Import pre-release gcc-5.0 to new vendor branch
[dragonfly.git] / contrib / gcc-5.0 / libstdc++-v3 / config / cpu / i486 / opt / bits / opt_random.h
1 // Optimizations for random number functions, x86 version -*- C++ -*-
2
3 // Copyright (C) 2012-2015 Free Software Foundation, Inc.
4 //
5 // This file is part of the GNU ISO C++ Library.  This library is free
6 // software; you can redistribute it and/or modify it under the
7 // terms of the GNU General Public License as published by the
8 // Free Software Foundation; either version 3, or (at your option)
9 // any later version.
10
11 // This library is distributed in the hope that it will be useful,
12 // but WITHOUT ANY WARRANTY; without even the implied warranty of
13 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14 // GNU General Public License for more details.
15
16 // Under Section 7 of GPL version 3, you are granted additional
17 // permissions described in the GCC Runtime Library Exception, version
18 // 3.1, as published by the Free Software Foundation.
19
20 // You should have received a copy of the GNU General Public License and
21 // a copy of the GCC Runtime Library Exception along with this program;
22 // see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
23 // <http://www.gnu.org/licenses/>.
24
25 /** @file bits/opt_random.h
26  *  This is an internal header file, included by other library headers.
27  *  Do not attempt to use it directly. @headername{random}
28  */
29
30 #ifndef _BITS_OPT_RANDOM_H
31 #define _BITS_OPT_RANDOM_H 1
32
33 #include <x86intrin.h>
34
35
36 #pragma GCC system_header
37
38
39 namespace std _GLIBCXX_VISIBILITY(default)
40 {
41 _GLIBCXX_BEGIN_NAMESPACE_VERSION
42
43 #ifdef __SSE3__
44   template<>
45     template<typename _UniformRandomNumberGenerator>
46       void
47       normal_distribution<double>::
48       __generate(typename normal_distribution<double>::result_type* __f,
49                  typename normal_distribution<double>::result_type* __t,
50                  _UniformRandomNumberGenerator& __urng,
51                  const param_type& __param)
52       {
53         typedef uint64_t __uctype;
54
55         if (__f == __t)
56           return;
57
58         if (_M_saved_available)
59           {
60             _M_saved_available = false;
61             *__f++ = _M_saved * __param.stddev() + __param.mean();
62
63             if (__f == __t)
64               return;
65           }
66
67         constexpr uint64_t __maskval = 0xfffffffffffffull;
68         static const __m128i __mask = _mm_set1_epi64x(__maskval);
69         static const __m128i __two = _mm_set1_epi64x(0x4000000000000000ull);
70         static const __m128d __three = _mm_set1_pd(3.0);
71         const __m128d __av = _mm_set1_pd(__param.mean());
72
73         const __uctype __urngmin = __urng.min();
74         const __uctype __urngmax = __urng.max();
75         const __uctype __urngrange = __urngmax - __urngmin;
76         const __uctype __uerngrange = __urngrange + 1;
77
78         while (__f + 1 < __t)
79           {
80             double __le;
81             __m128d __x;
82             do
83               {
84                 union
85                 {
86                   __m128i __i;
87                   __m128d __d;
88                 } __v;
89
90                 if (__urngrange > __maskval)
91                   {
92                     if (__detail::_Power_of_2(__uerngrange))
93                       __v.__i = _mm_and_si128(_mm_set_epi64x(__urng(),
94                                                              __urng()),
95                                               __mask);
96                     else
97                       {
98                         const __uctype __uerange = __maskval + 1;
99                         const __uctype __scaling = __urngrange / __uerange;
100                         const __uctype __past = __uerange * __scaling;
101                         uint64_t __v1;
102                         do
103                           __v1 = __uctype(__urng()) - __urngmin;
104                         while (__v1 >= __past);
105                         __v1 /= __scaling;
106                         uint64_t __v2;
107                         do
108                           __v2 = __uctype(__urng()) - __urngmin;
109                         while (__v2 >= __past);
110                         __v2 /= __scaling;
111
112                         __v.__i = _mm_set_epi64x(__v1, __v2);
113                       }
114                   }
115                 else if (__urngrange == __maskval)
116                   __v.__i = _mm_set_epi64x(__urng(), __urng());
117                 else if ((__urngrange + 2) * __urngrange >= __maskval
118                          && __detail::_Power_of_2(__uerngrange))
119                   {
120                     uint64_t __v1 = __urng() * __uerngrange + __urng();
121                     uint64_t __v2 = __urng() * __uerngrange + __urng();
122
123                     __v.__i = _mm_and_si128(_mm_set_epi64x(__v1, __v2),
124                                             __mask);
125                   }
126                 else
127                   {
128                     size_t __nrng = 2;
129                     __uctype __high = __maskval / __uerngrange / __uerngrange;
130                     while (__high > __uerngrange)
131                       {
132                         ++__nrng;
133                         __high /= __uerngrange;
134                       }
135                     const __uctype __highrange = __high + 1;
136                     const __uctype __scaling = __urngrange / __highrange;
137                     const __uctype __past = __highrange * __scaling;
138                     __uctype __tmp;
139
140                     uint64_t __v1;
141                     do
142                       {
143                         do
144                           __tmp = __uctype(__urng()) - __urngmin;
145                         while (__tmp >= __past);
146                         __v1 = __tmp / __scaling;
147                         for (size_t __cnt = 0; __cnt < __nrng; ++__cnt)
148                           {
149                             __tmp = __v1;
150                             __v1 *= __uerngrange;
151                             __v1 += __uctype(__urng()) - __urngmin;
152                           }
153                       }
154                     while (__v1 > __maskval || __v1 < __tmp);
155
156                     uint64_t __v2;
157                     do
158                       {
159                         do
160                           __tmp = __uctype(__urng()) - __urngmin;
161                         while (__tmp >= __past);
162                         __v2 = __tmp / __scaling;
163                         for (size_t __cnt = 0; __cnt < __nrng; ++__cnt)
164                           {
165                             __tmp = __v2;
166                             __v2 *= __uerngrange;
167                             __v2 += __uctype(__urng()) - __urngmin;
168                           }
169                       }
170                     while (__v2 > __maskval || __v2 < __tmp);
171                     
172                     __v.__i = _mm_set_epi64x(__v1, __v2);
173                   }
174
175                 __v.__i = _mm_or_si128(__v.__i, __two);
176                 __x = _mm_sub_pd(__v.__d, __three);
177                 __m128d __m = _mm_mul_pd(__x, __x);
178                 __le = _mm_cvtsd_f64(_mm_hadd_pd (__m, __m));
179               }
180             while (__le == 0.0 || __le >= 1.0);
181
182             double __mult = (std::sqrt(-2.0 * std::log(__le) / __le)
183                              * __param.stddev());
184
185             __x = _mm_add_pd(_mm_mul_pd(__x, _mm_set1_pd(__mult)), __av);
186
187             _mm_storeu_pd(__f, __x);
188             __f += 2;
189           }
190
191         if (__f != __t)
192           {
193             result_type __x, __y, __r2;
194
195             __detail::_Adaptor<_UniformRandomNumberGenerator, result_type>
196               __aurng(__urng);
197
198             do
199               {
200                 __x = result_type(2.0) * __aurng() - 1.0;
201                 __y = result_type(2.0) * __aurng() - 1.0;
202                 __r2 = __x * __x + __y * __y;
203               }
204             while (__r2 > 1.0 || __r2 == 0.0);
205
206             const result_type __mult = std::sqrt(-2 * std::log(__r2) / __r2);
207             _M_saved = __x * __mult;
208             _M_saved_available = true;
209             *__f = __y * __mult * __param.stddev() + __param.mean();
210           }
211       }
212 #endif
213
214
215 _GLIBCXX_END_NAMESPACE_VERSION
216 } // namespace
217
218
219 #endif // _BITS_OPT_RANDOM_H