summaryrefslogtreecommitdiff
path: root/boost/math/special_functions/next.hpp
blob: a63983e1c3d76a813d1133cb532152d6538188d8 (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
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
//  (C) Copyright John Maddock 2008.
//  Use, modification and distribution are subject to the
//  Boost Software License, Version 1.0. (See accompanying file
//  LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

#ifndef BOOST_MATH_SPECIAL_NEXT_HPP
#define BOOST_MATH_SPECIAL_NEXT_HPP

#ifdef _MSC_VER
#pragma once
#endif

#include <boost/math/special_functions/math_fwd.hpp>
#include <boost/math/policies/error_handling.hpp>
#include <boost/math/special_functions/fpclassify.hpp>
#include <boost/math/special_functions/sign.hpp>
#include <boost/math/special_functions/trunc.hpp>

#include <float.h>

#if !defined(_CRAYC) && !defined(__CUDACC__) && (!defined(__GNUC__) || (__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ > 3)))
#if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || defined(__SSE2__)
#include "xmmintrin.h"
#define BOOST_MATH_CHECK_SSE2
#endif
#endif

namespace boost{ namespace math{

   namespace concepts {

      class real_concept;
      class std_real_concept;

   }

namespace detail{

template <class T>
struct has_hidden_guard_digits;
template <>
struct has_hidden_guard_digits<float> : public mpl::false_ {};
template <>
struct has_hidden_guard_digits<double> : public mpl::false_ {};
template <>
struct has_hidden_guard_digits<long double> : public mpl::false_ {};
#ifdef BOOST_HAS_FLOAT128
template <>
struct has_hidden_guard_digits<__float128> : public mpl::false_ {};
#endif
template <>
struct has_hidden_guard_digits<boost::math::concepts::real_concept> : public mpl::false_ {};
template <>
struct has_hidden_guard_digits<boost::math::concepts::std_real_concept> : public mpl::false_ {};

template <class T, bool b>
struct has_hidden_guard_digits_10 : public mpl::false_ {};
template <class T>
struct has_hidden_guard_digits_10<T, true> : public mpl::bool_<(std::numeric_limits<T>::digits10 != std::numeric_limits<T>::max_digits10)> {};

template <class T>
struct has_hidden_guard_digits 
   : public has_hidden_guard_digits_10<T, 
   std::numeric_limits<T>::is_specialized
   && (std::numeric_limits<T>::radix == 10) >
{};

template <class T>
inline const T& normalize_value(const T& val, const mpl::false_&) { return val; }
template <class T>
inline T normalize_value(const T& val, const mpl::true_&) 
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   boost::intmax_t shift = std::numeric_limits<T>::digits - ilogb(val) - 1;
   T result = scalbn(val, shift);
   result = round(result);
   return scalbn(result, -shift); 
}

template <class T>
inline T get_smallest_value(mpl::true_ const&)
{
   //
   // numeric_limits lies about denorms being present - particularly
   // when this can be turned on or off at runtime, as is the case
   // when using the SSE2 registers in DAZ or FTZ mode.
   //
   static const T m = std::numeric_limits<T>::denorm_min();
#ifdef BOOST_MATH_CHECK_SSE2
   return (_mm_getcsr() & (_MM_FLUSH_ZERO_ON | 0x40)) ? tools::min_value<T>() : m;;
#else
   return ((tools::min_value<T>() / 2) == 0) ? tools::min_value<T>() : m;
#endif
}

template <class T>
inline T get_smallest_value(mpl::false_ const&)
{
   return tools::min_value<T>();
}

template <class T>
inline T get_smallest_value()
{
#if defined(BOOST_MSVC) && (BOOST_MSVC <= 1310)
   return get_smallest_value<T>(mpl::bool_<std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == 1)>());
#else
   return get_smallest_value<T>(mpl::bool_<std::numeric_limits<T>::is_specialized && (std::numeric_limits<T>::has_denorm == std::denorm_present)>());
#endif
}

//
// Returns the smallest value that won't generate denorms when
// we calculate the value of the least-significant-bit:
//
template <class T>
T get_min_shift_value();

template <class T>
struct min_shift_initializer
{
   struct init
   {
      init()
      {
         do_init();
      }
      static void do_init()
      {
         get_min_shift_value<T>();
      }
      void force_instantiate()const{}
   };
   static const init initializer;
   static void force_instantiate()
   {
      initializer.force_instantiate();
   }
};

template <class T>
const typename min_shift_initializer<T>::init min_shift_initializer<T>::initializer;

template <class T>
inline T calc_min_shifted(const mpl::true_&)
{
   BOOST_MATH_STD_USING
   return ldexp(tools::min_value<T>(), tools::digits<T>() + 1);
}
template <class T>
inline T calc_min_shifted(const mpl::false_&)
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   return scalbn(tools::min_value<T>(), std::numeric_limits<T>::digits + 1);
}


template <class T>
inline T get_min_shift_value()
{
   static const T val = calc_min_shifted<T>(mpl::bool_<!std::numeric_limits<T>::is_specialized || std::numeric_limits<T>::radix == 2>());
   min_shift_initializer<T>::force_instantiate();

   return val;
}

template <class T, class Policy>
T float_next_imp(const T& val, const mpl::true_&, const Policy& pol)
{
   BOOST_MATH_STD_USING
   int expon;
   static const char* function = "float_next<%1%>(%1%)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
   {
      if(val < 0)
         return -tools::max_value<T>();
      return policies::raise_domain_error<T>(
         function,
         "Argument must be finite, but got %1%", val, pol);
   }

   if(val >= tools::max_value<T>())
      return policies::raise_overflow_error<T>(function, 0, pol);

   if(val == 0)
      return detail::get_smallest_value<T>();

   if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>()))
   {
      //
      // Special case: if the value of the least significant bit is a denorm, and the result
      // would not be a denorm, then shift the input, increment, and shift back.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      return ldexp(float_next(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>());
   }

   if(-0.5f == frexp(val, &expon))
      --expon; // reduce exponent when val is a power of two, and negative.
   T diff = ldexp(T(1), expon - tools::digits<T>());
   if(diff == 0)
      diff = detail::get_smallest_value<T>();
   return val + diff;
} // float_next_imp
//
// Special version for some base other than 2:
//
template <class T, class Policy>
T float_next_imp(const T& val, const mpl::false_&, const Policy& pol)
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   BOOST_MATH_STD_USING
   boost::intmax_t expon;
   static const char* function = "float_next<%1%>(%1%)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
   {
      if(val < 0)
         return -tools::max_value<T>();
      return policies::raise_domain_error<T>(
         function,
         "Argument must be finite, but got %1%", val, pol);
   }

   if(val >= tools::max_value<T>())
      return policies::raise_overflow_error<T>(function, 0, pol);

   if(val == 0)
      return detail::get_smallest_value<T>();

   if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != -tools::min_value<T>()))
   {
      //
      // Special case: if the value of the least significant bit is a denorm, and the result
      // would not be a denorm, then shift the input, increment, and shift back.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      return scalbn(float_next(T(scalbn(val, 2 * std::numeric_limits<T>::digits)), pol), -2 * std::numeric_limits<T>::digits);
   }

   expon = 1 + ilogb(val);
   if(-1 == scalbn(val, -expon) * std::numeric_limits<T>::radix)
      --expon; // reduce exponent when val is a power of base, and negative.
   T diff = scalbn(T(1), expon - std::numeric_limits<T>::digits);
   if(diff == 0)
      diff = detail::get_smallest_value<T>();
   return val + diff;
} // float_next_imp

} // namespace detail

template <class T, class Policy>
inline typename tools::promote_args<T>::type float_next(const T& val, const Policy& pol)
{
   typedef typename tools::promote_args<T>::type result_type;
   return detail::float_next_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), mpl::bool_<!std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol);
}

#if 0 //def BOOST_MSVC
//
// We used to use ::_nextafter here, but doing so fails when using
// the SSE2 registers if the FTZ or DAZ flags are set, so use our own
// - albeit slower - code instead as at least that gives the correct answer.
//
template <class Policy>
inline double float_next(const double& val, const Policy& pol)
{
   static const char* function = "float_next<%1%>(%1%)";

   if(!(boost::math::isfinite)(val) && (val > 0))
      return policies::raise_domain_error<double>(
         function,
         "Argument must be finite, but got %1%", val, pol);

   if(val >= tools::max_value<double>())
      return policies::raise_overflow_error<double>(function, 0, pol);

   return ::_nextafter(val, tools::max_value<double>());
}
#endif

template <class T>
inline typename tools::promote_args<T>::type float_next(const T& val)
{
   return float_next(val, policies::policy<>());
}

namespace detail{

template <class T, class Policy>
T float_prior_imp(const T& val, const mpl::true_&, const Policy& pol)
{
   BOOST_MATH_STD_USING
   int expon;
   static const char* function = "float_prior<%1%>(%1%)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
   {
      if(val > 0)
         return tools::max_value<T>();
      return policies::raise_domain_error<T>(
         function,
         "Argument must be finite, but got %1%", val, pol);
   }

   if(val <= -tools::max_value<T>())
      return -policies::raise_overflow_error<T>(function, 0, pol);

   if(val == 0)
      return -detail::get_smallest_value<T>();

   if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>()))
   {
      //
      // Special case: if the value of the least significant bit is a denorm, and the result
      // would not be a denorm, then shift the input, increment, and shift back.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      return ldexp(float_prior(T(ldexp(val, 2 * tools::digits<T>())), pol), -2 * tools::digits<T>());
   }

   T remain = frexp(val, &expon);
   if(remain == 0.5f)
      --expon; // when val is a power of two we must reduce the exponent
   T diff = ldexp(T(1), expon - tools::digits<T>());
   if(diff == 0)
      diff = detail::get_smallest_value<T>();
   return val - diff;
} // float_prior_imp
//
// Special version for bases other than 2:
//
template <class T, class Policy>
T float_prior_imp(const T& val, const mpl::false_&, const Policy& pol)
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   BOOST_MATH_STD_USING
   boost::intmax_t expon;
   static const char* function = "float_prior<%1%>(%1%)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
   {
      if(val > 0)
         return tools::max_value<T>();
      return policies::raise_domain_error<T>(
         function,
         "Argument must be finite, but got %1%", val, pol);
   }

   if(val <= -tools::max_value<T>())
      return -policies::raise_overflow_error<T>(function, 0, pol);

   if(val == 0)
      return -detail::get_smallest_value<T>();

   if((fpclass != (int)FP_SUBNORMAL) && (fpclass != (int)FP_ZERO) && (fabs(val) < detail::get_min_shift_value<T>()) && (val != tools::min_value<T>()))
   {
      //
      // Special case: if the value of the least significant bit is a denorm, and the result
      // would not be a denorm, then shift the input, increment, and shift back.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      return scalbn(float_prior(T(scalbn(val, 2 * std::numeric_limits<T>::digits)), pol), -2 * std::numeric_limits<T>::digits);
   }

   expon = 1 + ilogb(val);
   T remain = scalbn(val, -expon);
   if(remain * std::numeric_limits<T>::radix == 1)
      --expon; // when val is a power of two we must reduce the exponent
   T diff = scalbn(T(1), expon - std::numeric_limits<T>::digits);
   if(diff == 0)
      diff = detail::get_smallest_value<T>();
   return val - diff;
} // float_prior_imp

} // namespace detail

template <class T, class Policy>
inline typename tools::promote_args<T>::type float_prior(const T& val, const Policy& pol)
{
   typedef typename tools::promote_args<T>::type result_type;
   return detail::float_prior_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), mpl::bool_<!std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol);
}

#if 0 //def BOOST_MSVC
//
// We used to use ::_nextafter here, but doing so fails when using
// the SSE2 registers if the FTZ or DAZ flags are set, so use our own
// - albeit slower - code instead as at least that gives the correct answer.
//
template <class Policy>
inline double float_prior(const double& val, const Policy& pol)
{
   static const char* function = "float_prior<%1%>(%1%)";

   if(!(boost::math::isfinite)(val) && (val < 0))
      return policies::raise_domain_error<double>(
         function,
         "Argument must be finite, but got %1%", val, pol);

   if(val <= -tools::max_value<double>())
      return -policies::raise_overflow_error<double>(function, 0, pol);

   return ::_nextafter(val, -tools::max_value<double>());
}
#endif

template <class T>
inline typename tools::promote_args<T>::type float_prior(const T& val)
{
   return float_prior(val, policies::policy<>());
}

template <class T, class U, class Policy>
inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction, const Policy& pol)
{
   typedef typename tools::promote_args<T, U>::type result_type;
   return val < direction ? boost::math::float_next<result_type>(val, pol) : val == direction ? val : boost::math::float_prior<result_type>(val, pol);
}

template <class T, class U>
inline typename tools::promote_args<T, U>::type nextafter(const T& val, const U& direction)
{
   return nextafter(val, direction, policies::policy<>());
}

namespace detail{

template <class T, class Policy>
T float_distance_imp(const T& a, const T& b, const mpl::true_&, const Policy& pol)
{
   BOOST_MATH_STD_USING
   //
   // Error handling:
   //
   static const char* function = "float_distance<%1%>(%1%, %1%)";
   if(!(boost::math::isfinite)(a))
      return policies::raise_domain_error<T>(
         function,
         "Argument a must be finite, but got %1%", a, pol);
   if(!(boost::math::isfinite)(b))
      return policies::raise_domain_error<T>(
         function,
         "Argument b must be finite, but got %1%", b, pol);
   //
   // Special cases:
   //
   if(a > b)
      return -float_distance(b, a, pol);
   if(a == b)
      return T(0);
   if(a == 0)
      return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol));
   if(b == 0)
      return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
   if(boost::math::sign(a) != boost::math::sign(b))
      return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol))
         + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
   //
   // By the time we get here, both a and b must have the same sign, we want
   // b > a and both postive for the following logic:
   //
   if(a < 0)
      return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol);

   BOOST_ASSERT(a >= 0);
   BOOST_ASSERT(b >= a);

   int expon;
   //
   // Note that if a is a denorm then the usual formula fails
   // because we actually have fewer than tools::digits<T>()
   // significant bits in the representation:
   //
   frexp(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) ? tools::min_value<T>() : a, &expon);
   T upper = ldexp(T(1), expon);
   T result = T(0);
   //
   // If b is greater than upper, then we *must* split the calculation
   // as the size of the ULP changes with each order of magnitude change:
   //
   if(b > upper)
   {
      int expon2;
      frexp(b, &expon2);
      T upper2 = ldexp(T(0.5), expon2);
      result = float_distance(upper2, b);
      result += (expon2 - expon - 1) * ldexp(T(1), tools::digits<T>() - 1);
   }
   //
   // Use compensated double-double addition to avoid rounding
   // errors in the subtraction:
   //
   expon = tools::digits<T>() - expon;
   T mb, x, y, z;
   if(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) || (b - a < tools::min_value<T>()))
   {
      //
      // Special case - either one end of the range is a denormal, or else the difference is.
      // The regular code will fail if we're using the SSE2 registers on Intel and either
      // the FTZ or DAZ flags are set.
      //
      T a2 = ldexp(a, tools::digits<T>());
      T b2 = ldexp(b, tools::digits<T>());
      mb = -(std::min)(T(ldexp(upper, tools::digits<T>())), b2);
      x = a2 + mb;
      z = x - a2;
      y = (a2 - (x - z)) + (mb - z);

      expon -= tools::digits<T>();
   }
   else
   {
      mb = -(std::min)(upper, b);
      x = a + mb;
      z = x - a;
      y = (a - (x - z)) + (mb - z);
   }
   if(x < 0)
   {
      x = -x;
      y = -y;
   }
   result += ldexp(x, expon) + ldexp(y, expon);
   //
   // Result must be an integer:
   //
   BOOST_ASSERT(result == floor(result));
   return result;
} // float_distance_imp
//
// Special versions for bases other than 2:
//
template <class T, class Policy>
T float_distance_imp(const T& a, const T& b, const mpl::false_&, const Policy& pol)
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   BOOST_MATH_STD_USING
   //
   // Error handling:
   //
   static const char* function = "float_distance<%1%>(%1%, %1%)";
   if(!(boost::math::isfinite)(a))
      return policies::raise_domain_error<T>(
         function,
         "Argument a must be finite, but got %1%", a, pol);
   if(!(boost::math::isfinite)(b))
      return policies::raise_domain_error<T>(
         function,
         "Argument b must be finite, but got %1%", b, pol);
   //
   // Special cases:
   //
   if(a > b)
      return -float_distance(b, a, pol);
   if(a == b)
      return T(0);
   if(a == 0)
      return 1 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol));
   if(b == 0)
      return 1 + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
   if(boost::math::sign(a) != boost::math::sign(b))
      return 2 + fabs(float_distance(static_cast<T>((b < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), b, pol))
         + fabs(float_distance(static_cast<T>((a < 0) ? T(-detail::get_smallest_value<T>()) : detail::get_smallest_value<T>()), a, pol));
   //
   // By the time we get here, both a and b must have the same sign, we want
   // b > a and both postive for the following logic:
   //
   if(a < 0)
      return float_distance(static_cast<T>(-b), static_cast<T>(-a), pol);

   BOOST_ASSERT(a >= 0);
   BOOST_ASSERT(b >= a);

   boost::intmax_t expon;
   //
   // Note that if a is a denorm then the usual formula fails
   // because we actually have fewer than tools::digits<T>()
   // significant bits in the representation:
   //
   expon = 1 + ilogb(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) ? tools::min_value<T>() : a);
   T upper = scalbn(T(1), expon);
   T result = T(0);
   //
   // If b is greater than upper, then we *must* split the calculation
   // as the size of the ULP changes with each order of magnitude change:
   //
   if(b > upper)
   {
      boost::intmax_t expon2 = 1 + ilogb(b);
      T upper2 = scalbn(T(1), expon2 - 1);
      result = float_distance(upper2, b);
      result += (expon2 - expon - 1) * scalbn(T(1), std::numeric_limits<T>::digits - 1);
   }
   //
   // Use compensated double-double addition to avoid rounding
   // errors in the subtraction:
   //
   expon = std::numeric_limits<T>::digits - expon;
   T mb, x, y, z;
   if(((boost::math::fpclassify)(a) == (int)FP_SUBNORMAL) || (b - a < tools::min_value<T>()))
   {
      //
      // Special case - either one end of the range is a denormal, or else the difference is.
      // The regular code will fail if we're using the SSE2 registers on Intel and either
      // the FTZ or DAZ flags are set.
      //
      T a2 = scalbn(a, std::numeric_limits<T>::digits);
      T b2 = scalbn(b, std::numeric_limits<T>::digits);
      mb = -(std::min)(T(scalbn(upper, std::numeric_limits<T>::digits)), b2);
      x = a2 + mb;
      z = x - a2;
      y = (a2 - (x - z)) + (mb - z);

      expon -= std::numeric_limits<T>::digits;
   }
   else
   {
      mb = -(std::min)(upper, b);
      x = a + mb;
      z = x - a;
      y = (a - (x - z)) + (mb - z);
   }
   if(x < 0)
   {
      x = -x;
      y = -y;
   }
   result += scalbn(x, expon) + scalbn(y, expon);
   //
   // Result must be an integer:
   //
   BOOST_ASSERT(result == floor(result));
   return result;
} // float_distance_imp

} // namespace detail

template <class T, class U, class Policy>
inline typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b, const Policy& pol)
{
   typedef typename tools::promote_args<T, U>::type result_type;
   return detail::float_distance_imp(detail::normalize_value(static_cast<result_type>(a), typename detail::has_hidden_guard_digits<result_type>::type()), detail::normalize_value(static_cast<result_type>(b), typename detail::has_hidden_guard_digits<result_type>::type()), mpl::bool_<!std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol);
}

template <class T, class U>
typename tools::promote_args<T, U>::type float_distance(const T& a, const U& b)
{
   return boost::math::float_distance(a, b, policies::policy<>());
}

namespace detail{

template <class T, class Policy>
T float_advance_imp(T val, int distance, const mpl::true_&, const Policy& pol)
{
   BOOST_MATH_STD_USING
   //
   // Error handling:
   //
   static const char* function = "float_advance<%1%>(%1%, int)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
      return policies::raise_domain_error<T>(
         function,
         "Argument val must be finite, but got %1%", val, pol);

   if(val < 0)
      return -float_advance(-val, -distance, pol);
   if(distance == 0)
      return val;
   if(distance == 1)
      return float_next(val, pol);
   if(distance == -1)
      return float_prior(val, pol);

   if(fabs(val) < detail::get_min_shift_value<T>())
   {
      //
      // Special case: if the value of the least significant bit is a denorm,
      // implement in terms of float_next/float_prior.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      if(distance > 0)
      {
         do{ val = float_next(val, pol); } while(--distance);
      }
      else
      {
         do{ val = float_prior(val, pol); } while(++distance);
      }
      return val;
   }

   int expon;
   frexp(val, &expon);
   T limit = ldexp((distance < 0 ? T(0.5f) : T(1)), expon);
   if(val <= tools::min_value<T>())
   {
      limit = sign(T(distance)) * tools::min_value<T>();
   }
   T limit_distance = float_distance(val, limit);
   while(fabs(limit_distance) < abs(distance))
   {
      distance -= itrunc(limit_distance);
      val = limit;
      if(distance < 0)
      {
         limit /= 2;
         expon--;
      }
      else
      {
         limit *= 2;
         expon++;
      }
      limit_distance = float_distance(val, limit);
      if(distance && (limit_distance == 0))
      {
         return policies::raise_evaluation_error<T>(function, "Internal logic failed while trying to increment floating point value %1%: most likely your FPU is in non-IEEE conforming mode.", val, pol);
      }
   }
   if((0.5f == frexp(val, &expon)) && (distance < 0))
      --expon;
   T diff = 0;
   if(val != 0)
      diff = distance * ldexp(T(1), expon - tools::digits<T>());
   if(diff == 0)
      diff = distance * detail::get_smallest_value<T>();
   return val += diff;
} // float_advance_imp
//
// Special version for bases other than 2:
//
template <class T, class Policy>
T float_advance_imp(T val, int distance, const mpl::false_&, const Policy& pol)
{
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::is_specialized);
   BOOST_STATIC_ASSERT(std::numeric_limits<T>::radix != 2);

   BOOST_MATH_STD_USING
   //
   // Error handling:
   //
   static const char* function = "float_advance<%1%>(%1%, int)";

   int fpclass = (boost::math::fpclassify)(val);

   if((fpclass == (int)FP_NAN) || (fpclass == (int)FP_INFINITE))
      return policies::raise_domain_error<T>(
         function,
         "Argument val must be finite, but got %1%", val, pol);

   if(val < 0)
      return -float_advance(-val, -distance, pol);
   if(distance == 0)
      return val;
   if(distance == 1)
      return float_next(val, pol);
   if(distance == -1)
      return float_prior(val, pol);

   if(fabs(val) < detail::get_min_shift_value<T>())
   {
      //
      // Special case: if the value of the least significant bit is a denorm,
      // implement in terms of float_next/float_prior.
      // This avoids issues with the Intel SSE2 registers when the FTZ or DAZ flags are set.
      //
      if(distance > 0)
      {
         do{ val = float_next(val, pol); } while(--distance);
      }
      else
      {
         do{ val = float_prior(val, pol); } while(++distance);
      }
      return val;
   }

   boost::intmax_t expon = 1 + ilogb(val);
   T limit = scalbn(T(1), distance < 0 ? expon - 1 : expon);
   if(val <= tools::min_value<T>())
   {
      limit = sign(T(distance)) * tools::min_value<T>();
   }
   T limit_distance = float_distance(val, limit);
   while(fabs(limit_distance) < abs(distance))
   {
      distance -= itrunc(limit_distance);
      val = limit;
      if(distance < 0)
      {
         limit /= std::numeric_limits<T>::radix;
         expon--;
      }
      else
      {
         limit *= std::numeric_limits<T>::radix;
         expon++;
      }
      limit_distance = float_distance(val, limit);
      if(distance && (limit_distance == 0))
      {
         return policies::raise_evaluation_error<T>(function, "Internal logic failed while trying to increment floating point value %1%: most likely your FPU is in non-IEEE conforming mode.", val, pol);
      }
   }
   /*expon = 1 + ilogb(val);
   if((1 == scalbn(val, 1 + expon)) && (distance < 0))
      --expon;*/
   T diff = 0;
   if(val != 0)
      diff = distance * scalbn(T(1), expon - std::numeric_limits<T>::digits);
   if(diff == 0)
      diff = distance * detail::get_smallest_value<T>();
   return val += diff;
} // float_advance_imp

} // namespace detail

template <class T, class Policy>
inline typename tools::promote_args<T>::type float_advance(T val, int distance, const Policy& pol)
{
   typedef typename tools::promote_args<T>::type result_type;
   return detail::float_advance_imp(detail::normalize_value(static_cast<result_type>(val), typename detail::has_hidden_guard_digits<result_type>::type()), distance, mpl::bool_<!std::numeric_limits<result_type>::is_specialized || (std::numeric_limits<result_type>::radix == 2)>(), pol);
}

template <class T>
inline typename tools::promote_args<T>::type float_advance(const T& val, int distance)
{
   return boost::math::float_advance(val, distance, policies::policy<>());
}

}} // boost math namespaces

#endif // BOOST_MATH_SPECIAL_NEXT_HPP