Mercurial > mplayer.hg
annotate mp3lib/dct64_k7.c @ 5979:6cc15e96d709
rpm/deb/licensing stuff removed/changed
| author | gabucino |
|---|---|
| date | Sun, 05 May 2002 05:04:56 +0000 |
| parents | d6eab895c742 |
| children | 82a03c17f314 |
| rev | line source |
|---|---|
| 4148 | 1 /* |
| 2 * This code was taken from http://www.mpg123.org | |
| 3 * See ChangeLog of mpg123-0.59s-pre.1 for detail | |
| 4 * Applied to mplayer by Nick Kurshev <nickols_k@mail.ru> | |
| 5 * Partial 3dnowex-DSP! optimization by Nick Kurshev | |
| 6 * | |
| 7 * TODO: optimize scalar 3dnow! code | |
| 8 * Warning: Phases 7 & 8 are not tested | |
| 9 */ | |
| 10 #define real float /* ugly - but only way */ | |
| 11 | |
| 5291 | 12 #include "../mangle.h" |
| 13 | |
| 4148 | 14 static unsigned long long int __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL; |
| 15 static float plus_1f = 1.0; | |
| 16 | |
|
5842
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
17 #ifdef __CYGWIN__ |
|
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
18 /* will probably cause sig11 with debuggingsymbols, but otherwise undef. ref */ |
|
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
19 void dct64_MMX_3dnowex(real *a,real *b,real *c) |
|
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
20 #else |
| 4148 | 21 void __attribute__ (( __stdcall__ )) dct64_MMX_3dnowex(real *a,real *b,real *c) |
|
5842
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
22 #endif |
| 4148 | 23 { |
| 24 char tmp[256]; | |
| 25 __asm __volatile( | |
| 26 " movl %2,%%eax\n\t" | |
| 27 | |
| 28 " leal 128+%3,%%edx\n\t" | |
| 29 " movl %0,%%esi\n\t" | |
| 30 " movl %1,%%edi\n\t" | |
| 5291 | 31 " movl $"MANGLE(costab_mmx)",%%ebx\n\t" |
| 4148 | 32 " leal %3,%%ecx\n\t" |
| 33 | |
| 34 /* Phase 1*/ | |
| 35 " movq (%%eax), %%mm0\n\t" | |
| 36 " movq 8(%%eax), %%mm4\n\t" | |
| 37 " movq %%mm0, %%mm3\n\t" | |
| 38 " movq %%mm4, %%mm7\n\t" | |
| 39 " movq 120(%%eax), %%mm1\n\t" | |
| 40 " movq 112(%%eax), %%mm5\n\t" | |
| 41 " pswapd %%mm1, %%mm1\n\t" | |
| 42 " pswapd %%mm5, %%mm5\n\t" | |
| 43 " pfadd %%mm1, %%mm0\n\t" | |
| 44 " pfadd %%mm5, %%mm4\n\t" | |
| 45 " movq %%mm0, (%%edx)\n\t" | |
| 46 " movq %%mm4, 8(%%edx)\n\t" | |
| 47 " pfsub %%mm1, %%mm3\n\t" | |
| 48 " pfsub %%mm5, %%mm7\n\t" | |
| 49 " pfmul (%%ebx), %%mm3\n\t" | |
| 50 " pfmul 8(%%ebx), %%mm7\n\t" | |
| 51 " pswapd %%mm3, %%mm3\n\t" | |
| 52 " pswapd %%mm7, %%mm7\n\t" | |
| 53 " movq %%mm3, 120(%%edx)\n\t" | |
| 54 " movq %%mm7, 112(%%edx)\n\t" | |
| 55 | |
| 56 " movq 16(%%eax), %%mm0\n\t" | |
| 57 " movq 24(%%eax), %%mm4\n\t" | |
| 58 " movq %%mm0, %%mm3\n\t" | |
| 59 " movq %%mm4, %%mm7\n\t" | |
| 60 " movq 104(%%eax), %%mm1\n\t" | |
| 61 " movq 96(%%eax), %%mm5\n\t" | |
| 62 " pswapd %%mm1, %%mm1\n\t" | |
| 63 " pswapd %%mm5, %%mm5\n\t" | |
| 64 " pfadd %%mm1, %%mm0\n\t" | |
| 65 " pfadd %%mm5, %%mm4\n\t" | |
| 66 " movq %%mm0, 16(%%edx)\n\t" | |
| 67 " movq %%mm4, 24(%%edx)\n\t" | |
| 68 " pfsub %%mm1, %%mm3\n\t" | |
| 69 " pfsub %%mm5, %%mm7\n\t" | |
| 70 " pfmul 16(%%ebx), %%mm3\n\t" | |
| 71 " pfmul 24(%%ebx), %%mm7\n\t" | |
| 72 " pswapd %%mm3, %%mm3\n\t" | |
| 73 " pswapd %%mm7, %%mm7\n\t" | |
| 74 " movq %%mm3, 104(%%edx)\n\t" | |
| 75 " movq %%mm7, 96(%%edx)\n\t" | |
| 76 | |
| 77 " movq 32(%%eax), %%mm0\n\t" | |
| 78 " movq 40(%%eax), %%mm4\n\t" | |
| 79 " movq %%mm0, %%mm3\n\t" | |
| 80 " movq %%mm4, %%mm7\n\t" | |
| 81 " movq 88(%%eax), %%mm1\n\t" | |
| 82 " movq 80(%%eax), %%mm5\n\t" | |
| 83 " pswapd %%mm1, %%mm1\n\t" | |
| 84 " pswapd %%mm5, %%mm5\n\t" | |
| 85 " pfadd %%mm1, %%mm0\n\t" | |
| 86 " pfadd %%mm5, %%mm4\n\t" | |
| 87 " movq %%mm0, 32(%%edx)\n\t" | |
| 88 " movq %%mm4, 40(%%edx)\n\t" | |
| 89 " pfsub %%mm1, %%mm3\n\t" | |
| 90 " pfsub %%mm5, %%mm7\n\t" | |
| 91 " pfmul 32(%%ebx), %%mm3\n\t" | |
| 92 " pfmul 40(%%ebx), %%mm7\n\t" | |
| 93 " pswapd %%mm3, %%mm3\n\t" | |
| 94 " pswapd %%mm7, %%mm7\n\t" | |
| 95 " movq %%mm3, 88(%%edx)\n\t" | |
| 96 " movq %%mm7, 80(%%edx)\n\t" | |
| 97 | |
| 98 " movq 48(%%eax), %%mm0\n\t" | |
| 99 " movq 56(%%eax), %%mm4\n\t" | |
| 100 " movq %%mm0, %%mm3\n\t" | |
| 101 " movq %%mm4, %%mm7\n\t" | |
| 102 " movq 72(%%eax), %%mm1\n\t" | |
| 103 " movq 64(%%eax), %%mm5\n\t" | |
| 104 " pswapd %%mm1, %%mm1\n\t" | |
| 105 " pswapd %%mm5, %%mm5\n\t" | |
| 106 " pfadd %%mm1, %%mm0\n\t" | |
| 107 " pfadd %%mm5, %%mm4\n\t" | |
| 108 " movq %%mm0, 48(%%edx)\n\t" | |
| 109 " movq %%mm4, 56(%%edx)\n\t" | |
| 110 " pfsub %%mm1, %%mm3\n\t" | |
| 111 " pfsub %%mm5, %%mm7\n\t" | |
| 112 " pfmul 48(%%ebx), %%mm3\n\t" | |
| 113 " pfmul 56(%%ebx), %%mm7\n\t" | |
| 114 " pswapd %%mm3, %%mm3\n\t" | |
| 115 " pswapd %%mm7, %%mm7\n\t" | |
| 116 " movq %%mm3, 72(%%edx)\n\t" | |
| 117 " movq %%mm7, 64(%%edx)\n\t" | |
| 118 | |
| 119 /* Phase 2*/ | |
| 120 | |
| 121 " movq (%%edx), %%mm0\n\t" | |
| 122 " movq 8(%%edx), %%mm4\n\t" | |
| 123 " movq %%mm0, %%mm3\n\t" | |
| 124 " movq %%mm4, %%mm7\n\t" | |
| 125 " movq 56(%%edx), %%mm1\n\t" | |
| 126 " movq 48(%%edx), %%mm5\n\t" | |
| 127 " pswapd %%mm1, %%mm1\n\t" | |
| 128 " pswapd %%mm5, %%mm5\n\t" | |
| 129 " pfadd %%mm1, %%mm0\n\t" | |
| 130 " pfadd %%mm5, %%mm4\n\t" | |
| 131 " movq %%mm0, (%%ecx)\n\t" | |
| 132 " movq %%mm4, 8(%%ecx)\n\t" | |
| 133 " pfsub %%mm1, %%mm3\n\t" | |
| 134 " pfsub %%mm5, %%mm7\n\t" | |
| 135 " pfmul 64(%%ebx), %%mm3\n\t" | |
| 136 " pfmul 72(%%ebx), %%mm7\n\t" | |
| 137 " pswapd %%mm3, %%mm3\n\t" | |
| 138 " pswapd %%mm7, %%mm7\n\t" | |
| 139 " movq %%mm3, 56(%%ecx)\n\t" | |
| 140 " movq %%mm7, 48(%%ecx)\n\t" | |
| 141 | |
| 142 " movq 16(%%edx), %%mm0\n\t" | |
| 143 " movq 24(%%edx), %%mm4\n\t" | |
| 144 " movq %%mm0, %%mm3\n\t" | |
| 145 " movq %%mm4, %%mm7\n\t" | |
| 146 " movq 40(%%edx), %%mm1\n\t" | |
| 147 " movq 32(%%edx), %%mm5\n\t" | |
| 148 " pswapd %%mm1, %%mm1\n\t" | |
| 149 " pswapd %%mm5, %%mm5\n\t" | |
| 150 " pfadd %%mm1, %%mm0\n\t" | |
| 151 " pfadd %%mm5, %%mm4\n\t" | |
| 152 " movq %%mm0, 16(%%ecx)\n\t" | |
| 153 " movq %%mm4, 24(%%ecx)\n\t" | |
| 154 " pfsub %%mm1, %%mm3\n\t" | |
| 155 " pfsub %%mm5, %%mm7\n\t" | |
| 156 " pfmul 80(%%ebx), %%mm3\n\t" | |
| 157 " pfmul 88(%%ebx), %%mm7\n\t" | |
| 158 " pswapd %%mm3, %%mm3\n\t" | |
| 159 " pswapd %%mm7, %%mm7\n\t" | |
| 160 " movq %%mm3, 40(%%ecx)\n\t" | |
| 161 " movq %%mm7, 32(%%ecx)\n\t" | |
| 162 | |
| 163 /* Phase 3*/ | |
| 164 | |
| 165 " movq 64(%%edx), %%mm0\n\t" | |
| 166 " movq 72(%%edx), %%mm4\n\t" | |
| 167 " movq %%mm0, %%mm3\n\t" | |
| 168 " movq %%mm4, %%mm7\n\t" | |
| 169 " movq 120(%%edx), %%mm1\n\t" | |
| 170 " movq 112(%%edx), %%mm5\n\t" | |
| 171 " pswapd %%mm1, %%mm1\n\t" | |
| 172 " pswapd %%mm5, %%mm5\n\t" | |
| 173 " pfadd %%mm1, %%mm0\n\t" | |
| 174 " pfadd %%mm5, %%mm4\n\t" | |
| 175 " movq %%mm0, 64(%%ecx)\n\t" | |
| 176 " movq %%mm4, 72(%%ecx)\n\t" | |
| 177 " pfsubr %%mm1, %%mm3\n\t" | |
| 178 " pfsubr %%mm5, %%mm7\n\t" | |
| 179 " pfmul 64(%%ebx), %%mm3\n\t" | |
| 180 " pfmul 72(%%ebx), %%mm7\n\t" | |
| 181 " pswapd %%mm3, %%mm3\n\t" | |
| 182 " pswapd %%mm7, %%mm7\n\t" | |
| 183 " movq %%mm3, 120(%%ecx)\n\t" | |
| 184 " movq %%mm7, 112(%%ecx)\n\t" | |
| 185 | |
| 186 " movq 80(%%edx), %%mm0\n\t" | |
| 187 " movq 88(%%edx), %%mm4\n\t" | |
| 188 " movq %%mm0, %%mm3\n\t" | |
| 189 " movq %%mm4, %%mm7\n\t" | |
| 190 " movq 104(%%edx), %%mm1\n\t" | |
| 191 " movq 96(%%edx), %%mm5\n\t" | |
| 192 " pswapd %%mm1, %%mm1\n\t" | |
| 193 " pswapd %%mm5, %%mm5\n\t" | |
| 194 " pfadd %%mm1, %%mm0\n\t" | |
| 195 " pfadd %%mm5, %%mm4\n\t" | |
| 196 " movq %%mm0, 80(%%ecx)\n\t" | |
| 197 " movq %%mm4, 88(%%ecx)\n\t" | |
| 198 " pfsubr %%mm1, %%mm3\n\t" | |
| 199 " pfsubr %%mm5, %%mm7\n\t" | |
| 200 " pfmul 80(%%ebx), %%mm3\n\t" | |
| 201 " pfmul 88(%%ebx), %%mm7\n\t" | |
| 202 " pswapd %%mm3, %%mm3\n\t" | |
| 203 " pswapd %%mm7, %%mm7\n\t" | |
| 204 " movq %%mm3, 104(%%ecx)\n\t" | |
| 205 " movq %%mm7, 96(%%ecx)\n\t" | |
| 206 | |
| 207 /* Phase 4*/ | |
| 208 | |
| 209 " movq 96(%%ebx), %%mm2\n\t" | |
| 210 " movq 104(%%ebx), %%mm6\n\t" | |
| 211 | |
| 212 " movq (%%ecx), %%mm0\n\t" | |
| 213 " movq 8(%%ecx), %%mm4\n\t" | |
| 214 " movq %%mm0, %%mm3\n\t" | |
| 215 " movq %%mm4, %%mm7\n\t" | |
| 216 " movq 24(%%ecx), %%mm1\n\t" | |
| 217 " movq 16(%%ecx), %%mm5\n\t" | |
| 218 " pswapd %%mm1, %%mm1\n\t" | |
| 219 " pswapd %%mm5, %%mm5\n\t" | |
| 220 " pfadd %%mm1, %%mm0\n\t" | |
| 221 " pfadd %%mm5, %%mm4\n\t" | |
| 222 " movq %%mm0, (%%edx)\n\t" | |
| 223 " movq %%mm4, 8(%%edx)\n\t" | |
| 224 " pfsub %%mm1, %%mm3\n\t" | |
| 225 " pfsub %%mm5, %%mm7\n\t" | |
| 226 " pfmul %%mm2, %%mm3\n\t" | |
| 227 " pfmul %%mm6, %%mm7\n\t" | |
| 228 " pswapd %%mm3, %%mm3\n\t" | |
| 229 " pswapd %%mm7, %%mm7\n\t" | |
| 230 " movq %%mm3, 24(%%edx)\n\t" | |
| 231 " movq %%mm7, 16(%%edx)\n\t" | |
| 232 | |
| 233 " movq 32(%%ecx), %%mm0\n\t" | |
| 234 " movq 40(%%ecx), %%mm4\n\t" | |
| 235 " movq %%mm0, %%mm3\n\t" | |
| 236 " movq %%mm4, %%mm7\n\t" | |
| 237 " movq 56(%%ecx), %%mm1\n\t" | |
| 238 " movq 48(%%ecx), %%mm5\n\t" | |
| 239 " pswapd %%mm1, %%mm1\n\t" | |
| 240 " pswapd %%mm5, %%mm5\n\t" | |
| 241 " pfadd %%mm1, %%mm0\n\t" | |
| 242 " pfadd %%mm5, %%mm4\n\t" | |
| 243 " movq %%mm0, 32(%%edx)\n\t" | |
| 244 " movq %%mm4, 40(%%edx)\n\t" | |
| 245 " pfsubr %%mm1, %%mm3\n\t" | |
| 246 " pfsubr %%mm5, %%mm7\n\t" | |
| 247 " pfmul %%mm2, %%mm3\n\t" | |
| 248 " pfmul %%mm6, %%mm7\n\t" | |
| 249 " pswapd %%mm3, %%mm3\n\t" | |
| 250 " pswapd %%mm7, %%mm7\n\t" | |
| 251 " movq %%mm3, 56(%%edx)\n\t" | |
| 252 " movq %%mm7, 48(%%edx)\n\t" | |
| 253 | |
| 254 " movq 64(%%ecx), %%mm0\n\t" | |
| 255 " movq 72(%%ecx), %%mm4\n\t" | |
| 256 " movq %%mm0, %%mm3\n\t" | |
| 257 " movq %%mm4, %%mm7\n\t" | |
| 258 " movq 88(%%ecx), %%mm1\n\t" | |
| 259 " movq 80(%%ecx), %%mm5\n\t" | |
| 260 " pswapd %%mm1, %%mm1\n\t" | |
| 261 " pswapd %%mm5, %%mm5\n\t" | |
| 262 " pfadd %%mm1, %%mm0\n\t" | |
| 263 " pfadd %%mm5, %%mm4\n\t" | |
| 264 " movq %%mm0, 64(%%edx)\n\t" | |
| 265 " movq %%mm4, 72(%%edx)\n\t" | |
| 266 " pfsub %%mm1, %%mm3\n\t" | |
| 267 " pfsub %%mm5, %%mm7\n\t" | |
| 268 " pfmul %%mm2, %%mm3\n\t" | |
| 269 " pfmul %%mm6, %%mm7\n\t" | |
| 270 " pswapd %%mm3, %%mm3\n\t" | |
| 271 " pswapd %%mm7, %%mm7\n\t" | |
| 272 " movq %%mm3, 88(%%edx)\n\t" | |
| 273 " movq %%mm7, 80(%%edx)\n\t" | |
| 274 | |
| 275 " movq 96(%%ecx), %%mm0\n\t" | |
| 276 " movq 104(%%ecx), %%mm4\n\t" | |
| 277 " movq %%mm0, %%mm3\n\t" | |
| 278 " movq %%mm4, %%mm7\n\t" | |
| 279 " movq 120(%%ecx), %%mm1\n\t" | |
| 280 " movq 112(%%ecx), %%mm5\n\t" | |
| 281 " pswapd %%mm1, %%mm1\n\t" | |
| 282 " pswapd %%mm5, %%mm5\n\t" | |
| 283 " pfadd %%mm1, %%mm0\n\t" | |
| 284 " pfadd %%mm5, %%mm4\n\t" | |
| 285 " movq %%mm0, 96(%%edx)\n\t" | |
| 286 " movq %%mm4, 104(%%edx)\n\t" | |
| 287 " pfsubr %%mm1, %%mm3\n\t" | |
| 288 " pfsubr %%mm5, %%mm7\n\t" | |
| 289 " pfmul %%mm2, %%mm3\n\t" | |
| 290 " pfmul %%mm6, %%mm7\n\t" | |
| 291 " pswapd %%mm3, %%mm3\n\t" | |
| 292 " pswapd %%mm7, %%mm7\n\t" | |
| 293 " movq %%mm3, 120(%%edx)\n\t" | |
| 294 " movq %%mm7, 112(%%edx)\n\t" | |
| 295 | |
| 296 /* Phase 5 */ | |
| 297 | |
| 298 " movq 112(%%ebx), %%mm2\n\t" | |
| 299 | |
| 300 " movq (%%edx), %%mm0\n\t" | |
| 301 " movq 16(%%edx), %%mm4\n\t" | |
| 302 " movq %%mm0, %%mm3\n\t" | |
| 303 " movq %%mm4, %%mm7\n\t" | |
| 304 " movq 8(%%edx), %%mm1\n\t" | |
| 305 " movq 24(%%edx), %%mm5\n\t" | |
| 306 " pswapd %%mm1, %%mm1\n\t" | |
| 307 " pswapd %%mm5, %%mm5\n\t" | |
| 308 " pfadd %%mm1, %%mm0\n\t" | |
| 309 " pfadd %%mm5, %%mm4\n\t" | |
| 310 " movq %%mm0, (%%ecx)\n\t" | |
| 311 " movq %%mm4, 16(%%ecx)\n\t" | |
| 312 " pfsub %%mm1, %%mm3\n\t" | |
| 313 " pfsubr %%mm5, %%mm7\n\t" | |
| 314 " pfmul %%mm2, %%mm3\n\t" | |
| 315 " pfmul %%mm2, %%mm7\n\t" | |
| 316 " pswapd %%mm3, %%mm3\n\t" | |
| 317 " pswapd %%mm7, %%mm7\n\t" | |
| 318 " movq %%mm3, 8(%%ecx)\n\t" | |
| 319 " movq %%mm7, 24(%%ecx)\n\t" | |
| 320 | |
| 321 " movq 32(%%edx), %%mm0\n\t" | |
| 322 " movq 48(%%edx), %%mm4\n\t" | |
| 323 " movq %%mm0, %%mm3\n\t" | |
| 324 " movq %%mm4, %%mm7\n\t" | |
| 325 " movq 40(%%edx), %%mm1\n\t" | |
| 326 " movq 56(%%edx), %%mm5\n\t" | |
| 327 " pswapd %%mm1, %%mm1\n\t" | |
| 328 " pswapd %%mm5, %%mm5\n\t" | |
| 329 " pfadd %%mm1, %%mm0\n\t" | |
| 330 " pfadd %%mm5, %%mm4\n\t" | |
| 331 " movq %%mm0, 32(%%ecx)\n\t" | |
| 332 " movq %%mm4, 48(%%ecx)\n\t" | |
| 333 " pfsub %%mm1, %%mm3\n\t" | |
| 334 " pfsubr %%mm5, %%mm7\n\t" | |
| 335 " pfmul %%mm2, %%mm3\n\t" | |
| 336 " pfmul %%mm2, %%mm7\n\t" | |
| 337 " pswapd %%mm3, %%mm3\n\t" | |
| 338 " pswapd %%mm7, %%mm7\n\t" | |
| 339 " movq %%mm3, 40(%%ecx)\n\t" | |
| 340 " movq %%mm7, 56(%%ecx)\n\t" | |
| 341 | |
| 342 " movq 64(%%edx), %%mm0\n\t" | |
| 343 " movq 80(%%edx), %%mm4\n\t" | |
| 344 " movq %%mm0, %%mm3\n\t" | |
| 345 " movq %%mm4, %%mm7\n\t" | |
| 346 " movq 72(%%edx), %%mm1\n\t" | |
| 347 " movq 88(%%edx), %%mm5\n\t" | |
| 348 " pswapd %%mm1, %%mm1\n\t" | |
| 349 " pswapd %%mm5, %%mm5\n\t" | |
| 350 " pfadd %%mm1, %%mm0\n\t" | |
| 351 " pfadd %%mm5, %%mm4\n\t" | |
| 352 " movq %%mm0, 64(%%ecx)\n\t" | |
| 353 " movq %%mm4, 80(%%ecx)\n\t" | |
| 354 " pfsub %%mm1, %%mm3\n\t" | |
| 355 " pfsubr %%mm5, %%mm7\n\t" | |
| 356 " pfmul %%mm2, %%mm3\n\t" | |
| 357 " pfmul %%mm2, %%mm7\n\t" | |
| 358 " pswapd %%mm3, %%mm3\n\t" | |
| 359 " pswapd %%mm7, %%mm7\n\t" | |
| 360 " movq %%mm3, 72(%%ecx)\n\t" | |
| 361 " movq %%mm7, 88(%%ecx)\n\t" | |
| 362 | |
| 363 " movq 96(%%edx), %%mm0\n\t" | |
| 364 " movq 112(%%edx), %%mm4\n\t" | |
| 365 " movq %%mm0, %%mm3\n\t" | |
| 366 " movq %%mm4, %%mm7\n\t" | |
| 367 " movq 104(%%edx), %%mm1\n\t" | |
| 368 " movq 120(%%edx), %%mm5\n\t" | |
| 369 " pswapd %%mm1, %%mm1\n\t" | |
| 370 " pswapd %%mm5, %%mm5\n\t" | |
| 371 " pfadd %%mm1, %%mm0\n\t" | |
| 372 " pfadd %%mm5, %%mm4\n\t" | |
| 373 " movq %%mm0, 96(%%ecx)\n\t" | |
| 374 " movq %%mm4, 112(%%ecx)\n\t" | |
| 375 " pfsub %%mm1, %%mm3\n\t" | |
| 376 " pfsubr %%mm5, %%mm7\n\t" | |
| 377 " pfmul %%mm2, %%mm3\n\t" | |
| 378 " pfmul %%mm2, %%mm7\n\t" | |
| 379 " pswapd %%mm3, %%mm3\n\t" | |
| 380 " pswapd %%mm7, %%mm7\n\t" | |
| 381 " movq %%mm3, 104(%%ecx)\n\t" | |
| 382 " movq %%mm7, 120(%%ecx)\n\t" | |
| 383 | |
| 384 | |
| 385 /* Phase 6. This is the end of easy road. */ | |
| 386 /* Code below is coded in scalar mode. Should be optimized */ | |
| 387 | |
| 5291 | 388 " movd "MANGLE(plus_1f)", %%mm6\n\t" |
| 4148 | 389 " punpckldq 120(%%ebx), %%mm6\n\t" /* mm6 = 1.0 | 120(%%ebx)*/ |
| 390 " movq x_plus_minus_3dnow, %%mm7\n\t" /* mm7 = +1 | -1 */ | |
| 391 | |
| 392 " movq 32(%%ecx), %%mm0\n\t" | |
| 393 " movq 64(%%ecx), %%mm2\n\t" | |
| 394 " movq %%mm0, %%mm1\n\t" | |
| 395 " movq %%mm2, %%mm3\n\t" | |
| 396 " pxor %%mm7, %%mm1\n\t" | |
| 397 " pxor %%mm7, %%mm3\n\t" | |
| 398 " pfacc %%mm1, %%mm0\n\t" | |
| 399 " pfacc %%mm3, %%mm2\n\t" | |
| 400 " pfmul %%mm6, %%mm0\n\t" | |
| 401 " pfmul %%mm6, %%mm2\n\t" | |
| 402 " movq %%mm0, 32(%%edx)\n\t" | |
| 403 " movq %%mm2, 64(%%edx)\n\t" | |
| 404 | |
| 405 " movd 44(%%ecx), %%mm0\n\t" | |
| 406 " movd 40(%%ecx), %%mm2\n\t" | |
| 407 " movd 120(%%ebx), %%mm3\n\t" | |
| 408 " punpckldq 76(%%ecx), %%mm0\n\t" | |
| 409 " punpckldq 72(%%ecx), %%mm2\n\t" | |
| 410 " punpckldq %%mm3, %%mm3\n\t" | |
| 411 " movq %%mm0, %%mm4\n\t" | |
| 412 " movq %%mm2, %%mm5\n\t" | |
| 413 " pfsub %%mm2, %%mm0\n\t" | |
| 414 " pfmul %%mm3, %%mm0\n\t" | |
| 415 " movq %%mm0, %%mm1\n\t" | |
| 416 " pfadd %%mm5, %%mm0\n\t" | |
| 417 " pfadd %%mm4, %%mm0\n\t" | |
| 418 " movq %%mm0, %%mm2\n\t" | |
| 419 " punpckldq %%mm1, %%mm0\n\t" | |
| 420 " punpckhdq %%mm1, %%mm2\n\t" | |
| 421 " movq %%mm0, 40(%%edx)\n\t" | |
| 422 " movq %%mm2, 72(%%edx)\n\t" | |
| 423 | |
| 424 " movd 48(%%ecx), %%mm3\n\t" | |
| 425 " movd 60(%%ecx), %%mm2\n\t" | |
| 426 " pfsub 52(%%ecx), %%mm3\n\t" | |
| 427 " pfsub 56(%%ecx), %%mm2\n\t" | |
| 428 " pfmul 120(%%ebx), %%mm3\n\t" | |
| 429 " pfmul 120(%%ebx), %%mm2\n\t" | |
| 430 " movq %%mm2, %%mm1\n\t" | |
| 431 | |
| 432 " pfadd 56(%%ecx), %%mm1\n\t" | |
| 433 " pfadd 60(%%ecx), %%mm1\n\t" | |
| 434 " movq %%mm1, %%mm0\n\t" | |
| 435 | |
| 436 " pfadd 48(%%ecx), %%mm0\n\t" | |
| 437 " pfadd 52(%%ecx), %%mm0\n\t" | |
| 438 " pfadd %%mm3, %%mm1\n\t" | |
| 439 " punpckldq %%mm2, %%mm1\n\t" | |
| 440 " pfadd %%mm3, %%mm2\n\t" | |
| 441 " punpckldq %%mm2, %%mm0\n\t" | |
| 442 " movq %%mm1, 56(%%edx)\n\t" | |
| 443 " movq %%mm0, 48(%%edx)\n\t" | |
| 444 | |
| 445 /*---*/ | |
| 446 | |
| 447 " movd 92(%%ecx), %%mm1\n\t" | |
| 448 " pfsub 88(%%ecx), %%mm1\n\t" | |
| 449 " pfmul 120(%%ebx), %%mm1\n\t" | |
| 450 " movd %%mm1, 92(%%edx)\n\t" | |
| 451 " pfadd 92(%%ecx), %%mm1\n\t" | |
| 452 " pfadd 88(%%ecx), %%mm1\n\t" | |
| 453 " movq %%mm1, %%mm0\n\t" | |
| 454 | |
| 455 " pfadd 80(%%ecx), %%mm0\n\t" | |
| 456 " pfadd 84(%%ecx), %%mm0\n\t" | |
| 457 " movd %%mm0, 80(%%edx)\n\t" | |
| 458 | |
| 459 " movd 80(%%ecx), %%mm0\n\t" | |
| 460 " pfsub 84(%%ecx), %%mm0\n\t" | |
| 461 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 462 " pfadd %%mm0, %%mm1\n\t" | |
| 463 " pfadd 92(%%edx), %%mm0\n\t" | |
| 464 " punpckldq %%mm1, %%mm0\n\t" | |
| 465 " movq %%mm0, 84(%%edx)\n\t" | |
| 466 | |
| 467 " movq 96(%%ecx), %%mm0\n\t" | |
| 468 " movq %%mm0, %%mm1\n\t" | |
| 469 " pxor %%mm7, %%mm1\n\t" | |
| 470 " pfacc %%mm1, %%mm0\n\t" | |
| 471 " pfmul %%mm6, %%mm0\n\t" | |
| 472 " movq %%mm0, 96(%%edx)\n\t" | |
| 473 | |
| 474 " movd 108(%%ecx), %%mm0\n\t" | |
| 475 " pfsub 104(%%ecx), %%mm0\n\t" | |
| 476 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 477 " movd %%mm0, 108(%%edx)\n\t" | |
| 478 " pfadd 104(%%ecx), %%mm0\n\t" | |
| 479 " pfadd 108(%%ecx), %%mm0\n\t" | |
| 480 " movd %%mm0, 104(%%edx)\n\t" | |
| 481 | |
| 482 " movd 124(%%ecx), %%mm1\n\t" | |
| 483 " pfsub 120(%%ecx), %%mm1\n\t" | |
| 484 " pfmul 120(%%ebx), %%mm1\n\t" | |
| 485 " movd %%mm1, 124(%%edx)\n\t" | |
| 486 " pfadd 120(%%ecx), %%mm1\n\t" | |
| 487 " pfadd 124(%%ecx), %%mm1\n\t" | |
| 488 " movq %%mm1, %%mm0\n\t" | |
| 489 | |
| 490 " pfadd 112(%%ecx), %%mm0\n\t" | |
| 491 " pfadd 116(%%ecx), %%mm0\n\t" | |
| 492 " movd %%mm0, 112(%%edx)\n\t" | |
| 493 | |
| 494 " movd 112(%%ecx), %%mm0\n\t" | |
| 495 " pfsub 116(%%ecx), %%mm0\n\t" | |
| 496 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 497 " pfadd %%mm0,%%mm1\n\t" | |
| 498 " pfadd 124(%%edx), %%mm0\n\t" | |
| 499 " punpckldq %%mm1, %%mm0\n\t" | |
| 500 " movq %%mm0, 116(%%edx)\n\t" | |
| 501 | |
| 502 " jnz .L01\n\t" | |
| 503 | |
| 504 /* Phase 7*/ | |
| 505 /* Code below is coded in scalar mode. Should be optimized */ | |
| 506 | |
| 507 " movd (%%ecx), %%mm0\n\t" | |
| 508 " pfadd 4(%%ecx), %%mm0\n\t" | |
| 509 " movd %%mm0, 1024(%%esi)\n\t" | |
| 510 | |
| 511 " movd (%%ecx), %%mm0\n\t" | |
| 512 " pfsub 4(%%ecx), %%mm0\n\t" | |
| 513 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 514 " movd %%mm0, (%%esi)\n\t" | |
| 515 " movd %%mm0, (%%edi)\n\t" | |
| 516 | |
| 517 " movd 12(%%ecx), %%mm0\n\t" | |
| 518 " pfsub 8(%%ecx), %%mm0\n\t" | |
| 519 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 520 " movd %%mm0, 512(%%edi)\n\t" | |
| 521 " pfadd 12(%%ecx), %%mm0\n\t" | |
| 522 " pfadd 8(%%ecx), %%mm0\n\t" | |
| 523 " movd %%mm0, 512(%%esi)\n\t" | |
| 524 | |
| 525 " movd 16(%%ecx), %%mm0\n\t" | |
| 526 " pfsub 20(%%ecx), %%mm0\n\t" | |
| 527 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 528 " movq %%mm0, %%mm3\n\t" | |
| 529 | |
| 530 " movd 28(%%ecx), %%mm0\n\t" | |
| 531 " pfsub 24(%%ecx), %%mm0\n\t" | |
| 532 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 533 " movd %%mm0, 768(%%edi)\n\t" | |
| 534 " movq %%mm0, %%mm2\n\t" | |
| 535 | |
| 536 " pfadd 24(%%ecx), %%mm0\n\t" | |
| 537 " pfadd 28(%%ecx), %%mm0\n\t" | |
| 538 " movq %%mm0, %%mm1\n\t" | |
| 539 | |
| 540 " pfadd 16(%%ecx), %%mm0\n\t" | |
| 541 " pfadd 20(%%ecx), %%mm0\n\t" | |
| 542 " movd %%mm0, 768(%%esi)\n\t" | |
| 543 " pfadd %%mm3, %%mm1\n\t" | |
| 544 " movd %%mm1, 256(%%esi)\n\t" | |
| 545 " pfadd %%mm3, %%mm2\n\t" | |
| 546 " movd %%mm2, 256(%%edi)\n\t" | |
| 547 | |
| 548 /* Phase 8*/ | |
| 549 | |
| 550 " movq 32(%%edx), %%mm0\n\t" | |
| 551 " movq 48(%%edx), %%mm1\n\t" | |
| 552 " pfadd 48(%%edx), %%mm0\n\t" | |
| 553 " pfadd 40(%%edx), %%mm1\n\t" | |
| 554 " movd %%mm0, 896(%%esi)\n\t" | |
| 555 " movd %%mm1, 640(%%esi)\n\t" | |
| 556 " psrlq $32, %%mm0\n\t" | |
| 557 " psrlq $32, %%mm1\n\t" | |
| 558 " movd %%mm0, 128(%%edi)\n\t" | |
| 559 " movd %%mm1, 384(%%edi)\n\t" | |
| 560 | |
| 561 " movd 40(%%edx), %%mm0\n\t" | |
| 562 " pfadd 56(%%edx), %%mm0\n\t" | |
| 563 " movd %%mm0, 384(%%esi)\n\t" | |
| 564 | |
| 565 " movd 56(%%edx), %%mm0\n\t" | |
| 566 " pfadd 36(%%edx), %%mm0\n\t" | |
| 567 " movd %%mm0, 128(%%esi)\n\t" | |
| 568 | |
| 569 " movd 60(%%edx), %%mm0\n\t" | |
| 570 " movd %%mm0, 896(%%edi)\n\t" | |
| 571 " pfadd 44(%%edx), %%mm0\n\t" | |
| 572 " movd %%mm0, 640(%%edi)\n\t" | |
| 573 | |
| 574 " movq 96(%%edx), %%mm0\n\t" | |
| 575 " movq 112(%%edx), %%mm2\n\t" | |
| 576 " movq 104(%%edx), %%mm4\n\t" | |
| 577 " pfadd 112(%%edx), %%mm0\n\t" | |
| 578 " pfadd 104(%%edx), %%mm2\n\t" | |
| 579 " pfadd 120(%%edx), %%mm4\n\t" | |
| 580 " movq %%mm0, %%mm1\n\t" | |
| 581 " movq %%mm2, %%mm3\n\t" | |
| 582 " movq %%mm4, %%mm5\n\t" | |
| 583 " pfadd 64(%%edx), %%mm0\n\t" | |
| 584 " pfadd 80(%%edx), %%mm2\n\t" | |
| 585 " pfadd 72(%%edx), %%mm4\n\t" | |
| 586 " movd %%mm0, 960(%%esi)\n\t" | |
| 587 " movd %%mm2, 704(%%esi)\n\t" | |
| 588 " movd %%mm4, 448(%%esi)\n\t" | |
| 589 " psrlq $32, %%mm0\n\t" | |
| 590 " psrlq $32, %%mm2\n\t" | |
| 591 " psrlq $32, %%mm4\n\t" | |
| 592 " movd %%mm0, 64(%%edi)\n\t" | |
| 593 " movd %%mm2, 320(%%edi)\n\t" | |
| 594 " movd %%mm4, 576(%%edi)\n\t" | |
| 595 " pfadd 80(%%edx), %%mm1\n\t" | |
| 596 " pfadd 72(%%edx), %%mm3\n\t" | |
| 597 " pfadd 88(%%edx), %%mm5\n\t" | |
| 598 " movd %%mm1, 832(%%esi)\n\t" | |
| 599 " movd %%mm3, 576(%%esi)\n\t" | |
| 600 " movd %%mm5, 320(%%esi)\n\t" | |
| 601 " psrlq $32, %%mm1\n\t" | |
| 602 " psrlq $32, %%mm3\n\t" | |
| 603 " psrlq $32, %%mm5\n\t" | |
| 604 " movd %%mm1, 192(%%edi)\n\t" | |
| 605 " movd %%mm3, 448(%%edi)\n\t" | |
| 606 " movd %%mm5, 704(%%edi)\n\t" | |
| 607 | |
| 608 " movd 120(%%edx), %%mm0\n\t" | |
| 609 " pfadd 100(%%edx), %%mm0\n\t" | |
| 610 " movq %%mm0, %%mm1\n\t" | |
| 611 " pfadd 88(%%edx), %%mm0\n\t" | |
| 612 " movd %%mm0, 192(%%esi)\n\t" | |
| 613 " pfadd 68(%%edx), %%mm1\n\t" | |
| 614 " movd %%mm1, 64(%%esi)\n\t" | |
| 615 | |
| 616 " movd 124(%%edx), %%mm0\n\t" | |
| 617 " movd %%mm0, 960(%%edi)\n\t" | |
| 618 " pfadd 92(%%edx), %%mm0\n\t" | |
| 619 " movd %%mm0, 832(%%edi)\n\t" | |
| 620 | |
| 621 " jmp .L_bye\n\t" | |
| 622 ".L01: \n\t" | |
| 623 /* Phase 9*/ | |
| 624 | |
| 625 " movq (%%ecx), %%mm0\n\t" | |
| 626 " movq %%mm0, %%mm1\n\t" | |
| 627 " pxor %%mm7, %%mm1\n\t" | |
| 628 " pfacc %%mm1, %%mm0\n\t" | |
| 629 " pfmul %%mm6, %%mm0\n\t" | |
| 630 " pf2id %%mm0, %%mm0\n\t" | |
| 631 " movd %%mm0, %%eax\n\t" | |
| 632 " movw %%ax, 512(%%esi)\n\t" | |
| 633 " psrlq $32, %%mm0\n\t" | |
| 634 " movd %%mm0, %%eax\n\t" | |
| 635 " movw %%ax, (%%esi)\n\t" | |
| 636 | |
| 637 " movd 12(%%ecx), %%mm0\n\t" | |
| 638 " pfsub 8(%%ecx), %%mm0\n\t" | |
| 639 " pfmul 120(%%ebx), %%mm0\n\t" | |
| 640 " pf2id %%mm0, %%mm7\n\t" | |
| 641 " movd %%mm7, %%eax\n\t" | |
| 642 " movw %%ax, 256(%%edi)\n\t" | |
| 643 " pfadd 12(%%ecx), %%mm0\n\t" | |
| 644 " pfadd 8(%%ecx), %%mm0\n\t" | |
| 645 " pf2id %%mm0, %%mm0\n\t" | |
| 646 " movd %%mm0, %%eax\n\t" | |
| 647 " movw %%ax, 256(%%esi)\n\t" | |
| 648 | |
| 649 " movd 16(%%ecx), %%mm3\n\t" | |
| 650 " pfsub 20(%%ecx), %%mm3\n\t" | |
| 651 " pfmul 120(%%ebx), %%mm3\n\t" | |
| 652 " movq %%mm3, %%mm2\n\t" | |
| 653 | |
| 654 " movd 28(%%ecx), %%mm2\n\t" | |
| 655 " pfsub 24(%%ecx), %%mm2\n\t" | |
| 656 " pfmul 120(%%ebx), %%mm2\n\t" | |
| 657 " movq %%mm2, %%mm1\n\t" | |
| 658 | |
| 659 " pf2id %%mm2, %%mm7\n\t" | |
| 660 " movd %%mm7, %%eax\n\t" | |
| 661 " movw %%ax, 384(%%edi)\n\t" | |
| 662 | |
| 663 " pfadd 24(%%ecx), %%mm1\n\t" | |
| 664 " pfadd 28(%%ecx), %%mm1\n\t" | |
| 665 " movq %%mm1, %%mm0\n\t" | |
| 666 | |
| 667 " pfadd 16(%%ecx), %%mm0\n\t" | |
| 668 " pfadd 20(%%ecx), %%mm0\n\t" | |
| 669 " pf2id %%mm0, %%mm0\n\t" | |
| 670 " movd %%mm0, %%eax\n\t" | |
| 671 " movw %%ax, 384(%%esi)\n\t" | |
| 672 " pfadd %%mm3, %%mm1\n\t" | |
| 673 " pf2id %%mm1, %%mm1\n\t" | |
| 674 " movd %%mm1, %%eax\n\t" | |
| 675 " movw %%ax, 128(%%esi)\n\t" | |
| 676 " pfadd %%mm3, %%mm2\n\t" | |
| 677 " pf2id %%mm2, %%mm2\n\t" | |
| 678 " movd %%mm2, %%eax\n\t" | |
| 679 " movw %%ax, 128(%%edi)\n\t" | |
| 680 | |
| 681 /* Phase 10*/ | |
| 682 | |
| 683 " movq 32(%%edx), %%mm0\n\t" | |
| 684 " movq 48(%%edx), %%mm1\n\t" | |
| 685 " pfadd 48(%%edx), %%mm0\n\t" | |
| 686 " pfadd 40(%%edx), %%mm1\n\t" | |
| 687 " pf2id %%mm0, %%mm0\n\t" | |
| 688 " pf2id %%mm1, %%mm1\n\t" | |
| 689 " movd %%mm0, %%eax\n\t" | |
| 690 " movd %%mm1, %%ecx\n\t" | |
| 691 " movw %%ax, 448(%%esi)\n\t" | |
| 692 " movw %%cx, 320(%%esi)\n\t" | |
| 693 " psrlq $32, %%mm0\n\t" | |
| 694 " psrlq $32, %%mm1\n\t" | |
| 695 " movd %%mm0, %%eax\n\t" | |
| 696 " movd %%mm1, %%ecx\n\t" | |
| 697 " movw %%ax, 64(%%edi)\n\t" | |
| 698 " movw %%cx, 192(%%edi)\n\t" | |
| 699 | |
| 700 " movd 40(%%edx), %%mm3\n\t" | |
| 701 " movd 56(%%edx), %%mm4\n\t" | |
| 702 " movd 60(%%edx), %%mm0\n\t" | |
| 703 " movd 44(%%edx), %%mm2\n\t" | |
| 704 " movd 120(%%edx), %%mm5\n\t" | |
| 705 " punpckldq %%mm4, %%mm3\n\t" | |
| 706 " punpckldq 124(%%edx), %%mm0\n\t" | |
| 707 " pfadd 100(%%edx), %%mm5\n\t" | |
| 708 " punpckldq 36(%%edx), %%mm4\n\t" | |
| 709 " punpckldq 92(%%edx), %%mm2\n\t" | |
| 710 " movq %%mm5, %%mm6\n\t" | |
| 711 " pfadd %%mm4, %%mm3\n\t" | |
| 712 " pf2id %%mm0, %%mm1\n\t" | |
| 713 " pf2id %%mm3, %%mm3\n\t" | |
| 714 " pfadd 88(%%edx), %%mm5\n\t" | |
| 715 " movd %%mm1, %%eax\n\t" | |
| 716 " movd %%mm3, %%ecx\n\t" | |
| 717 " movw %%ax, 448(%%edi)\n\t" | |
| 718 " movw %%cx, 192(%%esi)\n\t" | |
| 719 " pf2id %%mm5, %%mm5\n\t" | |
| 720 " psrlq $32, %%mm1\n\t" | |
| 721 " psrlq $32, %%mm3\n\t" | |
| 722 " movd %%mm5, %%ebx\n\t" | |
| 723 " movd %%mm1, %%eax\n\t" | |
| 724 " movd %%mm3, %%ecx\n\t" | |
| 725 " movw %%bx, 96(%%esi)\n\t" | |
| 726 " movw %%ax, 480(%%edi)\n\t" | |
| 727 " movw %%cx, 64(%%esi)\n\t" | |
| 728 " pfadd %%mm2, %%mm0\n\t" | |
| 729 " pf2id %%mm0, %%mm0\n\t" | |
| 730 " movd %%mm0, %%eax\n\t" | |
| 731 " pfadd 68(%%edx), %%mm6\n\t" | |
| 732 " movw %%ax, 320(%%edi)\n\t" | |
| 733 " psrlq $32, %%mm0\n\t" | |
| 734 " pf2id %%mm6, %%mm6\n\t" | |
| 735 " movd %%mm0, %%eax\n\t" | |
| 736 " movd %%mm6, %%ebx\n\t" | |
| 737 " movw %%ax, 416(%%edi)\n\t" | |
| 738 " movw %%bx, 32(%%esi)\n\t" | |
| 739 | |
| 740 " movq 96(%%edx), %%mm0\n\t" | |
| 741 " movq 112(%%edx), %%mm2\n\t" | |
| 742 " movq 104(%%edx), %%mm4\n\t" | |
| 743 " pfadd %%mm2, %%mm0\n\t" | |
| 744 " pfadd %%mm4, %%mm2\n\t" | |
| 745 " pfadd 120(%%edx), %%mm4\n\t" | |
| 746 " movq %%mm0, %%mm1\n\t" | |
| 747 " movq %%mm2, %%mm3\n\t" | |
| 748 " movq %%mm4, %%mm5\n\t" | |
| 749 " pfadd 64(%%edx), %%mm0\n\t" | |
| 750 " pfadd 80(%%edx), %%mm2\n\t" | |
| 751 " pfadd 72(%%edx), %%mm4\n\t" | |
| 752 " pf2id %%mm0, %%mm0\n\t" | |
| 753 " pf2id %%mm2, %%mm2\n\t" | |
| 754 " pf2id %%mm4, %%mm4\n\t" | |
| 755 " movd %%mm0, %%eax\n\t" | |
| 756 " movd %%mm2, %%ecx\n\t" | |
| 757 " movd %%mm4, %%ebx\n\t" | |
| 758 " movw %%ax, 480(%%esi)\n\t" | |
| 759 " movw %%cx, 352(%%esi)\n\t" | |
| 760 " movw %%bx, 224(%%esi)\n\t" | |
| 761 " psrlq $32, %%mm0\n\t" | |
| 762 " psrlq $32, %%mm2\n\t" | |
| 763 " psrlq $32, %%mm4\n\t" | |
| 764 " movd %%mm0, %%eax\n\t" | |
| 765 " movd %%mm2, %%ecx\n\t" | |
| 766 " movd %%mm4, %%ebx\n\t" | |
| 767 " movw %%ax, 32(%%edi)\n\t" | |
| 768 " movw %%cx, 160(%%edi)\n\t" | |
| 769 " movw %%bx, 288(%%edi)\n\t" | |
| 770 " pfadd 80(%%edx), %%mm1\n\t" | |
| 771 " pfadd 72(%%edx), %%mm3\n\t" | |
| 772 " pfadd 88(%%edx), %%mm5\n\t" | |
| 773 " pf2id %%mm1, %%mm1\n\t" | |
| 774 " pf2id %%mm3, %%mm3\n\t" | |
| 775 " pf2id %%mm5, %%mm5\n\t" | |
| 776 " movd %%mm1, %%eax\n\t" | |
| 777 " movd %%mm3, %%ecx\n\t" | |
| 778 " movd %%mm5, %%ebx\n\t" | |
| 779 " movw %%ax, 416(%%esi)\n\t" | |
| 780 " movw %%cx, 288(%%esi)\n\t" | |
| 781 " movw %%bx, 160(%%esi)\n\t" | |
| 782 " psrlq $32, %%mm1\n\t" | |
| 783 " psrlq $32, %%mm3\n\t" | |
| 784 " psrlq $32, %%mm5\n\t" | |
| 785 " movd %%mm1, %%eax\n\t" | |
| 786 " movd %%mm3, %%ecx\n\t" | |
| 787 " movd %%mm5, %%ebx\n\t" | |
| 788 " movw %%ax, 96(%%edi)\n\t" | |
| 789 " movw %%cx, 224(%%edi)\n\t" | |
| 790 " movw %%bx, 352(%%edi)\n\t" | |
| 791 | |
| 792 " movsw\n\t" | |
| 793 | |
| 794 ".L_bye:\n\t" | |
| 795 " femms\n\t" | |
| 796 : | |
| 797 :"m"(a),"m"(b),"m"(c),"m"(tmp[0]) | |
| 798 :"memory","%ebx","%esi","%edi"); | |
|
5842
d6eab895c742
Avoid stdcall on cygwin, it causes undefined ref, code needs testing as I have no athlon or k6-2.
atmos4
parents:
5291
diff
changeset
|
799 } |