No bug, Manual HPKP preload list update. a=hpkp-update
[gecko.git] / mfbt / FloatingPoint.h
blob75b75df3f13945ec11bb8cb56e4b925d8a152c8d
1 /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
2 /* vim: set ts=8 sts=2 et sw=2 tw=80: */
3 /* This Source Code Form is subject to the terms of the Mozilla Public
4 * License, v. 2.0. If a copy of the MPL was not distributed with this
5 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
7 /* Various predicates and operations on IEEE-754 floating point types. */
9 #ifndef mozilla_FloatingPoint_h
10 #define mozilla_FloatingPoint_h
12 #include "mozilla/Assertions.h"
13 #include "mozilla/Attributes.h"
14 #include "mozilla/Casting.h"
15 #include "mozilla/MathAlgorithms.h"
16 #include "mozilla/Types.h"
18 #include <stdint.h>
20 namespace mozilla {
23 * It's reasonable to ask why we have this header at all. Don't isnan,
24 * copysign, the built-in comparison operators, and the like solve these
25 * problems? Unfortunately, they don't. We've found that various compilers
26 * (MSVC, MSVC when compiling with PGO, and GCC on OS X, at least) miscompile
27 * the standard methods in various situations, so we can't use them. Some of
28 * these compilers even have problems compiling seemingly reasonable bitwise
29 * algorithms! But with some care we've found algorithms that seem to not
30 * trigger those compiler bugs.
32 * For the aforementioned reasons, be very wary of making changes to any of
33 * these algorithms. If you must make changes, keep a careful eye out for
34 * compiler bustage, particularly PGO-specific bustage.
37 struct FloatTypeTraits
39 typedef uint32_t Bits;
41 static const unsigned kExponentBias = 127;
42 static const unsigned kExponentShift = 23;
44 static const Bits kSignBit = 0x80000000UL;
45 static const Bits kExponentBits = 0x7F800000UL;
46 static const Bits kSignificandBits = 0x007FFFFFUL;
49 struct DoubleTypeTraits
51 typedef uint64_t Bits;
53 static const unsigned kExponentBias = 1023;
54 static const unsigned kExponentShift = 52;
56 static const Bits kSignBit = 0x8000000000000000ULL;
57 static const Bits kExponentBits = 0x7ff0000000000000ULL;
58 static const Bits kSignificandBits = 0x000fffffffffffffULL;
61 template<typename T> struct SelectTrait;
62 template<> struct SelectTrait<float> : public FloatTypeTraits {};
63 template<> struct SelectTrait<double> : public DoubleTypeTraits {};
66 * This struct contains details regarding the encoding of floating-point
67 * numbers that can be useful for direct bit manipulation. As of now, the
68 * template parameter has to be float or double.
70 * The nested typedef |Bits| is the unsigned integral type with the same size
71 * as T: uint32_t for float and uint64_t for double (static assertions
72 * double-check these assumptions).
74 * kExponentBias is the offset that is subtracted from the exponent when
75 * computing the value, i.e. one plus the opposite of the mininum possible
76 * exponent.
77 * kExponentShift is the shift that one needs to apply to retrieve the
78 * exponent component of the value.
80 * kSignBit contains a bits mask. Bit-and-ing with this mask will result in
81 * obtaining the sign bit.
82 * kExponentBits contains the mask needed for obtaining the exponent bits and
83 * kSignificandBits contains the mask needed for obtaining the significand
84 * bits.
86 * Full details of how floating point number formats are encoded are beyond
87 * the scope of this comment. For more information, see
88 * http://en.wikipedia.org/wiki/IEEE_floating_point
89 * http://en.wikipedia.org/wiki/Floating_point#IEEE_754:_floating_point_in_modern_computers
91 template<typename T>
92 struct FloatingPoint : public SelectTrait<T>
94 typedef SelectTrait<T> Base;
95 typedef typename Base::Bits Bits;
97 static_assert((Base::kSignBit & Base::kExponentBits) == 0,
98 "sign bit shouldn't overlap exponent bits");
99 static_assert((Base::kSignBit & Base::kSignificandBits) == 0,
100 "sign bit shouldn't overlap significand bits");
101 static_assert((Base::kExponentBits & Base::kSignificandBits) == 0,
102 "exponent bits shouldn't overlap significand bits");
104 static_assert((Base::kSignBit | Base::kExponentBits | Base::kSignificandBits) ==
105 ~Bits(0),
106 "all bits accounted for");
109 * These implementations assume float/double are 32/64-bit single/double
110 * format number types compatible with the IEEE-754 standard. C++ don't
111 * require this to be the case. But we required this in implementations of
112 * these algorithms that preceded this header, so we shouldn't break anything
113 * if we keep doing so.
115 static_assert(sizeof(T) == sizeof(Bits), "Bits must be same size as T");
118 /** Determines whether a double is NaN. */
119 template<typename T>
120 static MOZ_ALWAYS_INLINE bool
121 IsNaN(T aValue)
124 * A float/double is NaN if all exponent bits are 1 and the significand
125 * contains at least one non-zero bit.
127 typedef FloatingPoint<T> Traits;
128 typedef typename Traits::Bits Bits;
129 Bits bits = BitwiseCast<Bits>(aValue);
130 return (bits & Traits::kExponentBits) == Traits::kExponentBits &&
131 (bits & Traits::kSignificandBits) != 0;
134 /** Determines whether a float/double is +Infinity or -Infinity. */
135 template<typename T>
136 static MOZ_ALWAYS_INLINE bool
137 IsInfinite(T aValue)
139 /* Infinities have all exponent bits set to 1 and an all-0 significand. */
140 typedef FloatingPoint<T> Traits;
141 typedef typename Traits::Bits Bits;
142 Bits bits = BitwiseCast<Bits>(aValue);
143 return (bits & ~Traits::kSignBit) == Traits::kExponentBits;
146 /** Determines whether a float/double is not NaN or infinite. */
147 template<typename T>
148 static MOZ_ALWAYS_INLINE bool
149 IsFinite(T aValue)
152 * NaN and Infinities are the only non-finite floats/doubles, and both have
153 * all exponent bits set to 1.
155 typedef FloatingPoint<T> Traits;
156 typedef typename Traits::Bits Bits;
157 Bits bits = BitwiseCast<Bits>(aValue);
158 return (bits & Traits::kExponentBits) != Traits::kExponentBits;
162 * Determines whether a float/double is negative. It is an error to call this
163 * method on a float/double which is NaN.
165 template<typename T>
166 static MOZ_ALWAYS_INLINE bool
167 IsNegative(T aValue)
169 MOZ_ASSERT(!IsNaN(aValue), "NaN does not have a sign");
171 /* The sign bit is set if the double is negative. */
172 typedef FloatingPoint<T> Traits;
173 typedef typename Traits::Bits Bits;
174 Bits bits = BitwiseCast<Bits>(aValue);
175 return (bits & Traits::kSignBit) != 0;
178 /** Determines whether a float/double represents -0. */
179 template<typename T>
180 static MOZ_ALWAYS_INLINE bool
181 IsNegativeZero(T aValue)
183 /* Only the sign bit is set if the value is -0. */
184 typedef FloatingPoint<T> Traits;
185 typedef typename Traits::Bits Bits;
186 Bits bits = BitwiseCast<Bits>(aValue);
187 return bits == Traits::kSignBit;
191 * Returns 0 if a float/double is NaN or infinite;
192 * otherwise, the float/double is returned.
194 template<typename T>
195 static MOZ_ALWAYS_INLINE T
196 ToZeroIfNonfinite(T aValue)
198 return IsFinite(aValue) ? aValue : 0;
202 * Returns the exponent portion of the float/double.
204 * Zero is not special-cased, so ExponentComponent(0.0) is
205 * -int_fast16_t(Traits::kExponentBias).
207 template<typename T>
208 static MOZ_ALWAYS_INLINE int_fast16_t
209 ExponentComponent(T aValue)
212 * The exponent component of a float/double is an unsigned number, biased
213 * from its actual value. Subtract the bias to retrieve the actual exponent.
215 typedef FloatingPoint<T> Traits;
216 typedef typename Traits::Bits Bits;
217 Bits bits = BitwiseCast<Bits>(aValue);
218 return int_fast16_t((bits & Traits::kExponentBits) >> Traits::kExponentShift) -
219 int_fast16_t(Traits::kExponentBias);
222 /** Returns +Infinity. */
223 template<typename T>
224 static MOZ_ALWAYS_INLINE T
225 PositiveInfinity()
228 * Positive infinity has all exponent bits set, sign bit set to 0, and no
229 * significand.
231 typedef FloatingPoint<T> Traits;
232 return BitwiseCast<T>(Traits::kExponentBits);
235 /** Returns -Infinity. */
236 template<typename T>
237 static MOZ_ALWAYS_INLINE T
238 NegativeInfinity()
241 * Negative infinity has all exponent bits set, sign bit set to 1, and no
242 * significand.
244 typedef FloatingPoint<T> Traits;
245 return BitwiseCast<T>(Traits::kSignBit | Traits::kExponentBits);
249 /** Constructs a NaN value with the specified sign bit and significand bits. */
250 template<typename T>
251 static MOZ_ALWAYS_INLINE T
252 SpecificNaN(int signbit, typename FloatingPoint<T>::Bits significand)
254 typedef FloatingPoint<T> Traits;
255 MOZ_ASSERT(signbit == 0 || signbit == 1);
256 MOZ_ASSERT((significand & ~Traits::kSignificandBits) == 0);
257 MOZ_ASSERT(significand & Traits::kSignificandBits);
259 T t = BitwiseCast<T>((signbit ? Traits::kSignBit : 0) |
260 Traits::kExponentBits |
261 significand);
262 MOZ_ASSERT(IsNaN(t));
263 return t;
266 /** Computes the smallest non-zero positive float/double value. */
267 template<typename T>
268 static MOZ_ALWAYS_INLINE T
269 MinNumberValue()
271 typedef FloatingPoint<T> Traits;
272 typedef typename Traits::Bits Bits;
273 return BitwiseCast<T>(Bits(1));
277 * If aValue is equal to some int32_t value, set *aInt32 to that value and
278 * return true; otherwise return false.
280 * Note that negative zero is "equal" to zero here. To test whether a value can
281 * be losslessly converted to int32_t and back, use NumberIsInt32 instead.
283 template<typename T>
284 static MOZ_ALWAYS_INLINE bool
285 NumberEqualsInt32(T aValue, int32_t* aInt32)
288 * XXX Casting a floating-point value that doesn't truncate to int32_t, to
289 * int32_t, induces undefined behavior. We should definitely fix this
290 * (bug 744965), but as apparently it "works" in practice, it's not a
291 * pressing concern now.
293 return aValue == (*aInt32 = int32_t(aValue));
297 * If d can be converted to int32_t and back to an identical double value,
298 * set *aInt32 to that value and return true; otherwise return false.
300 * The difference between this and NumberEqualsInt32 is that this method returns
301 * false for negative zero.
303 template<typename T>
304 static MOZ_ALWAYS_INLINE bool
305 NumberIsInt32(T aValue, int32_t* aInt32)
307 return !IsNegativeZero(aValue) && NumberEqualsInt32(aValue, aInt32);
311 * Computes a NaN value. Do not use this method if you depend upon a particular
312 * NaN value being returned.
314 template<typename T>
315 static MOZ_ALWAYS_INLINE T
316 UnspecifiedNaN()
319 * If we can use any quiet NaN, we might as well use the all-ones NaN,
320 * since it's cheap to materialize on common platforms (such as x64, where
321 * this value can be represented in a 32-bit signed immediate field, allowing
322 * it to be stored to memory in a single instruction).
324 typedef FloatingPoint<T> Traits;
325 return SpecificNaN<T>(1, Traits::kSignificandBits);
329 * Compare two doubles for equality, *without* equating -0 to +0, and equating
330 * any NaN value to any other NaN value. (The normal equality operators equate
331 * -0 with +0, and they equate NaN to no other value.)
333 template<typename T>
334 static inline bool
335 NumbersAreIdentical(T aValue1, T aValue2)
337 typedef FloatingPoint<T> Traits;
338 typedef typename Traits::Bits Bits;
339 if (IsNaN(aValue1)) {
340 return IsNaN(aValue2);
342 return BitwiseCast<Bits>(aValue1) == BitwiseCast<Bits>(aValue2);
345 namespace detail {
347 template<typename T>
348 struct FuzzyEqualsEpsilon;
350 template<>
351 struct FuzzyEqualsEpsilon<float>
353 // A number near 1e-5 that is exactly representable in a float.
354 static float value() { return 1.0f / (1 << 17); }
357 template<>
358 struct FuzzyEqualsEpsilon<double>
360 // A number near 1e-12 that is exactly representable in a double.
361 static double value() { return 1.0 / (1LL << 40); }
364 } // namespace detail
367 * Compare two floating point values for equality, modulo rounding error. That
368 * is, the two values are considered equal if they are both not NaN and if they
369 * are less than or equal to aEpsilon apart. The default value of aEpsilon is
370 * near 1e-5.
372 * For most scenarios you will want to use FuzzyEqualsMultiplicative instead,
373 * as it is more reasonable over the entire range of floating point numbers.
374 * This additive version should only be used if you know the range of the
375 * numbers you are dealing with is bounded and stays around the same order of
376 * magnitude.
378 template<typename T>
379 static MOZ_ALWAYS_INLINE bool
380 FuzzyEqualsAdditive(T aValue1, T aValue2,
381 T aEpsilon = detail::FuzzyEqualsEpsilon<T>::value())
383 static_assert(IsFloatingPoint<T>::value, "floating point type required");
384 return Abs(aValue1 - aValue2) <= aEpsilon;
388 * Compare two floating point values for equality, allowing for rounding error
389 * relative to the magnitude of the values. That is, the two values are
390 * considered equal if they are both not NaN and they are less than or equal to
391 * some aEpsilon apart, where the aEpsilon is scaled by the smaller of the two
392 * argument values.
394 * In most cases you will want to use this rather than FuzzyEqualsAdditive, as
395 * this function effectively masks out differences in the bottom few bits of
396 * the floating point numbers being compared, regardless of what order of
397 * magnitude those numbers are at.
399 template<typename T>
400 static MOZ_ALWAYS_INLINE bool
401 FuzzyEqualsMultiplicative(T aValue1, T aValue2,
402 T aEpsilon = detail::FuzzyEqualsEpsilon<T>::value())
404 static_assert(IsFloatingPoint<T>::value, "floating point type required");
405 // can't use std::min because of bug 965340
406 T smaller = Abs(aValue1) < Abs(aValue2) ? Abs(aValue1) : Abs(aValue2);
407 return Abs(aValue1 - aValue2) <= aEpsilon * smaller;
411 * Returns true if the given value can be losslessly represented as an IEEE-754
412 * single format number, false otherwise. All NaN values are considered
413 * representable (notwithstanding that the exact bit pattern of a double format
414 * NaN value can't be exactly represented in single format).
416 * This function isn't inlined to avoid buggy optimizations by MSVC.
418 MOZ_WARN_UNUSED_RESULT
419 extern MFBT_API bool
420 IsFloat32Representable(double aFloat32);
422 } /* namespace mozilla */
424 #endif /* mozilla_FloatingPoint_h */