mp3lib/dct64_3dnow.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 * Partial 3dnow! 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
  12 #include "../config.h"
  13 #include "../mangle.h"
  14
  15 static unsigned long long int attribute_used __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
  16 static float attribute_used plus_1f = 1.0;
  17
  18 void dct64_MMX_3dnow(real *a,real *b,real *c)
  19 {
  20   char tmp[256];
  21     __asm __volatile(
  22 "       movl %2,%%eax\n\t"
  23
  24 "       leal 128+%3,%%edx\n\t"
  25 "       movl %0,%%esi\n\t"
  26 "       movl %1,%%edi\n\t"
  27 "       movl $"MANGLE(costab_mmx)",%%ebx\n\t"
  28 "       leal %3,%%ecx\n\t"
  29
  30 /* Phase 1*/
  31 "       movq    (%%eax), %%mm0\n\t"
  32 "       movq    8(%%eax), %%mm4\n\t"
  33 "       movq    %%mm0, %%mm3\n\t"
  34 "       movq    %%mm4, %%mm7\n\t"
  35 "       movq    120(%%eax), %%mm1\n\t"
  36 "       movq    112(%%eax), %%mm5\n\t"
  37 /* n.b.: pswapd*/
  38 "       movq    %%mm1, %%mm2\n\t"
  39 "       movq    %%mm5, %%mm6\n\t"
  40 "       psrlq   $32, %%mm1\n\t"
  41 "       psrlq   $32, %%mm5\n\t"
  42 "       punpckldq %%mm2, %%mm1\n\t"
  43 "       punpckldq %%mm6, %%mm5\n\t"
  44 /**/
  45 "       pfadd   %%mm1, %%mm0\n\t"
  46 "       pfadd   %%mm5, %%mm4\n\t"
  47 "       movq    %%mm0, (%%edx)\n\t"
  48 "       movq    %%mm4, 8(%%edx)\n\t"
  49 "       pfsub   %%mm1, %%mm3\n\t"
  50 "       pfsub   %%mm5, %%mm7\n\t"
  51 "       pfmul   (%%ebx), %%mm3\n\t"
  52 "       pfmul   8(%%ebx), %%mm7\n\t"
  53 "       movd    %%mm3, 124(%%edx)\n\t"
  54 "       movd    %%mm7, 116(%%edx)\n\t"
  55 "       psrlq   $32, %%mm3\n\t"
  56 "       psrlq   $32, %%mm7\n\t"
  57 "       movd    %%mm3, 120(%%edx)\n\t"
  58 "       movd    %%mm7, 112(%%edx)\n\t"
  59
  60 "       movq    16(%%eax), %%mm0\n\t"
  61 "       movq    24(%%eax), %%mm4\n\t"
  62 "       movq    %%mm0, %%mm3\n\t"
  63 "       movq    %%mm4, %%mm7\n\t"
  64 "       movq    104(%%eax), %%mm1\n\t"
  65 "       movq    96(%%eax), %%mm5\n\t"
  66 /* n.b.: pswapd*/
  67 "       movq    %%mm1, %%mm2\n\t"
  68 "       movq    %%mm5, %%mm6\n\t"
  69 "       psrlq   $32, %%mm1\n\t"
  70 "       psrlq   $32, %%mm5\n\t"
  71 "       punpckldq %%mm2, %%mm1\n\t"
  72 "       punpckldq %%mm6, %%mm5\n\t"
  73 /**/
  74 "       pfadd   %%mm1, %%mm0\n\t"
  75 "       pfadd   %%mm5, %%mm4\n\t"
  76 "       movq    %%mm0, 16(%%edx)\n\t"
  77 "       movq    %%mm4, 24(%%edx)\n\t"
  78 "       pfsub   %%mm1, %%mm3\n\t"
  79 "       pfsub   %%mm5, %%mm7\n\t"
  80 "       pfmul   16(%%ebx), %%mm3\n\t"
  81 "       pfmul   24(%%ebx), %%mm7\n\t"
  82 "       movd    %%mm3, 108(%%edx)\n\t"
  83 "       movd    %%mm7, 100(%%edx)\n\t"
  84 "       psrlq   $32, %%mm3\n\t"
  85 "       psrlq   $32, %%mm7\n\t"
  86 "       movd    %%mm3, 104(%%edx)\n\t"
  87 "       movd    %%mm7, 96(%%edx)\n\t"
  88
  89 "       movq    32(%%eax), %%mm0\n\t"
  90 "       movq    40(%%eax), %%mm4\n\t"
  91 "       movq    %%mm0, %%mm3\n\t"
  92 "       movq    %%mm4, %%mm7\n\t"
  93 "       movq    88(%%eax), %%mm1\n\t"
  94 "       movq    80(%%eax), %%mm5\n\t"
  95 /* n.b.: pswapd*/
  96 "       movq    %%mm1, %%mm2\n\t"
  97 "       movq    %%mm5, %%mm6\n\t"
  98 "       psrlq   $32, %%mm1\n\t"
  99 "       psrlq   $32, %%mm5\n\t"
 100 "       punpckldq %%mm2, %%mm1\n\t"
 101 "       punpckldq %%mm6, %%mm5\n\t"
 102 /**/
 103 "       pfadd   %%mm1, %%mm0\n\t"
 104 "       pfadd   %%mm5, %%mm4\n\t"
 105 "       movq    %%mm0, 32(%%edx)\n\t"
 106 "       movq    %%mm4, 40(%%edx)\n\t"
 107 "       pfsub   %%mm1, %%mm3\n\t"
 108 "       pfsub   %%mm5, %%mm7\n\t"
 109 "       pfmul   32(%%ebx), %%mm3\n\t"
 110 "       pfmul   40(%%ebx), %%mm7\n\t"
 111 "       movd    %%mm3, 92(%%edx)\n\t"
 112 "       movd    %%mm7, 84(%%edx)\n\t"
 113 "       psrlq   $32, %%mm3\n\t"
 114 "       psrlq   $32, %%mm7\n\t"
 115 "       movd    %%mm3, 88(%%edx)\n\t"
 116 "       movd    %%mm7, 80(%%edx)\n\t"
 117
 118 "       movq    48(%%eax), %%mm0\n\t"
 119 "       movq    56(%%eax), %%mm4\n\t"
 120 "       movq    %%mm0, %%mm3\n\t"
 121 "       movq    %%mm4, %%mm7\n\t"
 122 "       movq    72(%%eax), %%mm1\n\t"
 123 "       movq    64(%%eax), %%mm5\n\t"
 124 /* n.b.: pswapd*/
 125 "       movq    %%mm1, %%mm2\n\t"
 126 "       movq    %%mm5, %%mm6\n\t"
 127 "       psrlq   $32, %%mm1\n\t"
 128 "       psrlq   $32, %%mm5\n\t"
 129 "       punpckldq %%mm2, %%mm1\n\t"
 130 "       punpckldq %%mm6, %%mm5\n\t"
 131 /**/
 132 "       pfadd   %%mm1, %%mm0\n\t"
 133 "       pfadd   %%mm5, %%mm4\n\t"
 134 "       movq    %%mm0, 48(%%edx)\n\t"
 135 "       movq    %%mm4, 56(%%edx)\n\t"
 136 "       pfsub   %%mm1, %%mm3\n\t"
 137 "       pfsub   %%mm5, %%mm7\n\t"
 138 "       pfmul   48(%%ebx), %%mm3\n\t"
 139 "       pfmul   56(%%ebx), %%mm7\n\t"
 140 "       movd    %%mm3, 76(%%edx)\n\t"
 141 "       movd    %%mm7, 68(%%edx)\n\t"
 142 "       psrlq   $32, %%mm3\n\t"
 143 "       psrlq   $32, %%mm7\n\t"
 144 "       movd    %%mm3, 72(%%edx)\n\t"
 145 "       movd    %%mm7, 64(%%edx)\n\t"
 146
 147 /* Phase 2*/
 148
 149 "       movq    (%%edx), %%mm0\n\t"
 150 "       movq    8(%%edx), %%mm4\n\t"
 151 "       movq    %%mm0, %%mm3\n\t"
 152 "       movq    %%mm4, %%mm7\n\t"
 153 "       movq    56(%%edx), %%mm1\n\t"
 154 "       movq    48(%%edx), %%mm5\n\t"
 155 /* n.b.: pswapd*/
 156 "       movq    %%mm1, %%mm2\n\t"
 157 "       movq    %%mm5, %%mm6\n\t"
 158 "       psrlq   $32, %%mm1\n\t"
 159 "       psrlq   $32, %%mm5\n\t"
 160 "       punpckldq %%mm2, %%mm1\n\t"
 161 "       punpckldq %%mm6, %%mm5\n\t"
 162 /**/
 163 "       pfadd   %%mm1, %%mm0\n\t"
 164 "       pfadd   %%mm5, %%mm4\n\t"
 165 "       movq    %%mm0, (%%ecx)\n\t"
 166 "       movq    %%mm4, 8(%%ecx)\n\t"
 167 "       pfsub   %%mm1, %%mm3\n\t"
 168 "       pfsub   %%mm5, %%mm7\n\t"
 169 "       pfmul   64(%%ebx), %%mm3\n\t"
 170 "       pfmul   72(%%ebx), %%mm7\n\t"
 171 "       movd    %%mm3, 60(%%ecx)\n\t"
 172 "       movd    %%mm7, 52(%%ecx)\n\t"
 173 "       psrlq   $32, %%mm3\n\t"
 174 "       psrlq   $32, %%mm7\n\t"
 175 "       movd    %%mm3, 56(%%ecx)\n\t"
 176 "       movd    %%mm7, 48(%%ecx)\n\t"
 177
 178 "       movq    16(%%edx), %%mm0\n\t"
 179 "       movq    24(%%edx), %%mm4\n\t"
 180 "       movq    %%mm0, %%mm3\n\t"
 181 "       movq    %%mm4, %%mm7\n\t"
 182 "       movq    40(%%edx), %%mm1\n\t"
 183 "       movq    32(%%edx), %%mm5\n\t"
 184 /* n.b.: pswapd*/
 185 "       movq    %%mm1, %%mm2\n\t"
 186 "       movq    %%mm5, %%mm6\n\t"
 187 "       psrlq   $32, %%mm1\n\t"
 188 "       psrlq   $32, %%mm5\n\t"
 189 "       punpckldq %%mm2, %%mm1\n\t"
 190 "       punpckldq %%mm6, %%mm5\n\t"
 191 /**/
 192 "       pfadd   %%mm1, %%mm0\n\t"
 193 "       pfadd   %%mm5, %%mm4\n\t"
 194 "       movq    %%mm0, 16(%%ecx)\n\t"
 195 "       movq    %%mm4, 24(%%ecx)\n\t"
 196 "       pfsub   %%mm1, %%mm3\n\t"
 197 "       pfsub   %%mm5, %%mm7\n\t"
 198 "       pfmul   80(%%ebx), %%mm3\n\t"
 199 "       pfmul   88(%%ebx), %%mm7\n\t"
 200 "       movd    %%mm3, 44(%%ecx)\n\t"
 201 "       movd    %%mm7, 36(%%ecx)\n\t"
 202 "       psrlq   $32, %%mm3\n\t"
 203 "       psrlq   $32, %%mm7\n\t"
 204 "       movd    %%mm3, 40(%%ecx)\n\t"
 205 "       movd    %%mm7, 32(%%ecx)\n\t"
 206
 207 /* Phase 3*/
 208
 209 "       movq    64(%%edx), %%mm0\n\t"
 210 "       movq    72(%%edx), %%mm4\n\t"
 211 "       movq    %%mm0, %%mm3\n\t"
 212 "       movq    %%mm4, %%mm7\n\t"
 213 "       movq    120(%%edx), %%mm1\n\t"
 214 "       movq    112(%%edx), %%mm5\n\t"
 215 /* n.b.: pswapd*/
 216 "       movq    %%mm1, %%mm2\n\t"
 217 "       movq    %%mm5, %%mm6\n\t"
 218 "       psrlq   $32, %%mm1\n\t"
 219 "       psrlq   $32, %%mm5\n\t"
 220 "       punpckldq %%mm2, %%mm1\n\t"
 221 "       punpckldq %%mm6, %%mm5\n\t"
 222 /**/
 223 "       pfadd   %%mm1, %%mm0\n\t"
 224 "       pfadd   %%mm5, %%mm4\n\t"
 225 "       movq    %%mm0, 64(%%ecx)\n\t"
 226 "       movq    %%mm4, 72(%%ecx)\n\t"
 227 "       pfsubr  %%mm1, %%mm3\n\t"
 228 "       pfsubr  %%mm5, %%mm7\n\t"
 229 "       pfmul   64(%%ebx), %%mm3\n\t"
 230 "       pfmul   72(%%ebx), %%mm7\n\t"
 231 "       movd    %%mm3, 124(%%ecx)\n\t"
 232 "       movd    %%mm7, 116(%%ecx)\n\t"
 233 "       psrlq   $32, %%mm3\n\t"
 234 "       psrlq   $32, %%mm7\n\t"
 235 "       movd    %%mm3, 120(%%ecx)\n\t"
 236 "       movd    %%mm7, 112(%%ecx)\n\t"
 237
 238 "       movq    80(%%edx), %%mm0\n\t"
 239 "       movq    88(%%edx), %%mm4\n\t"
 240 "       movq    %%mm0, %%mm3\n\t"
 241 "       movq    %%mm4, %%mm7\n\t"
 242 "       movq    104(%%edx), %%mm1\n\t"
 243 "       movq    96(%%edx), %%mm5\n\t"
 244 /* n.b.: pswapd*/
 245 "       movq    %%mm1, %%mm2\n\t"
 246 "       movq    %%mm5, %%mm6\n\t"
 247 "       psrlq   $32, %%mm1\n\t"
 248 "       psrlq   $32, %%mm5\n\t"
 249 "       punpckldq %%mm2, %%mm1\n\t"
 250 "       punpckldq %%mm6, %%mm5\n\t"
 251 /**/
 252 "       pfadd   %%mm1, %%mm0\n\t"
 253 "       pfadd   %%mm5, %%mm4\n\t"
 254 "       movq    %%mm0, 80(%%ecx)\n\t"
 255 "       movq    %%mm4, 88(%%ecx)\n\t"
 256 "       pfsubr  %%mm1, %%mm3\n\t"
 257 "       pfsubr  %%mm5, %%mm7\n\t"
 258 "       pfmul   80(%%ebx), %%mm3\n\t"
 259 "       pfmul   88(%%ebx), %%mm7\n\t"
 260 "       movd    %%mm3, 108(%%ecx)\n\t"
 261 "       movd    %%mm7, 100(%%ecx)\n\t"
 262 "       psrlq   $32, %%mm3\n\t"
 263 "       psrlq   $32, %%mm7\n\t"
 264 "       movd    %%mm3, 104(%%ecx)\n\t"
 265 "       movd    %%mm7, 96(%%ecx)\n\t"
 266
 267 /* Phase 4*/
 268
 269 "       movq    (%%ecx), %%mm0\n\t"
 270 "       movq    8(%%ecx), %%mm4\n\t"
 271 "       movq    %%mm0, %%mm3\n\t"
 272 "       movq    %%mm4, %%mm7\n\t"
 273 "       movq    24(%%ecx), %%mm1\n\t"
 274 "       movq    16(%%ecx), %%mm5\n\t"
 275 /* n.b.: pswapd*/
 276 "       movq    %%mm1, %%mm2\n\t"
 277 "       movq    %%mm5, %%mm6\n\t"
 278 "       psrlq   $32, %%mm1\n\t"
 279 "       psrlq   $32, %%mm5\n\t"
 280 "       punpckldq %%mm2, %%mm1\n\t"
 281 "       punpckldq %%mm6, %%mm5\n\t"
 282 /**/
 283 "       pfadd   %%mm1, %%mm0\n\t"
 284 "       pfadd   %%mm5, %%mm4\n\t"
 285 "       movq    %%mm0, (%%edx)\n\t"
 286 "       movq    %%mm4, 8(%%edx)\n\t"
 287 "       pfsub   %%mm1, %%mm3\n\t"
 288 "       pfsub   %%mm5, %%mm7\n\t"
 289 "       pfmul   96(%%ebx), %%mm3\n\t"
 290 "       pfmul   104(%%ebx), %%mm7\n\t"
 291 "       movd    %%mm3, 28(%%edx)\n\t"
 292 "       movd    %%mm7, 20(%%edx)\n\t"
 293 "       psrlq   $32, %%mm3\n\t"
 294 "       psrlq   $32, %%mm7\n\t"
 295 "       movd    %%mm3, 24(%%edx)\n\t"
 296 "       movd    %%mm7, 16(%%edx)\n\t"
 297
 298 "       movq    32(%%ecx), %%mm0\n\t"
 299 "       movq    40(%%ecx), %%mm4\n\t"
 300 "       movq    %%mm0, %%mm3\n\t"
 301 "       movq    %%mm4, %%mm7\n\t"
 302 "       movq    56(%%ecx), %%mm1\n\t"
 303 "       movq    48(%%ecx), %%mm5\n\t"
 304 /* n.b.: pswapd*/
 305 "       movq    %%mm1, %%mm2\n\t"
 306 "       movq    %%mm5, %%mm6\n\t"
 307 "       psrlq   $32, %%mm1\n\t"
 308 "       psrlq   $32, %%mm5\n\t"
 309 "       punpckldq %%mm2, %%mm1\n\t"
 310 "       punpckldq %%mm6, %%mm5\n\t"
 311 /**/
 312 "       pfadd   %%mm1, %%mm0\n\t"
 313 "       pfadd   %%mm5, %%mm4\n\t"
 314 "       movq    %%mm0, 32(%%edx)\n\t"
 315 "       movq    %%mm4, 40(%%edx)\n\t"
 316 "       pfsubr  %%mm1, %%mm3\n\t"
 317 "       pfsubr  %%mm5, %%mm7\n\t"
 318 "       pfmul   96(%%ebx), %%mm3\n\t"
 319 "       pfmul   104(%%ebx), %%mm7\n\t"
 320 "       movd    %%mm3, 60(%%edx)\n\t"
 321 "       movd    %%mm7, 52(%%edx)\n\t"
 322 "       psrlq   $32, %%mm3\n\t"
 323 "       psrlq   $32, %%mm7\n\t"
 324 "       movd    %%mm3, 56(%%edx)\n\t"
 325 "       movd    %%mm7, 48(%%edx)\n\t"
 326
 327 "       movq    64(%%ecx), %%mm0\n\t"
 328 "       movq    72(%%ecx), %%mm4\n\t"
 329 "       movq    %%mm0, %%mm3\n\t"
 330 "       movq    %%mm4, %%mm7\n\t"
 331 "       movq    88(%%ecx), %%mm1\n\t"
 332 "       movq    80(%%ecx), %%mm5\n\t"
 333 /* n.b.: pswapd*/
 334 "       movq    %%mm1, %%mm2\n\t"
 335 "       movq    %%mm5, %%mm6\n\t"
 336 "       psrlq   $32, %%mm1\n\t"
 337 "       psrlq   $32, %%mm5\n\t"
 338 "       punpckldq %%mm2, %%mm1\n\t"
 339 "       punpckldq %%mm6, %%mm5\n\t"
 340 /**/
 341 "       pfadd   %%mm1, %%mm0\n\t"
 342 "       pfadd   %%mm5, %%mm4\n\t"
 343 "       movq    %%mm0, 64(%%edx)\n\t"
 344 "       movq    %%mm4, 72(%%edx)\n\t"
 345 "       pfsub   %%mm1, %%mm3\n\t"
 346 "       pfsub   %%mm5, %%mm7\n\t"
 347 "       pfmul   96(%%ebx), %%mm3\n\t"
 348 "       pfmul   104(%%ebx), %%mm7\n\t"
 349 "       movd    %%mm3, 92(%%edx)\n\t"
 350 "       movd    %%mm7, 84(%%edx)\n\t"
 351 "       psrlq   $32, %%mm3\n\t"
 352 "       psrlq   $32, %%mm7\n\t"
 353 "       movd    %%mm3, 88(%%edx)\n\t"
 354 "       movd    %%mm7, 80(%%edx)\n\t"
 355
 356 "       movq    96(%%ecx), %%mm0\n\t"
 357 "       movq    104(%%ecx), %%mm4\n\t"
 358 "       movq    %%mm0, %%mm3\n\t"
 359 "       movq    %%mm4, %%mm7\n\t"
 360 "       movq    120(%%ecx), %%mm1\n\t"
 361 "       movq    112(%%ecx), %%mm5\n\t"
 362 /* n.b.: pswapd*/
 363 "       movq    %%mm1, %%mm2\n\t"
 364 "       movq    %%mm5, %%mm6\n\t"
 365 "       psrlq   $32, %%mm1\n\t"
 366 "       psrlq   $32, %%mm5\n\t"
 367 "       punpckldq %%mm2, %%mm1\n\t"
 368 "       punpckldq %%mm6, %%mm5\n\t"
 369 /**/
 370 "       pfadd   %%mm1, %%mm0\n\t"
 371 "       pfadd   %%mm5, %%mm4\n\t"
 372 "       movq    %%mm0, 96(%%edx)\n\t"
 373 "       movq    %%mm4, 104(%%edx)\n\t"
 374 "       pfsubr  %%mm1, %%mm3\n\t"
 375 "       pfsubr  %%mm5, %%mm7\n\t"
 376 "       pfmul   96(%%ebx), %%mm3\n\t"
 377 "       pfmul   104(%%ebx), %%mm7\n\t"
 378 "       movd    %%mm3, 124(%%edx)\n\t"
 379 "       movd    %%mm7, 116(%%edx)\n\t"
 380 "       psrlq   $32, %%mm3\n\t"
 381 "       psrlq   $32, %%mm7\n\t"
 382 "       movd    %%mm3, 120(%%edx)\n\t"
 383 "       movd    %%mm7, 112(%%edx)\n\t"
 384
 385 /* Phase 5 */
 386
 387 "       movq    (%%edx), %%mm0\n\t"
 388 "       movq    16(%%edx), %%mm4\n\t"
 389 "       movq    %%mm0, %%mm3\n\t"
 390 "       movq    %%mm4, %%mm7\n\t"
 391 "       movq    8(%%edx), %%mm1\n\t"
 392 "       movq    24(%%edx), %%mm5\n\t"
 393 /* n.b.: pswapd*/
 394 "       movq    %%mm1, %%mm2\n\t"
 395 "       movq    %%mm5, %%mm6\n\t"
 396 "       psrlq   $32, %%mm1\n\t"
 397 "       psrlq   $32, %%mm5\n\t"
 398 "       punpckldq %%mm2, %%mm1\n\t"
 399 "       punpckldq %%mm6, %%mm5\n\t"
 400 /**/
 401 "       pfadd   %%mm1, %%mm0\n\t"
 402 "       pfadd   %%mm5, %%mm4\n\t"
 403 "       movq    %%mm0, (%%ecx)\n\t"
 404 "       movq    %%mm4, 16(%%ecx)\n\t"
 405 "       pfsub   %%mm1, %%mm3\n\t"
 406 "       pfsubr  %%mm5, %%mm7\n\t"
 407 "       pfmul   112(%%ebx), %%mm3\n\t"
 408 "       pfmul   112(%%ebx), %%mm7\n\t"
 409 "       movd    %%mm3, 12(%%ecx)\n\t"
 410 "       movd    %%mm7, 28(%%ecx)\n\t"
 411 "       psrlq   $32, %%mm3\n\t"
 412 "       psrlq   $32, %%mm7\n\t"
 413 "       movd    %%mm3, 8(%%ecx)\n\t"
 414 "       movd    %%mm7, 24(%%ecx)\n\t"
 415
 416 "       movq    32(%%edx), %%mm0\n\t"
 417 "       movq    48(%%edx), %%mm4\n\t"
 418 "       movq    %%mm0, %%mm3\n\t"
 419 "       movq    %%mm4, %%mm7\n\t"
 420 "       movq    40(%%edx), %%mm1\n\t"
 421 "       movq    56(%%edx), %%mm5\n\t"
 422 /* n.b.: pswapd*/
 423 "       movq    %%mm1, %%mm2\n\t"
 424 "       movq    %%mm5, %%mm6\n\t"
 425 "       psrlq   $32, %%mm1\n\t"
 426 "       psrlq   $32, %%mm5\n\t"
 427 "       punpckldq %%mm2, %%mm1\n\t"
 428 "       punpckldq %%mm6, %%mm5\n\t"
 429 /**/
 430 "       pfadd   %%mm1, %%mm0\n\t"
 431 "       pfadd   %%mm5, %%mm4\n\t"
 432 "       movq    %%mm0, 32(%%ecx)\n\t"
 433 "       movq    %%mm4, 48(%%ecx)\n\t"
 434 "       pfsub   %%mm1, %%mm3\n\t"
 435 "       pfsubr  %%mm5, %%mm7\n\t"
 436 "       pfmul   112(%%ebx), %%mm3\n\t"
 437 "       pfmul   112(%%ebx), %%mm7\n\t"
 438 "       movd    %%mm3, 44(%%ecx)\n\t"
 439 "       movd    %%mm7, 60(%%ecx)\n\t"
 440 "       psrlq   $32, %%mm3\n\t"
 441 "       psrlq   $32, %%mm7\n\t"
 442 "       movd    %%mm3, 40(%%ecx)\n\t"
 443 "       movd    %%mm7, 56(%%ecx)\n\t"
 444
 445 "       movq    64(%%edx), %%mm0\n\t"
 446 "       movq    80(%%edx), %%mm4\n\t"
 447 "       movq    %%mm0, %%mm3\n\t"
 448 "       movq    %%mm4, %%mm7\n\t"
 449 "       movq    72(%%edx), %%mm1\n\t"
 450 "       movq    88(%%edx), %%mm5\n\t"
 451 /* n.b.: pswapd*/
 452 "       movq    %%mm1, %%mm2\n\t"
 453 "       movq    %%mm5, %%mm6\n\t"
 454 "       psrlq   $32, %%mm1\n\t"
 455 "       psrlq   $32, %%mm5\n\t"
 456 "       punpckldq %%mm2, %%mm1\n\t"
 457 "       punpckldq %%mm6, %%mm5\n\t"
 458 /**/
 459 "       pfadd   %%mm1, %%mm0\n\t"
 460 "       pfadd   %%mm5, %%mm4\n\t"
 461 "       movq    %%mm0, 64(%%ecx)\n\t"
 462 "       movq    %%mm4, 80(%%ecx)\n\t"
 463 "       pfsub   %%mm1, %%mm3\n\t"
 464 "       pfsubr  %%mm5, %%mm7\n\t"
 465 "       pfmul   112(%%ebx), %%mm3\n\t"
 466 "       pfmul   112(%%ebx), %%mm7\n\t"
 467 "       movd    %%mm3, 76(%%ecx)\n\t"
 468 "       movd    %%mm7, 92(%%ecx)\n\t"
 469 "       psrlq   $32, %%mm3\n\t"
 470 "       psrlq   $32, %%mm7\n\t"
 471 "       movd    %%mm3, 72(%%ecx)\n\t"
 472 "       movd    %%mm7, 88(%%ecx)\n\t"
 473
 474 "       movq    96(%%edx), %%mm0\n\t"
 475 "       movq    112(%%edx), %%mm4\n\t"
 476 "       movq    %%mm0, %%mm3\n\t"
 477 "       movq    %%mm4, %%mm7\n\t"
 478 "       movq    104(%%edx), %%mm1\n\t"
 479 "       movq    120(%%edx), %%mm5\n\t"
 480 /* n.b.: pswapd*/
 481 "       movq    %%mm1, %%mm2\n\t"
 482 "       movq    %%mm5, %%mm6\n\t"
 483 "       psrlq   $32, %%mm1\n\t"
 484 "       psrlq   $32, %%mm5\n\t"
 485 "       punpckldq %%mm2, %%mm1\n\t"
 486 "       punpckldq %%mm6, %%mm5\n\t"
 487 /**/
 488 "       pfadd   %%mm1, %%mm0\n\t"
 489 "       pfadd   %%mm5, %%mm4\n\t"
 490 "       movq    %%mm0, 96(%%ecx)\n\t"
 491 "       movq    %%mm4, 112(%%ecx)\n\t"
 492 "       pfsub   %%mm1, %%mm3\n\t"
 493 "       pfsubr  %%mm5, %%mm7\n\t"
 494 "       pfmul   112(%%ebx), %%mm3\n\t"
 495 "       pfmul   112(%%ebx), %%mm7\n\t"
 496 "       movd    %%mm3, 108(%%ecx)\n\t"
 497 "       movd    %%mm7, 124(%%ecx)\n\t"
 498 "       psrlq   $32, %%mm3\n\t"
 499 "       psrlq   $32, %%mm7\n\t"
 500 "       movd    %%mm3, 104(%%ecx)\n\t"
 501 "       movd    %%mm7, 120(%%ecx)\n\t"
 502
 503 /* Phase 6. This is the end of easy road. */
 504 /* Code below is coded in scalar mode. Should be optimized */
 505
 506 "       movd    "MANGLE(plus_1f)", %%mm6\n\t"
 507 "       punpckldq 120(%%ebx), %%mm6\n\t"      /* mm6 = 1.0 | 120(%%ebx)*/
 508 "       movq    "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */
 509
 510 "       movq    32(%%ecx), %%mm0\n\t"
 511 "       movq    64(%%ecx), %%mm2\n\t"
 512 "       movq    %%mm0, %%mm1\n\t"
 513 "       movq    %%mm2, %%mm3\n\t"
 514 "       pxor    %%mm7, %%mm1\n\t"
 515 "       pxor    %%mm7, %%mm3\n\t"
 516 "       pfacc   %%mm1, %%mm0\n\t"
 517 "       pfacc   %%mm3, %%mm2\n\t"
 518 "       pfmul   %%mm6, %%mm0\n\t"
 519 "       pfmul   %%mm6, %%mm2\n\t"
 520 "       movq    %%mm0, 32(%%edx)\n\t"
 521 "       movq    %%mm2, 64(%%edx)\n\t"
 522
 523 "       movd    44(%%ecx), %%mm0\n\t"
 524 "       movd    40(%%ecx), %%mm2\n\t"
 525 "       movd    120(%%ebx), %%mm3\n\t"
 526 "       punpckldq 76(%%ecx), %%mm0\n\t"
 527 "       punpckldq 72(%%ecx), %%mm2\n\t"
 528 "       punpckldq %%mm3, %%mm3\n\t"
 529 "       movq    %%mm0, %%mm4\n\t"
 530 "       movq    %%mm2, %%mm5\n\t"
 531 "       pfsub   %%mm2, %%mm0\n\t"
 532 "       pfmul   %%mm3, %%mm0\n\t"
 533 "       movq    %%mm0, %%mm1\n\t"
 534 "       pfadd   %%mm5, %%mm0\n\t"
 535 "       pfadd   %%mm4, %%mm0\n\t"
 536 "       movq    %%mm0, %%mm2\n\t"
 537 "       punpckldq %%mm1, %%mm0\n\t"
 538 "       punpckhdq %%mm1, %%mm2\n\t"
 539 "       movq    %%mm0, 40(%%edx)\n\t"
 540 "       movq    %%mm2, 72(%%edx)\n\t"
 541
 542 "       movd   48(%%ecx), %%mm3\n\t"
 543 "       movd   60(%%ecx), %%mm2\n\t"
 544 "       pfsub  52(%%ecx), %%mm3\n\t"
 545 "       pfsub  56(%%ecx), %%mm2\n\t"
 546 "       pfmul 120(%%ebx), %%mm3\n\t"
 547 "       pfmul 120(%%ebx), %%mm2\n\t"
 548 "       movq    %%mm2, %%mm1\n\t"
 549
 550 "       pfadd  56(%%ecx), %%mm1\n\t"
 551 "       pfadd  60(%%ecx), %%mm1\n\t"
 552 "       movq    %%mm1, %%mm0\n\t"
 553
 554 "       pfadd  48(%%ecx), %%mm0\n\t"
 555 "       pfadd  52(%%ecx), %%mm0\n\t"
 556 "       pfadd   %%mm3, %%mm1\n\t"
 557 "       punpckldq %%mm2, %%mm1\n\t"
 558 "       pfadd   %%mm3, %%mm2\n\t"
 559 "       punpckldq %%mm2, %%mm0\n\t"
 560 "       movq    %%mm1, 56(%%edx)\n\t"
 561 "       movq    %%mm0, 48(%%edx)\n\t"
 562
 563 /*---*/
 564
 565 "       movd   92(%%ecx), %%mm1\n\t"
 566 "       pfsub  88(%%ecx), %%mm1\n\t"
 567 "       pfmul 120(%%ebx), %%mm1\n\t"
 568 "       movd   %%mm1, 92(%%edx)\n\t"
 569 "       pfadd  92(%%ecx), %%mm1\n\t"
 570 "       pfadd  88(%%ecx), %%mm1\n\t"
 571 "       movq   %%mm1, %%mm0\n\t"
 572
 573 "       pfadd  80(%%ecx), %%mm0\n\t"
 574 "       pfadd  84(%%ecx), %%mm0\n\t"
 575 "       movd   %%mm0, 80(%%edx)\n\t"
 576
 577 "       movd   80(%%ecx), %%mm0\n\t"
 578 "       pfsub  84(%%ecx), %%mm0\n\t"
 579 "       pfmul 120(%%ebx), %%mm0\n\t"
 580 "       pfadd  %%mm0, %%mm1\n\t"
 581 "       pfadd  92(%%edx), %%mm0\n\t"
 582 "       punpckldq %%mm1, %%mm0\n\t"
 583 "       movq   %%mm0, 84(%%edx)\n\t"
 584
 585 "       movq    96(%%ecx), %%mm0\n\t"
 586 "       movq    %%mm0, %%mm1\n\t"
 587 "       pxor    %%mm7, %%mm1\n\t"
 588 "       pfacc   %%mm1, %%mm0\n\t"
 589 "       pfmul   %%mm6, %%mm0\n\t"
 590 "       movq    %%mm0, 96(%%edx)\n\t"
 591
 592 "       movd  108(%%ecx), %%mm0\n\t"
 593 "       pfsub 104(%%ecx), %%mm0\n\t"
 594 "       pfmul 120(%%ebx), %%mm0\n\t"
 595 "       movd  %%mm0, 108(%%edx)\n\t"
 596 "       pfadd 104(%%ecx), %%mm0\n\t"
 597 "       pfadd 108(%%ecx), %%mm0\n\t"
 598 "       movd  %%mm0, 104(%%edx)\n\t"
 599
 600 "       movd  124(%%ecx), %%mm1\n\t"
 601 "       pfsub 120(%%ecx), %%mm1\n\t"
 602 "       pfmul 120(%%ebx), %%mm1\n\t"
 603 "       movd  %%mm1, 124(%%edx)\n\t"
 604 "       pfadd 120(%%ecx), %%mm1\n\t"
 605 "       pfadd 124(%%ecx), %%mm1\n\t"
 606 "       movq  %%mm1, %%mm0\n\t"
 607
 608 "       pfadd 112(%%ecx), %%mm0\n\t"
 609 "       pfadd 116(%%ecx), %%mm0\n\t"
 610 "       movd  %%mm0, 112(%%edx)\n\t"
 611
 612 "       movd  112(%%ecx), %%mm0\n\t"
 613 "       pfsub 116(%%ecx), %%mm0\n\t"
 614 "       pfmul 120(%%ebx), %%mm0\n\t"
 615 "       pfadd %%mm0,%%mm1\n\t"
 616 "       pfadd 124(%%edx), %%mm0\n\t"
 617 "       punpckldq %%mm1, %%mm0\n\t"
 618 "       movq  %%mm0, 116(%%edx)\n\t"
 619
 620 "       jnz .L01\n\t"
 621
 622 /* Phase 7*/
 623 /* Code below is coded in scalar mode. Should be optimized */
 624
 625 "       movd      (%%ecx), %%mm0\n\t"
 626 "       pfadd    4(%%ecx), %%mm0\n\t"
 627 "       movd     %%mm0, 1024(%%esi)\n\t"
 628
 629 "       movd      (%%ecx), %%mm0\n\t"
 630 "       pfsub    4(%%ecx), %%mm0\n\t"
 631 "       pfmul  120(%%ebx), %%mm0\n\t"
 632 "       movd      %%mm0, (%%esi)\n\t"
 633 "       movd      %%mm0, (%%edi)\n\t"
 634
 635 "       movd   12(%%ecx), %%mm0\n\t"
 636 "       pfsub   8(%%ecx), %%mm0\n\t"
 637 "       pfmul 120(%%ebx), %%mm0\n\t"
 638 "       movd    %%mm0, 512(%%edi)\n\t"
 639 "       pfadd   12(%%ecx), %%mm0\n\t"
 640 "       pfadd   8(%%ecx), %%mm0\n\t"
 641 "       movd    %%mm0, 512(%%esi)\n\t"
 642
 643 "       movd   16(%%ecx), %%mm0\n\t"
 644 "       pfsub  20(%%ecx), %%mm0\n\t"
 645 "       pfmul 120(%%ebx), %%mm0\n\t"
 646 "       movq    %%mm0, %%mm3\n\t"
 647
 648 "       movd   28(%%ecx), %%mm0\n\t"
 649 "       pfsub  24(%%ecx), %%mm0\n\t"
 650 "       pfmul 120(%%ebx), %%mm0\n\t"
 651 "       movd    %%mm0, 768(%%edi)\n\t"
 652 "       movq    %%mm0, %%mm2\n\t"
 653
 654 "       pfadd  24(%%ecx), %%mm0\n\t"
 655 "       pfadd  28(%%ecx), %%mm0\n\t"
 656 "       movq    %%mm0, %%mm1\n\t"
 657
 658 "       pfadd  16(%%ecx), %%mm0\n\t"
 659 "       pfadd  20(%%ecx), %%mm0\n\t"
 660 "       movd   %%mm0, 768(%%esi)\n\t"
 661 "       pfadd  %%mm3, %%mm1\n\t"
 662 "       movd   %%mm1, 256(%%esi)\n\t"
 663 "       pfadd  %%mm3, %%mm2\n\t"
 664 "       movd   %%mm2, 256(%%edi)\n\t"
 665
 666 /* Phase 8*/
 667
 668 "       movq   32(%%edx), %%mm0\n\t"
 669 "       movq   48(%%edx), %%mm1\n\t"
 670 "       pfadd  48(%%edx), %%mm0\n\t"
 671 "       pfadd  40(%%edx), %%mm1\n\t"
 672 "       movd   %%mm0, 896(%%esi)\n\t"
 673 "       movd   %%mm1, 640(%%esi)\n\t"
 674 "       psrlq  $32, %%mm0\n\t"
 675 "       psrlq  $32, %%mm1\n\t"
 676 "       movd   %%mm0, 128(%%edi)\n\t"
 677 "       movd   %%mm1, 384(%%edi)\n\t"
 678
 679 "       movd   40(%%edx), %%mm0\n\t"
 680 "       pfadd  56(%%edx), %%mm0\n\t"
 681 "       movd   %%mm0, 384(%%esi)\n\t"
 682
 683 "       movd   56(%%edx), %%mm0\n\t"
 684 "       pfadd  36(%%edx), %%mm0\n\t"
 685 "       movd   %%mm0, 128(%%esi)\n\t"
 686
 687 "       movd   60(%%edx), %%mm0\n\t"
 688 "       movd   %%mm0, 896(%%edi)\n\t"
 689 "       pfadd  44(%%edx), %%mm0\n\t"
 690 "       movd   %%mm0, 640(%%edi)\n\t"
 691
 692 "       movq   96(%%edx), %%mm0\n\t"
 693 "       movq   112(%%edx), %%mm2\n\t"
 694 "       movq   104(%%edx), %%mm4\n\t"
 695 "       pfadd  112(%%edx), %%mm0\n\t"
 696 "       pfadd  104(%%edx), %%mm2\n\t"
 697 "       pfadd  120(%%edx), %%mm4\n\t"
 698 "       movq   %%mm0, %%mm1\n\t"
 699 "       movq   %%mm2, %%mm3\n\t"
 700 "       movq   %%mm4, %%mm5\n\t"
 701 "       pfadd  64(%%edx), %%mm0\n\t"
 702 "       pfadd  80(%%edx), %%mm2\n\t"
 703 "       pfadd  72(%%edx), %%mm4\n\t"
 704 "       movd   %%mm0, 960(%%esi)\n\t"
 705 "       movd   %%mm2, 704(%%esi)\n\t"
 706 "       movd   %%mm4, 448(%%esi)\n\t"
 707 "       psrlq  $32, %%mm0\n\t"
 708 "       psrlq  $32, %%mm2\n\t"
 709 "       psrlq  $32, %%mm4\n\t"
 710 "       movd   %%mm0, 64(%%edi)\n\t"
 711 "       movd   %%mm2, 320(%%edi)\n\t"
 712 "       movd   %%mm4, 576(%%edi)\n\t"
 713 "       pfadd  80(%%edx), %%mm1\n\t"
 714 "       pfadd  72(%%edx), %%mm3\n\t"
 715 "       pfadd  88(%%edx), %%mm5\n\t"
 716 "       movd   %%mm1, 832(%%esi)\n\t"
 717 "       movd   %%mm3, 576(%%esi)\n\t"
 718 "       movd   %%mm5, 320(%%esi)\n\t"
 719 "       psrlq  $32, %%mm1\n\t"
 720 "       psrlq  $32, %%mm3\n\t"
 721 "       psrlq  $32, %%mm5\n\t"
 722 "       movd   %%mm1, 192(%%edi)\n\t"
 723 "       movd   %%mm3, 448(%%edi)\n\t"
 724 "       movd   %%mm5, 704(%%edi)\n\t"
 725
 726 "       movd   120(%%edx), %%mm0\n\t"
 727 "       pfadd  100(%%edx), %%mm0\n\t"
 728 "       movq   %%mm0, %%mm1\n\t"
 729 "       pfadd  88(%%edx), %%mm0\n\t"
 730 "       movd   %%mm0, 192(%%esi)\n\t"
 731 "       pfadd  68(%%edx), %%mm1\n\t"
 732 "       movd   %%mm1, 64(%%esi)\n\t"
 733
 734 "       movd  124(%%edx), %%mm0\n\t"
 735 "       movd  %%mm0, 960(%%edi)\n\t"
 736 "       pfadd  92(%%edx), %%mm0\n\t"
 737 "       movd  %%mm0, 832(%%edi)\n\t"
 738
 739 "       jmp     .L_bye\n\t"
 740 ".L01:\n\t"
 741 /* Phase 9*/
 742
 743 "       movq    (%%ecx), %%mm0\n\t"
 744 "       movq    %%mm0, %%mm1\n\t"
 745 "       pxor    %%mm7, %%mm1\n\t"
 746 "       pfacc   %%mm1, %%mm0\n\t"
 747 "       pfmul   %%mm6, %%mm0\n\t"
 748 "       pf2id   %%mm0, %%mm0\n\t"
 749 "       movd    %%mm0, %%eax\n\t"
 750 "       movw    %%ax, 512(%%esi)\n\t"
 751 "       psrlq   $32, %%mm0\n\t"
 752 "       movd    %%mm0, %%eax\n\t"
 753 "       movw    %%ax, (%%esi)\n\t"
 754
 755 "       movd    12(%%ecx), %%mm0\n\t"
 756 "       pfsub    8(%%ecx), %%mm0\n\t"
 757 "       pfmul  120(%%ebx), %%mm0\n\t"
 758 "       pf2id    %%mm0, %%mm7\n\t"
 759 "       movd     %%mm7, %%eax\n\t"
 760 "       movw     %%ax, 256(%%edi)\n\t"
 761 "       pfadd   12(%%ecx), %%mm0\n\t"
 762 "       pfadd    8(%%ecx), %%mm0\n\t"
 763 "       pf2id    %%mm0, %%mm0\n\t"
 764 "       movd     %%mm0, %%eax\n\t"
 765 "       movw     %%ax, 256(%%esi)\n\t"
 766
 767 "       movd   16(%%ecx), %%mm3\n\t"
 768 "       pfsub  20(%%ecx), %%mm3\n\t"
 769 "       pfmul  120(%%ebx), %%mm3\n\t"
 770 "       movq   %%mm3, %%mm2\n\t"
 771
 772 "       movd   28(%%ecx), %%mm2\n\t"
 773 "       pfsub  24(%%ecx), %%mm2\n\t"
 774 "       pfmul 120(%%ebx), %%mm2\n\t"
 775 "       movq   %%mm2, %%mm1\n\t"
 776
 777 "       pf2id  %%mm2, %%mm7\n\t"
 778 "       movd   %%mm7, %%eax\n\t"
 779 "       movw   %%ax, 384(%%edi)\n\t"
 780
 781 "       pfadd  24(%%ecx), %%mm1\n\t"
 782 "       pfadd  28(%%ecx), %%mm1\n\t"
 783 "       movq   %%mm1, %%mm0\n\t"
 784
 785 "       pfadd  16(%%ecx), %%mm0\n\t"
 786 "       pfadd  20(%%ecx), %%mm0\n\t"
 787 "       pf2id  %%mm0, %%mm0\n\t"
 788 "       movd   %%mm0, %%eax\n\t"
 789 "       movw   %%ax, 384(%%esi)\n\t"
 790 "       pfadd  %%mm3, %%mm1\n\t"
 791 "       pf2id  %%mm1, %%mm1\n\t"
 792 "       movd   %%mm1, %%eax\n\t"
 793 "       movw   %%ax, 128(%%esi)\n\t"
 794 "       pfadd  %%mm3, %%mm2\n\t"
 795 "       pf2id  %%mm2, %%mm2\n\t"
 796 "       movd   %%mm2, %%eax\n\t"
 797 "       movw   %%ax, 128(%%edi)\n\t"
 798
 799 /* Phase 10*/
 800
 801 "       movq    32(%%edx), %%mm0\n\t"
 802 "       movq    48(%%edx), %%mm1\n\t"
 803 "       pfadd   48(%%edx), %%mm0\n\t"
 804 "       pfadd   40(%%edx), %%mm1\n\t"
 805 "       pf2id   %%mm0, %%mm0\n\t"
 806 "       pf2id   %%mm1, %%mm1\n\t"
 807 "       movd    %%mm0, %%eax\n\t"
 808 "       movd    %%mm1, %%ecx\n\t"
 809 "       movw    %%ax, 448(%%esi)\n\t"
 810 "       movw    %%cx, 320(%%esi)\n\t"
 811 "       psrlq   $32, %%mm0\n\t"
 812 "       psrlq   $32, %%mm1\n\t"
 813 "       movd    %%mm0, %%eax\n\t"
 814 "       movd    %%mm1, %%ecx\n\t"
 815 "       movw    %%ax, 64(%%edi)\n\t"
 816 "       movw    %%cx, 192(%%edi)\n\t"
 817
 818 "       movd   40(%%edx), %%mm3\n\t"
 819 "       movd   56(%%edx), %%mm4\n\t"
 820 "       movd   60(%%edx), %%mm0\n\t"
 821 "       movd   44(%%edx), %%mm2\n\t"
 822 "       movd  120(%%edx), %%mm5\n\t"
 823 "       punpckldq %%mm4, %%mm3\n\t"
 824 "       punpckldq 124(%%edx), %%mm0\n\t"
 825 "       pfadd 100(%%edx), %%mm5\n\t"
 826 "       punpckldq 36(%%edx), %%mm4\n\t"
 827 "       punpckldq 92(%%edx), %%mm2\n\t"
 828 "       movq  %%mm5, %%mm6\n\t"
 829 "       pfadd  %%mm4, %%mm3\n\t"
 830 "       pf2id  %%mm0, %%mm1\n\t"
 831 "       pf2id  %%mm3, %%mm3\n\t"
 832 "       pfadd  88(%%edx), %%mm5\n\t"
 833 "       movd   %%mm1, %%eax\n\t"
 834 "       movd   %%mm3, %%ecx\n\t"
 835 "       movw   %%ax, 448(%%edi)\n\t"
 836 "       movw   %%cx, 192(%%esi)\n\t"
 837 "       pf2id  %%mm5, %%mm5\n\t"
 838 "       psrlq  $32, %%mm1\n\t"
 839 "       psrlq  $32, %%mm3\n\t"
 840 "       movd   %%mm5, %%ebx\n\t"
 841 "       movd   %%mm1, %%eax\n\t"
 842 "       movd   %%mm3, %%ecx\n\t"
 843 "       movw   %%bx, 96(%%esi)\n\t"
 844 "       movw   %%ax, 480(%%edi)\n\t"
 845 "       movw   %%cx, 64(%%esi)\n\t"
 846 "       pfadd  %%mm2, %%mm0\n\t"
 847 "       pf2id  %%mm0, %%mm0\n\t"
 848 "       movd   %%mm0, %%eax\n\t"
 849 "       pfadd  68(%%edx), %%mm6\n\t"
 850 "       movw   %%ax, 320(%%edi)\n\t"
 851 "       psrlq  $32, %%mm0\n\t"
 852 "       pf2id  %%mm6, %%mm6\n\t"
 853 "       movd   %%mm0, %%eax\n\t"
 854 "       movd   %%mm6, %%ebx\n\t"
 855 "       movw   %%ax, 416(%%edi)\n\t"
 856 "       movw   %%bx, 32(%%esi)\n\t"
 857
 858 "       movq   96(%%edx), %%mm0\n\t"
 859 "       movq  112(%%edx), %%mm2\n\t"
 860 "       movq  104(%%edx), %%mm4\n\t"
 861 "       pfadd %%mm2, %%mm0\n\t"
 862 "       pfadd %%mm4, %%mm2\n\t"
 863 "       pfadd 120(%%edx), %%mm4\n\t"
 864 "       movq  %%mm0, %%mm1\n\t"
 865 "       movq  %%mm2, %%mm3\n\t"
 866 "       movq  %%mm4, %%mm5\n\t"
 867 "       pfadd  64(%%edx), %%mm0\n\t"
 868 "       pfadd  80(%%edx), %%mm2\n\t"
 869 "       pfadd  72(%%edx), %%mm4\n\t"
 870 "       pf2id  %%mm0, %%mm0\n\t"
 871 "       pf2id  %%mm2, %%mm2\n\t"
 872 "       pf2id  %%mm4, %%mm4\n\t"
 873 "       movd   %%mm0, %%eax\n\t"
 874 "       movd   %%mm2, %%ecx\n\t"
 875 "       movd   %%mm4, %%ebx\n\t"
 876 "       movw   %%ax, 480(%%esi)\n\t"
 877 "       movw   %%cx, 352(%%esi)\n\t"
 878 "       movw   %%bx, 224(%%esi)\n\t"
 879 "       psrlq  $32, %%mm0\n\t"
 880 "       psrlq  $32, %%mm2\n\t"
 881 "       psrlq  $32, %%mm4\n\t"
 882 "       movd   %%mm0, %%eax\n\t"
 883 "       movd   %%mm2, %%ecx\n\t"
 884 "       movd   %%mm4, %%ebx\n\t"
 885 "       movw   %%ax, 32(%%edi)\n\t"
 886 "       movw   %%cx, 160(%%edi)\n\t"
 887 "       movw   %%bx, 288(%%edi)\n\t"
 888 "       pfadd  80(%%edx), %%mm1\n\t"
 889 "       pfadd  72(%%edx), %%mm3\n\t"
 890 "       pfadd  88(%%edx), %%mm5\n\t"
 891 "       pf2id  %%mm1, %%mm1\n\t"
 892 "       pf2id  %%mm3, %%mm3\n\t"
 893 "       pf2id  %%mm5, %%mm5\n\t"
 894 "       movd   %%mm1, %%eax\n\t"
 895 "       movd   %%mm3, %%ecx\n\t"
 896 "       movd   %%mm5, %%ebx\n\t"
 897 "       movw   %%ax, 416(%%esi)\n\t"
 898 "       movw   %%cx, 288(%%esi)\n\t"
 899 "       movw   %%bx, 160(%%esi)\n\t"
 900 "       psrlq  $32, %%mm1\n\t"
 901 "       psrlq  $32, %%mm3\n\t"
 902 "       psrlq  $32, %%mm5\n\t"
 903 "       movd   %%mm1, %%eax\n\t"
 904 "       movd   %%mm3, %%ecx\n\t"
 905 "       movd   %%mm5, %%ebx\n\t"
 906 "       movw   %%ax, 96(%%edi)\n\t"
 907 "       movw   %%cx, 224(%%edi)\n\t"
 908 "       movw   %%bx, 352(%%edi)\n\t"
 909
 910 "       movsw\n\t"
 911
 912 ".L_bye:\n\t"
 913 "       femms\n\t"
 914         :
 915         :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
 916         :"memory","%ebx","%esi","%edi");
 917 }