Emeric Vigier | eebea67 | 2012-08-06 17:36:30 -0400 | [diff] [blame] | 1 | /* |
| 2 | ** Copyright (C) 2001-2011 Erik de Castro Lopo <erikd@mega-nerd.com> |
| 3 | ** |
| 4 | ** This program is free software; you can redistribute it and/or modify |
| 5 | ** it under the terms of the GNU Lesser General Public License as published by |
| 6 | ** the Free Software Foundation; either version 2.1 of the License, or |
| 7 | ** (at your option) any later version. |
| 8 | ** |
| 9 | ** This program is distributed in the hope that it will be useful, |
| 10 | ** but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 11 | ** MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 12 | ** GNU Lesser General Public License for more details. |
| 13 | ** |
| 14 | ** You should have received a copy of the GNU Lesser General Public License |
| 15 | ** along with this program; if not, write to the Free Software |
| 16 | ** Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. |
| 17 | */ |
| 18 | |
| 19 | /* Version 1.5 */ |
| 20 | |
| 21 | #ifndef FLOAT_CAST_HEADER |
| 22 | #define FLOAT_CAST_HEADER |
| 23 | |
| 24 | /*============================================================================ |
| 25 | ** On Intel Pentium processors (especially PIII and probably P4), converting |
| 26 | ** from float to int is very slow. To meet the C specs, the code produced by |
| 27 | ** most C compilers targeting Pentium needs to change the FPU rounding mode |
| 28 | ** before the float to int conversion is performed. |
| 29 | ** |
| 30 | ** Changing the FPU rounding mode causes the FPU pipeline to be flushed. It |
| 31 | ** is this flushing of the pipeline which is so slow. |
| 32 | ** |
| 33 | ** Fortunately the ISO C99 specifications define the functions lrint, lrintf, |
| 34 | ** llrint and llrintf which fix this problem as a side effect. |
| 35 | ** |
| 36 | ** On Unix-like systems, the configure process should have detected the |
| 37 | ** presence of these functions. If they weren't found we have to replace them |
| 38 | ** here with a standard C cast. |
| 39 | */ |
| 40 | |
| 41 | /* |
| 42 | ** The C99 prototypes for lrint and lrintf are as follows: |
| 43 | ** |
| 44 | ** long int lrintf (float x) ; |
| 45 | ** long int lrint (double x) ; |
| 46 | */ |
| 47 | |
| 48 | #include "config.h" |
| 49 | |
| 50 | /* |
| 51 | ** The presence of the required functions are detected during the configure |
| 52 | ** process and the values HAVE_LRINT and HAVE_LRINTF are set accordingly in |
| 53 | ** the config.h file. |
| 54 | */ |
| 55 | |
| 56 | #define HAVE_LRINT_REPLACEMENT 0 |
| 57 | |
| 58 | #if (HAVE_LRINT && HAVE_LRINTF) |
| 59 | |
| 60 | /* |
| 61 | ** These defines enable functionality introduced with the 1999 ISO C |
| 62 | ** standard. They must be defined before the inclusion of math.h to |
| 63 | ** engage them. If optimisation is enabled, these functions will be |
| 64 | ** inlined. With optimisation switched off, you have to link in the |
| 65 | ** maths library using -lm. |
| 66 | */ |
| 67 | |
| 68 | #define _ISOC9X_SOURCE 1 |
| 69 | #define _ISOC99_SOURCE 1 |
| 70 | |
| 71 | #define __USE_ISOC9X 1 |
| 72 | #define __USE_ISOC99 1 |
| 73 | |
| 74 | #include <math.h> |
| 75 | |
| 76 | #elif (defined (__CYGWIN__)) |
| 77 | |
| 78 | #include <math.h> |
| 79 | |
| 80 | #undef HAVE_LRINT_REPLACEMENT |
| 81 | #define HAVE_LRINT_REPLACEMENT 1 |
| 82 | |
| 83 | #undef lrint |
| 84 | #undef lrintf |
| 85 | |
| 86 | #define lrint double2int |
| 87 | #define lrintf float2int |
| 88 | |
| 89 | /* |
| 90 | ** The native CYGWIN lrint and lrintf functions are buggy: |
| 91 | ** http://sourceware.org/ml/cygwin/2005-06/msg00153.html |
| 92 | ** http://sourceware.org/ml/cygwin/2005-09/msg00047.html |
| 93 | ** and slow. |
| 94 | ** These functions (pulled from the Public Domain MinGW math.h header) |
| 95 | ** replace the native versions. |
| 96 | */ |
| 97 | |
| 98 | static inline long double2int (double in) |
| 99 | { long retval ; |
| 100 | |
| 101 | __asm__ __volatile__ |
| 102 | ( "fistpl %0" |
| 103 | : "=m" (retval) |
| 104 | : "t" (in) |
| 105 | : "st" |
| 106 | ) ; |
| 107 | |
| 108 | return retval ; |
| 109 | } /* double2int */ |
| 110 | |
| 111 | static inline long float2int (float in) |
| 112 | { long retval ; |
| 113 | |
| 114 | __asm__ __volatile__ |
| 115 | ( "fistpl %0" |
| 116 | : "=m" (retval) |
| 117 | : "t" (in) |
| 118 | : "st" |
| 119 | ) ; |
| 120 | |
| 121 | return retval ; |
| 122 | } /* float2int */ |
| 123 | |
| 124 | #elif (defined (WIN64) || defined(_WIN64)) |
| 125 | |
| 126 | /* Win64 section should be places before Win32 one, because |
| 127 | ** most likely both WIN32 and WIN64 will be defined in 64-bit case. |
| 128 | */ |
| 129 | |
| 130 | #include <math.h> |
| 131 | |
| 132 | /* Win64 doesn't seem to have these functions, nor inline assembly. |
| 133 | ** Therefore implement inline versions of these functions here. |
| 134 | */ |
| 135 | #include <emmintrin.h> |
| 136 | #include <mmintrin.h> |
| 137 | |
| 138 | __inline long int |
| 139 | lrint(double flt) |
| 140 | { |
| 141 | return _mm_cvtsd_si32(_mm_load_sd(&flt)); |
| 142 | } |
| 143 | |
| 144 | __inline long int |
| 145 | lrintf(float flt) |
| 146 | { |
| 147 | return _mm_cvtss_si32(_mm_load_ss(&flt)); |
| 148 | } |
| 149 | |
| 150 | #elif (defined (WIN32) || defined (_WIN32)) |
| 151 | |
| 152 | #undef HAVE_LRINT_REPLACEMENT |
| 153 | #define HAVE_LRINT_REPLACEMENT 1 |
| 154 | |
| 155 | #include <math.h> |
| 156 | |
| 157 | /* |
| 158 | ** Win32 doesn't seem to have these functions. |
| 159 | ** Therefore implement inline versions of these functions here. |
| 160 | */ |
| 161 | |
| 162 | __inline long int |
| 163 | lrint (double flt) |
| 164 | { int intgr ; |
| 165 | |
| 166 | _asm |
| 167 | { fld flt |
| 168 | fistp intgr |
| 169 | } ; |
| 170 | |
| 171 | return intgr ; |
| 172 | } |
| 173 | |
| 174 | __inline long int |
| 175 | lrintf (float flt) |
| 176 | { int intgr ; |
| 177 | |
| 178 | _asm |
| 179 | { fld flt |
| 180 | fistp intgr |
| 181 | } ; |
| 182 | |
| 183 | return intgr ; |
| 184 | } |
| 185 | |
| 186 | #elif (defined (__MWERKS__) && defined (macintosh)) |
| 187 | |
| 188 | /* This MacOS 9 solution was provided by Stephane Letz */ |
| 189 | |
| 190 | #undef HAVE_LRINT_REPLACEMENT |
| 191 | #define HAVE_LRINT_REPLACEMENT 1 |
| 192 | #include <math.h> |
| 193 | |
| 194 | #undef lrint |
| 195 | #undef lrintf |
| 196 | |
| 197 | #define lrint double2int |
| 198 | #define lrintf float2int |
| 199 | |
| 200 | inline int |
| 201 | float2int (register float in) |
| 202 | { long res [2] ; |
| 203 | |
| 204 | asm |
| 205 | { fctiw in, in |
| 206 | stfd in, res |
| 207 | } |
| 208 | return res [1] ; |
| 209 | } /* float2int */ |
| 210 | |
| 211 | inline int |
| 212 | double2int (register double in) |
| 213 | { long res [2] ; |
| 214 | |
| 215 | asm |
| 216 | { fctiw in, in |
| 217 | stfd in, res |
| 218 | } |
| 219 | return res [1] ; |
| 220 | } /* double2int */ |
| 221 | |
| 222 | #elif (defined (__MACH__) && defined (__APPLE__)) |
| 223 | |
| 224 | /* For Apple MacOSX. */ |
| 225 | |
| 226 | #undef HAVE_LRINT_REPLACEMENT |
| 227 | #define HAVE_LRINT_REPLACEMENT 1 |
| 228 | #include <math.h> |
| 229 | |
| 230 | #undef lrint |
| 231 | #undef lrintf |
| 232 | |
| 233 | #define lrint double2int |
| 234 | #define lrintf float2int |
| 235 | |
| 236 | inline static long |
| 237 | float2int (register float in) |
| 238 | { int res [2] ; |
| 239 | |
| 240 | __asm__ __volatile__ |
| 241 | ( "fctiw %1, %1\n\t" |
| 242 | "stfd %1, %0" |
| 243 | : "=m" (res) /* Output */ |
| 244 | : "f" (in) /* Input */ |
| 245 | : "memory" |
| 246 | ) ; |
| 247 | |
| 248 | return res [1] ; |
| 249 | } /* lrintf */ |
| 250 | |
| 251 | inline static long |
| 252 | double2int (register double in) |
| 253 | { int res [2] ; |
| 254 | |
| 255 | __asm__ __volatile__ |
| 256 | ( "fctiw %1, %1\n\t" |
| 257 | "stfd %1, %0" |
| 258 | : "=m" (res) /* Output */ |
| 259 | : "f" (in) /* Input */ |
| 260 | : "memory" |
| 261 | ) ; |
| 262 | |
| 263 | return res [1] ; |
| 264 | } /* lrint */ |
| 265 | |
| 266 | #else |
| 267 | #ifndef __sgi |
| 268 | #warning "Don't have the functions lrint() and lrintf()." |
| 269 | #warning "Replacing these functions with a standard C cast." |
| 270 | #endif |
| 271 | |
| 272 | #include <math.h> |
| 273 | |
| 274 | #define lrint(dbl) ((long) (dbl)) |
| 275 | #define lrintf(flt) ((long) (flt)) |
| 276 | |
| 277 | #endif |
| 278 | |
| 279 | |
| 280 | #endif /* FLOAT_CAST_HEADER */ |
| 281 | |