
result.value.word_ = InfinityWord(isNegative);
|
|
}
|
|
} else {
|
|
auto product{GetFraction().MultiplyUnsigned(y.GetFraction())};
|
|
std::int64_t exponent{CombineExponents(y, false)};
|
|
if (exponent < 1) {
|
|
int rshift = 1 - exponent;
|
|
exponent = 1;
|
|
bool sticky{false};
|
|
if (rshift >= product.upper.bits + product.lower.bits) {
|
|
sticky = !product.lower.IsZero() || !product.upper.IsZero();
|
|
} else if (rshift >= product.lower.bits) {
|
|
sticky = !product.lower.IsZero() ||
|
|
!product.upper
|
|
.IAND(product.upper.MASKR(rshift - product.lower.bits))
|
|
.IsZero();
|
|
} else {
|
|
sticky = !product.lower.IAND(product.lower.MASKR(rshift)).IsZero();
|
|
}
|
|
product.lower = product.lower.DSHIFTR(product.upper, rshift);
|
|
product.upper = product.upper.SHIFTR(rshift);
|
|
if (sticky) {
|
|
product.lower = product.lower.IBSET(0);
|
|
}
|
|
}
|
|
int leadz{product.upper.LEADZ()};
|
|
if (leadz >= product.upper.bits) {
|
|
leadz += product.lower.LEADZ();
|
|
}
|
|
int lshift{leadz};
|
|
if (lshift > exponent - 1) {
|
|
lshift = exponent - 1;
|
|
}
|
|
exponent -= lshift;
|
|
product.upper = product.upper.DSHIFTL(product.lower, lshift);
|
|
product.lower = product.lower.SHIFTL(lshift);
|
|
RoundingBits roundingBits{product.lower, product.lower.bits};
|
|
NormalizeAndRound(result, isNegative, exponent, product.upper, rounding,
|
|
roundingBits, true /*multiply*/);
|
|
}
|
|
}
|
|
return result;
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
ValueWithRealFlags<Real<W, P, IM>> Real<W, P, IM>::Divide(
|
|
const Real &y, Rounding rounding) const {
|
|
ValueWithRealFlags<Real> result;
|
|
if (IsNotANumber() || y.IsNotANumber()) {
|
|
result.value.word_ = NaNWord(); // NaN / x -> NaN, x / NaN -> NaN
|
|
if (IsSignalingNaN() || y.IsSignalingNaN()) {
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
}
|
|
} else {
|
|
bool isNegative{IsNegative() != y.IsNegative()};
|
|
if (IsInfinite()) {
|
|
if (y.IsInfinite()) {
|
|
result.value.word_ = NaNWord(); // Inf/Inf -> NaN
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
} else { // Inf/x -> Inf, Inf/0 -> Inf
|
|
result.value.word_ = InfinityWord(isNegative);
|
|
}
|
|
} else if (y.IsZero()) {
|
|
if (IsZero()) { // 0/0 -> NaN
|
|
result.value.word_ = NaNWord();
|
|
result.flags.set(RealFlag::InvalidArgument);
|
|
} else { // x/0 -> Inf, Inf/0 -> Inf
|
|
result.value.word_ = InfinityWord(isNegative);
|
|
result.flags.set(RealFlag::DivideByZero);
|
|
}
|
|
} else if (IsZero() || y.IsInfinite()) { // 0/x, x/Inf -> 0
|
|
if (isNegative) {
|
|
result.value.word_ = result.value.word_.IBSET(bits - 1);
|
|
}
|
|
} else {
|
|
// dividend and divisor are both finite and nonzero numbers
|
|
Fraction top{GetFraction()}, divisor{y.GetFraction()};
|
|
std::int64_t exponent{CombineExponents(y, true)};
|
|
Fraction quotient;
|
|
bool msb{false};
|
|
if (!top.BTEST(top.bits - 1) || !divisor.BTEST(divisor.bits - 1)) {
|
|
// One or two denormals
|
|
int topLshift{top.LEADZ()};
|
|
top = top.SHIFTL(topLshift);
|
|
int divisorLshift{divisor.LEADZ()};
|
|
divisor = divisor.SHIFTL(divisorLshift);
|
|
exponent += divisorLshift - topLshift;
|
|
}
|
|
for (int j{1}; j <= quotient.bits; ++j) {
|
|
if (NextQuotientBit(top, msb, divisor)) {
|
|
quotient = quotient.IBSET(quotient.bits - j);
|
|
}
|
|
}
|
|
bool guard{NextQuotientBit(top, msb, divisor)};
|
|
bool round{NextQuotientBit(top, msb, divisor)};
|
|
bool sticky{msb || !top.IsZero()};
|
|
RoundingBits roundingBits{guard, round, sticky};
|
|
if (exponent < 1) {
|
|
std::int64_t rshift{1 - exponent};
|
|
for (; rshift > 0; --rshift) {
|
|
roundingBits.ShiftRight(quotient.BTEST(0));
|
|
quotient = quotient.SHIFTR(1);
|
|
}
|
|
exponent = 1;
|
|
}
|
|
NormalizeAndRound(
|
|
result, isNegative, exponent, quotient, rounding, roundingBits);
|
|
}
|
|
}
|
|
return result;
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
RealFlags Real<W, P, IM>::Normalize(bool negative, std::uint64_t exponent,
|
|
const Fraction &fraction, Rounding rounding, RoundingBits *roundingBits) {
|
|
std::uint64_t lshift = fraction.LEADZ();
|
|
if (lshift == fraction.bits /* fraction is zero */ &&
|
|
(roundingBits == nullptr || roundingBits->empty())) {
|
|
// No fraction, no rounding bits -> +/-0.0
|
|
exponent = lshift = 0;
|
|
} else if (lshift < exponent) {
|
|
exponent -= lshift;
|
|
} else if (exponent > 0) {
|
|
lshift = exponent - 1;
|
|
exponent = 0;
|
|
} else if (lshift == 0) {
|
|
exponent = 1;
|
|
} else {
|
|
lshift = 0;
|
|
}
|
|
if (exponent >= maxExponent) {
|
|
// Infinity or overflow
|
|
if (rounding == Rounding::TiesToEven ||
|
|
rounding == Rounding::TiesAwayFromZero ||
|
|
(rounding == Rounding::Up && !negative) ||
|
|
(rounding == Rounding::Down && negative)) {
|
|
word_ = Word{maxExponent}.SHIFTL(significandBits); // Inf
|
|
} else {
|
|
// directed rounding: round to largest finite value rather than infinity
|
|
// (x86 does this, not sure whether it's standard behavior)
|
|
word_ = Word{word_.MASKR(word_.bits - 1)}.IBCLR(significandBits);
|
|
}
|
|
if (negative) {
|
|
word_ = word_.IBSET(bits - 1);
|
|
}
|
|
RealFlags flags{RealFlag::Overflow};
|
|
if (!fraction.IsZero()) {
|
|
flags.set(RealFlag::Inexact);
|
|
}
|
|
return flags;
|
|
}
|
|
word_ = Word::ConvertUnsigned(fraction).value;
|
|
if (lshift > 0) {
|
|
word_ = word_.SHIFTL(lshift);
|
|
if (roundingBits != nullptr) {
|
|
for (; lshift > 0; --lshift) {
|
|
if (roundingBits->ShiftLeft()) {
|
|
word_ = word_.IBSET(lshift - 1);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
if constexpr (implicitMSB) {
|
|
word_ = word_.IBCLR(significandBits);
|
|
}
|
|
word_ = word_.IOR(Word{exponent}.SHIFTL(significandBits));
|
|
if (negative) {
|
|
word_ = word_.IBSET(bits - 1);
|
|
}
|
|
return {};
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
RealFlags Real<W, P, IM>::Round(
|
|
Rounding rounding, const RoundingBits &bits, bool multiply) {
|
|
std::uint64_t origExponent{Exponent()};
|
|
RealFlags flags;
|
|
bool inexact{!bits.empty()};
|
|
if (inexact) {
|
|
flags.set(RealFlag::Inexact);
|
|
}
|
|
if (origExponent < maxExponent &&
|
|
bits.MustRound(rounding, IsNegative(), word_.BTEST(0) /* is odd */)) {
|
|
typename Fraction::ValueWithCarry sum{
|
|
GetFraction().AddUnsigned(Fraction{}, true)};
|
|
std::uint64_t newExponent{origExponent};
|
|
if (sum.carry) {
|
|
// The fraction was all ones before rounding; sum.value is now zero
|
|
sum.value = sum.value.IBSET(precision - 1);
|
|
if (++newExponent >= maxExponent) {
|
|
flags.set(RealFlag::Overflow); // rounded away to an infinity
|
|
}
|
|
}
|
|
flags |= Normalize(IsNegative(), newExponent, sum.value);
|
|
}
|
|
if (inexact && origExponent == 0) {
|
|
// inexact denormal input
|
|
if (Exponent() == 0) {
|
|
flags.set(RealFlag::Underflow); // output still denormal -> Underflow
|
|
} else {
|
|
// Rounding went up to the smallest normal number.
|
|
// Still signal Underflow unless we're in a weird x86 edge case with
|
|
// multiplication: if the sticky bit is set (i.e., the lower half of
|
|
// the full product had bits below the top 2), Underflow gets set in
|
|
// a directed rounding mode only if the guard bit was also set.
|
|
if (multiply && bits.sticky() &&
|
|
(bits.guard() ||
|
|
!(rounding == Rounding::Up || rounding == Rounding::Down))) {
|
|
} else {
|
|
flags.set(RealFlag::Underflow);
|
|
}
|
|
}
|
|
}
|
|
return flags;
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
void Real<W, P, IM>::NormalizeAndRound(ValueWithRealFlags<Real> &result,
|
|
bool isNegative, std::uint64_t exponent, const Fraction &fraction,
|
|
Rounding rounding, RoundingBits roundingBits, bool multiply) {
|
|
result.flags |= result.value.Normalize(
|
|
isNegative, exponent, fraction, rounding, &roundingBits);
|
|
result.flags |= result.value.Round(rounding, roundingBits, multiply);
|
|
}
|
|
|
|
template<typename W, int P, bool IM>
|
|
std::string Real<W, P, IM>::DumpHexadecimal() const {
|
|
if (IsNotANumber()) {
|
|
return "NaN 0x"s + word_.Hexadecimal();
|
|
} else if (IsNegative()) {
|
|
return "-"s + Negate().DumpHexadecimal();
|
|
} else if (IsInfinite()) {
|
|
return "Inf"s;
|
|
} else if (IsZero()) {
|
|
return "0.0"s;
|
|
} else {
|
|
Fraction frac{GetFraction()};
|
|
std::string result{"0x"};
|
|
char intPart = '0' + frac.BTEST(frac.bits - 1);
|
|
result += intPart;
|
|
result += '.';
|
|
int trailz{frac.TRAILZ()};
|
|
if (trailz >= frac.bits - 1) {
|
|
result += '0';
|
|
} else {
|
|
int remainingBits{frac.bits - 1 - trailz};
|
|
int wholeNybbles{remainingBits / 4};
|
|
int lostBits{remainingBits - 4 * wholeNybbles};
|
|
if (wholeNybbles > 0) {
|
|
std::string fracHex{frac.SHIFTR(trailz + lostBits)
|
|
.IAND(frac.MASKR(4 * wholeNybbles))
|
|
.Hexadecimal()};
|
|
std::size_t field = wholeNybbles;
|
|
if (fracHex.size() < field) {
|
|
result += std::string(field - fracHex.size(), '0');
|
|
}
|
|
result += fracHex;
|
|
}
|
|
if (lostBits > 0) {
|
|
result += frac.SHIFTR(trailz)
|
|
.IAND(frac.MASKR(lostBits))
|
|
.SHIFTL(4 - lostBits)
|
|
.Hexadecimal();
|
|
}
|
|
}
|
|
result += 'p';
|
|
int exponent = Exponent() - exponentBias;
|
|
result += Integer<32>{exponent}.SignedDecimal();
|
|
return result;
|
|
}
|
|
}
|
|
|
|
template class Real<Integer<16>, 11>;
|
|
template class Real<Integer<32>, 24>;
|
|
template class Real<Integer<64>, 53>;
|
|
template class Real<Integer<80>, 64, false>;
|
|
template class Real<Integer<128>, 112>;
|
|
} // namespace Fortran::evaluate::value
|