summaryrefslogtreecommitdiff
path: root/mfbt/FloatingPoint.h
blob: a2846ce298c4278f3160ab3b077d17b796722ed9 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* 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 <stdint.h>

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.
 */

namespace detail {

/*
 * These implementations assume float/double are 32/64-bit single/double
 * format number types compatible with the IEEE-754 standard.  C++ doesn't
 * require this, but we required it in implementations of these algorithms that
 * preceded this header, so we shouldn't break anything to continue doing so.
 */
template<typename T>
struct FloatingPointTrait;

template<>
struct FloatingPointTrait<float>
{
protected:
  typedef uint32_t Bits;

  static constexpr unsigned kExponentWidth = 8;
  static constexpr unsigned kSignificandWidth = 23;
};

template<>
struct FloatingPointTrait<double>
{
protected:
  typedef uint64_t Bits;

  static constexpr unsigned kExponentWidth = 11;
  static constexpr unsigned kSignificandWidth = 52;
};

} // namespace detail

/*
 *  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).
 *
 *  kExponentBias is the offset that is subtracted from the exponent when
 *  computing the value, i.e. one plus the opposite of the mininum possible
 *  exponent.
 *  kExponentShift is the shift that one needs to apply to retrieve the
 *  exponent component of the value.
 *
 *  kSignBit contains a bits mask. Bit-and-ing with this mask will result in
 *  obtaining the sign bit.
 *  kExponentBits contains the mask needed for obtaining the exponent bits and
 *  kSignificandBits 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<typename T>
struct FloatingPoint final : private detail::FloatingPointTrait<T>
{
private:
  using Base = detail::FloatingPointTrait<T>;

public:
  /**
   * An unsigned integral type suitable for accessing the bitwise representation
   * of T.
   */
  using Bits = typename Base::Bits;

  static_assert(sizeof(T) == sizeof(Bits), "Bits must be same size as T");

  /** The bit-width of the exponent component of T. */
  using Base::kExponentWidth;

  /** The bit-width of the significand component of T. */
  using Base::kSignificandWidth;

  static_assert(1 + kExponentWidth + kSignificandWidth ==
                CHAR_BIT * sizeof(T),
                "sign bit plus bit widths should sum to overall bit width");

  /**
   * The exponent field in an IEEE-754 floating point number consists of bits
   * encoding an unsigned number.  The *actual* represented exponent (for all
   * values finite and not denormal) is that value, minus a bias |kExponentBias|
   * so that a useful range of numbers is represented.
   */
  static constexpr unsigned kExponentBias = (1U << (kExponentWidth - 1)) - 1;

  /**
   * The amount by which the bits of the exponent-field in an IEEE-754 floating
   * point number are shifted from the LSB of the floating point type.
   */
  static constexpr unsigned kExponentShift = kSignificandWidth;

  /** The sign bit in the floating point representation. */
  static constexpr Bits kSignBit =
    static_cast<Bits>(1) << (CHAR_BIT * sizeof(Bits) - 1);

  /** The exponent bits in the floating point representation. */
  static constexpr Bits kExponentBits =
    ((static_cast<Bits>(1) << kExponentWidth) - 1) << kSignificandWidth;

  /** The significand bits in the floating point representation. */
  static constexpr Bits kSignificandBits =
    (static_cast<Bits>(1) << kSignificandWidth) - 1;

  static_assert((kSignBit & kExponentBits) == 0,
                "sign bit shouldn't overlap exponent bits");
  static_assert((kSignBit & kSignificandBits) == 0,
                "sign bit shouldn't overlap significand bits");
  static_assert((kExponentBits & kSignificandBits) == 0,
                "exponent bits shouldn't overlap significand bits");

  static_assert((kSignBit | kExponentBits | kSignificandBits) == ~Bits(0),
                "all bits accounted for");
};

/** Determines whether a float/double is NaN. */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsNaN(T aValue)
{
  /*
   * A float/double is NaN if all exponent bits are 1 and the significand
   * contains at least one non-zero bit.
   */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  return (BitwiseCast<Bits>(aValue) & Traits::kExponentBits) == Traits::kExponentBits &&
         (BitwiseCast<Bits>(aValue) & Traits::kSignificandBits) != 0;
}

/** Determines whether a float/double is +Infinity or -Infinity. */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsInfinite(T aValue)
{
  /* Infinities have all exponent bits set to 1 and an all-0 significand. */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return (bits & ~Traits::kSignBit) == Traits::kExponentBits;
}

/** Determines whether a float/double is not NaN or infinite. */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsFinite(T aValue)
{
  /*
   * NaN and Infinities are the only non-finite floats/doubles, and both have
   * all exponent bits set to 1.
   */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return (bits & Traits::kExponentBits) != Traits::kExponentBits;
}

/**
 * Determines whether a float/double is negative or -0.  It is an error
 * to call this method on a float/double which is NaN.
 */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsNegative(T aValue)
{
  MOZ_ASSERT(!IsNaN(aValue), "NaN does not have a sign");

  /* The sign bit is set if the double is negative. */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return (bits & Traits::kSignBit) != 0;
}

/** Determines whether a float/double represents -0. */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsNegativeZero(T aValue)
{
  /* Only the sign bit is set if the value is -0. */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return bits == Traits::kSignBit;
}

/** Determines wether a float/double represents +0. */
template<typename T>
static MOZ_ALWAYS_INLINE bool
IsPositiveZero(T aValue)
{
  /* All bits are zero if the value is +0. */
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return bits == 0;
}

/**
 * Returns 0 if a float/double is NaN or infinite;
 * otherwise, the float/double is returned.
 */
template<typename T>
static MOZ_ALWAYS_INLINE T
ToZeroIfNonfinite(T aValue)
{
  return IsFinite(aValue) ? aValue : 0;
}

/**
 * Returns the exponent portion of the float/double.
 *
 * Zero is not special-cased, so ExponentComponent(0.0) is
 * -int_fast16_t(Traits::kExponentBias).
 */
template<typename T>
static MOZ_ALWAYS_INLINE int_fast16_t
ExponentComponent(T aValue)
{
  /*
   * 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<T> Traits;
  typedef typename Traits::Bits Bits;
  Bits bits = BitwiseCast<Bits>(aValue);
  return int_fast16_t((bits & Traits::kExponentBits) >> Traits::kExponentShift) -
         int_fast16_t(Traits::kExponentBias);
}

/** Returns +Infinity. */
template<typename T>
static MOZ_ALWAYS_INLINE T
PositiveInfinity()
{
  /*
   * Positive infinity has all exponent bits set, sign bit set to 0, and no
   * significand.
   */
  typedef FloatingPoint<T> Traits;
  return BitwiseCast<T>(Traits::kExponentBits);
}

/** Returns -Infinity. */
template<typename T>
static MOZ_ALWAYS_INLINE T
NegativeInfinity()
{
  /*
   * Negative infinity has all exponent bits set, sign bit set to 1, and no
   * significand.
   */
  typedef FloatingPoint<T> Traits;
  return BitwiseCast<T>(Traits::kSignBit | Traits::kExponentBits);
}

/**
 * Computes the bit pattern for a NaN with the specified sign bit and
 * significand bits.
 */
template<typename T,
         int SignBit,
         typename FloatingPoint<T>::Bits Significand>
struct SpecificNaNBits
{
  using Traits = FloatingPoint<T>;

  static_assert(SignBit == 0 || SignBit == 1, "bad sign bit");
  static_assert((Significand & ~Traits::kSignificandBits) == 0,
                "significand must only have significand bits set");
  static_assert(Significand & Traits::kSignificandBits,
                "significand must be nonzero");

  static constexpr typename Traits::Bits value =
    (SignBit * Traits::kSignBit) | Traits::kExponentBits | Significand;
};

/**
 * Constructs a NaN value with the specified sign bit and significand bits.
 *
 * There is also a variant that returns the value directly.  In most cases, the
 * two variants should be identical.  However, in the specific case of x86
 * chips, the behavior differs: returning floating-point values directly is done
 * through the x87 stack, and x87 loads and stores turn signaling NaNs into
 * quiet NaNs... silently.  Returning floating-point values via outparam,
 * however, is done entirely within the SSE registers when SSE2 floating-point
 * is enabled in the compiler, which has semantics-preserving behavior you would
 * expect.
 *
 * If preserving the distinction between signaling NaNs and quiet NaNs is
 * important to you, you should use the outparam version.  In all other cases,
 * you should use the direct return version.
 */
template<typename T>
static MOZ_ALWAYS_INLINE void
SpecificNaN(int signbit, typename FloatingPoint<T>::Bits significand, T* result)
{
  typedef FloatingPoint<T> Traits;
  MOZ_ASSERT(signbit == 0 || signbit == 1);
  MOZ_ASSERT((significand & ~Traits::kSignificandBits) == 0);
  MOZ_ASSERT(significand & Traits::kSignificandBits);

  BitwiseCast<T>((signbit ? Traits::kSignBit : 0) |
                  Traits::kExponentBits |
                  significand,
                  result);
  MOZ_ASSERT(IsNaN(*result));
}

template<typename T>
static MOZ_ALWAYS_INLINE T
SpecificNaN(int signbit, typename FloatingPoint<T>::Bits significand)
{
  T t;
  SpecificNaN(signbit, significand, &t);
  return t;
}

/** Computes the smallest non-zero positive float/double value. */
template<typename T>
static MOZ_ALWAYS_INLINE T
MinNumberValue()
{
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  return BitwiseCast<T>(Bits(1));
}

/**
 * If aValue is equal to some int32_t value, set *aInt32 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<typename T>
static MOZ_ALWAYS_INLINE bool
NumberEqualsInt32(T aValue, int32_t* aInt32)
{
  /*
   * 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 aValue == (*aInt32 = int32_t(aValue));
}

/**
 * If d can be converted to int32_t and back to an identical double value,
 * set *aInt32 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<typename T>
static MOZ_ALWAYS_INLINE bool
NumberIsInt32(T aValue, int32_t* aInt32)
{
  return !IsNegativeZero(aValue) && NumberEqualsInt32(aValue, aInt32);
}

/**
 * Computes a NaN value.  Do not use this method if you depend upon a particular
 * NaN value being returned.
 */
template<typename T>
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<T> Traits;
  return SpecificNaN<T>(1, Traits::kSignificandBits);
}

/**
 * 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<typename T>
static inline bool
NumbersAreIdentical(T aValue1, T aValue2)
{
  typedef FloatingPoint<T> Traits;
  typedef typename Traits::Bits Bits;
  if (IsNaN(aValue1)) {
    return IsNaN(aValue2);
  }
  return BitwiseCast<Bits>(aValue1) == BitwiseCast<Bits>(aValue2);
}

/**
 * Return true if |aValue| and |aValue2| are equal (ignoring sign if both are
 * zero) or both NaN.
 */
template <typename T>
static inline bool
EqualOrBothNaN(T aValue1, T aValue2)
{
  if (IsNaN(aValue1)) {
    return IsNaN(aValue2);
  }
  return aValue1 == aValue2;
}

namespace detail {

template<typename T>
struct FuzzyEqualsEpsilon;

template<>
struct FuzzyEqualsEpsilon<float>
{
  // A number near 1e-5 that is exactly representable in a float.
  static float value() { return 1.0f / (1 << 17); }
};

template<>
struct FuzzyEqualsEpsilon<double>
{
  // A number near 1e-12 that is exactly representable in a double.
  static 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 aEpsilon apart. The default value of aEpsilon 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<typename T>
static MOZ_ALWAYS_INLINE bool
FuzzyEqualsAdditive(T aValue1, T aValue2,
                    T aEpsilon = detail::FuzzyEqualsEpsilon<T>::value())
{
  static_assert(IsFloatingPoint<T>::value, "floating point type required");
  return Abs(aValue1 - aValue2) <= aEpsilon;
}

/**
 * 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 aEpsilon apart, where the aEpsilon 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<typename T>
static MOZ_ALWAYS_INLINE bool
FuzzyEqualsMultiplicative(T aValue1, T aValue2,
                          T aEpsilon = detail::FuzzyEqualsEpsilon<T>::value())
{
  static_assert(IsFloatingPoint<T>::value, "floating point type required");
  // can't use std::min because of bug 965340
  T smaller = Abs(aValue1) < Abs(aValue2) ? Abs(aValue1) : Abs(aValue2);
  return Abs(aValue1 - aValue2) <= aEpsilon * 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_MUST_USE
extern MFBT_API bool
IsFloat32Representable(double aFloat32);

} /* namespace mozilla */

#endif /* mozilla_FloatingPoint_h */