mp3lib/dct64_mmx.c

   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 */
   6 #include "mangle.h"
   7 #define real float /* ugly - but only way */
   8
   9 void dct64_MMX(short *a,short *b,real *c)
  10 {
  11     char tmp[256];
  12     __asm __volatile(
  13 "       movl %2,%%eax\n\t"
  14 /* Phase 1*/
  15 "       flds     (%%eax)\n\t"
  16 "       leal 128+%3,%%edx\n\t"
  17 "       fadds 124(%%eax)\n\t"
  18 "       movl %0,%%esi\n\t"
  19 "       fstps    (%%edx)\n\t"
  20 "       movl %1,%%edi\n\t"
  21
  22 "       flds    4(%%eax)\n\t"
  23 "       movl $"MANGLE(costab_mmx)",%%ebx\n\t"
  24 "       fadds 120(%%eax)\n\t"
  25 "       orl %%ecx,%%ecx\n\t"
  26 "       fstps   4(%%edx)\n\t"
  27
  28 "       flds     (%%eax)\n\t"
  29 "       leal %3,%%ecx\n\t"
  30 "       fsubs 124(%%eax)\n\t"
  31 "       fmuls    (%%ebx)\n\t"
  32 "       fstps 124(%%edx)\n\t"
  33
  34 "       flds    4(%%eax)\n\t"
  35 "       fsubs 120(%%eax)\n\t"
  36 "       fmuls   4(%%ebx)\n\t"
  37 "       fstps 120(%%edx)\n\t"
  38
  39 "       flds    8(%%eax)\n\t"
  40 "       fadds 116(%%eax)\n\t"
  41 "       fstps   8(%%edx)\n\t"
  42
  43 "       flds   12(%%eax)\n\t"
  44 "       fadds 112(%%eax)\n\t"
  45 "       fstps  12(%%edx)\n\t"
  46
  47 "       flds    8(%%eax)\n\t"
  48 "       fsubs 116(%%eax)\n\t"
  49 "       fmuls   8(%%ebx)\n\t"
  50 "       fstps 116(%%edx)\n\t"
  51
  52 "       flds   12(%%eax)\n\t"
  53 "       fsubs 112(%%eax)\n\t"
  54 "       fmuls  12(%%ebx)\n\t"
  55 "       fstps 112(%%edx)\n\t"
  56
  57 "       flds   16(%%eax)\n\t"
  58 "       fadds 108(%%eax)\n\t"
  59 "       fstps  16(%%edx)\n\t"
  60
  61 "       flds   20(%%eax)\n\t"
  62 "       fadds 104(%%eax)\n\t"
  63 "       fstps  20(%%edx)\n\t"
  64
  65 "       flds   16(%%eax)\n\t"
  66 "       fsubs 108(%%eax)\n\t"
  67 "       fmuls  16(%%ebx)\n\t"
  68 "       fstps 108(%%edx)\n\t"
  69
  70 "       flds   20(%%eax)\n\t"
  71 "       fsubs 104(%%eax)\n\t"
  72 "       fmuls  20(%%ebx)\n\t"
  73 "       fstps 104(%%edx)\n\t"
  74
  75 "       flds   24(%%eax)\n\t"
  76 "       fadds 100(%%eax)\n\t"
  77 "       fstps  24(%%edx)\n\t"
  78
  79 "       flds   28(%%eax)\n\t"
  80 "       fadds  96(%%eax)\n\t"
  81 "       fstps  28(%%edx)\n\t"
  82
  83 "       flds   24(%%eax)\n\t"
  84 "       fsubs 100(%%eax)\n\t"
  85 "       fmuls  24(%%ebx)\n\t"
  86 "       fstps 100(%%edx)\n\t"
  87
  88 "       flds   28(%%eax)\n\t"
  89 "       fsubs  96(%%eax)\n\t"
  90 "       fmuls  28(%%ebx)\n\t"
  91 "       fstps  96(%%edx)\n\t"
  92
  93 "       flds   32(%%eax)\n\t"
  94 "       fadds  92(%%eax)\n\t"
  95 "       fstps  32(%%edx)\n\t"
  96
  97 "       flds   36(%%eax)\n\t"
  98 "       fadds  88(%%eax)\n\t"
  99 "       fstps  36(%%edx)\n\t"
 100
 101 "       flds   32(%%eax)\n\t"
 102 "       fsubs  92(%%eax)\n\t"
 103 "       fmuls  32(%%ebx)\n\t"
 104 "       fstps  92(%%edx)\n\t"
 105
 106 "       flds   36(%%eax)\n\t"
 107 "       fsubs  88(%%eax)\n\t"
 108 "       fmuls  36(%%ebx)\n\t"
 109 "       fstps  88(%%edx)\n\t"
 110
 111 "       flds   40(%%eax)\n\t"
 112 "       fadds  84(%%eax)\n\t"
 113 "       fstps  40(%%edx)\n\t"
 114
 115 "       flds   44(%%eax)\n\t"
 116 "       fadds  80(%%eax)\n\t"
 117 "       fstps  44(%%edx)\n\t"
 118
 119 "       flds   40(%%eax)\n\t"
 120 "       fsubs  84(%%eax)\n\t"
 121 "       fmuls  40(%%ebx)\n\t"
 122 "       fstps  84(%%edx)\n\t"
 123
 124 "       flds   44(%%eax)\n\t"
 125 "       fsubs  80(%%eax)\n\t"
 126 "       fmuls  44(%%ebx)\n\t"
 127 "       fstps  80(%%edx)\n\t"
 128
 129 "       flds   48(%%eax)\n\t"
 130 "       fadds  76(%%eax)\n\t"
 131 "       fstps  48(%%edx)\n\t"
 132
 133 "       flds   52(%%eax)\n\t"
 134 "       fadds  72(%%eax)\n\t"
 135 "       fstps  52(%%edx)\n\t"
 136
 137 "       flds   48(%%eax)\n\t"
 138 "       fsubs  76(%%eax)\n\t"
 139 "       fmuls  48(%%ebx)\n\t"
 140 "       fstps  76(%%edx)\n\t"
 141
 142 "       flds   52(%%eax)\n\t"
 143 "       fsubs  72(%%eax)\n\t"
 144 "       fmuls  52(%%ebx)\n\t"
 145 "       fstps  72(%%edx)\n\t"
 146
 147 "       flds   56(%%eax)\n\t"
 148 "       fadds  68(%%eax)\n\t"
 149 "       fstps  56(%%edx)\n\t"
 150
 151 "       flds   60(%%eax)\n\t"
 152 "       fadds  64(%%eax)\n\t"
 153 "       fstps  60(%%edx)\n\t"
 154
 155 "       flds   56(%%eax)\n\t"
 156 "       fsubs  68(%%eax)\n\t"
 157 "       fmuls  56(%%ebx)\n\t"
 158 "       fstps  68(%%edx)\n\t"
 159
 160 "       flds   60(%%eax)\n\t"
 161 "       fsubs  64(%%eax)\n\t"
 162 "       fmuls  60(%%ebx)\n\t"
 163 "       fstps  64(%%edx)\n\t"
 164
 165 /* Phase 2*/
 166
 167 "       flds     (%%edx)\n\t"
 168 "       fadds  60(%%edx)\n\t"
 169 "       fstps    (%%ecx)\n\t"
 170
 171 "       flds    4(%%edx)\n\t"
 172 "       fadds  56(%%edx)\n\t"
 173 "       fstps   4(%%ecx)\n\t"
 174
 175 "       flds     (%%edx)\n\t"
 176 "       fsubs  60(%%edx)\n\t"
 177 "       fmuls  64(%%ebx)\n\t"
 178 "       fstps  60(%%ecx)\n\t"
 179
 180 "       flds    4(%%edx)\n\t"
 181 "       fsubs  56(%%edx)\n\t"
 182 "       fmuls  68(%%ebx)\n\t"
 183 "       fstps  56(%%ecx)\n\t"
 184
 185 "       flds    8(%%edx)\n\t"
 186 "       fadds  52(%%edx)\n\t"
 187 "       fstps   8(%%ecx)\n\t"
 188
 189 "       flds   12(%%edx)\n\t"
 190 "       fadds  48(%%edx)\n\t"
 191 "       fstps  12(%%ecx)\n\t"
 192
 193 "       flds    8(%%edx)\n\t"
 194 "       fsubs  52(%%edx)\n\t"
 195 "       fmuls  72(%%ebx)\n\t"
 196 "       fstps  52(%%ecx)\n\t"
 197
 198 "       flds   12(%%edx)\n\t"
 199 "       fsubs  48(%%edx)\n\t"
 200 "       fmuls  76(%%ebx)\n\t"
 201 "       fstps  48(%%ecx)\n\t"
 202
 203 "       flds   16(%%edx)\n\t"
 204 "       fadds  44(%%edx)\n\t"
 205 "       fstps  16(%%ecx)\n\t"
 206
 207 "       flds   20(%%edx)\n\t"
 208 "       fadds  40(%%edx)\n\t"
 209 "       fstps  20(%%ecx)\n\t"
 210
 211 "       flds   16(%%edx)\n\t"
 212 "       fsubs  44(%%edx)\n\t"
 213 "       fmuls  80(%%ebx)\n\t"
 214 "       fstps  44(%%ecx)\n\t"
 215
 216 "       flds   20(%%edx)\n\t"
 217 "       fsubs  40(%%edx)\n\t"
 218 "       fmuls  84(%%ebx)\n\t"
 219 "       fstps  40(%%ecx)\n\t"
 220
 221 "       flds   24(%%edx)\n\t"
 222 "       fadds  36(%%edx)\n\t"
 223 "       fstps  24(%%ecx)\n\t"
 224
 225 "       flds   28(%%edx)\n\t"
 226 "       fadds  32(%%edx)\n\t"
 227 "       fstps  28(%%ecx)\n\t"
 228
 229 "       flds   24(%%edx)\n\t"
 230 "       fsubs  36(%%edx)\n\t"
 231 "       fmuls  88(%%ebx)\n\t"
 232 "       fstps  36(%%ecx)\n\t"
 233
 234 "       flds   28(%%edx)\n\t"
 235 "       fsubs  32(%%edx)\n\t"
 236 "       fmuls  92(%%ebx)\n\t"
 237 "       fstps  32(%%ecx)\n\t"
 238
 239 /* Phase 3*/
 240
 241 "       flds   64(%%edx)\n\t"
 242 "       fadds 124(%%edx)\n\t"
 243 "       fstps  64(%%ecx)\n\t"
 244
 245 "       flds   68(%%edx)\n\t"
 246 "       fadds 120(%%edx)\n\t"
 247 "       fstps  68(%%ecx)\n\t"
 248
 249 "       flds  124(%%edx)\n\t"
 250 "       fsubs  64(%%edx)\n\t"
 251 "       fmuls  64(%%ebx)\n\t"
 252 "       fstps 124(%%ecx)\n\t"
 253
 254 "       flds  120(%%edx)\n\t"
 255 "       fsubs  68(%%edx)\n\t"
 256 "       fmuls  68(%%ebx)\n\t"
 257 "       fstps 120(%%ecx)\n\t"
 258
 259 "       flds   72(%%edx)\n\t"
 260 "       fadds 116(%%edx)\n\t"
 261 "       fstps  72(%%ecx)\n\t"
 262
 263 "       flds   76(%%edx)\n\t"
 264 "       fadds 112(%%edx)\n\t"
 265 "       fstps  76(%%ecx)\n\t"
 266
 267 "       flds  116(%%edx)\n\t"
 268 "       fsubs  72(%%edx)\n\t"
 269 "       fmuls  72(%%ebx)\n\t"
 270 "       fstps 116(%%ecx)\n\t"
 271
 272 "       flds  112(%%edx)\n\t"
 273 "       fsubs  76(%%edx)\n\t"
 274 "       fmuls  76(%%ebx)\n\t"
 275 "       fstps 112(%%ecx)\n\t"
 276
 277 "       flds   80(%%edx)\n\t"
 278 "       fadds 108(%%edx)\n\t"
 279 "       fstps  80(%%ecx)\n\t"
 280
 281 "       flds   84(%%edx)\n\t"
 282 "       fadds 104(%%edx)\n\t"
 283 "       fstps  84(%%ecx)\n\t"
 284
 285 "       flds  108(%%edx)\n\t"
 286 "       fsubs  80(%%edx)\n\t"
 287 "       fmuls  80(%%ebx)\n\t"
 288 "       fstps 108(%%ecx)\n\t"
 289
 290 "       flds  104(%%edx)\n\t"
 291 "       fsubs  84(%%edx)\n\t"
 292 "       fmuls  84(%%ebx)\n\t"
 293 "       fstps 104(%%ecx)\n\t"
 294
 295 "       flds   88(%%edx)\n\t"
 296 "       fadds 100(%%edx)\n\t"
 297 "       fstps  88(%%ecx)\n\t"
 298
 299 "       flds   92(%%edx)\n\t"
 300 "       fadds  96(%%edx)\n\t"
 301 "       fstps  92(%%ecx)\n\t"
 302
 303 "       flds  100(%%edx)\n\t"
 304 "       fsubs  88(%%edx)\n\t"
 305 "       fmuls  88(%%ebx)\n\t"
 306 "       fstps 100(%%ecx)\n\t"
 307
 308 "       flds   96(%%edx)\n\t"
 309 "       fsubs  92(%%edx)\n\t"
 310 "       fmuls  92(%%ebx)\n\t"
 311 "       fstps  96(%%ecx)\n\t"
 312
 313 /* Phase 4*/
 314
 315 "       flds     (%%ecx)\n\t"
 316 "       fadds  28(%%ecx)\n\t"
 317 "       fstps    (%%edx)\n\t"
 318
 319 "       flds     (%%ecx)\n\t"
 320 "       fsubs  28(%%ecx)\n\t"
 321 "       fmuls  96(%%ebx)\n\t"
 322 "       fstps  28(%%edx)\n\t"
 323
 324 "       flds    4(%%ecx)\n\t"
 325 "       fadds  24(%%ecx)\n\t"
 326 "       fstps   4(%%edx)\n\t"
 327
 328 "       flds    4(%%ecx)\n\t"
 329 "       fsubs  24(%%ecx)\n\t"
 330 "       fmuls 100(%%ebx)\n\t"
 331 "       fstps  24(%%edx)\n\t"
 332
 333 "       flds    8(%%ecx)\n\t"
 334 "       fadds  20(%%ecx)\n\t"
 335 "       fstps   8(%%edx)\n\t"
 336
 337 "       flds    8(%%ecx)\n\t"
 338 "       fsubs  20(%%ecx)\n\t"
 339 "       fmuls 104(%%ebx)\n\t"
 340 "       fstps  20(%%edx)\n\t"
 341
 342 "       flds   12(%%ecx)\n\t"
 343 "       fadds  16(%%ecx)\n\t"
 344 "       fstps  12(%%edx)\n\t"
 345
 346 "       flds   12(%%ecx)\n\t"
 347 "       fsubs  16(%%ecx)\n\t"
 348 "       fmuls 108(%%ebx)\n\t"
 349 "       fstps  16(%%edx)\n\t"
 350
 351 "       flds   32(%%ecx)\n\t"
 352 "       fadds  60(%%ecx)\n\t"
 353 "       fstps  32(%%edx)\n\t"
 354
 355 "       flds   60(%%ecx)\n\t"
 356 "       fsubs  32(%%ecx)\n\t"
 357 "       fmuls  96(%%ebx)\n\t"
 358 "       fstps  60(%%edx)\n\t"
 359
 360 "       flds   36(%%ecx)\n\t"
 361 "       fadds  56(%%ecx)\n\t"
 362 "       fstps  36(%%edx)\n\t"
 363
 364 "       flds   56(%%ecx)\n\t"
 365 "       fsubs  36(%%ecx)\n\t"
 366 "       fmuls 100(%%ebx)\n\t"
 367 "       fstps  56(%%edx)\n\t"
 368
 369 "       flds   40(%%ecx)\n\t"
 370 "       fadds  52(%%ecx)\n\t"
 371 "       fstps  40(%%edx)\n\t"
 372
 373 "       flds   52(%%ecx)\n\t"
 374 "       fsubs  40(%%ecx)\n\t"
 375 "       fmuls 104(%%ebx)\n\t"
 376 "       fstps  52(%%edx)\n\t"
 377
 378 "       flds   44(%%ecx)\n\t"
 379 "       fadds  48(%%ecx)\n\t"
 380 "       fstps  44(%%edx)\n\t"
 381
 382 "       flds   48(%%ecx)\n\t"
 383 "       fsubs  44(%%ecx)\n\t"
 384 "       fmuls 108(%%ebx)\n\t"
 385 "       fstps  48(%%edx)\n\t"
 386
 387 "       flds   64(%%ecx)\n\t"
 388 "       fadds  92(%%ecx)\n\t"
 389 "       fstps  64(%%edx)\n\t"
 390
 391 "       flds   64(%%ecx)\n\t"
 392 "       fsubs  92(%%ecx)\n\t"
 393 "       fmuls  96(%%ebx)\n\t"
 394 "       fstps  92(%%edx)\n\t"
 395
 396 "       flds   68(%%ecx)\n\t"
 397 "       fadds  88(%%ecx)\n\t"
 398 "       fstps  68(%%edx)\n\t"
 399
 400 "       flds   68(%%ecx)\n\t"
 401 "       fsubs  88(%%ecx)\n\t"
 402 "       fmuls 100(%%ebx)\n\t"
 403 "       fstps  88(%%edx)\n\t"
 404
 405 "       flds   72(%%ecx)\n\t"
 406 "       fadds  84(%%ecx)\n\t"
 407 "       fstps  72(%%edx)\n\t"
 408
 409 "       flds   72(%%ecx)\n\t"
 410 "       fsubs  84(%%ecx)\n\t"
 411 "       fmuls 104(%%ebx)\n\t"
 412 "       fstps  84(%%edx)\n\t"
 413
 414 "       flds   76(%%ecx)\n\t"
 415 "       fadds  80(%%ecx)\n\t"
 416 "       fstps  76(%%edx)\n\t"
 417
 418 "       flds   76(%%ecx)\n\t"
 419 "       fsubs  80(%%ecx)\n\t"
 420 "       fmuls 108(%%ebx)\n\t"
 421 "       fstps  80(%%edx)\n\t"
 422
 423 "       flds   96(%%ecx)\n\t"
 424 "       fadds 124(%%ecx)\n\t"
 425 "       fstps  96(%%edx)\n\t"
 426
 427 "       flds  124(%%ecx)\n\t"
 428 "       fsubs  96(%%ecx)\n\t"
 429 "       fmuls  96(%%ebx)\n\t"
 430 "       fstps 124(%%edx)\n\t"
 431
 432 "       flds  100(%%ecx)\n\t"
 433 "       fadds 120(%%ecx)\n\t"
 434 "       fstps 100(%%edx)\n\t"
 435
 436 "       flds  120(%%ecx)\n\t"
 437 "       fsubs 100(%%ecx)\n\t"
 438 "       fmuls 100(%%ebx)\n\t"
 439 "       fstps 120(%%edx)\n\t"
 440
 441 "       flds  104(%%ecx)\n\t"
 442 "       fadds 116(%%ecx)\n\t"
 443 "       fstps 104(%%edx)\n\t"
 444
 445 "       flds  116(%%ecx)\n\t"
 446 "       fsubs 104(%%ecx)\n\t"
 447 "       fmuls 104(%%ebx)\n\t"
 448 "       fstps 116(%%edx)\n\t"
 449
 450 "       flds  108(%%ecx)\n\t"
 451 "       fadds 112(%%ecx)\n\t"
 452 "       fstps 108(%%edx)\n\t"
 453
 454 "       flds  112(%%ecx)\n\t"
 455 "       fsubs 108(%%ecx)\n\t"
 456 "       fmuls 108(%%ebx)\n\t"
 457 "       fstps 112(%%edx)\n\t"
 458
 459 "       flds     (%%edx)\n\t"
 460 "       fadds  12(%%edx)\n\t"
 461 "       fstps    (%%ecx)\n\t"
 462
 463 "       flds     (%%edx)\n\t"
 464 "       fsubs  12(%%edx)\n\t"
 465 "       fmuls 112(%%ebx)\n\t"
 466 "       fstps  12(%%ecx)\n\t"
 467
 468 "       flds    4(%%edx)\n\t"
 469 "       fadds   8(%%edx)\n\t"
 470 "       fstps   4(%%ecx)\n\t"
 471
 472 "       flds    4(%%edx)\n\t"
 473 "       fsubs   8(%%edx)\n\t"
 474 "       fmuls 116(%%ebx)\n\t"
 475 "       fstps   8(%%ecx)\n\t"
 476
 477 "       flds   16(%%edx)\n\t"
 478 "       fadds  28(%%edx)\n\t"
 479 "       fstps  16(%%ecx)\n\t"
 480
 481 "       flds   28(%%edx)\n\t"
 482 "       fsubs  16(%%edx)\n\t"
 483 "       fmuls 112(%%ebx)\n\t"
 484 "       fstps  28(%%ecx)\n\t"
 485
 486 "       flds   20(%%edx)\n\t"
 487 "       fadds  24(%%edx)\n\t"
 488 "       fstps  20(%%ecx)\n\t"
 489
 490 "       flds   24(%%edx)\n\t"
 491 "       fsubs  20(%%edx)\n\t"
 492 "       fmuls 116(%%ebx)\n\t"
 493 "       fstps  24(%%ecx)\n\t"
 494
 495 "       flds   32(%%edx)\n\t"
 496 "       fadds  44(%%edx)\n\t"
 497 "       fstps  32(%%ecx)\n\t"
 498
 499 "       flds   32(%%edx)\n\t"
 500 "       fsubs  44(%%edx)\n\t"
 501 "       fmuls 112(%%ebx)\n\t"
 502 "       fstps  44(%%ecx)\n\t"
 503
 504 "       flds   36(%%edx)\n\t"
 505 "       fadds  40(%%edx)\n\t"
 506 "       fstps  36(%%ecx)\n\t"
 507
 508 "       flds   36(%%edx)\n\t"
 509 "       fsubs  40(%%edx)\n\t"
 510 "       fmuls 116(%%ebx)\n\t"
 511 "       fstps  40(%%ecx)\n\t"
 512
 513 "       flds   48(%%edx)\n\t"
 514 "       fadds  60(%%edx)\n\t"
 515 "       fstps  48(%%ecx)\n\t"
 516
 517 "       flds   60(%%edx)\n\t"
 518 "       fsubs  48(%%edx)\n\t"
 519 "       fmuls 112(%%ebx)\n\t"
 520 "       fstps  60(%%ecx)\n\t"
 521
 522 "       flds   52(%%edx)\n\t"
 523 "       fadds  56(%%edx)\n\t"
 524 "       fstps  52(%%ecx)\n\t"
 525
 526 "       flds   56(%%edx)\n\t"
 527 "       fsubs  52(%%edx)\n\t"
 528 "       fmuls 116(%%ebx)\n\t"
 529 "       fstps  56(%%ecx)\n\t"
 530
 531 "       flds   64(%%edx)\n\t"
 532 "       fadds  76(%%edx)\n\t"
 533 "       fstps  64(%%ecx)\n\t"
 534
 535 "       flds   64(%%edx)\n\t"
 536 "       fsubs  76(%%edx)\n\t"
 537 "       fmuls 112(%%ebx)\n\t"
 538 "       fstps  76(%%ecx)\n\t"
 539
 540 "       flds   68(%%edx)\n\t"
 541 "       fadds  72(%%edx)\n\t"
 542 "       fstps  68(%%ecx)\n\t"
 543
 544 "       flds   68(%%edx)\n\t"
 545 "       fsubs  72(%%edx)\n\t"
 546 "       fmuls 116(%%ebx)\n\t"
 547 "       fstps  72(%%ecx)\n\t"
 548
 549 "       flds   80(%%edx)\n\t"
 550 "       fadds  92(%%edx)\n\t"
 551 "       fstps  80(%%ecx)\n\t"
 552
 553 "       flds   92(%%edx)\n\t"
 554 "       fsubs  80(%%edx)\n\t"
 555 "       fmuls 112(%%ebx)\n\t"
 556 "       fstps  92(%%ecx)\n\t"
 557
 558 "       flds   84(%%edx)\n\t"
 559 "       fadds  88(%%edx)\n\t"
 560 "       fstps  84(%%ecx)\n\t"
 561
 562 "       flds   88(%%edx)\n\t"
 563 "       fsubs  84(%%edx)\n\t"
 564 "       fmuls 116(%%ebx)\n\t"
 565 "       fstps  88(%%ecx)\n\t"
 566
 567 "       flds   96(%%edx)\n\t"
 568 "       fadds 108(%%edx)\n\t"
 569 "       fstps  96(%%ecx)\n\t"
 570
 571 "       flds   96(%%edx)\n\t"
 572 "       fsubs 108(%%edx)\n\t"
 573 "       fmuls 112(%%ebx)\n\t"
 574 "       fstps 108(%%ecx)\n\t"
 575
 576 "       flds  100(%%edx)\n\t"
 577 "       fadds 104(%%edx)\n\t"
 578 "       fstps 100(%%ecx)\n\t"
 579
 580 "       flds  100(%%edx)\n\t"
 581 "       fsubs 104(%%edx)\n\t"
 582 "       fmuls 116(%%ebx)\n\t"
 583 "       fstps 104(%%ecx)\n\t"
 584
 585 "       flds  112(%%edx)\n\t"
 586 "       fadds 124(%%edx)\n\t"
 587 "       fstps 112(%%ecx)\n\t"
 588
 589 "       flds  124(%%edx)\n\t"
 590 "       fsubs 112(%%edx)\n\t"
 591 "       fmuls 112(%%ebx)\n\t"
 592 "       fstps 124(%%ecx)\n\t"
 593
 594 "       flds  116(%%edx)\n\t"
 595 "       fadds 120(%%edx)\n\t"
 596 "       fstps 116(%%ecx)\n\t"
 597
 598 "       flds  120(%%edx)\n\t"
 599 "       fsubs 116(%%edx)\n\t"
 600 "       fmuls 116(%%ebx)\n\t"
 601 "       fstps 120(%%ecx)\n\t"
 602
 603 /* Phase 5*/
 604
 605 "       flds   32(%%ecx)\n\t"
 606 "       fadds  36(%%ecx)\n\t"
 607 "       fstps  32(%%edx)\n\t"
 608
 609 "       flds   32(%%ecx)\n\t"
 610 "       fsubs  36(%%ecx)\n\t"
 611 "       fmuls 120(%%ebx)\n\t"
 612 "       fstps  36(%%edx)\n\t"
 613
 614 "       flds   44(%%ecx)\n\t"
 615 "       fsubs  40(%%ecx)\n\t"
 616 "       fmuls 120(%%ebx)\n\t"
 617 "       fsts   44(%%edx)\n\t"
 618 "       fadds  40(%%ecx)\n\t"
 619 "       fadds  44(%%ecx)\n\t"
 620 "       fstps  40(%%edx)\n\t"
 621
 622 "       flds   48(%%ecx)\n\t"
 623 "       fsubs  52(%%ecx)\n\t"
 624 "       fmuls 120(%%ebx)\n\t"
 625
 626 "       flds   60(%%ecx)\n\t"
 627 "       fsubs  56(%%ecx)\n\t"
 628 "       fmuls 120(%%ebx)\n\t"
 629 "       fld      %%st(0)\n\t"
 630 "       fadds  56(%%ecx)\n\t"
 631 "       fadds  60(%%ecx)\n\t"
 632 "       fld      %%st(0)\n\t"
 633 "       fadds  48(%%ecx)\n\t"
 634 "       fadds  52(%%ecx)\n\t"
 635 "       fstps  48(%%edx)\n\t"
 636 "       fadd     %%st(2)\n\t"
 637 "       fstps  56(%%edx)\n\t"
 638 "       fsts   60(%%edx)\n\t"
 639 "       faddp    %%st(1)\n\t"
 640 "       fstps  52(%%edx)\n\t"
 641
 642 "       flds   64(%%ecx)\n\t"
 643 "       fadds  68(%%ecx)\n\t"
 644 "       fstps  64(%%edx)\n\t"
 645
 646 "       flds   64(%%ecx)\n\t"
 647 "       fsubs  68(%%ecx)\n\t"
 648 "       fmuls 120(%%ebx)\n\t"
 649 "       fstps  68(%%edx)\n\t"
 650
 651 "       flds   76(%%ecx)\n\t"
 652 "       fsubs  72(%%ecx)\n\t"
 653 "       fmuls 120(%%ebx)\n\t"
 654 "       fsts   76(%%edx)\n\t"
 655 "       fadds  72(%%ecx)\n\t"
 656 "       fadds  76(%%ecx)\n\t"
 657 "       fstps  72(%%edx)\n\t"
 658
 659 "       flds   92(%%ecx)\n\t"
 660 "       fsubs  88(%%ecx)\n\t"
 661 "       fmuls 120(%%ebx)\n\t"
 662 "       fsts   92(%%edx)\n\t"
 663 "       fadds  92(%%ecx)\n\t"
 664 "       fadds  88(%%ecx)\n\t"
 665 "       fld      %%st(0)\n\t"
 666 "       fadds  80(%%ecx)\n\t"
 667 "       fadds  84(%%ecx)\n\t"
 668 "       fstps  80(%%edx)\n\t"
 669
 670 "       flds   80(%%ecx)\n\t"
 671 "       fsubs  84(%%ecx)\n\t"
 672 "       fmuls 120(%%ebx)\n\t"
 673 "       fadd  %%st(0), %%st(1)\n\t"
 674 "       fadds 92(%%edx)\n\t"
 675 "       fstps 84(%%edx)\n\t"
 676 "       fstps 88(%%edx)\n\t"
 677
 678 "       flds   96(%%ecx)\n\t"
 679 "       fadds 100(%%ecx)\n\t"
 680 "       fstps  96(%%edx)\n\t"
 681
 682 "       flds   96(%%ecx)\n\t"
 683 "       fsubs 100(%%ecx)\n\t"
 684 "       fmuls 120(%%ebx)\n\t"
 685 "       fstps 100(%%edx)\n\t"
 686
 687 "       flds  108(%%ecx)\n\t"
 688 "       fsubs 104(%%ecx)\n\t"
 689 "       fmuls 120(%%ebx)\n\t"
 690 "       fsts  108(%%edx)\n\t"
 691 "       fadds 104(%%ecx)\n\t"
 692 "       fadds 108(%%ecx)\n\t"
 693 "       fstps 104(%%edx)\n\t"
 694
 695 "       flds  124(%%ecx)\n\t"
 696 "       fsubs 120(%%ecx)\n\t"
 697 "       fmuls 120(%%ebx)\n\t"
 698 "       fsts  124(%%edx)\n\t"
 699 "       fadds 120(%%ecx)\n\t"
 700 "       fadds 124(%%ecx)\n\t"
 701 "       fld      %%st(0)\n\t"
 702 "       fadds 112(%%ecx)\n\t"
 703 "       fadds 116(%%ecx)\n\t"
 704 "       fstps 112(%%edx)\n\t"
 705
 706 "       flds  112(%%ecx)\n\t"
 707 "       fsubs 116(%%ecx)\n\t"
 708 "       fmuls 120(%%ebx)\n\t"
 709 "       fadd  %%st(0),%%st(1)\n\t"
 710 "       fadds 124(%%edx)\n\t"
 711 "       fstps 116(%%edx)\n\t"
 712 "       fstps 120(%%edx)\n\t"
 713 "       jnz .L01\n\t"
 714
 715 /* Phase 6*/
 716
 717 "       flds      (%%ecx)\n\t"
 718 "       fadds    4(%%ecx)\n\t"
 719 "       fstps 1024(%%esi)\n\t"
 720
 721 "       flds      (%%ecx)\n\t"
 722 "       fsubs    4(%%ecx)\n\t"
 723 "       fmuls  120(%%ebx)\n\t"
 724 "       fsts      (%%esi)\n\t"
 725 "       fstps     (%%edi)\n\t"
 726
 727 "       flds   12(%%ecx)\n\t"
 728 "       fsubs   8(%%ecx)\n\t"
 729 "       fmuls 120(%%ebx)\n\t"
 730 "       fsts  512(%%edi)\n\t"
 731 "       fadds  12(%%ecx)\n\t"
 732 "       fadds   8(%%ecx)\n\t"
 733 "       fstps 512(%%esi)\n\t"
 734
 735 "       flds   16(%%ecx)\n\t"
 736 "       fsubs  20(%%ecx)\n\t"
 737 "       fmuls 120(%%ebx)\n\t"
 738
 739 "       flds   28(%%ecx)\n\t"
 740 "       fsubs  24(%%ecx)\n\t"
 741 "       fmuls 120(%%ebx)\n\t"
 742 "       fsts  768(%%edi)\n\t"
 743 "       fld      %%st(0)\n\t"
 744 "       fadds  24(%%ecx)\n\t"
 745 "       fadds  28(%%ecx)\n\t"
 746 "       fld      %%st(0)\n\t"
 747 "       fadds  16(%%ecx)\n\t"
 748 "       fadds  20(%%ecx)\n\t"
 749 "       fstps 768(%%esi)\n\t"
 750 "       fadd     %%st(2)\n\t"
 751 "       fstps 256(%%esi)\n\t"
 752 "       faddp    %%st(1)\n\t"
 753 "       fstps 256(%%edi)\n\t"
 754
 755 /* Phase 7*/
 756
 757 "       flds   32(%%edx)\n\t"
 758 "       fadds  48(%%edx)\n\t"
 759 "       fstps 896(%%esi)\n\t"
 760
 761 "       flds   48(%%edx)\n\t"
 762 "       fadds  40(%%edx)\n\t"
 763 "       fstps 640(%%esi)\n\t"
 764
 765 "       flds   40(%%edx)\n\t"
 766 "       fadds  56(%%edx)\n\t"
 767 "       fstps 384(%%esi)\n\t"
 768
 769 "       flds   56(%%edx)\n\t"
 770 "       fadds  36(%%edx)\n\t"
 771 "       fstps 128(%%esi)\n\t"
 772
 773 "       flds   36(%%edx)\n\t"
 774 "       fadds  52(%%edx)\n\t"
 775 "       fstps 128(%%edi)\n\t"
 776
 777 "       flds   52(%%edx)\n\t"
 778 "       fadds  44(%%edx)\n\t"
 779 "       fstps 384(%%edi)\n\t"
 780
 781 "       flds   60(%%edx)\n\t"
 782 "       fsts  896(%%edi)\n\t"
 783 "       fadds  44(%%edx)\n\t"
 784 "       fstps 640(%%edi)\n\t"
 785
 786 "       flds   96(%%edx)\n\t"
 787 "       fadds 112(%%edx)\n\t"
 788 "       fld      %%st(0)\n\t"
 789 "       fadds  64(%%edx)\n\t"
 790 "       fstps 960(%%esi)\n\t"
 791 "       fadds  80(%%edx)\n\t"
 792 "       fstps 832(%%esi)\n\t"
 793
 794 "       flds  112(%%edx)\n\t"
 795 "       fadds 104(%%edx)\n\t"
 796 "       fld      %%st(0)\n\t"
 797 "       fadds  80(%%edx)\n\t"
 798 "       fstps 704(%%esi)\n\t"
 799 "       fadds  72(%%edx)\n\t"
 800 "       fstps 576(%%esi)\n\t"
 801
 802 "       flds  104(%%edx)\n\t"
 803 "       fadds 120(%%edx)\n\t"
 804 "       fld      %%st(0)\n\t"
 805 "       fadds  72(%%edx)\n\t"
 806 "       fstps 448(%%esi)\n\t"
 807 "       fadds  88(%%edx)\n\t"
 808 "       fstps 320(%%esi)\n\t"
 809
 810 "       flds  120(%%edx)\n\t"
 811 "       fadds 100(%%edx)\n\t"
 812 "       fld      %%st(0)\n\t"
 813 "       fadds  88(%%edx)\n\t"
 814 "       fstps 192(%%esi)\n\t"
 815 "       fadds  68(%%edx)\n\t"
 816 "       fstps  64(%%esi)\n\t"
 817
 818 "       flds  100(%%edx)\n\t"
 819 "       fadds 116(%%edx)\n\t"
 820 "       fld      %%st(0)\n\t"
 821 "       fadds  68(%%edx)\n\t"
 822 "       fstps  64(%%edi)\n\t"
 823 "       fadds  84(%%edx)\n\t"
 824 "       fstps 192(%%edi)\n\t"
 825
 826 "       flds  116(%%edx)\n\t"
 827 "       fadds 108(%%edx)\n\t"
 828 "       fld      %%st(0)\n\t"
 829 "       fadds  84(%%edx)\n\t"
 830 "       fstps 320(%%edi)\n\t"
 831 "       fadds  76(%%edx)\n\t"
 832 "       fstps 448(%%edi)\n\t"
 833
 834 "       flds  108(%%edx)\n\t"
 835 "       fadds 124(%%edx)\n\t"
 836 "       fld      %%st(0)\n\t"
 837 "       fadds  76(%%edx)\n\t"
 838 "       fstps 576(%%edi)\n\t"
 839 "       fadds  92(%%edx)\n\t"
 840 "       fstps 704(%%edi)\n\t"
 841
 842 "       flds  124(%%edx)\n\t"
 843 "       fsts  960(%%edi)\n\t"
 844 "       fadds  92(%%edx)\n\t"
 845 "       fstps 832(%%edi)\n\t"
 846 "       jmp     .L_bye\n\t"
 847 ".L01:\n\t"
 848 /* Phase 8*/
 849
 850 "       flds      (%%ecx)\n\t"
 851 "       fadds    4(%%ecx)\n\t"
 852 "       fistp  512(%%esi)\n\t"
 853
 854 "       flds      (%%ecx)\n\t"
 855 "       fsubs    4(%%ecx)\n\t"
 856 "       fmuls  120(%%ebx)\n\t"
 857
 858 "       fistp     (%%esi)\n\t"
 859
 860
 861 "       flds    12(%%ecx)\n\t"
 862 "       fsubs    8(%%ecx)\n\t"
 863 "       fmuls  120(%%ebx)\n\t"
 864 "       fist   256(%%edi)\n\t"
 865 "       fadds   12(%%ecx)\n\t"
 866 "       fadds    8(%%ecx)\n\t"
 867 "       fistp  256(%%esi)\n\t"
 868
 869 "       flds   16(%%ecx)\n\t"
 870 "       fsubs  20(%%ecx)\n\t"
 871 "       fmuls 120(%%ebx)\n\t"
 872
 873 "       flds   28(%%ecx)\n\t"
 874 "       fsubs  24(%%ecx)\n\t"
 875 "       fmuls 120(%%ebx)\n\t"
 876 "       fist  384(%%edi)\n\t"
 877 "       fld      %%st(0)\n\t"
 878 "       fadds  24(%%ecx)\n\t"
 879 "       fadds  28(%%ecx)\n\t"
 880 "       fld      %%st(0)\n\t"
 881 "       fadds  16(%%ecx)\n\t"
 882 "       fadds  20(%%ecx)\n\t"
 883 "       fistp  384(%%esi)\n\t"
 884 "       fadd     %%st(2)\n\t"
 885 "       fistp  128(%%esi)\n\t"
 886 "       faddp    %%st(1)\n\t"
 887 "       fistp  128(%%edi)\n\t"
 888
 889 /* Phase 9*/
 890
 891 "       flds    32(%%edx)\n\t"
 892 "       fadds   48(%%edx)\n\t"
 893 "       fistp  448(%%esi)\n\t"
 894
 895 "       flds   48(%%edx)\n\t"
 896 "       fadds  40(%%edx)\n\t"
 897 "       fistp 320(%%esi)\n\t"
 898
 899 "       flds   40(%%edx)\n\t"
 900 "       fadds  56(%%edx)\n\t"
 901 "       fistp 192(%%esi)\n\t"
 902
 903 "       flds   56(%%edx)\n\t"
 904 "       fadds  36(%%edx)\n\t"
 905 "       fistp  64(%%esi)\n\t"
 906
 907 "       flds   36(%%edx)\n\t"
 908 "       fadds  52(%%edx)\n\t"
 909 "       fistp  64(%%edi)\n\t"
 910
 911 "       flds   52(%%edx)\n\t"
 912 "       fadds  44(%%edx)\n\t"
 913 "       fistp 192(%%edi)\n\t"
 914
 915 "       flds   60(%%edx)\n\t"
 916 "       fist   448(%%edi)\n\t"
 917 "       fadds  44(%%edx)\n\t"
 918 "       fistp 320(%%edi)\n\t"
 919
 920 "       flds   96(%%edx)\n\t"
 921 "       fadds 112(%%edx)\n\t"
 922 "       fld      %%st(0)\n\t"
 923 "       fadds  64(%%edx)\n\t"
 924 "       fistp 480(%%esi)\n\t"
 925 "       fadds  80(%%edx)\n\t"
 926 "       fistp 416(%%esi)\n\t"
 927
 928 "       flds  112(%%edx)\n\t"
 929 "       fadds 104(%%edx)\n\t"
 930 "       fld      %%st(0)\n\t"
 931 "       fadds  80(%%edx)\n\t"
 932 "       fistp 352(%%esi)\n\t"
 933 "       fadds  72(%%edx)\n\t"
 934 "       fistp 288(%%esi)\n\t"
 935
 936 "       flds  104(%%edx)\n\t"
 937 "       fadds 120(%%edx)\n\t"
 938 "       fld      %%st(0)\n\t"
 939 "       fadds  72(%%edx)\n\t"
 940 "       fistp 224(%%esi)\n\t"
 941 "       fadds  88(%%edx)\n\t"
 942 "       fistp 160(%%esi)\n\t"
 943
 944 "       flds  120(%%edx)\n\t"
 945 "       fadds 100(%%edx)\n\t"
 946 "       fld      %%st(0)\n\t"
 947 "       fadds  88(%%edx)\n\t"
 948 "       fistp  96(%%esi)\n\t"
 949 "       fadds  68(%%edx)\n\t"
 950 "       fistp  32(%%esi)\n\t"
 951
 952 "       flds  100(%%edx)\n\t"
 953 "       fadds 116(%%edx)\n\t"
 954 "       fld      %%st(0)\n\t"
 955 "       fadds  68(%%edx)\n\t"
 956 "       fistp  32(%%edi)\n\t"
 957 "       fadds  84(%%edx)\n\t"
 958 "       fistp  96(%%edi)\n\t"
 959
 960 "       flds  116(%%edx)\n\t"
 961 "       fadds 108(%%edx)\n\t"
 962 "       fld      %%st(0)\n\t"
 963 "       fadds  84(%%edx)\n\t"
 964 "       fistp 160(%%edi)\n\t"
 965 "       fadds  76(%%edx)\n\t"
 966 "       fistp 224(%%edi)\n\t"
 967
 968 "       flds  108(%%edx)\n\t"
 969 "       fadds 124(%%edx)\n\t"
 970 "       fld      %%st(0)\n\t"
 971 "       fadds  76(%%edx)\n\t"
 972 "       fistp 288(%%edi)\n\t"
 973 "       fadds  92(%%edx)\n\t"
 974 "       fistp 352(%%edi)\n\t"
 975
 976 "       flds  124(%%edx)\n\t"
 977 "       fist  480(%%edi)\n\t"
 978 "       fadds  92(%%edx)\n\t"
 979 "       fistp 416(%%edi)\n\t"
 980 "       movsw\n\t"
 981 ".L_bye:"
 982         :
 983         :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
 984         :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi");
 985 }