mp3lib/dct64_k7.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 3dnowex-DSP! optimization by Nick Kurshev
   6 *
   7 * TODO: optimize scalar 3dnow! code
   8 * Warning: Phases 7 & 8 are not tested
   9 */
  10
  11 #include "config.h"
  12 #include "mangle.h"
  13 #include "mpg123.h"
  14 #include "libavutil/x86_cpu.h"
  15
  16 static unsigned long long int attribute_used __attribute__((aligned(8))) x_plus_minus_3dnow = 0x8000000000000000ULL;
  17 static float attribute_used plus_1f = 1.0;
  18
  19 void dct64_MMX_3dnowex(short *a,short *b,real *c)
  20 {
  21   char tmp[256];
  22     __asm__ volatile(
  23 "       mov %2,%%"REG_a"\n\t"
  24
  25 "       lea 128+%3,%%"REG_d"\n\t"
  26 "       mov %0,%%"REG_S"\n\t"
  27 "       mov %1,%%"REG_D"\n\t"
  28 "       mov $"MANGLE(costab_mmx)",%%"REG_b"\n\t"
  29 "       lea %3,%%"REG_c"\n\t"
  30
  31 /* Phase 1*/
  32 "       movq    (%%"REG_a"), %%mm0\n\t"
  33 "       movq    8(%%"REG_a"), %%mm4\n\t"
  34 "       movq    %%mm0, %%mm3\n\t"
  35 "       movq    %%mm4, %%mm7\n\t"
  36 "       pswapd  120(%%"REG_a"), %%mm1\n\t"
  37 "       pswapd  112(%%"REG_a"), %%mm5\n\t"
  38 "       pfadd   %%mm1, %%mm0\n\t"
  39 "       pfadd   %%mm5, %%mm4\n\t"
  40 "       movq    %%mm0, (%%"REG_d")\n\t"
  41 "       movq    %%mm4, 8(%%"REG_d")\n\t"
  42 "       pfsub   %%mm1, %%mm3\n\t"
  43 "       pfsub   %%mm5, %%mm7\n\t"
  44 "       pfmul   (%%"REG_b"), %%mm3\n\t"
  45 "       pfmul   8(%%"REG_b"), %%mm7\n\t"
  46 "       pswapd  %%mm3, %%mm3\n\t"
  47 "       pswapd  %%mm7, %%mm7\n\t"
  48 "       movq    %%mm3, 120(%%"REG_d")\n\t"
  49 "       movq    %%mm7, 112(%%"REG_d")\n\t"
  50
  51 "       movq    16(%%"REG_a"), %%mm0\n\t"
  52 "       movq    24(%%"REG_a"), %%mm4\n\t"
  53 "       movq    %%mm0, %%mm3\n\t"
  54 "       movq    %%mm4, %%mm7\n\t"
  55 "       pswapd  104(%%"REG_a"), %%mm1\n\t"
  56 "       pswapd  96(%%"REG_a"), %%mm5\n\t"
  57 "       pfadd   %%mm1, %%mm0\n\t"
  58 "       pfadd   %%mm5, %%mm4\n\t"
  59 "       movq    %%mm0, 16(%%"REG_d")\n\t"
  60 "       movq    %%mm4, 24(%%"REG_d")\n\t"
  61 "       pfsub   %%mm1, %%mm3\n\t"
  62 "       pfsub   %%mm5, %%mm7\n\t"
  63 "       pfmul   16(%%"REG_b"), %%mm3\n\t"
  64 "       pfmul   24(%%"REG_b"), %%mm7\n\t"
  65 "       pswapd  %%mm3, %%mm3\n\t"
  66 "       pswapd  %%mm7, %%mm7\n\t"
  67 "       movq    %%mm3, 104(%%"REG_d")\n\t"
  68 "       movq    %%mm7, 96(%%"REG_d")\n\t"
  69
  70 "       movq    32(%%"REG_a"), %%mm0\n\t"
  71 "       movq    40(%%"REG_a"), %%mm4\n\t"
  72 "       movq    %%mm0, %%mm3\n\t"
  73 "       movq    %%mm4, %%mm7\n\t"
  74 "       pswapd  88(%%"REG_a"), %%mm1\n\t"
  75 "       pswapd  80(%%"REG_a"), %%mm5\n\t"
  76 "       pfadd   %%mm1, %%mm0\n\t"
  77 "       pfadd   %%mm5, %%mm4\n\t"
  78 "       movq    %%mm0, 32(%%"REG_d")\n\t"
  79 "       movq    %%mm4, 40(%%"REG_d")\n\t"
  80 "       pfsub   %%mm1, %%mm3\n\t"
  81 "       pfsub   %%mm5, %%mm7\n\t"
  82 "       pfmul   32(%%"REG_b"), %%mm3\n\t"
  83 "       pfmul   40(%%"REG_b"), %%mm7\n\t"
  84 "       pswapd  %%mm3, %%mm3\n\t"
  85 "       pswapd  %%mm7, %%mm7\n\t"
  86 "       movq    %%mm3, 88(%%"REG_d")\n\t"
  87 "       movq    %%mm7, 80(%%"REG_d")\n\t"
  88
  89 "       movq    48(%%"REG_a"), %%mm0\n\t"
  90 "       movq    56(%%"REG_a"), %%mm4\n\t"
  91 "       movq    %%mm0, %%mm3\n\t"
  92 "       movq    %%mm4, %%mm7\n\t"
  93 "       pswapd  72(%%"REG_a"), %%mm1\n\t"
  94 "       pswapd  64(%%"REG_a"), %%mm5\n\t"
  95 "       pfadd   %%mm1, %%mm0\n\t"
  96 "       pfadd   %%mm5, %%mm4\n\t"
  97 "       movq    %%mm0, 48(%%"REG_d")\n\t"
  98 "       movq    %%mm4, 56(%%"REG_d")\n\t"
  99 "       pfsub   %%mm1, %%mm3\n\t"
 100 "       pfsub   %%mm5, %%mm7\n\t"
 101 "       pfmul   48(%%"REG_b"), %%mm3\n\t"
 102 "       pfmul   56(%%"REG_b"), %%mm7\n\t"
 103 "       pswapd  %%mm3, %%mm3\n\t"
 104 "       pswapd  %%mm7, %%mm7\n\t"
 105 "       movq    %%mm3, 72(%%"REG_d")\n\t"
 106 "       movq    %%mm7, 64(%%"REG_d")\n\t"
 107
 108 /* Phase 2*/
 109
 110 "       movq    (%%"REG_d"), %%mm0\n\t"
 111 "       movq    8(%%"REG_d"), %%mm4\n\t"
 112 "       movq    %%mm0, %%mm3\n\t"
 113 "       movq    %%mm4, %%mm7\n\t"
 114 "       pswapd  56(%%"REG_d"), %%mm1\n\t"
 115 "       pswapd  48(%%"REG_d"), %%mm5\n\t"
 116 "       pfadd   %%mm1, %%mm0\n\t"
 117 "       pfadd   %%mm5, %%mm4\n\t"
 118 "       movq    %%mm0, (%%"REG_c")\n\t"
 119 "       movq    %%mm4, 8(%%"REG_c")\n\t"
 120 "       pfsub   %%mm1, %%mm3\n\t"
 121 "       pfsub   %%mm5, %%mm7\n\t"
 122 "       pfmul   64(%%"REG_b"), %%mm3\n\t"
 123 "       pfmul   72(%%"REG_b"), %%mm7\n\t"
 124 "       pswapd  %%mm3, %%mm3\n\t"
 125 "       pswapd  %%mm7, %%mm7\n\t"
 126 "       movq    %%mm3, 56(%%"REG_c")\n\t"
 127 "       movq    %%mm7, 48(%%"REG_c")\n\t"
 128
 129 "       movq    16(%%"REG_d"), %%mm0\n\t"
 130 "       movq    24(%%"REG_d"), %%mm4\n\t"
 131 "       movq    %%mm0, %%mm3\n\t"
 132 "       movq    %%mm4, %%mm7\n\t"
 133 "       pswapd  40(%%"REG_d"), %%mm1\n\t"
 134 "       pswapd  32(%%"REG_d"), %%mm5\n\t"
 135 "       pfadd   %%mm1, %%mm0\n\t"
 136 "       pfadd   %%mm5, %%mm4\n\t"
 137 "       movq    %%mm0, 16(%%"REG_c")\n\t"
 138 "       movq    %%mm4, 24(%%"REG_c")\n\t"
 139 "       pfsub   %%mm1, %%mm3\n\t"
 140 "       pfsub   %%mm5, %%mm7\n\t"
 141 "       pfmul   80(%%"REG_b"), %%mm3\n\t"
 142 "       pfmul   88(%%"REG_b"), %%mm7\n\t"
 143 "       pswapd  %%mm3, %%mm3\n\t"
 144 "       pswapd  %%mm7, %%mm7\n\t"
 145 "       movq    %%mm3, 40(%%"REG_c")\n\t"
 146 "       movq    %%mm7, 32(%%"REG_c")\n\t"
 147
 148 /* Phase 3*/
 149
 150 "       movq    64(%%"REG_d"), %%mm0\n\t"
 151 "       movq    72(%%"REG_d"), %%mm4\n\t"
 152 "       movq    %%mm0, %%mm3\n\t"
 153 "       movq    %%mm4, %%mm7\n\t"
 154 "       pswapd  120(%%"REG_d"), %%mm1\n\t"
 155 "       pswapd  112(%%"REG_d"), %%mm5\n\t"
 156 "       pfadd   %%mm1, %%mm0\n\t"
 157 "       pfadd   %%mm5, %%mm4\n\t"
 158 "       movq    %%mm0, 64(%%"REG_c")\n\t"
 159 "       movq    %%mm4, 72(%%"REG_c")\n\t"
 160 "       pfsubr  %%mm1, %%mm3\n\t"
 161 "       pfsubr  %%mm5, %%mm7\n\t"
 162 "       pfmul   64(%%"REG_b"), %%mm3\n\t"
 163 "       pfmul   72(%%"REG_b"), %%mm7\n\t"
 164 "       pswapd  %%mm3, %%mm3\n\t"
 165 "       pswapd  %%mm7, %%mm7\n\t"
 166 "       movq    %%mm3, 120(%%"REG_c")\n\t"
 167 "       movq    %%mm7, 112(%%"REG_c")\n\t"
 168
 169 "       movq    80(%%"REG_d"), %%mm0\n\t"
 170 "       movq    88(%%"REG_d"), %%mm4\n\t"
 171 "       movq    %%mm0, %%mm3\n\t"
 172 "       movq    %%mm4, %%mm7\n\t"
 173 "       pswapd  104(%%"REG_d"), %%mm1\n\t"
 174 "       pswapd  96(%%"REG_d"), %%mm5\n\t"
 175 "       pfadd   %%mm1, %%mm0\n\t"
 176 "       pfadd   %%mm5, %%mm4\n\t"
 177 "       movq    %%mm0, 80(%%"REG_c")\n\t"
 178 "       movq    %%mm4, 88(%%"REG_c")\n\t"
 179 "       pfsubr  %%mm1, %%mm3\n\t"
 180 "       pfsubr  %%mm5, %%mm7\n\t"
 181 "       pfmul   80(%%"REG_b"), %%mm3\n\t"
 182 "       pfmul   88(%%"REG_b"), %%mm7\n\t"
 183 "       pswapd  %%mm3, %%mm3\n\t"
 184 "       pswapd  %%mm7, %%mm7\n\t"
 185 "       movq    %%mm3, 104(%%"REG_c")\n\t"
 186 "       movq    %%mm7, 96(%%"REG_c")\n\t"
 187
 188 /* Phase 4*/
 189
 190 "       movq    96(%%"REG_b"), %%mm2\n\t"
 191 "       movq    104(%%"REG_b"), %%mm6\n\t"
 192
 193 "       movq    (%%"REG_c"), %%mm0\n\t"
 194 "       movq    8(%%"REG_c"), %%mm4\n\t"
 195 "       movq    %%mm0, %%mm3\n\t"
 196 "       movq    %%mm4, %%mm7\n\t"
 197 "       pswapd  24(%%"REG_c"), %%mm1\n\t"
 198 "       pswapd  16(%%"REG_c"), %%mm5\n\t"
 199 "       pfadd   %%mm1, %%mm0\n\t"
 200 "       pfadd   %%mm5, %%mm4\n\t"
 201 "       movq    %%mm0, (%%"REG_d")\n\t"
 202 "       movq    %%mm4, 8(%%"REG_d")\n\t"
 203 "       pfsub   %%mm1, %%mm3\n\t"
 204 "       pfsub   %%mm5, %%mm7\n\t"
 205 "       pfmul   %%mm2, %%mm3\n\t"
 206 "       pfmul   %%mm6, %%mm7\n\t"
 207 "       pswapd  %%mm3, %%mm3\n\t"
 208 "       pswapd  %%mm7, %%mm7\n\t"
 209 "       movq    %%mm3, 24(%%"REG_d")\n\t"
 210 "       movq    %%mm7, 16(%%"REG_d")\n\t"
 211
 212 "       movq    32(%%"REG_c"), %%mm0\n\t"
 213 "       movq    40(%%"REG_c"), %%mm4\n\t"
 214 "       movq    %%mm0, %%mm3\n\t"
 215 "       movq    %%mm4, %%mm7\n\t"
 216 "       pswapd  56(%%"REG_c"), %%mm1\n\t"
 217 "       pswapd  48(%%"REG_c"), %%mm5\n\t"
 218 "       pfadd   %%mm1, %%mm0\n\t"
 219 "       pfadd   %%mm5, %%mm4\n\t"
 220 "       movq    %%mm0, 32(%%"REG_d")\n\t"
 221 "       movq    %%mm4, 40(%%"REG_d")\n\t"
 222 "       pfsubr  %%mm1, %%mm3\n\t"
 223 "       pfsubr  %%mm5, %%mm7\n\t"
 224 "       pfmul   %%mm2, %%mm3\n\t"
 225 "       pfmul   %%mm6, %%mm7\n\t"
 226 "       pswapd  %%mm3, %%mm3\n\t"
 227 "       pswapd  %%mm7, %%mm7\n\t"
 228 "       movq    %%mm3, 56(%%"REG_d")\n\t"
 229 "       movq    %%mm7, 48(%%"REG_d")\n\t"
 230
 231 "       movq    64(%%"REG_c"), %%mm0\n\t"
 232 "       movq    72(%%"REG_c"), %%mm4\n\t"
 233 "       movq    %%mm0, %%mm3\n\t"
 234 "       movq    %%mm4, %%mm7\n\t"
 235 "       pswapd  88(%%"REG_c"), %%mm1\n\t"
 236 "       pswapd  80(%%"REG_c"), %%mm5\n\t"
 237 "       pfadd   %%mm1, %%mm0\n\t"
 238 "       pfadd   %%mm5, %%mm4\n\t"
 239 "       movq    %%mm0, 64(%%"REG_d")\n\t"
 240 "       movq    %%mm4, 72(%%"REG_d")\n\t"
 241 "       pfsub   %%mm1, %%mm3\n\t"
 242 "       pfsub   %%mm5, %%mm7\n\t"
 243 "       pfmul   %%mm2, %%mm3\n\t"
 244 "       pfmul   %%mm6, %%mm7\n\t"
 245 "       pswapd  %%mm3, %%mm3\n\t"
 246 "       pswapd  %%mm7, %%mm7\n\t"
 247 "       movq    %%mm3, 88(%%"REG_d")\n\t"
 248 "       movq    %%mm7, 80(%%"REG_d")\n\t"
 249
 250 "       movq    96(%%"REG_c"), %%mm0\n\t"
 251 "       movq    104(%%"REG_c"), %%mm4\n\t"
 252 "       movq    %%mm0, %%mm3\n\t"
 253 "       movq    %%mm4, %%mm7\n\t"
 254 "       pswapd  120(%%"REG_c"), %%mm1\n\t"
 255 "       pswapd  112(%%"REG_c"), %%mm5\n\t"
 256 "       pfadd   %%mm1, %%mm0\n\t"
 257 "       pfadd   %%mm5, %%mm4\n\t"
 258 "       movq    %%mm0, 96(%%"REG_d")\n\t"
 259 "       movq    %%mm4, 104(%%"REG_d")\n\t"
 260 "       pfsubr  %%mm1, %%mm3\n\t"
 261 "       pfsubr  %%mm5, %%mm7\n\t"
 262 "       pfmul   %%mm2, %%mm3\n\t"
 263 "       pfmul   %%mm6, %%mm7\n\t"
 264 "       pswapd  %%mm3, %%mm3\n\t"
 265 "       pswapd  %%mm7, %%mm7\n\t"
 266 "       movq    %%mm3, 120(%%"REG_d")\n\t"
 267 "       movq    %%mm7, 112(%%"REG_d")\n\t"
 268
 269 /* Phase 5 */
 270
 271 "       movq    112(%%"REG_b"), %%mm2\n\t"
 272
 273 "       movq    (%%"REG_d"), %%mm0\n\t"
 274 "       movq    16(%%"REG_d"), %%mm4\n\t"
 275 "       movq    %%mm0, %%mm3\n\t"
 276 "       movq    %%mm4, %%mm7\n\t"
 277 "       pswapd  8(%%"REG_d"), %%mm1\n\t"
 278 "       pswapd  24(%%"REG_d"), %%mm5\n\t"
 279 "       pfadd   %%mm1, %%mm0\n\t"
 280 "       pfadd   %%mm5, %%mm4\n\t"
 281 "       movq    %%mm0, (%%"REG_c")\n\t"
 282 "       movq    %%mm4, 16(%%"REG_c")\n\t"
 283 "       pfsub   %%mm1, %%mm3\n\t"
 284 "       pfsubr  %%mm5, %%mm7\n\t"
 285 "       pfmul   %%mm2, %%mm3\n\t"
 286 "       pfmul   %%mm2, %%mm7\n\t"
 287 "       pswapd  %%mm3, %%mm3\n\t"
 288 "       pswapd  %%mm7, %%mm7\n\t"
 289 "       movq    %%mm3, 8(%%"REG_c")\n\t"
 290 "       movq    %%mm7, 24(%%"REG_c")\n\t"
 291
 292 "       movq    32(%%"REG_d"), %%mm0\n\t"
 293 "       movq    48(%%"REG_d"), %%mm4\n\t"
 294 "       movq    %%mm0, %%mm3\n\t"
 295 "       movq    %%mm4, %%mm7\n\t"
 296 "       pswapd  40(%%"REG_d"), %%mm1\n\t"
 297 "       pswapd  56(%%"REG_d"), %%mm5\n\t"
 298 "       pfadd   %%mm1, %%mm0\n\t"
 299 "       pfadd   %%mm5, %%mm4\n\t"
 300 "       movq    %%mm0, 32(%%"REG_c")\n\t"
 301 "       movq    %%mm4, 48(%%"REG_c")\n\t"
 302 "       pfsub   %%mm1, %%mm3\n\t"
 303 "       pfsubr  %%mm5, %%mm7\n\t"
 304 "       pfmul   %%mm2, %%mm3\n\t"
 305 "       pfmul   %%mm2, %%mm7\n\t"
 306 "       pswapd  %%mm3, %%mm3\n\t"
 307 "       pswapd  %%mm7, %%mm7\n\t"
 308 "       movq    %%mm3, 40(%%"REG_c")\n\t"
 309 "       movq    %%mm7, 56(%%"REG_c")\n\t"
 310
 311 "       movq    64(%%"REG_d"), %%mm0\n\t"
 312 "       movq    80(%%"REG_d"), %%mm4\n\t"
 313 "       movq    %%mm0, %%mm3\n\t"
 314 "       movq    %%mm4, %%mm7\n\t"
 315 "       pswapd  72(%%"REG_d"), %%mm1\n\t"
 316 "       pswapd  88(%%"REG_d"), %%mm5\n\t"
 317 "       pfadd   %%mm1, %%mm0\n\t"
 318 "       pfadd   %%mm5, %%mm4\n\t"
 319 "       movq    %%mm0, 64(%%"REG_c")\n\t"
 320 "       movq    %%mm4, 80(%%"REG_c")\n\t"
 321 "       pfsub   %%mm1, %%mm3\n\t"
 322 "       pfsubr  %%mm5, %%mm7\n\t"
 323 "       pfmul   %%mm2, %%mm3\n\t"
 324 "       pfmul   %%mm2, %%mm7\n\t"
 325 "       pswapd  %%mm3, %%mm3\n\t"
 326 "       pswapd  %%mm7, %%mm7\n\t"
 327 "       movq    %%mm3, 72(%%"REG_c")\n\t"
 328 "       movq    %%mm7, 88(%%"REG_c")\n\t"
 329
 330 "       movq    96(%%"REG_d"), %%mm0\n\t"
 331 "       movq    112(%%"REG_d"), %%mm4\n\t"
 332 "       movq    %%mm0, %%mm3\n\t"
 333 "       movq    %%mm4, %%mm7\n\t"
 334 "       pswapd  104(%%"REG_d"), %%mm1\n\t"
 335 "       pswapd  120(%%"REG_d"), %%mm5\n\t"
 336 "       pfadd   %%mm1, %%mm0\n\t"
 337 "       pfadd   %%mm5, %%mm4\n\t"
 338 "       movq    %%mm0, 96(%%"REG_c")\n\t"
 339 "       movq    %%mm4, 112(%%"REG_c")\n\t"
 340 "       pfsub   %%mm1, %%mm3\n\t"
 341 "       pfsubr  %%mm5, %%mm7\n\t"
 342 "       pfmul   %%mm2, %%mm3\n\t"
 343 "       pfmul   %%mm2, %%mm7\n\t"
 344 "       pswapd  %%mm3, %%mm3\n\t"
 345 "       pswapd  %%mm7, %%mm7\n\t"
 346 "       movq    %%mm3, 104(%%"REG_c")\n\t"
 347 "       movq    %%mm7, 120(%%"REG_c")\n\t"
 348
 349
 350 /* Phase 6. This is the end of easy road. */
 351 /* Code below is coded in scalar mode. Should be optimized */
 352
 353 "       movd    "MANGLE(plus_1f)", %%mm6\n\t"
 354 "       punpckldq 120(%%"REG_b"), %%mm6\n\t"      /* mm6 = 1.0 | 120(%%"REG_b")*/
 355 "       movq    "MANGLE(x_plus_minus_3dnow)", %%mm7\n\t" /* mm7 = +1 | -1 */
 356
 357 "       movq    32(%%"REG_c"), %%mm0\n\t"
 358 "       movq    64(%%"REG_c"), %%mm2\n\t"
 359 "       movq    %%mm0, %%mm1\n\t"
 360 "       movq    %%mm2, %%mm3\n\t"
 361 "       pxor    %%mm7, %%mm1\n\t"
 362 "       pxor    %%mm7, %%mm3\n\t"
 363 "       pfacc   %%mm1, %%mm0\n\t"
 364 "       pfacc   %%mm3, %%mm2\n\t"
 365 "       pfmul   %%mm6, %%mm0\n\t"
 366 "       pfmul   %%mm6, %%mm2\n\t"
 367 "       movq    %%mm0, 32(%%"REG_d")\n\t"
 368 "       movq    %%mm2, 64(%%"REG_d")\n\t"
 369
 370 "       movd    44(%%"REG_c"), %%mm0\n\t"
 371 "       movd    40(%%"REG_c"), %%mm2\n\t"
 372 "       movd    120(%%"REG_b"), %%mm3\n\t"
 373 "       punpckldq 76(%%"REG_c"), %%mm0\n\t"
 374 "       punpckldq 72(%%"REG_c"), %%mm2\n\t"
 375 "       punpckldq %%mm3, %%mm3\n\t"
 376 "       movq    %%mm0, %%mm4\n\t"
 377 "       movq    %%mm2, %%mm5\n\t"
 378 "       pfsub   %%mm2, %%mm0\n\t"
 379 "       pfmul   %%mm3, %%mm0\n\t"
 380 "       movq    %%mm0, %%mm1\n\t"
 381 "       pfadd   %%mm5, %%mm0\n\t"
 382 "       pfadd   %%mm4, %%mm0\n\t"
 383 "       movq    %%mm0, %%mm2\n\t"
 384 "       punpckldq %%mm1, %%mm0\n\t"
 385 "       punpckhdq %%mm1, %%mm2\n\t"
 386 "       movq    %%mm0, 40(%%"REG_d")\n\t"
 387 "       movq    %%mm2, 72(%%"REG_d")\n\t"
 388
 389 "       movd   48(%%"REG_c"), %%mm3\n\t"
 390 "       movd   60(%%"REG_c"), %%mm2\n\t"
 391 "       pfsub  52(%%"REG_c"), %%mm3\n\t"
 392 "       pfsub  56(%%"REG_c"), %%mm2\n\t"
 393 "       pfmul 120(%%"REG_b"), %%mm3\n\t"
 394 "       pfmul 120(%%"REG_b"), %%mm2\n\t"
 395 "       movq    %%mm2, %%mm1\n\t"
 396
 397 "       pfadd  56(%%"REG_c"), %%mm1\n\t"
 398 "       pfadd  60(%%"REG_c"), %%mm1\n\t"
 399 "       movq    %%mm1, %%mm0\n\t"
 400
 401 "       pfadd  48(%%"REG_c"), %%mm0\n\t"
 402 "       pfadd  52(%%"REG_c"), %%mm0\n\t"
 403 "       pfadd   %%mm3, %%mm1\n\t"
 404 "       punpckldq %%mm2, %%mm1\n\t"
 405 "       pfadd   %%mm3, %%mm2\n\t"
 406 "       punpckldq %%mm2, %%mm0\n\t"
 407 "       movq    %%mm1, 56(%%"REG_d")\n\t"
 408 "       movq    %%mm0, 48(%%"REG_d")\n\t"
 409
 410 /*---*/
 411
 412 "       movd   92(%%"REG_c"), %%mm1\n\t"
 413 "       pfsub  88(%%"REG_c"), %%mm1\n\t"
 414 "       pfmul 120(%%"REG_b"), %%mm1\n\t"
 415 "       movd   %%mm1, 92(%%"REG_d")\n\t"
 416 "       pfadd  92(%%"REG_c"), %%mm1\n\t"
 417 "       pfadd  88(%%"REG_c"), %%mm1\n\t"
 418 "       movq   %%mm1, %%mm0\n\t"
 419
 420 "       pfadd  80(%%"REG_c"), %%mm0\n\t"
 421 "       pfadd  84(%%"REG_c"), %%mm0\n\t"
 422 "       movd   %%mm0, 80(%%"REG_d")\n\t"
 423
 424 "       movd   80(%%"REG_c"), %%mm0\n\t"
 425 "       pfsub  84(%%"REG_c"), %%mm0\n\t"
 426 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 427 "       pfadd  %%mm0, %%mm1\n\t"
 428 "       pfadd  92(%%"REG_d"), %%mm0\n\t"
 429 "       punpckldq %%mm1, %%mm0\n\t"
 430 "       movq   %%mm0, 84(%%"REG_d")\n\t"
 431
 432 "       movq    96(%%"REG_c"), %%mm0\n\t"
 433 "       movq    %%mm0, %%mm1\n\t"
 434 "       pxor    %%mm7, %%mm1\n\t"
 435 "       pfacc   %%mm1, %%mm0\n\t"
 436 "       pfmul   %%mm6, %%mm0\n\t"
 437 "       movq    %%mm0, 96(%%"REG_d")\n\t"
 438
 439 "       movd  108(%%"REG_c"), %%mm0\n\t"
 440 "       pfsub 104(%%"REG_c"), %%mm0\n\t"
 441 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 442 "       movd  %%mm0, 108(%%"REG_d")\n\t"
 443 "       pfadd 104(%%"REG_c"), %%mm0\n\t"
 444 "       pfadd 108(%%"REG_c"), %%mm0\n\t"
 445 "       movd  %%mm0, 104(%%"REG_d")\n\t"
 446
 447 "       movd  124(%%"REG_c"), %%mm1\n\t"
 448 "       pfsub 120(%%"REG_c"), %%mm1\n\t"
 449 "       pfmul 120(%%"REG_b"), %%mm1\n\t"
 450 "       movd  %%mm1, 124(%%"REG_d")\n\t"
 451 "       pfadd 120(%%"REG_c"), %%mm1\n\t"
 452 "       pfadd 124(%%"REG_c"), %%mm1\n\t"
 453 "       movq  %%mm1, %%mm0\n\t"
 454
 455 "       pfadd 112(%%"REG_c"), %%mm0\n\t"
 456 "       pfadd 116(%%"REG_c"), %%mm0\n\t"
 457 "       movd  %%mm0, 112(%%"REG_d")\n\t"
 458
 459 "       movd  112(%%"REG_c"), %%mm0\n\t"
 460 "       pfsub 116(%%"REG_c"), %%mm0\n\t"
 461 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 462 "       pfadd %%mm0,%%mm1\n\t"
 463 "       pfadd 124(%%"REG_d"), %%mm0\n\t"
 464 "       punpckldq %%mm1, %%mm0\n\t"
 465 "       movq  %%mm0, 116(%%"REG_d")\n\t"
 466
 467 // this code is broken, there is nothing modifying the z flag above.
 468 #if 0
 469 "       jnz .L01\n\t"
 470
 471 /* Phase 7*/
 472 /* Code below is coded in scalar mode. Should be optimized */
 473
 474 "       movd      (%%"REG_c"), %%mm0\n\t"
 475 "       pfadd    4(%%"REG_c"), %%mm0\n\t"
 476 "       movd     %%mm0, 1024(%%"REG_S")\n\t"
 477
 478 "       movd      (%%"REG_c"), %%mm0\n\t"
 479 "       pfsub    4(%%"REG_c"), %%mm0\n\t"
 480 "       pfmul  120(%%"REG_b"), %%mm0\n\t"
 481 "       movd      %%mm0, (%%"REG_S")\n\t"
 482 "       movd      %%mm0, (%%"REG_D")\n\t"
 483
 484 "       movd   12(%%"REG_c"), %%mm0\n\t"
 485 "       pfsub   8(%%"REG_c"), %%mm0\n\t"
 486 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 487 "       movd    %%mm0, 512(%%"REG_D")\n\t"
 488 "       pfadd   12(%%"REG_c"), %%mm0\n\t"
 489 "       pfadd   8(%%"REG_c"), %%mm0\n\t"
 490 "       movd    %%mm0, 512(%%"REG_S")\n\t"
 491
 492 "       movd   16(%%"REG_c"), %%mm0\n\t"
 493 "       pfsub  20(%%"REG_c"), %%mm0\n\t"
 494 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 495 "       movq    %%mm0, %%mm3\n\t"
 496
 497 "       movd   28(%%"REG_c"), %%mm0\n\t"
 498 "       pfsub  24(%%"REG_c"), %%mm0\n\t"
 499 "       pfmul 120(%%"REG_b"), %%mm0\n\t"
 500 "       movd    %%mm0, 768(%%"REG_D")\n\t"
 501 "       movq    %%mm0, %%mm2\n\t"
 502
 503 "       pfadd  24(%%"REG_c"), %%mm0\n\t"
 504 "       pfadd  28(%%"REG_c"), %%mm0\n\t"
 505 "       movq    %%mm0, %%mm1\n\t"
 506
 507 "       pfadd  16(%%"REG_c"), %%mm0\n\t"
 508 "       pfadd  20(%%"REG_c"), %%mm0\n\t"
 509 "       movd   %%mm0, 768(%%"REG_S")\n\t"
 510 "       pfadd  %%mm3, %%mm1\n\t"
 511 "       movd   %%mm1, 256(%%"REG_S")\n\t"
 512 "       pfadd  %%mm3, %%mm2\n\t"
 513 "       movd   %%mm2, 256(%%"REG_D")\n\t"
 514
 515 /* Phase 8*/
 516
 517 "       movq   32(%%"REG_d"), %%mm0\n\t"
 518 "       movq   48(%%"REG_d"), %%mm1\n\t"
 519 "       pfadd  48(%%"REG_d"), %%mm0\n\t"
 520 "       pfadd  40(%%"REG_d"), %%mm1\n\t"
 521 "       movd   %%mm0, 896(%%"REG_S")\n\t"
 522 "       movd   %%mm1, 640(%%"REG_S")\n\t"
 523 "       psrlq  $32, %%mm0\n\t"
 524 "       psrlq  $32, %%mm1\n\t"
 525 "       movd   %%mm0, 128(%%"REG_D")\n\t"
 526 "       movd   %%mm1, 384(%%"REG_D")\n\t"
 527
 528 "       movd   40(%%"REG_d"), %%mm0\n\t"
 529 "       pfadd  56(%%"REG_d"), %%mm0\n\t"
 530 "       movd   %%mm0, 384(%%"REG_S")\n\t"
 531
 532 "       movd   56(%%"REG_d"), %%mm0\n\t"
 533 "       pfadd  36(%%"REG_d"), %%mm0\n\t"
 534 "       movd   %%mm0, 128(%%"REG_S")\n\t"
 535
 536 "       movd   60(%%"REG_d"), %%mm0\n\t"
 537 "       movd   %%mm0, 896(%%"REG_D")\n\t"
 538 "       pfadd  44(%%"REG_d"), %%mm0\n\t"
 539 "       movd   %%mm0, 640(%%"REG_D")\n\t"
 540
 541 "       movq   96(%%"REG_d"), %%mm0\n\t"
 542 "       movq   112(%%"REG_d"), %%mm2\n\t"
 543 "       movq   104(%%"REG_d"), %%mm4\n\t"
 544 "       pfadd  112(%%"REG_d"), %%mm0\n\t"
 545 "       pfadd  104(%%"REG_d"), %%mm2\n\t"
 546 "       pfadd  120(%%"REG_d"), %%mm4\n\t"
 547 "       movq   %%mm0, %%mm1\n\t"
 548 "       movq   %%mm2, %%mm3\n\t"
 549 "       movq   %%mm4, %%mm5\n\t"
 550 "       pfadd  64(%%"REG_d"), %%mm0\n\t"
 551 "       pfadd  80(%%"REG_d"), %%mm2\n\t"
 552 "       pfadd  72(%%"REG_d"), %%mm4\n\t"
 553 "       movd   %%mm0, 960(%%"REG_S")\n\t"
 554 "       movd   %%mm2, 704(%%"REG_S")\n\t"
 555 "       movd   %%mm4, 448(%%"REG_S")\n\t"
 556 "       psrlq  $32, %%mm0\n\t"
 557 "       psrlq  $32, %%mm2\n\t"
 558 "       psrlq  $32, %%mm4\n\t"
 559 "       movd   %%mm0, 64(%%"REG_D")\n\t"
 560 "       movd   %%mm2, 320(%%"REG_D")\n\t"
 561 "       movd   %%mm4, 576(%%"REG_D")\n\t"
 562 "       pfadd  80(%%"REG_d"), %%mm1\n\t"
 563 "       pfadd  72(%%"REG_d"), %%mm3\n\t"
 564 "       pfadd  88(%%"REG_d"), %%mm5\n\t"
 565 "       movd   %%mm1, 832(%%"REG_S")\n\t"
 566 "       movd   %%mm3, 576(%%"REG_S")\n\t"
 567 "       movd   %%mm5, 320(%%"REG_S")\n\t"
 568 "       psrlq  $32, %%mm1\n\t"
 569 "       psrlq  $32, %%mm3\n\t"
 570 "       psrlq  $32, %%mm5\n\t"
 571 "       movd   %%mm1, 192(%%"REG_D")\n\t"
 572 "       movd   %%mm3, 448(%%"REG_D")\n\t"
 573 "       movd   %%mm5, 704(%%"REG_D")\n\t"
 574
 575 "       movd   120(%%"REG_d"), %%mm0\n\t"
 576 "       pfadd  100(%%"REG_d"), %%mm0\n\t"
 577 "       movq   %%mm0, %%mm1\n\t"
 578 "       pfadd  88(%%"REG_d"), %%mm0\n\t"
 579 "       movd   %%mm0, 192(%%"REG_S")\n\t"
 580 "       pfadd  68(%%"REG_d"), %%mm1\n\t"
 581 "       movd   %%mm1, 64(%%"REG_S")\n\t"
 582
 583 "       movd  124(%%"REG_d"), %%mm0\n\t"
 584 "       movd  %%mm0, 960(%%"REG_D")\n\t"
 585 "       pfadd  92(%%"REG_d"), %%mm0\n\t"
 586 "       movd  %%mm0, 832(%%"REG_D")\n\t"
 587
 588 "       jmp     .L_bye\n\t"
 589 ".L01:  \n\t"
 590 #endif
 591 /* Phase 9*/
 592
 593 "       movq    (%%"REG_c"), %%mm0\n\t"
 594 "       movq    %%mm0, %%mm1\n\t"
 595 "       pxor    %%mm7, %%mm1\n\t"
 596 "       pfacc   %%mm1, %%mm0\n\t"
 597 "       pfmul   %%mm6, %%mm0\n\t"
 598 "       pf2iw   %%mm0, %%mm0\n\t"
 599 "       movd    %%mm0, %%"REG_a"\n\t"
 600 "       movw    %%ax, 512(%%"REG_S")\n\t"
 601 "       psrlq   $32, %%mm0\n\t"
 602 "       movd    %%mm0, %%"REG_a"\n\t"
 603 "       movw    %%ax, (%%"REG_S")\n\t"
 604
 605 "       movd    12(%%"REG_c"), %%mm0\n\t"
 606 "       pfsub    8(%%"REG_c"), %%mm0\n\t"
 607 "       pfmul  120(%%"REG_b"), %%mm0\n\t"
 608 "       pf2iw    %%mm0, %%mm7\n\t"
 609 "       movd     %%mm7, %%"REG_a"\n\t"
 610 "       movw     %%ax, 256(%%"REG_D")\n\t"
 611 "       pfadd   12(%%"REG_c"), %%mm0\n\t"
 612 "       pfadd    8(%%"REG_c"), %%mm0\n\t"
 613 "       pf2iw    %%mm0, %%mm0\n\t"
 614 "       movd     %%mm0, %%"REG_a"\n\t"
 615 "       movw     %%ax, 256(%%"REG_S")\n\t"
 616
 617 "       movd   16(%%"REG_c"), %%mm3\n\t"
 618 "       pfsub  20(%%"REG_c"), %%mm3\n\t"
 619 "       pfmul  120(%%"REG_b"), %%mm3\n\t"
 620 "       movq   %%mm3, %%mm2\n\t"
 621
 622 "       movd   28(%%"REG_c"), %%mm2\n\t"
 623 "       pfsub  24(%%"REG_c"), %%mm2\n\t"
 624 "       pfmul 120(%%"REG_b"), %%mm2\n\t"
 625 "       movq   %%mm2, %%mm1\n\t"
 626
 627 "       pf2iw  %%mm2, %%mm7\n\t"
 628 "       movd   %%mm7, %%"REG_a"\n\t"
 629 "       movw   %%ax, 384(%%"REG_D")\n\t"
 630
 631 "       pfadd  24(%%"REG_c"), %%mm1\n\t"
 632 "       pfadd  28(%%"REG_c"), %%mm1\n\t"
 633 "       movq   %%mm1, %%mm0\n\t"
 634
 635 "       pfadd  16(%%"REG_c"), %%mm0\n\t"
 636 "       pfadd  20(%%"REG_c"), %%mm0\n\t"
 637 "       pf2iw  %%mm0, %%mm0\n\t"
 638 "       movd   %%mm0, %%"REG_a"\n\t"
 639 "       movw   %%ax, 384(%%"REG_S")\n\t"
 640 "       pfadd  %%mm3, %%mm1\n\t"
 641 "       pf2iw  %%mm1, %%mm1\n\t"
 642 "       movd   %%mm1, %%"REG_a"\n\t"
 643 "       movw   %%ax, 128(%%"REG_S")\n\t"
 644 "       pfadd  %%mm3, %%mm2\n\t"
 645 "       pf2iw  %%mm2, %%mm2\n\t"
 646 "       movd   %%mm2, %%"REG_a"\n\t"
 647 "       movw   %%ax, 128(%%"REG_D")\n\t"
 648
 649 /* Phase 10*/
 650
 651 "       movq    32(%%"REG_d"), %%mm0\n\t"
 652 "       movq    48(%%"REG_d"), %%mm1\n\t"
 653 "       pfadd   48(%%"REG_d"), %%mm0\n\t"
 654 "       pfadd   40(%%"REG_d"), %%mm1\n\t"
 655 "       pf2iw   %%mm0, %%mm0\n\t"
 656 "       pf2iw   %%mm1, %%mm1\n\t"
 657 "       movd    %%mm0, %%"REG_a"\n\t"
 658 "       movd    %%mm1, %%"REG_c"\n\t"
 659 "       movw    %%ax, 448(%%"REG_S")\n\t"
 660 "       movw    %%cx, 320(%%"REG_S")\n\t"
 661 "       psrlq   $32, %%mm0\n\t"
 662 "       psrlq   $32, %%mm1\n\t"
 663 "       movd    %%mm0, %%"REG_a"\n\t"
 664 "       movd    %%mm1, %%"REG_c"\n\t"
 665 "       movw    %%ax, 64(%%"REG_D")\n\t"
 666 "       movw    %%cx, 192(%%"REG_D")\n\t"
 667
 668 "       movd   40(%%"REG_d"), %%mm3\n\t"
 669 "       movd   56(%%"REG_d"), %%mm4\n\t"
 670 "       movd   60(%%"REG_d"), %%mm0\n\t"
 671 "       movd   44(%%"REG_d"), %%mm2\n\t"
 672 "       movd  120(%%"REG_d"), %%mm5\n\t"
 673 "       punpckldq %%mm4, %%mm3\n\t"
 674 "       punpckldq 124(%%"REG_d"), %%mm0\n\t"
 675 "       pfadd 100(%%"REG_d"), %%mm5\n\t"
 676 "       punpckldq 36(%%"REG_d"), %%mm4\n\t"
 677 "       punpckldq 92(%%"REG_d"), %%mm2\n\t"
 678 "       movq  %%mm5, %%mm6\n\t"
 679 "       pfadd  %%mm4, %%mm3\n\t"
 680 "       pf2iw  %%mm0, %%mm1\n\t"
 681 "       pf2iw  %%mm3, %%mm3\n\t"
 682 "       pfadd  88(%%"REG_d"), %%mm5\n\t"
 683 "       movd   %%mm1, %%"REG_a"\n\t"
 684 "       movd   %%mm3, %%"REG_c"\n\t"
 685 "       movw   %%ax, 448(%%"REG_D")\n\t"
 686 "       movw   %%cx, 192(%%"REG_S")\n\t"
 687 "       pf2iw  %%mm5, %%mm5\n\t"
 688 "       psrlq  $32, %%mm1\n\t"
 689 "       psrlq  $32, %%mm3\n\t"
 690 "       movd   %%mm5, %%"REG_b"\n\t"
 691 "       movd   %%mm1, %%"REG_a"\n\t"
 692 "       movd   %%mm3, %%"REG_c"\n\t"
 693 "       movw   %%bx, 96(%%"REG_S")\n\t"
 694 "       movw   %%ax, 480(%%"REG_D")\n\t"
 695 "       movw   %%cx, 64(%%"REG_S")\n\t"
 696 "       pfadd  %%mm2, %%mm0\n\t"
 697 "       pf2iw  %%mm0, %%mm0\n\t"
 698 "       movd   %%mm0, %%"REG_a"\n\t"
 699 "       pfadd  68(%%"REG_d"), %%mm6\n\t"
 700 "       movw   %%ax, 320(%%"REG_D")\n\t"
 701 "       psrlq  $32, %%mm0\n\t"
 702 "       pf2iw  %%mm6, %%mm6\n\t"
 703 "       movd   %%mm0, %%"REG_a"\n\t"
 704 "       movd   %%mm6, %%"REG_b"\n\t"
 705 "       movw   %%ax, 416(%%"REG_D")\n\t"
 706 "       movw   %%bx, 32(%%"REG_S")\n\t"
 707
 708 "       movq   96(%%"REG_d"), %%mm0\n\t"
 709 "       movq  112(%%"REG_d"), %%mm2\n\t"
 710 "       movq  104(%%"REG_d"), %%mm4\n\t"
 711 "       pfadd %%mm2, %%mm0\n\t"
 712 "       pfadd %%mm4, %%mm2\n\t"
 713 "       pfadd 120(%%"REG_d"), %%mm4\n\t"
 714 "       movq  %%mm0, %%mm1\n\t"
 715 "       movq  %%mm2, %%mm3\n\t"
 716 "       movq  %%mm4, %%mm5\n\t"
 717 "       pfadd  64(%%"REG_d"), %%mm0\n\t"
 718 "       pfadd  80(%%"REG_d"), %%mm2\n\t"
 719 "       pfadd  72(%%"REG_d"), %%mm4\n\t"
 720 "       pf2iw  %%mm0, %%mm0\n\t"
 721 "       pf2iw  %%mm2, %%mm2\n\t"
 722 "       pf2iw  %%mm4, %%mm4\n\t"
 723 "       movd   %%mm0, %%"REG_a"\n\t"
 724 "       movd   %%mm2, %%"REG_c"\n\t"
 725 "       movd   %%mm4, %%"REG_b"\n\t"
 726 "       movw   %%ax, 480(%%"REG_S")\n\t"
 727 "       movw   %%cx, 352(%%"REG_S")\n\t"
 728 "       movw   %%bx, 224(%%"REG_S")\n\t"
 729 "       psrlq  $32, %%mm0\n\t"
 730 "       psrlq  $32, %%mm2\n\t"
 731 "       psrlq  $32, %%mm4\n\t"
 732 "       movd   %%mm0, %%"REG_a"\n\t"
 733 "       movd   %%mm2, %%"REG_c"\n\t"
 734 "       movd   %%mm4, %%"REG_b"\n\t"
 735 "       movw   %%ax, 32(%%"REG_D")\n\t"
 736 "       movw   %%cx, 160(%%"REG_D")\n\t"
 737 "       movw   %%bx, 288(%%"REG_D")\n\t"
 738 "       pfadd  80(%%"REG_d"), %%mm1\n\t"
 739 "       pfadd  72(%%"REG_d"), %%mm3\n\t"
 740 "       pfadd  88(%%"REG_d"), %%mm5\n\t"
 741 "       pf2iw  %%mm1, %%mm1\n\t"
 742 "       pf2iw  %%mm3, %%mm3\n\t"
 743 "       pf2iw  %%mm5, %%mm5\n\t"
 744 "       movd   %%mm1, %%"REG_a"\n\t"
 745 "       movd   %%mm3, %%"REG_c"\n\t"
 746 "       movd   %%mm5, %%"REG_b"\n\t"
 747 "       movw   %%ax, 416(%%"REG_S")\n\t"
 748 "       movw   %%cx, 288(%%"REG_S")\n\t"
 749 "       movw   %%bx, 160(%%"REG_S")\n\t"
 750 "       psrlq  $32, %%mm1\n\t"
 751 "       psrlq  $32, %%mm3\n\t"
 752 "       psrlq  $32, %%mm5\n\t"
 753 "       movd   %%mm1, %%"REG_a"\n\t"
 754 "       movd   %%mm3, %%"REG_c"\n\t"
 755 "       movd   %%mm5, %%"REG_b"\n\t"
 756 "       movw   %%ax, 96(%%"REG_D")\n\t"
 757 "       movw   %%cx, 224(%%"REG_D")\n\t"
 758 "       movw   %%bx, 352(%%"REG_D")\n\t"
 759
 760 "       movsw\n\t"
 761
 762 ".L_bye:\n\t"
 763 "       femms\n\t"
 764         :
 765         :"m"(a),"m"(b),"m"(c),"m"(tmp[0])
 766         :"memory","%eax","%ebx","%ecx","%edx","%esi","%edi");
 767 }