/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */ /* vim: set ts=8 sts=2 et sw=2 tw=80: */ /* This Source Code Form is subject to the terms of the Mozilla Public * License, v. 2.0. If a copy of the MPL was not distributed with this * file, You can obtain one at http://mozilla.org/MPL/2.0/. */ /* Various predicates and operations on IEEE-754 floating point types. */ #ifndef mozilla_FloatingPoint_h #define mozilla_FloatingPoint_h #include "mozilla/Assertions.h" #include "mozilla/Attributes.h" #include "mozilla/Casting.h" #include "mozilla/MathAlgorithms.h" #include "mozilla/Types.h" #include namespace mozilla { /* * It's reasonable to ask why we have this header at all. Don't isnan, * copysign, the built-in comparison operators, and the like solve these * problems? Unfortunately, they don't. We've found that various compilers * (MSVC, MSVC when compiling with PGO, and GCC on OS X, at least) miscompile * the standard methods in various situations, so we can't use them. Some of * these compilers even have problems compiling seemingly reasonable bitwise * algorithms! But with some care we've found algorithms that seem to not * trigger those compiler bugs. * * For the aforementioned reasons, be very wary of making changes to any of * these algorithms. If you must make changes, keep a careful eye out for * compiler bustage, particularly PGO-specific bustage. */ struct FloatTypeTraits { typedef uint32_t Bits; static const unsigned ExponentBias = 127; static const unsigned ExponentShift = 23; static const Bits SignBit = 0x80000000UL; static const Bits ExponentBits = 0x7F800000UL; static const Bits SignificandBits = 0x007FFFFFUL; }; struct DoubleTypeTraits { typedef uint64_t Bits; static const unsigned ExponentBias = 1023; static const unsigned ExponentShift = 52; static const Bits SignBit = 0x8000000000000000ULL; static const Bits ExponentBits = 0x7ff0000000000000ULL; static const Bits SignificandBits = 0x000fffffffffffffULL; }; template struct SelectTrait; template<> struct SelectTrait : public FloatTypeTraits {}; template<> struct SelectTrait : public DoubleTypeTraits {}; /* * This struct contains details regarding the encoding of floating-point * numbers that can be useful for direct bit manipulation. As of now, the * template parameter has to be float or double. * * The nested typedef |Bits| is the unsigned integral type with the same size * as T: uint32_t for float and uint64_t for double (static assertions * double-check these assumptions). * * ExponentBias is the offset that is subtracted from the exponent when * computing the value, i.e. one plus the opposite of the mininum possible * exponent. * ExponentShift is the shift that one needs to apply to retrieve the exponent * component of the value. * * SignBit contains a bits mask. Bit-and-ing with this mask will result in * obtaining the sign bit. * ExponentBits contains the mask needed for obtaining the exponent bits and * SignificandBits contains the mask needed for obtaining the significand bits. * * Full details of how floating point number formats are encoded are beyond the * scope of this comment. For more information, see * http://en.wikipedia.org/wiki/IEEE_floating_point * http://en.wikipedia.org/wiki/Floating_point#IEEE_754:_floating_point_in_modern_computers */ template struct FloatingPoint : public SelectTrait { typedef SelectTrait Base; typedef typename Base::Bits Bits; static_assert((Base::SignBit & Base::ExponentBits) == 0, "sign bit shouldn't overlap exponent bits"); static_assert((Base::SignBit & Base::SignificandBits) == 0, "sign bit shouldn't overlap significand bits"); static_assert((Base::ExponentBits & Base::SignificandBits) == 0, "exponent bits shouldn't overlap significand bits"); static_assert((Base::SignBit | Base::ExponentBits | Base::SignificandBits) == ~Bits(0), "all bits accounted for"); /* * These implementations assume float/double are 32/64-bit single/double format * number types compatible with the IEEE-754 standard. C++ don't require this * to be the case. But we required this in implementations of these algorithms * that preceded this header, so we shouldn't break anything if we keep doing so. */ static_assert(sizeof(T) == sizeof(Bits), "Bits must be same size as T"); }; /** Determines whether a double is NaN. */ template static MOZ_ALWAYS_INLINE bool IsNaN(T t) { /* * A float/double is NaN if all exponent bits are 1 and the significand contains at * least one non-zero bit. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return (bits & Traits::ExponentBits) == Traits::ExponentBits && (bits & Traits::SignificandBits) != 0; } /** Determines whether a float/double is +Infinity or -Infinity. */ template static MOZ_ALWAYS_INLINE bool IsInfinite(T t) { /* Infinities have all exponent bits set to 1 and an all-0 significand. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return (bits & ~Traits::SignBit) == Traits::ExponentBits; } /** Determines whether a float/double is not NaN or infinite. */ template static MOZ_ALWAYS_INLINE bool IsFinite(T t) { /* * NaN and Infinities are the only non-finite floats/doubles, and both have all * exponent bits set to 1. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return (bits & Traits::ExponentBits) != Traits::ExponentBits; } /** * Determines whether a float/double is negative. It is an error to call this method * on a float/double which is NaN. */ template static MOZ_ALWAYS_INLINE bool IsNegative(T t) { MOZ_ASSERT(!IsNaN(t), "NaN does not have a sign"); /* The sign bit is set if the double is negative. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return (bits & Traits::SignBit) != 0; } /** Determines whether a float/double represents -0. */ template static MOZ_ALWAYS_INLINE bool IsNegativeZero(T t) { /* Only the sign bit is set if the value is -0. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return bits == Traits::SignBit; } /** * Returns the exponent portion of the float/double. * * Zero is not special-cased, so ExponentComponent(0.0) is * -int_fast16_t(Traits::ExponentBias). */ template static MOZ_ALWAYS_INLINE int_fast16_t ExponentComponent(T t) { /* * The exponent component of a float/double is an unsigned number, biased from its * actual value. Subtract the bias to retrieve the actual exponent. */ typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; Bits bits = BitwiseCast(t); return int_fast16_t((bits & Traits::ExponentBits) >> Traits::ExponentShift) - int_fast16_t(Traits::ExponentBias); } /** Returns +Infinity. */ template static MOZ_ALWAYS_INLINE T PositiveInfinity() { /* * Positive infinity has all exponent bits set, sign bit set to 0, and no * significand. */ typedef FloatingPoint Traits; return BitwiseCast(Traits::ExponentBits); } /** Returns -Infinity. */ template static MOZ_ALWAYS_INLINE T NegativeInfinity() { /* * Negative infinity has all exponent bits set, sign bit set to 1, and no * significand. */ typedef FloatingPoint Traits; return BitwiseCast(Traits::SignBit | Traits::ExponentBits); } /** Constructs a NaN value with the specified sign bit and significand bits. */ template static MOZ_ALWAYS_INLINE T SpecificNaN(int signbit, typename FloatingPoint::Bits significand) { typedef FloatingPoint Traits; MOZ_ASSERT(signbit == 0 || signbit == 1); MOZ_ASSERT((significand & ~Traits::SignificandBits) == 0); MOZ_ASSERT(significand & Traits::SignificandBits); T t = BitwiseCast((signbit ? Traits::SignBit : 0) | Traits::ExponentBits | significand); MOZ_ASSERT(IsNaN(t)); return t; } /** Computes the smallest non-zero positive float/double value. */ template static MOZ_ALWAYS_INLINE T MinNumberValue() { typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; return BitwiseCast(Bits(1)); } /** * If t is equal to some int32_t value, set *i to that value and return true; * otherwise return false. * * Note that negative zero is "equal" to zero here. To test whether a value can * be losslessly converted to int32_t and back, use NumberIsInt32 instead. */ template static MOZ_ALWAYS_INLINE bool NumberEqualsInt32(T t, int32_t* i) { /* * XXX Casting a floating-point value that doesn't truncate to int32_t, to * int32_t, induces undefined behavior. We should definitely fix this * (bug 744965), but as apparently it "works" in practice, it's not a * pressing concern now. */ return t == (*i = int32_t(t)); } /** * If d can be converted to int32_t and back to an identical double value, * set *i to that value and return true; otherwise return false. * * The difference between this and NumberEqualsInt32 is that this method returns * false for negative zero. */ template static MOZ_ALWAYS_INLINE bool NumberIsInt32(T t, int32_t* i) { return !IsNegativeZero(t) && NumberEqualsInt32(t, i); } /** * Computes a NaN value. Do not use this method if you depend upon a particular * NaN value being returned. */ template static MOZ_ALWAYS_INLINE T UnspecifiedNaN() { /* * If we can use any quiet NaN, we might as well use the all-ones NaN, * since it's cheap to materialize on common platforms (such as x64, where * this value can be represented in a 32-bit signed immediate field, allowing * it to be stored to memory in a single instruction). */ typedef FloatingPoint Traits; return SpecificNaN(1, Traits::SignificandBits); } /** * Compare two doubles for equality, *without* equating -0 to +0, and equating * any NaN value to any other NaN value. (The normal equality operators equate * -0 with +0, and they equate NaN to no other value.) */ template static inline bool NumbersAreIdentical(T t1, T t2) { typedef FloatingPoint Traits; typedef typename Traits::Bits Bits; if (IsNaN(t1)) return IsNaN(t2); return BitwiseCast(t1) == BitwiseCast(t2); } namespace detail { template struct FuzzyEqualsEpsilon; template<> struct FuzzyEqualsEpsilon { // A number near 1e-5 that is exactly representable in // floating point static const float value() { return 1.0f / (1 << 17); } }; template<> struct FuzzyEqualsEpsilon { // A number near 1e-12 that is exactly representable in // a double static const double value() { return 1.0 / (1LL << 40); } }; } // namespace detail /** * Compare two floating point values for equality, modulo rounding error. That * is, the two values are considered equal if they are both not NaN and if they * are less than or equal to epsilon apart. The default value of epsilon is near * 1e-5. * * For most scenarios you will want to use FuzzyEqualsMultiplicative instead, * as it is more reasonable over the entire range of floating point numbers. * This additive version should only be used if you know the range of the numbers * you are dealing with is bounded and stays around the same order of magnitude. */ template static MOZ_ALWAYS_INLINE bool FuzzyEqualsAdditive(T val1, T val2, T epsilon = detail::FuzzyEqualsEpsilon::value()) { static_assert(IsFloatingPoint::value, "floating point type required"); return Abs(val1 - val2) <= epsilon; } /** * Compare two floating point values for equality, allowing for rounding error * relative to the magnitude of the values. That is, the two values are * considered equal if they are both not NaN and they are less than or equal to * some epsilon apart, where the epsilon is scaled by the smaller of the two * argument values. * * In most cases you will want to use this rather than FuzzyEqualsAdditive, as * this function effectively masks out differences in the bottom few bits of * the floating point numbers being compared, regardless of what order of magnitude * those numbers are at. */ template static MOZ_ALWAYS_INLINE bool FuzzyEqualsMultiplicative(T val1, T val2, T epsilon = detail::FuzzyEqualsEpsilon::value()) { static_assert(IsFloatingPoint::value, "floating point type required"); // can't use std::min because of bug 965340 T smaller = Abs(val1) < Abs(val2) ? Abs(val1) : Abs(val2); return Abs(val1 - val2) <= epsilon * smaller; } /** * Returns true if the given value can be losslessly represented as an IEEE-754 * single format number, false otherwise. All NaN values are considered * representable (notwithstanding that the exact bit pattern of a double format * NaN value can't be exactly represented in single format). * * This function isn't inlined to avoid buggy optimizations by MSVC. */ MOZ_WARN_UNUSED_RESULT extern MFBT_API bool IsFloat32Representable(double x); } /* namespace mozilla */ #endif /* mozilla_FloatingPoint_h */