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 "config.h"
   7 #include "mangle.h"
   8 #define real float /* ugly - but only way */
   9
  10 void dct64_MMX(short *a,short *b,real *c)
  11 {
  12     char tmp[256];
  13     __asm__ volatile(
  14 "       movl %2,%%eax\n\t"
  15 /* Phase 1*/
  16 "       flds     (%%eax)\n\t"
  17 "       leal 128+%3,%%edx\n\t"
  18 "       fadds 124(%%eax)\n\t"
  19 "       movl %0,%%esi\n\t"
  20 "       fstps    (%%edx)\n\t"
  21 "       movl %1,%%edi\n\t"
  22
  23 "       flds    4(%%eax)\n\t"
  24 "       movl $"MANGLE(costab_mmx)",%%ebx\n\t"
  25 "       fadds 120(%%eax)\n\t"
  26 "       orl %%ecx,%%ecx\n\t"
  27 "       fstps   4(%%edx)\n\t"
  28
  29 "       flds     (%%eax)\n\t"
  30 "       leal %3,%%ecx\n\t"
  31 "       fsubs 124(%%eax)\n\t"
  32 "       fmuls    (%%ebx)\n\t"
  33 "       fstps 124(%%edx)\n\t"
  34
  35 "       flds    4(%%eax)\n\t"
  36 "       fsubs 120(%%eax)\n\t"
  37 "       fmuls   4(%%ebx)\n\t"
  38 "       fstps 120(%%edx)\n\t"
  39
  40 "       flds    8(%%eax)\n\t"
  41 "       fadds 116(%%eax)\n\t"
  42 "       fstps   8(%%edx)\n\t"
  43
  44 "       flds   12(%%eax)\n\t"
  45 "       fadds 112(%%eax)\n\t"
  46 "       fstps  12(%%edx)\n\t"
  47
  48 "       flds    8(%%eax)\n\t"
  49 "       fsubs 116(%%eax)\n\t"
  50 "       fmuls   8(%%ebx)\n\t"
  51 "       fstps 116(%%edx)\n\t"
  52
  53 "       flds   12(%%eax)\n\t"
  54 "       fsubs 112(%%eax)\n\t"
  55 "       fmuls  12(%%ebx)\n\t"
  56 "       fstps 112(%%edx)\n\t"
  57
  58 "       flds   16(%%eax)\n\t"
  59 "       fadds 108(%%eax)\n\t"
  60 "       fstps  16(%%edx)\n\t"
  61
  62 "       flds   20(%%eax)\n\t"
  63 "       fadds 104(%%eax)\n\t"
  64 "       fstps  20(%%edx)\n\t"
  65
  66 "       flds   16(%%eax)\n\t"
  67 "       fsubs 108(%%eax)\n\t"
  68 "       fmuls  16(%%ebx)\n\t"
  69 "       fstps 108(%%edx)\n\t"
  70
  71 "       flds   20(%%eax)\n\t"
  72 "       fsubs 104(%%eax)\n\t"
  73 "       fmuls  20(%%ebx)\n\t"
  74 "       fstps 104(%%edx)\n\t"
  75
  76 "       flds   24(%%eax)\n\t"
  77 "       fadds 100(%%eax)\n\t"
  78 "       fstps  24(%%edx)\n\t"
  79
  80 "       flds   28(%%eax)\n\t"
  81 "       fadds  96(%%eax)\n\t"
  82 "       fstps  28(%%edx)\n\t"
  83
  84 "       flds   24(%%eax)\n\t"
  85 "       fsubs 100(%%eax)\n\t"
  86 "       fmuls  24(%%ebx)\n\t"
  87 "       fstps 100(%%edx)\n\t"
  88
  89 "       flds   28(%%eax)\n\t"
  90 "       fsubs  96(%%eax)\n\t"
  91 "       fmuls  28(%%ebx)\n\t"
  92 "       fstps  96(%%edx)\n\t"
  93
  94 "       flds   32(%%eax)\n\t"
  95 "       fadds  92(%%eax)\n\t"
  96 "       fstps  32(%%edx)\n\t"
  97
  98 "       flds   36(%%eax)\n\t"
  99 "       fadds  88(%%eax)\n\t"
 100 "       fstps  36(%%edx)\n\t"
 101
 102 "       flds   32(%%eax)\n\t"
 103 "       fsubs  92(%%eax)\n\t"
 104 "       fmuls  32(%%ebx)\n\t"
 105 "       fstps  92(%%edx)\n\t"
 106
 107 "       flds   36(%%eax)\n\t"
 108 "       fsubs  88(%%eax)\n\t"
 109 "       fmuls  36(%%ebx)\n\t"
 110 "       fstps  88(%%edx)\n\t"
 111
 112 "       flds   40(%%eax)\n\t"
 113 "       fadds  84(%%eax)\n\t"
 114 "       fstps  40(%%edx)\n\t"
 115
 116 "       flds   44(%%eax)\n\t"
 117 "       fadds  80(%%eax)\n\t"
 118 "       fstps  44(%%edx)\n\t"
 119
 120 "       flds   40(%%eax)\n\t"
 121 "       fsubs  84(%%eax)\n\t"
 122 "       fmuls  40(%%ebx)\n\t"
 123 "       fstps  84(%%edx)\n\t"
 124
 125 "       flds   44(%%eax)\n\t"
 126 "       fsubs  80(%%eax)\n\t"
 127 "       fmuls  44(%%ebx)\n\t"
 128 "       fstps  80(%%edx)\n\t"
 129
 130 "       flds   48(%%eax)\n\t"
 131 "       fadds  76(%%eax)\n\t"
 132 "       fstps  48(%%edx)\n\t"
 133
 134 "       flds   52(%%eax)\n\t"
 135 "       fadds  72(%%eax)\n\t"
 136 "       fstps  52(%%edx)\n\t"
 137
 138 "       flds   48(%%eax)\n\t"
 139 "       fsubs  76(%%eax)\n\t"
 140 "       fmuls  48(%%ebx)\n\t"
 141 "       fstps  76(%%edx)\n\t"
 142
 143 "       flds   52(%%eax)\n\t"
 144 "       fsubs  72(%%eax)\n\t"
 145 "       fmuls  52(%%ebx)\n\t"
 146 "       fstps  72(%%edx)\n\t"
 147
 148 "       flds   56(%%eax)\n\t"
 149 "       fadds  68(%%eax)\n\t"
 150 "       fstps  56(%%edx)\n\t"
 151
 152 "       flds   60(%%eax)\n\t"
 153 "       fadds  64(%%eax)\n\t"
 154 "       fstps  60(%%edx)\n\t"
 155
 156 "       flds   56(%%eax)\n\t"
 157 "       fsubs  68(%%eax)\n\t"
 158 "       fmuls  56(%%ebx)\n\t"
 159 "       fstps  68(%%edx)\n\t"
 160
 161 "       flds   60(%%eax)\n\t"
 162 "       fsubs  64(%%eax)\n\t"
 163 "       fmuls  60(%%ebx)\n\t"
 164 "       fstps  64(%%edx)\n\t"
 165
 166 /* Phase 2*/
 167
 168 "       flds     (%%edx)\n\t"
 169 "       fadds  60(%%edx)\n\t"
 170 "       fstps    (%%ecx)\n\t"
 171
 172 "       flds    4(%%edx)\n\t"
 173 "       fadds  56(%%edx)\n\t"
 174 "       fstps   4(%%ecx)\n\t"
 175
 176 "       flds     (%%edx)\n\t"
 177 "       fsubs  60(%%edx)\n\t"
 178 "       fmuls  64(%%ebx)\n\t"
 179 "       fstps  60(%%ecx)\n\t"
 180
 181 "       flds    4(%%edx)\n\t"
 182 "       fsubs  56(%%edx)\n\t"
 183 "       fmuls  68(%%ebx)\n\t"
 184 "       fstps  56(%%ecx)\n\t"
 185
 186 "       flds    8(%%edx)\n\t"
 187 "       fadds  52(%%edx)\n\t"
 188 "       fstps   8(%%ecx)\n\t"
 189
 190 "       flds   12(%%edx)\n\t"
 191 "       fadds  48(%%edx)\n\t"
 192 "       fstps  12(%%ecx)\n\t"
 193
 194 "       flds    8(%%edx)\n\t"
 195 "       fsubs  52(%%edx)\n\t"
 196 "       fmuls  72(%%ebx)\n\t"
 197 "       fstps  52(%%ecx)\n\t"
 198
 199 "       flds   12(%%edx)\n\t"
 200 "       fsubs  48(%%edx)\n\t"
 201 "       fmuls  76(%%ebx)\n\t"
 202 "       fstps  48(%%ecx)\n\t"
 203
 204 "       flds   16(%%edx)\n\t"
 205 "       fadds  44(%%edx)\n\t"
 206 "       fstps  16(%%ecx)\n\t"
 207
 208 "       flds   20(%%edx)\n\t"
 209 "       fadds  40(%%edx)\n\t"
 210 "       fstps  20(%%ecx)\n\t"
 211
 212 "       flds   16(%%edx)\n\t"
 213 "       fsubs  44(%%edx)\n\t"
 214 "       fmuls  80(%%ebx)\n\t"
 215 "       fstps  44(%%ecx)\n\t"
 216
 217 "       flds   20(%%edx)\n\t"
 218 "       fsubs  40(%%edx)\n\t"
 219 "       fmuls  84(%%ebx)\n\t"
 220 "       fstps  40(%%ecx)\n\t"
 221
 222 "       flds   24(%%edx)\n\t"
 223 "       fadds  36(%%edx)\n\t"
 224 "       fstps  24(%%ecx)\n\t"
 225
 226 "       flds   28(%%edx)\n\t"
 227 "       fadds  32(%%edx)\n\t"
 228 "       fstps  28(%%ecx)\n\t"
 229
 230 "       flds   24(%%edx)\n\t"
 231 "       fsubs  36(%%edx)\n\t"
 232 "       fmuls  88(%%ebx)\n\t"
 233 "       fstps  36(%%ecx)\n\t"
 234
 235 "       flds   28(%%edx)\n\t"
 236 "       fsubs  32(%%edx)\n\t"
 237 "       fmuls  92(%%ebx)\n\t"
 238 "       fstps  32(%%ecx)\n\t"
 239
 240 /* Phase 3*/
 241
 242 "       flds   64(%%edx)\n\t"
 243 "       fadds 124(%%edx)\n\t"
 244 "       fstps  64(%%ecx)\n\t"
 245
 246 "       flds   68(%%edx)\n\t"
 247 "       fadds 120(%%edx)\n\t"
 248 "       fstps  68(%%ecx)\n\t"
 249
 250 "       flds  124(%%edx)\n\t"
 251 "       fsubs  64(%%edx)\n\t"
 252 "       fmuls  64(%%ebx)\n\t"
 253 "       fstps 124(%%ecx)\n\t"
 254
 255 "       flds  120(%%edx)\n\t"
 256 "       fsubs  68(%%edx)\n\t"
 257 "       fmuls  68(%%ebx)\n\t"
 258 "       fstps 120(%%ecx)\n\t"
 259
 260 "       flds   72(%%edx)\n\t"
 261 "       fadds 116(%%edx)\n\t"
 262 "       fstps  72(%%ecx)\n\t"
 263
 264 "       flds   76(%%edx)\n\t"
 265 "       fadds 112(%%edx)\n\t"
 266 "       fstps  76(%%ecx)\n\t"
 267
 268 "       flds  116(%%edx)\n\t"
 269 "       fsubs  72(%%edx)\n\t"
 270 "       fmuls  72(%%ebx)\n\t"
 271 "       fstps 116(%%ecx)\n\t"
 272
 273 "       flds  112(%%edx)\n\t"
 274 "       fsubs  76(%%edx)\n\t"
 275 "       fmuls  76(%%ebx)\n\t"
 276 "       fstps 112(%%ecx)\n\t"
 277
 278 "       flds   80(%%edx)\n\t"
 279 "       fadds 108(%%edx)\n\t"
 280 "       fstps  80(%%ecx)\n\t"
 281
 282 "       flds   84(%%edx)\n\t"
 283 "       fadds 104(%%edx)\n\t"
 284 "       fstps  84(%%ecx)\n\t"
 285
 286 "       flds  108(%%edx)\n\t"
 287 "       fsubs  80(%%edx)\n\t"
 288 "       fmuls  80(%%ebx)\n\t"
 289 "       fstps 108(%%ecx)\n\t"
 290
 291 "       flds  104(%%edx)\n\t"
 292 "       fsubs  84(%%edx)\n\t"
 293 "       fmuls  84(%%ebx)\n\t"
 294 "       fstps 104(%%ecx)\n\t"
 295
 296 "       flds   88(%%edx)\n\t"
 297 "       fadds 100(%%edx)\n\t"
 298 "       fstps  88(%%ecx)\n\t"
 299
 300 "       flds   92(%%edx)\n\t"
 301 "       fadds  96(%%edx)\n\t"
 302 "       fstps  92(%%ecx)\n\t"
 303
 304 "       flds  100(%%edx)\n\t"
 305 "       fsubs  88(%%edx)\n\t"
 306 "       fmuls  88(%%ebx)\n\t"
 307 "       fstps 100(%%ecx)\n\t"
 308
 309 "       flds   96(%%edx)\n\t"
 310 "       fsubs  92(%%edx)\n\t"
 311 "       fmuls  92(%%ebx)\n\t"
 312 "       fstps  96(%%ecx)\n\t"
 313
 314 /* Phase 4*/
 315
 316 "       flds     (%%ecx)\n\t"
 317 "       fadds  28(%%ecx)\n\t"
 318 "       fstps    (%%edx)\n\t"
 319
 320 "       flds     (%%ecx)\n\t"
 321 "       fsubs  28(%%ecx)\n\t"
 322 "       fmuls  96(%%ebx)\n\t"
 323 "       fstps  28(%%edx)\n\t"
 324
 325 "       flds    4(%%ecx)\n\t"
 326 "       fadds  24(%%ecx)\n\t"
 327 "       fstps   4(%%edx)\n\t"
 328
 329 "       flds    4(%%ecx)\n\t"
 330 "       fsubs  24(%%ecx)\n\t"
 331 "       fmuls 100(%%ebx)\n\t"
 332 "       fstps  24(%%edx)\n\t"
 333
 334 "       flds    8(%%ecx)\n\t"
 335 "       fadds  20(%%ecx)\n\t"
 336 "       fstps   8(%%edx)\n\t"
 337
 338 "       flds    8(%%ecx)\n\t"
 339 "       fsubs  20(%%ecx)\n\t"
 340 "       fmuls 104(%%ebx)\n\t"
 341 "       fstps  20(%%edx)\n\t"
 342
 343 "       flds   12(%%ecx)\n\t"
 344 "       fadds  16(%%ecx)\n\t"
 345 "       fstps  12(%%edx)\n\t"
 346
 347 "       flds   12(%%ecx)\n\t"
 348 "       fsubs  16(%%ecx)\n\t"
 349 "       fmuls 108(%%ebx)\n\t"
 350 "       fstps  16(%%edx)\n\t"
 351
 352 "       flds   32(%%ecx)\n\t"
 353 "       fadds  60(%%ecx)\n\t"
 354 "       fstps  32(%%edx)\n\t"
 355
 356 "       flds   60(%%ecx)\n\t"
 357 "       fsubs  32(%%ecx)\n\t"
 358 "       fmuls  96(%%ebx)\n\t"
 359 "       fstps  60(%%edx)\n\t"
 360
 361 "       flds   36(%%ecx)\n\t"
 362 "       fadds  56(%%ecx)\n\t"
 363 "       fstps  36(%%edx)\n\t"
 364
 365 "       flds   56(%%ecx)\n\t"
 366 "       fsubs  36(%%ecx)\n\t"
 367 "       fmuls 100(%%ebx)\n\t"
 368 "       fstps  56(%%edx)\n\t"
 369
 370 "       flds   40(%%ecx)\n\t"
 371 "       fadds  52(%%ecx)\n\t"
 372 "       fstps  40(%%edx)\n\t"
 373
 374 "       flds   52(%%ecx)\n\t"
 375 "       fsubs  40(%%ecx)\n\t"
 376 "       fmuls 104(%%ebx)\n\t"
 377 "       fstps  52(%%edx)\n\t"
 378
 379 "       flds   44(%%ecx)\n\t"
 380 "       fadds  48(%%ecx)\n\t"
 381 "       fstps  44(%%edx)\n\t"
 382
 383 "       flds   48(%%ecx)\n\t"
 384 "       fsubs  44(%%ecx)\n\t"
 385 "       fmuls 108(%%ebx)\n\t"
 386 "       fstps  48(%%edx)\n\t"
 387
 388 "       flds   64(%%ecx)\n\t"
 389 "       fadds  92(%%ecx)\n\t"
 390 "       fstps  64(%%edx)\n\t"
 391
 392 "       flds   64(%%ecx)\n\t"
 393 "       fsubs  92(%%ecx)\n\t"
 394 "       fmuls  96(%%ebx)\n\t"
 395 "       fstps  92(%%edx)\n\t"
 396
 397 "       flds   68(%%ecx)\n\t"
 398 "       fadds  88(%%ecx)\n\t"
 399 "       fstps  68(%%edx)\n\t"
 400
 401 "       flds   68(%%ecx)\n\t"
 402 "       fsubs  88(%%ecx)\n\t"
 403 "       fmuls 100(%%ebx)\n\t"
 404 "       fstps  88(%%edx)\n\t"
 405
 406 "       flds   72(%%ecx)\n\t"
 407 "       fadds  84(%%ecx)\n\t"
 408 "       fstps  72(%%edx)\n\t"
 409
 410 "       flds   72(%%ecx)\n\t"
 411 "       fsubs  84(%%ecx)\n\t"
 412 "       fmuls 104(%%ebx)\n\t"
 413 "       fstps  84(%%edx)\n\t"
 414
 415 "       flds   76(%%ecx)\n\t"
 416 "       fadds  80(%%ecx)\n\t"
 417 "       fstps  76(%%edx)\n\t"
 418
 419 "       flds   76(%%ecx)\n\t"
 420 "       fsubs  80(%%ecx)\n\t"
 421 "       fmuls 108(%%ebx)\n\t"
 422 "       fstps  80(%%edx)\n\t"
 423
 424 "       flds   96(%%ecx)\n\t"
 425 "       fadds 124(%%ecx)\n\t"
 426 "       fstps  96(%%edx)\n\t"
 427
 428 "       flds  124(%%ecx)\n\t"
 429 "       fsubs  96(%%ecx)\n\t"
 430 "       fmuls  96(%%ebx)\n\t"
 431 "       fstps 124(%%edx)\n\t"
 432
 433 "       flds  100(%%ecx)\n\t"
 434 "       fadds 120(%%ecx)\n\t"
 435 "       fstps 100(%%edx)\n\t"
 436
 437 "       flds  120(%%ecx)\n\t"
 438 "       fsubs 100(%%ecx)\n\t"
 439 "       fmuls 100(%%ebx)\n\t"
 440 "       fstps 120(%%edx)\n\t"
 441
 442 "       flds  104(%%ecx)\n\t"
 443 "       fadds 116(%%ecx)\n\t"
 444 "       fstps 104(%%edx)\n\t"
 445
 446 "       flds  116(%%ecx)\n\t"
 447 "       fsubs 104(%%ecx)\n\t"
 448 "       fmuls 104(%%ebx)\n\t"
 449 "       fstps 116(%%edx)\n\t"
 450
 451 "       flds  108(%%ecx)\n\t"
 452 "       fadds 112(%%ecx)\n\t"
 453 "       fstps 108(%%edx)\n\t"
 454
 455 "       flds  112(%%ecx)\n\t"
 456 "       fsubs 108(%%ecx)\n\t"
 457 "       fmuls 108(%%ebx)\n\t"
 458 "       fstps 112(%%edx)\n\t"
 459
 460 "       flds     (%%edx)\n\t"
 461 "       fadds  12(%%edx)\n\t"
 462 "       fstps    (%%ecx)\n\t"
 463
 464 "       flds     (%%edx)\n\t"
 465 "       fsubs  12(%%edx)\n\t"
 466 "       fmuls 112(%%ebx)\n\t"
 467 "       fstps  12(%%ecx)\n\t"
 468
 469 "       flds    4(%%edx)\n\t"
 470 "       fadds   8(%%edx)\n\t"
 471 "       fstps   4(%%ecx)\n\t"
 472
 473 "       flds    4(%%edx)\n\t"
 474 "       fsubs   8(%%edx)\n\t"
 475 "       fmuls 116(%%ebx)\n\t"
 476 "       fstps   8(%%ecx)\n\t"
 477
 478 "       flds   16(%%edx)\n\t"
 479 "       fadds  28(%%edx)\n\t"
 480 "       fstps  16(%%ecx)\n\t"
 481
 482 "       flds   28(%%edx)\n\t"
 483 "       fsubs  16(%%edx)\n\t"
 484 "       fmuls 112(%%ebx)\n\t"
 485 "       fstps  28(%%ecx)\n\t"
 486
 487 "       flds   20(%%edx)\n\t"
 488 "       fadds  24(%%edx)\n\t"
 489 "       fstps  20(%%ecx)\n\t"
 490
 491 "       flds   24(%%edx)\n\t"
 492 "       fsubs  20(%%edx)\n\t"
 493 "       fmuls 116(%%ebx)\n\t"
 494 "       fstps  24(%%ecx)\n\t"
 495
 496 "       flds   32(%%edx)\n\t"
 497 "       fadds  44(%%edx)\n\t"
 498 "       fstps  32(%%ecx)\n\t"
 499
 500 "       flds   32(%%edx)\n\t"
 501 "       fsubs  44(%%edx)\n\t"
 502 "       fmuls 112(%%ebx)\n\t"
 503 "       fstps  44(%%ecx)\n\t"
 504
 505 "       flds   36(%%edx)\n\t"
 506 "       fadds  40(%%edx)\n\t"
 507 "       fstps  36(%%ecx)\n\t"
 508
 509 "       flds   36(%%edx)\n\t"
 510 "       fsubs  40(%%edx)\n\t"
 511 "       fmuls 116(%%ebx)\n\t"
 512 "       fstps  40(%%ecx)\n\t"
 513
 514 "       flds   48(%%edx)\n\t"
 515 "       fadds  60(%%edx)\n\t"
 516 "       fstps  48(%%ecx)\n\t"
 517
 518 "       flds   60(%%edx)\n\t"
 519 "       fsubs  48(%%edx)\n\t"
 520 "       fmuls 112(%%ebx)\n\t"
 521 "       fstps  60(%%ecx)\n\t"
 522
 523 "       flds   52(%%edx)\n\t"
 524 "       fadds  56(%%edx)\n\t"
 525 "       fstps  52(%%ecx)\n\t"
 526
 527 "       flds   56(%%edx)\n\t"
 528 "       fsubs  52(%%edx)\n\t"
 529 "       fmuls 116(%%ebx)\n\t"
 530 "       fstps  56(%%ecx)\n\t"
 531
 532 "       flds   64(%%edx)\n\t"
 533 "       fadds  76(%%edx)\n\t"
 534 "       fstps  64(%%ecx)\n\t"
 535
 536 "       flds   64(%%edx)\n\t"
 537 "       fsubs  76(%%edx)\n\t"
 538 "       fmuls 112(%%ebx)\n\t"
 539 "       fstps  76(%%ecx)\n\t"
 540
 541 "       flds   68(%%edx)\n\t"
 542 "       fadds  72(%%edx)\n\t"
 543 "       fstps  68(%%ecx)\n\t"
 544
 545 "       flds   68(%%edx)\n\t"
 546 "       fsubs  72(%%edx)\n\t"
 547 "       fmuls 116(%%ebx)\n\t"
 548 "       fstps  72(%%ecx)\n\t"
 549
 550 "       flds   80(%%edx)\n\t"
 551 "       fadds  92(%%edx)\n\t"
 552 "       fstps  80(%%ecx)\n\t"
 553
 554 "       flds   92(%%edx)\n\t"
 555 "       fsubs  80(%%edx)\n\t"
 556 "       fmuls 112(%%ebx)\n\t"
 557 "       fstps  92(%%ecx)\n\t"
 558
 559 "       flds   84(%%edx)\n\t"
 560 "       fadds  88(%%edx)\n\t"
 561 "       fstps  84(%%ecx)\n\t"
 562
 563 "       flds   88(%%edx)\n\t"
 564 "       fsubs  84(%%edx)\n\t"
 565 "       fmuls 116(%%ebx)\n\t"
 566 "       fstps  88(%%ecx)\n\t"
 567
 568 "       flds   96(%%edx)\n\t"
 569 "       fadds 108(%%edx)\n\t"
 570 "       fstps  96(%%ecx)\n\t"
 571
 572 "       flds   96(%%edx)\n\t"
 573 "       fsubs 108(%%edx)\n\t"
 574 "       fmuls 112(%%ebx)\n\t"
 575 "       fstps 108(%%ecx)\n\t"
 576
 577 "       flds  100(%%edx)\n\t"
 578 "       fadds 104(%%edx)\n\t"
 579 "       fstps 100(%%ecx)\n\t"
 580
 581 "       flds  100(%%edx)\n\t"
 582 "       fsubs 104(%%edx)\n\t"
 583 "       fmuls 116(%%ebx)\n\t"
 584 "       fstps 104(%%ecx)\n\t"
 585
 586 "       flds  112(%%edx)\n\t"
 587 "       fadds 124(%%edx)\n\t"
 588 "       fstps 112(%%ecx)\n\t"
 589
 590 "       flds  124(%%edx)\n\t"
 591 "       fsubs 112(%%edx)\n\t"
 592 "       fmuls 112(%%ebx)\n\t"
 593 "       fstps 124(%%ecx)\n\t"
 594
 595 "       flds  116(%%edx)\n\t"
 596 "       fadds 120(%%edx)\n\t"
 597 "       fstps 116(%%ecx)\n\t"
 598
 599 "       flds  120(%%edx)\n\t"
 600 "       fsubs 116(%%edx)\n\t"
 601 "       fmuls 116(%%ebx)\n\t"
 602 "       fstps 120(%%ecx)\n\t"
 603
 604 /* Phase 5*/
 605
 606 "       flds   32(%%ecx)\n\t"
 607 "       fadds  36(%%ecx)\n\t"
 608 "       fstps  32(%%edx)\n\t"
 609
 610 "       flds   32(%%ecx)\n\t"
 611 "       fsubs  36(%%ecx)\n\t"
 612 "       fmuls 120(%%ebx)\n\t"
 613 "       fstps  36(%%edx)\n\t"
 614
 615 "       flds   44(%%ecx)\n\t"
 616 "       fsubs  40(%%ecx)\n\t"
 617 "       fmuls 120(%%ebx)\n\t"
 618 "       fsts   44(%%edx)\n\t"
 619 "       fadds  40(%%ecx)\n\t"
 620 "       fadds  44(%%ecx)\n\t"
 621 "       fstps  40(%%edx)\n\t"
 622
 623 "       flds   48(%%ecx)\n\t"
 624 "       fsubs  52(%%ecx)\n\t"
 625 "       fmuls 120(%%ebx)\n\t"
 626
 627 "       flds   60(%%ecx)\n\t"
 628 "       fsubs  56(%%ecx)\n\t"
 629 "       fmuls 120(%%ebx)\n\t"
 630 "       fld      %%st(0)\n\t"
 631 "       fadds  56(%%ecx)\n\t"
 632 "       fadds  60(%%ecx)\n\t"
 633 "       fld      %%st(0)\n\t"
 634 "       fadds  48(%%ecx)\n\t"
 635 "       fadds  52(%%ecx)\n\t"
 636 "       fstps  48(%%edx)\n\t"
 637 "       fadd     %%st(2)\n\t"
 638 "       fstps  56(%%edx)\n\t"
 639 "       fsts   60(%%edx)\n\t"
 640 "       faddp    %%st(1)\n\t"
 641 "       fstps  52(%%edx)\n\t"
 642
 643 "       flds   64(%%ecx)\n\t"
 644 "       fadds  68(%%ecx)\n\t"
 645 "       fstps  64(%%edx)\n\t"
 646
 647 "       flds   64(%%ecx)\n\t"
 648 "       fsubs  68(%%ecx)\n\t"
 649 "       fmuls 120(%%ebx)\n\t"
 650 "       fstps  68(%%edx)\n\t"
 651
 652 "       flds   76(%%ecx)\n\t"
 653 "       fsubs  72(%%ecx)\n\t"
 654 "       fmuls 120(%%ebx)\n\t"
 655 "       fsts   76(%%edx)\n\t"
 656 "       fadds  72(%%ecx)\n\t"
 657 "       fadds  76(%%ecx)\n\t"
 658 "       fstps  72(%%edx)\n\t"
 659
 660 "       flds   92(%%ecx)\n\t"
 661 "       fsubs  88(%%ecx)\n\t"
 662 "       fmuls 120(%%ebx)\n\t"
 663 "       fsts   92(%%edx)\n\t"
 664 "       fadds  92(%%ecx)\n\t"
 665 "       fadds  88(%%ecx)\n\t"
 666 "       fld      %%st(0)\n\t"
 667 "       fadds  80(%%ecx)\n\t"
 668 "       fadds  84(%%ecx)\n\t"
 669 "       fstps  80(%%edx)\n\t"
 670
 671 "       flds   80(%%ecx)\n\t"
 672 "       fsubs  84(%%ecx)\n\t"
 673 "       fmuls 120(%%ebx)\n\t"
 674 "       fadd  %%st(0), %%st(1)\n\t"
 675 "       fadds 92(%%edx)\n\t"
 676 "       fstps 84(%%edx)\n\t"
 677 "       fstps 88(%%edx)\n\t"
 678
 679 "       flds   96(%%ecx)\n\t"
 680 "       fadds 100(%%ecx)\n\t"
 681 "       fstps  96(%%edx)\n\t"
 682
 683 "       flds   96(%%ecx)\n\t"
 684 "       fsubs 100(%%ecx)\n\t"
 685 "       fmuls 120(%%ebx)\n\t"
 686 "       fstps 100(%%edx)\n\t"
 687
 688 "       flds  108(%%ecx)\n\t"
 689 "       fsubs 104(%%ecx)\n\t"
 690 "       fmuls 120(%%ebx)\n\t"
 691 "       fsts  108(%%edx)\n\t"
 692 "       fadds 104(%%ecx)\n\t"
 693 "       fadds 108(%%ecx)\n\t"
 694 "       fstps 104(%%edx)\n\t"
 695
 696 "       flds  124(%%ecx)\n\t"
 697 "       fsubs 120(%%ecx)\n\t"
 698 "       fmuls 120(%%ebx)\n\t"
 699 "       fsts  124(%%edx)\n\t"
 700 "       fadds 120(%%ecx)\n\t"
 701 "       fadds 124(%%ecx)\n\t"
 702 "       fld      %%st(0)\n\t"
 703 "       fadds 112(%%ecx)\n\t"
 704 "       fadds 116(%%ecx)\n\t"
 705 "       fstps 112(%%edx)\n\t"
 706
 707 "       flds  112(%%ecx)\n\t"
 708 "       fsubs 116(%%ecx)\n\t"
 709 "       fmuls 120(%%ebx)\n\t"
 710 "       fadd  %%st(0),%%st(1)\n\t"
 711 "       fadds 124(%%edx)\n\t"
 712 "       fstps 116(%%edx)\n\t"
 713 "       fstps 120(%%edx)\n\t"
 714 "       jnz .L01\n\t"
 715
 716 /* Phase 6*/
 717
 718 "       flds      (%%ecx)\n\t"
 719 "       fadds    4(%%ecx)\n\t"
 720 "       fstps 1024(%%esi)\n\t"
 721
 722 "       flds      (%%ecx)\n\t"
 723 "       fsubs    4(%%ecx)\n\t"
 724 "       fmuls  120(%%ebx)\n\t"
 725 "       fsts      (%%esi)\n\t"
 726 "       fstps     (%%edi)\n\t"
 727
 728 "       flds   12(%%ecx)\n\t"
 729 "       fsubs   8(%%ecx)\n\t"
 730 "       fmuls 120(%%ebx)\n\t"
 731 "       fsts  512(%%edi)\n\t"
 732 "       fadds  12(%%ecx)\n\t"
 733 "       fadds   8(%%ecx)\n\t"
 734 "       fstps 512(%%esi)\n\t"
 735
 736 "       flds   16(%%ecx)\n\t"
 737 "       fsubs  20(%%ecx)\n\t"
 738 "       fmuls 120(%%ebx)\n\t"
 739
 740 "       flds   28(%%ecx)\n\t"
 741 "       fsubs  24(%%ecx)\n\t"
 742 "       fmuls 120(%%ebx)\n\t"
 743 "       fsts  768(%%edi)\n\t"
 744 "       fld      %%st(0)\n\t"
 745 "       fadds  24(%%ecx)\n\t"
 746 "       fadds  28(%%ecx)\n\t"
 747 "       fld      %%st(0)\n\t"
 748 "       fadds  16(%%ecx)\n\t"
 749 "       fadds  20(%%ecx)\n\t"
 750 "       fstps 768(%%esi)\n\t"
 751 "       fadd     %%st(2)\n\t"
 752 "       fstps 256(%%esi)\n\t"
 753 "       faddp    %%st(1)\n\t"
 754 "       fstps 256(%%edi)\n\t"
 755
 756 /* Phase 7*/
 757
 758 "       flds   32(%%edx)\n\t"
 759 "       fadds  48(%%edx)\n\t"
 760 "       fstps 896(%%esi)\n\t"
 761
 762 "       flds   48(%%edx)\n\t"
 763 "       fadds  40(%%edx)\n\t"
 764 "       fstps 640(%%esi)\n\t"
 765
 766 "       flds   40(%%edx)\n\t"
 767 "       fadds  56(%%edx)\n\t"
 768 "       fstps 384(%%esi)\n\t"
 769
 770 "       flds   56(%%edx)\n\t"
 771 "       fadds  36(%%edx)\n\t"
 772 "       fstps 128(%%esi)\n\t"
 773
 774 "       flds   36(%%edx)\n\t"
 775 "       fadds  52(%%edx)\n\t"
 776 "       fstps 128(%%edi)\n\t"
 777
 778 "       flds   52(%%edx)\n\t"
 779 "       fadds  44(%%edx)\n\t"
 780 "       fstps 384(%%edi)\n\t"
 781
 782 "       flds   60(%%edx)\n\t"
 783 "       fsts  896(%%edi)\n\t"
 784 "       fadds  44(%%edx)\n\t"
 785 "       fstps 640(%%edi)\n\t"
 786
 787 "       flds   96(%%edx)\n\t"
 788 "       fadds 112(%%edx)\n\t"
 789 "       fld      %%st(0)\n\t"
 790 "       fadds  64(%%edx)\n\t"
 791 "       fstps 960(%%esi)\n\t"
 792 "       fadds  80(%%edx)\n\t"
 793 "       fstps 832(%%esi)\n\t"
 794
 795 "       flds  112(%%edx)\n\t"
 796 "       fadds 104(%%edx)\n\t"
 797 "       fld      %%st(0)\n\t"
 798 "       fadds  80(%%edx)\n\t"
 799 "       fstps 704(%%esi)\n\t"
 800 "       fadds  72(%%edx)\n\t"
 801 "       fstps 576(%%esi)\n\t"
 802
 803 "       flds  104(%%edx)\n\t"
 804 "       fadds 120(%%edx)\n\t"
 805 "       fld      %%st(0)\n\t"
 806 "       fadds  72(%%edx)\n\t"
 807 "       fstps 448(%%esi)\n\t"
 808 "       fadds  88(%%edx)\n\t"
 809 "       fstps 320(%%esi)\n\t"
 810
 811 "       flds  120(%%edx)\n\t"
 812 "       fadds 100(%%edx)\n\t"
 813 "       fld      %%st(0)\n\t"
 814 "       fadds  88(%%edx)\n\t"
 815 "       fstps 192(%%esi)\n\t"
 816 "       fadds  68(%%edx)\n\t"
 817 "       fstps  64(%%esi)\n\t"
 818
 819 "       flds  100(%%edx)\n\t"
 820 "       fadds 116(%%edx)\n\t"
 821 "       fld      %%st(0)\n\t"
 822 "       fadds  68(%%edx)\n\t"
 823 "       fstps  64(%%edi)\n\t"
 824 "       fadds  84(%%edx)\n\t"
 825 "       fstps 192(%%edi)\n\t"
 826
 827 "       flds  116(%%edx)\n\t"
 828 "       fadds 108(%%edx)\n\t"
 829 "       fld      %%st(0)\n\t"
 830 "       fadds  84(%%edx)\n\t"
 831 "       fstps 320(%%edi)\n\t"
 832 "       fadds  76(%%edx)\n\t"
 833 "       fstps 448(%%edi)\n\t"
 834
 835 "       flds  108(%%edx)\n\t"
 836 "       fadds 124(%%edx)\n\t"
 837 "       fld      %%st(0)\n\t"
 838 "       fadds  76(%%edx)\n\t"
 839 "       fstps 576(%%edi)\n\t"
 840 "       fadds  92(%%edx)\n\t"
 841 "       fstps 704(%%edi)\n\t"
 842
 843 "       flds  124(%%edx)\n\t"
 844 "       fsts  960(%%edi)\n\t"
 845 "       fadds  92(%%edx)\n\t"
 846 "       fstps 832(%%edi)\n\t"
 847 "       jmp     .L_bye\n\t"
 848 ".L01:\n\t"
 849 /* Phase 8*/
 850
 851 "       flds      (%%ecx)\n\t"
 852 "       fadds    4(%%ecx)\n\t"
 853 "       fistp  512(%%esi)\n\t"
 854
 855 "       flds      (%%ecx)\n\t"
 856 "       fsubs    4(%%ecx)\n\t"
 857 "       fmuls  120(%%ebx)\n\t"
 858
 859 "       fistp     (%%esi)\n\t"
 860
 861
 862 "       flds    12(%%ecx)\n\t"
 863 "       fsubs    8(%%ecx)\n\t"
 864 "       fmuls  120(%%ebx)\n\t"
 865 "       fist   256(%%edi)\n\t"
 866 "       fadds   12(%%ecx)\n\t"
 867 "       fadds    8(%%ecx)\n\t"
 868 "       fistp  256(%%esi)\n\t"
 869
 870 "       flds   16(%%ecx)\n\t"
 871 "       fsubs  20(%%ecx)\n\t"
 872 "       fmuls 120(%%ebx)\n\t"
 873
 874 "       flds   28(%%ecx)\n\t"
 875 "       fsubs  24(%%ecx)\n\t"
 876 "       fmuls 120(%%ebx)\n\t"
 877 "       fist  384(%%edi)\n\t"
 878 "       fld      %%st(0)\n\t"
 879 "       fadds  24(%%ecx)\n\t"
 880 "       fadds  28(%%ecx)\n\t"
 881 "       fld      %%st(0)\n\t"
 882 "       fadds  16(%%ecx)\n\t"
 883 "       fadds  20(%%ecx)\n\t"
 884 "       fistp  384(%%esi)\n\t"
 885 "       fadd     %%st(2)\n\t"
 886 "       fistp  128(%%esi)\n\t"
 887 "       faddp    %%st(1)\n\t"
 888 "       fistp  128(%%edi)\n\t"
 889
 890 /* Phase 9*/
 891
 892 "       flds    32(%%edx)\n\t"
 893 "       fadds   48(%%edx)\n\t"
 894 "       fistp  448(%%esi)\n\t"
 895
 896 "       flds   48(%%edx)\n\t"
 897 "       fadds  40(%%edx)\n\t"
 898 "       fistp 320(%%esi)\n\t"
 899
 900 "       flds   40(%%edx)\n\t"
 901 "       fadds  56(%%edx)\n\t"
 902 "       fistp 192(%%esi)\n\t"
 903
 904 "       flds   56(%%edx)\n\t"
 905 "       fadds  36(%%edx)\n\t"
 906 "       fistp  64(%%esi)\n\t"
 907
 908 "       flds   36(%%edx)\n\t"
 909 "       fadds  52(%%edx)\n\t"
 910 "       fistp  64(%%edi)\n\t"
 911
 912 "       flds   52(%%edx)\n\t"
 913 "       fadds  44(%%edx)\n\t"
 914 "       fistp 192(%%edi)\n\t"
 915
 916 "       flds   60(%%edx)\n\t"
 917 "       fist   448(%%edi)\n\t"
 918 "       fadds  44(%%edx)\n\t"
 919 "       fistp 320(%%edi)\n\t"
 920
 921 "       flds   96(%%edx)\n\t"
 922 "       fadds 112(%%edx)\n\t"
 923 "       fld      %%st(0)\n\t"
 924 "       fadds  64(%%edx)\n\t"
 925 "       fistp 480(%%esi)\n\t"
 926 "       fadds  80(%%edx)\n\t"
 927 "       fistp 416(%%esi)\n\t"
 928
 929 "       flds  112(%%edx)\n\t"
 930 "       fadds 104(%%edx)\n\t"
 931 "       fld      %%st(0)\n\t"
 932 "       fadds  80(%%edx)\n\t"
 933 "       fistp 352(%%esi)\n\t"
 934 "       fadds  72(%%edx)\n\t"
 935 "       fistp 288(%%esi)\n\t"
 936
 937 "       flds  104(%%edx)\n\t"
 938 "       fadds 120(%%edx)\n\t"
 939 "       fld      %%st(0)\n\t"
 940 "       fadds  72(%%edx)\n\t"
 941 "       fistp 224(%%esi)\n\t"
 942 "       fadds  88(%%edx)\n\t"
 943 "       fistp 160(%%esi)\n\t"
 944
 945 "       flds  120(%%edx)\n\t"
 946 "       fadds 100(%%edx)\n\t"
 947 "       fld      %%st(0)\n\t"
 948 "       fadds  88(%%edx)\n\t"
 949 "       fistp  96(%%esi)\n\t"
 950 "       fadds  68(%%edx)\n\t"
 951 "       fistp  32(%%esi)\n\t"
 952
 953 "       flds  100(%%edx)\n\t"
 954 "       fadds 116(%%edx)\n\t"
 955 "       fld      %%st(0)\n\t"
 956 "       fadds  68(%%edx)\n\t"
 957 "       fistp  32(%%edi)\n\t"
 958 "       fadds  84(%%edx)\n\t"
 959 "       fistp  96(%%edi)\n\t"
 960
 961 "       flds  116(%%edx)\n\t"
 962 "       fadds 108(%%edx)\n\t"
 963 "       fld      %%st(0)\n\t"
 964 "       fadds  84(%%edx)\n\t"
 965 "       fistp 160(%%edi)\n\t"
 966 "       fadds  76(%%edx)\n\t"
 967 "       fistp 224(%%edi)\n\t"
 968
 969 "       flds  108(%%edx)\n\t"
 970 "       fadds 124(%%edx)\n\t"
 971 "       fld      %%st(0)\n\t"
 972 "       fadds  76(%%edx)\n\t"
 973 "       fistp 288(%%edi)\n\t"
 974 "       fadds  92(%%edx)\n\t"
 975 "       fistp 352(%%edi)\n\t"
 976
 977 "       flds  124(%%edx)\n\t"
 978 "       fist  480(%%edi)\n\t"
 979 "       fadds  92(%%edx)\n\t"
 980 "       fistp 416(%%edi)\n\t"
 981 "       movsw\n\t"
 982 ".L_bye:"
 983         :
 984         :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
 985         :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi");
 986 }