mingw/lib/gcc/mingw32/4.4.0/include/emmintrin.h

   1 /* Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009
   2    Free Software Foundation, Inc.
   3
   4    This file is part of GCC.
   5
   6    GCC is free software; you can redistribute it and/or modify
   7    it under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3, or (at your option)
   9    any later version.
  10
  11    GCC is distributed in the hope that it will be useful,
  12    but WITHOUT ANY WARRANTY; without even the implied warranty of
  13    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14    GNU General Public License for more details.
  15
  16    Under Section 7 of GPL version 3, you are granted additional
  17    permissions described in the GCC Runtime Library Exception, version
  18    3.1, as published by the Free Software Foundation.
  19
  20    You should have received a copy of the GNU General Public License and
  21    a copy of the GCC Runtime Library Exception along with this program;
  22    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  23    <http://www.gnu.org/licenses/>.  */
  24
  25 /* Implemented from the specification included in the Intel C++ Compiler
  26    User Guide and Reference, version 9.0.  */
  27
  28 #ifndef _EMMINTRIN_H_INCLUDED
  29 #define _EMMINTRIN_H_INCLUDED
  30
  31 #ifndef __SSE2__
  32 # error "SSE2 instruction set not enabled"
  33 #else
  34
  35 /* We need definitions from the SSE header files*/
  36 #include <xmmintrin.h>
  37
  38 /* SSE2 */
  39 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  40 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  41 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  42 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  43 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  44
  45 /* The Intel API is flexible enough that we must allow aliasing with other
  46    vector types, and their scalar components.  */
  47 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  48 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  49
  50 /* Create a selector for use with the SHUFPD instruction.  */
  51 #define _MM_SHUFFLE2(fp1,fp0) \
  52  (((fp1) << 1) | (fp0))
  53
  54 /* Create a vector with element 0 as F and the rest zero.  */
  55 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  56 _mm_set_sd (double __F)
  57 {
  58   return __extension__ (__m128d){ __F, 0.0 };
  59 }
  60
  61 /* Create a vector with both elements equal to F.  */
  62 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  63 _mm_set1_pd (double __F)
  64 {
  65   return __extension__ (__m128d){ __F, __F };
  66 }
  67
  68 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  69 _mm_set_pd1 (double __F)
  70 {
  71   return _mm_set1_pd (__F);
  72 }
  73
  74 /* Create a vector with the lower value X and upper value W.  */
  75 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  76 _mm_set_pd (double __W, double __X)
  77 {
  78   return __extension__ (__m128d){ __X, __W };
  79 }
  80
  81 /* Create a vector with the lower value W and upper value X.  */
  82 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  83 _mm_setr_pd (double __W, double __X)
  84 {
  85   return __extension__ (__m128d){ __W, __X };
  86 }
  87
  88 /* Create a vector of zeros.  */
  89 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  90 _mm_setzero_pd (void)
  91 {
  92   return __extension__ (__m128d){ 0.0, 0.0 };
  93 }
  94
  95 /* Sets the low DPFP value of A from the low value of B.  */
  96 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  97 _mm_move_sd (__m128d __A, __m128d __B)
  98 {
  99   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 100 }
 101
 102 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 103 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 104 _mm_load_pd (double const *__P)
 105 {
 106   return *(__m128d *)__P;
 107 }
 108
 109 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 110 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 111 _mm_loadu_pd (double const *__P)
 112 {
 113   return __builtin_ia32_loadupd (__P);
 114 }
 115
 116 /* Create a vector with all two elements equal to *P.  */
 117 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 118 _mm_load1_pd (double const *__P)
 119 {
 120   return _mm_set1_pd (*__P);
 121 }
 122
 123 /* Create a vector with element 0 as *P and the rest zero.  */
 124 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 125 _mm_load_sd (double const *__P)
 126 {
 127   return _mm_set_sd (*__P);
 128 }
 129
 130 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 131 _mm_load_pd1 (double const *__P)
 132 {
 133   return _mm_load1_pd (__P);
 134 }
 135
 136 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 137 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 138 _mm_loadr_pd (double const *__P)
 139 {
 140   __m128d __tmp = _mm_load_pd (__P);
 141   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 142 }
 143
 144 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 145 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 146 _mm_store_pd (double *__P, __m128d __A)
 147 {
 148   *(__m128d *)__P = __A;
 149 }
 150
 151 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 152 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 153 _mm_storeu_pd (double *__P, __m128d __A)
 154 {
 155   __builtin_ia32_storeupd (__P, __A);
 156 }
 157
 158 /* Stores the lower DPFP value.  */
 159 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 160 _mm_store_sd (double *__P, __m128d __A)
 161 {
 162   *__P = __builtin_ia32_vec_ext_v2df (__A, 0);
 163 }
 164
 165 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 166 _mm_cvtsd_f64 (__m128d __A)
 167 {
 168   return __builtin_ia32_vec_ext_v2df (__A, 0);
 169 }
 170
 171 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 172 _mm_storel_pd (double *__P, __m128d __A)
 173 {
 174   _mm_store_sd (__P, __A);
 175 }
 176
 177 /* Stores the upper DPFP value.  */
 178 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 179 _mm_storeh_pd (double *__P, __m128d __A)
 180 {
 181   *__P = __builtin_ia32_vec_ext_v2df (__A, 1);
 182 }
 183
 184 /* Store the lower DPFP value across two words.
 185    The address must be 16-byte aligned.  */
 186 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 187 _mm_store1_pd (double *__P, __m128d __A)
 188 {
 189   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 190 }
 191
 192 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 193 _mm_store_pd1 (double *__P, __m128d __A)
 194 {
 195   _mm_store1_pd (__P, __A);
 196 }
 197
 198 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 199 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 200 _mm_storer_pd (double *__P, __m128d __A)
 201 {
 202   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 203 }
 204
 205 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 206 _mm_cvtsi128_si32 (__m128i __A)
 207 {
 208   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 209 }
 210
 211 #ifdef __x86_64__
 212 /* Intel intrinsic.  */
 213 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 214 _mm_cvtsi128_si64 (__m128i __A)
 215 {
 216   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 217 }
 218
 219 /* Microsoft intrinsic.  */
 220 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 221 _mm_cvtsi128_si64x (__m128i __A)
 222 {
 223   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 224 }
 225 #endif
 226
 227 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 228 _mm_add_pd (__m128d __A, __m128d __B)
 229 {
 230   return (__m128d)__builtin_ia32_addpd ((__v2df)__A, (__v2df)__B);
 231 }
 232
 233 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 234 _mm_add_sd (__m128d __A, __m128d __B)
 235 {
 236   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 237 }
 238
 239 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 240 _mm_sub_pd (__m128d __A, __m128d __B)
 241 {
 242   return (__m128d)__builtin_ia32_subpd ((__v2df)__A, (__v2df)__B);
 243 }
 244
 245 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 246 _mm_sub_sd (__m128d __A, __m128d __B)
 247 {
 248   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 249 }
 250
 251 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 252 _mm_mul_pd (__m128d __A, __m128d __B)
 253 {
 254   return (__m128d)__builtin_ia32_mulpd ((__v2df)__A, (__v2df)__B);
 255 }
 256
 257 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 258 _mm_mul_sd (__m128d __A, __m128d __B)
 259 {
 260   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 261 }
 262
 263 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 264 _mm_div_pd (__m128d __A, __m128d __B)
 265 {
 266   return (__m128d)__builtin_ia32_divpd ((__v2df)__A, (__v2df)__B);
 267 }
 268
 269 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 270 _mm_div_sd (__m128d __A, __m128d __B)
 271 {
 272   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 273 }
 274
 275 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 276 _mm_sqrt_pd (__m128d __A)
 277 {
 278   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 279 }
 280
 281 /* Return pair {sqrt (A[0), B[1]}.  */
 282 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 283 _mm_sqrt_sd (__m128d __A, __m128d __B)
 284 {
 285   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 286   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 287 }
 288
 289 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 290 _mm_min_pd (__m128d __A, __m128d __B)
 291 {
 292   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 293 }
 294
 295 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 296 _mm_min_sd (__m128d __A, __m128d __B)
 297 {
 298   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 299 }
 300
 301 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 302 _mm_max_pd (__m128d __A, __m128d __B)
 303 {
 304   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 305 }
 306
 307 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 308 _mm_max_sd (__m128d __A, __m128d __B)
 309 {
 310   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 311 }
 312
 313 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 314 _mm_and_pd (__m128d __A, __m128d __B)
 315 {
 316   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 317 }
 318
 319 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 320 _mm_andnot_pd (__m128d __A, __m128d __B)
 321 {
 322   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 323 }
 324
 325 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 326 _mm_or_pd (__m128d __A, __m128d __B)
 327 {
 328   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 329 }
 330
 331 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 332 _mm_xor_pd (__m128d __A, __m128d __B)
 333 {
 334   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 335 }
 336
 337 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 338 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 339 {
 340   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 341 }
 342
 343 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 344 _mm_cmplt_pd (__m128d __A, __m128d __B)
 345 {
 346   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 347 }
 348
 349 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 350 _mm_cmple_pd (__m128d __A, __m128d __B)
 351 {
 352   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 353 }
 354
 355 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 356 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 357 {
 358   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 359 }
 360
 361 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 362 _mm_cmpge_pd (__m128d __A, __m128d __B)
 363 {
 364   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 365 }
 366
 367 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 368 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 369 {
 370   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 371 }
 372
 373 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 374 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 375 {
 376   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 377 }
 378
 379 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 380 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 381 {
 382   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 383 }
 384
 385 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 386 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 387 {
 388   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 389 }
 390
 391 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 392 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 393 {
 394   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 395 }
 396
 397 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 398 _mm_cmpord_pd (__m128d __A, __m128d __B)
 399 {
 400   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 401 }
 402
 403 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 404 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 405 {
 406   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 407 }
 408
 409 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 410 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 411 {
 412   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 413 }
 414
 415 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 416 _mm_cmplt_sd (__m128d __A, __m128d __B)
 417 {
 418   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 419 }
 420
 421 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 422 _mm_cmple_sd (__m128d __A, __m128d __B)
 423 {
 424   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 425 }
 426
 427 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 428 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 429 {
 430   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 431                                          (__v2df)
 432                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 433                                                                  (__v2df)
 434                                                                  __A));
 435 }
 436
 437 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 438 _mm_cmpge_sd (__m128d __A, __m128d __B)
 439 {
 440   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 441                                          (__v2df)
 442                                          __builtin_ia32_cmplesd ((__v2df) __B,
 443                                                                  (__v2df)
 444                                                                  __A));
 445 }
 446
 447 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 448 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 449 {
 450   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 451 }
 452
 453 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 454 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 455 {
 456   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 457 }
 458
 459 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 460 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 461 {
 462   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 463 }
 464
 465 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 466 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 467 {
 468   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 469                                          (__v2df)
 470                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 471                                                                   (__v2df)
 472                                                                   __A));
 473 }
 474
 475 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 476 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 477 {
 478   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 479                                          (__v2df)
 480                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 481                                                                   (__v2df)
 482                                                                   __A));
 483 }
 484
 485 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 486 _mm_cmpord_sd (__m128d __A, __m128d __B)
 487 {
 488   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 489 }
 490
 491 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 492 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 493 {
 494   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 495 }
 496
 497 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 498 _mm_comieq_sd (__m128d __A, __m128d __B)
 499 {
 500   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 501 }
 502
 503 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 504 _mm_comilt_sd (__m128d __A, __m128d __B)
 505 {
 506   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 507 }
 508
 509 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 510 _mm_comile_sd (__m128d __A, __m128d __B)
 511 {
 512   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 513 }
 514
 515 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 516 _mm_comigt_sd (__m128d __A, __m128d __B)
 517 {
 518   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 519 }
 520
 521 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 522 _mm_comige_sd (__m128d __A, __m128d __B)
 523 {
 524   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 525 }
 526
 527 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 528 _mm_comineq_sd (__m128d __A, __m128d __B)
 529 {
 530   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 531 }
 532
 533 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 534 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 535 {
 536   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 537 }
 538
 539 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 540 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 541 {
 542   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 543 }
 544
 545 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 546 _mm_ucomile_sd (__m128d __A, __m128d __B)
 547 {
 548   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 549 }
 550
 551 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 552 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 553 {
 554   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 555 }
 556
 557 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 558 _mm_ucomige_sd (__m128d __A, __m128d __B)
 559 {
 560   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 561 }
 562
 563 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 564 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 565 {
 566   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 567 }
 568
 569 /* Create a vector of Qi, where i is the element number.  */
 570
 571 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 572 _mm_set_epi64x (long long __q1, long long __q0)
 573 {
 574   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 575 }
 576
 577 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 578 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 579 {
 580   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 581 }
 582
 583 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 584 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 585 {
 586   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 587 }
 588
 589 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 590 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 591                short __q3, short __q2, short __q1, short __q0)
 592 {
 593   return __extension__ (__m128i)(__v8hi){
 594     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 595 }
 596
 597 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 598 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 599               char __q11, char __q10, char __q09, char __q08,
 600               char __q07, char __q06, char __q05, char __q04,
 601               char __q03, char __q02, char __q01, char __q00)
 602 {
 603   return __extension__ (__m128i)(__v16qi){
 604     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 605     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 606   };
 607 }
 608
 609 /* Set all of the elements of the vector to A.  */
 610
 611 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 612 _mm_set1_epi64x (long long __A)
 613 {
 614   return _mm_set_epi64x (__A, __A);
 615 }
 616
 617 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 618 _mm_set1_epi64 (__m64 __A)
 619 {
 620   return _mm_set_epi64 (__A, __A);
 621 }
 622
 623 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 624 _mm_set1_epi32 (int __A)
 625 {
 626   return _mm_set_epi32 (__A, __A, __A, __A);
 627 }
 628
 629 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 630 _mm_set1_epi16 (short __A)
 631 {
 632   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 633 }
 634
 635 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 636 _mm_set1_epi8 (char __A)
 637 {
 638   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 639                        __A, __A, __A, __A, __A, __A, __A, __A);
 640 }
 641
 642 /* Create a vector of Qi, where i is the element number.
 643    The parameter order is reversed from the _mm_set_epi* functions.  */
 644
 645 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 646 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 647 {
 648   return _mm_set_epi64 (__q1, __q0);
 649 }
 650
 651 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 652 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 653 {
 654   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 655 }
 656
 657 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 658 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 659                 short __q4, short __q5, short __q6, short __q7)
 660 {
 661   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 662 }
 663
 664 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 665 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 666                char __q04, char __q05, char __q06, char __q07,
 667                char __q08, char __q09, char __q10, char __q11,
 668                char __q12, char __q13, char __q14, char __q15)
 669 {
 670   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 671                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 672 }
 673
 674 /* Create a vector with element 0 as *P and the rest zero.  */
 675
 676 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 677 _mm_load_si128 (__m128i const *__P)
 678 {
 679   return *__P;
 680 }
 681
 682 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 683 _mm_loadu_si128 (__m128i const *__P)
 684 {
 685   return (__m128i) __builtin_ia32_loaddqu ((char const *)__P);
 686 }
 687
 688 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 689 _mm_loadl_epi64 (__m128i const *__P)
 690 {
 691   return _mm_set_epi64 ((__m64)0LL, *(__m64 *)__P);
 692 }
 693
 694 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 695 _mm_store_si128 (__m128i *__P, __m128i __B)
 696 {
 697   *__P = __B;
 698 }
 699
 700 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 701 _mm_storeu_si128 (__m128i *__P, __m128i __B)
 702 {
 703   __builtin_ia32_storedqu ((char *)__P, (__v16qi)__B);
 704 }
 705
 706 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 707 _mm_storel_epi64 (__m128i *__P, __m128i __B)
 708 {
 709   *(long long *)__P = __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 710 }
 711
 712 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 713 _mm_movepi64_pi64 (__m128i __B)
 714 {
 715   return (__m64) __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 716 }
 717
 718 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 719 _mm_movpi64_epi64 (__m64 __A)
 720 {
 721   return _mm_set_epi64 ((__m64)0LL, __A);
 722 }
 723
 724 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 725 _mm_move_epi64 (__m128i __A)
 726 {
 727   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 728 }
 729
 730 /* Create a vector of zeros.  */
 731 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 732 _mm_setzero_si128 (void)
 733 {
 734   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 735 }
 736
 737 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 738 _mm_cvtepi32_pd (__m128i __A)
 739 {
 740   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 741 }
 742
 743 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 744 _mm_cvtepi32_ps (__m128i __A)
 745 {
 746   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 747 }
 748
 749 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 750 _mm_cvtpd_epi32 (__m128d __A)
 751 {
 752   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 753 }
 754
 755 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 756 _mm_cvtpd_pi32 (__m128d __A)
 757 {
 758   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 759 }
 760
 761 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 762 _mm_cvtpd_ps (__m128d __A)
 763 {
 764   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 765 }
 766
 767 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 768 _mm_cvttpd_epi32 (__m128d __A)
 769 {
 770   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 771 }
 772
 773 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 774 _mm_cvttpd_pi32 (__m128d __A)
 775 {
 776   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 777 }
 778
 779 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 780 _mm_cvtpi32_pd (__m64 __A)
 781 {
 782   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 783 }
 784
 785 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 786 _mm_cvtps_epi32 (__m128 __A)
 787 {
 788   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 789 }
 790
 791 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 792 _mm_cvttps_epi32 (__m128 __A)
 793 {
 794   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 795 }
 796
 797 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 798 _mm_cvtps_pd (__m128 __A)
 799 {
 800   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 801 }
 802
 803 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 804 _mm_cvtsd_si32 (__m128d __A)
 805 {
 806   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 807 }
 808
 809 #ifdef __x86_64__
 810 /* Intel intrinsic.  */
 811 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 812 _mm_cvtsd_si64 (__m128d __A)
 813 {
 814   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 815 }
 816
 817 /* Microsoft intrinsic.  */
 818 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 819 _mm_cvtsd_si64x (__m128d __A)
 820 {
 821   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 822 }
 823 #endif
 824
 825 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 826 _mm_cvttsd_si32 (__m128d __A)
 827 {
 828   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 829 }
 830
 831 #ifdef __x86_64__
 832 /* Intel intrinsic.  */
 833 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 834 _mm_cvttsd_si64 (__m128d __A)
 835 {
 836   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 837 }
 838
 839 /* Microsoft intrinsic.  */
 840 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 841 _mm_cvttsd_si64x (__m128d __A)
 842 {
 843   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 844 }
 845 #endif
 846
 847 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 848 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 849 {
 850   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 851 }
 852
 853 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 854 _mm_cvtsi32_sd (__m128d __A, int __B)
 855 {
 856   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 857 }
 858
 859 #ifdef __x86_64__
 860 /* Intel intrinsic.  */
 861 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 862 _mm_cvtsi64_sd (__m128d __A, long long __B)
 863 {
 864   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 865 }
 866
 867 /* Microsoft intrinsic.  */
 868 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 869 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 870 {
 871   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 872 }
 873 #endif
 874
 875 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 876 _mm_cvtss_sd (__m128d __A, __m128 __B)
 877 {
 878   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 879 }
 880
 881 #ifdef __OPTIMIZE__
 882 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 883 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 884 {
 885   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 886 }
 887 #else
 888 #define _mm_shuffle_pd(A, B, N)                                         \
 889   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 890                                    (__v2df)(__m128d)(B), (int)(N)))
 891 #endif
 892
 893 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 894 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 895 {
 896   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 897 }
 898
 899 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 900 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 901 {
 902   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 903 }
 904
 905 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 906 _mm_loadh_pd (__m128d __A, double const *__B)
 907 {
 908   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 909 }
 910
 911 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 912 _mm_loadl_pd (__m128d __A, double const *__B)
 913 {
 914   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 915 }
 916
 917 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 918 _mm_movemask_pd (__m128d __A)
 919 {
 920   return __builtin_ia32_movmskpd ((__v2df)__A);
 921 }
 922
 923 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 924 _mm_packs_epi16 (__m128i __A, __m128i __B)
 925 {
 926   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 927 }
 928
 929 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 930 _mm_packs_epi32 (__m128i __A, __m128i __B)
 931 {
 932   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 933 }
 934
 935 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 936 _mm_packus_epi16 (__m128i __A, __m128i __B)
 937 {
 938   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 939 }
 940
 941 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 942 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 943 {
 944   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 945 }
 946
 947 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 948 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 949 {
 950   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 951 }
 952
 953 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 954 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 955 {
 956   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 957 }
 958
 959 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 960 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 961 {
 962   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 963 }
 964
 965 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 966 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 967 {
 968   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 969 }
 970
 971 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 972 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 973 {
 974   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
 975 }
 976
 977 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 978 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
 979 {
 980   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
 981 }
 982
 983 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 984 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
 985 {
 986   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
 987 }
 988
 989 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 990 _mm_add_epi8 (__m128i __A, __m128i __B)
 991 {
 992   return (__m128i)__builtin_ia32_paddb128 ((__v16qi)__A, (__v16qi)__B);
 993 }
 994
 995 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 996 _mm_add_epi16 (__m128i __A, __m128i __B)
 997 {
 998   return (__m128i)__builtin_ia32_paddw128 ((__v8hi)__A, (__v8hi)__B);
 999 }
1000
1001 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1002 _mm_add_epi32 (__m128i __A, __m128i __B)
1003 {
1004   return (__m128i)__builtin_ia32_paddd128 ((__v4si)__A, (__v4si)__B);
1005 }
1006
1007 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1008 _mm_add_epi64 (__m128i __A, __m128i __B)
1009 {
1010   return (__m128i)__builtin_ia32_paddq128 ((__v2di)__A, (__v2di)__B);
1011 }
1012
1013 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1014 _mm_adds_epi8 (__m128i __A, __m128i __B)
1015 {
1016   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1017 }
1018
1019 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1020 _mm_adds_epi16 (__m128i __A, __m128i __B)
1021 {
1022   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1023 }
1024
1025 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1026 _mm_adds_epu8 (__m128i __A, __m128i __B)
1027 {
1028   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1029 }
1030
1031 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1032 _mm_adds_epu16 (__m128i __A, __m128i __B)
1033 {
1034   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1035 }
1036
1037 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1038 _mm_sub_epi8 (__m128i __A, __m128i __B)
1039 {
1040   return (__m128i)__builtin_ia32_psubb128 ((__v16qi)__A, (__v16qi)__B);
1041 }
1042
1043 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1044 _mm_sub_epi16 (__m128i __A, __m128i __B)
1045 {
1046   return (__m128i)__builtin_ia32_psubw128 ((__v8hi)__A, (__v8hi)__B);
1047 }
1048
1049 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1050 _mm_sub_epi32 (__m128i __A, __m128i __B)
1051 {
1052   return (__m128i)__builtin_ia32_psubd128 ((__v4si)__A, (__v4si)__B);
1053 }
1054
1055 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1056 _mm_sub_epi64 (__m128i __A, __m128i __B)
1057 {
1058   return (__m128i)__builtin_ia32_psubq128 ((__v2di)__A, (__v2di)__B);
1059 }
1060
1061 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1062 _mm_subs_epi8 (__m128i __A, __m128i __B)
1063 {
1064   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1065 }
1066
1067 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1068 _mm_subs_epi16 (__m128i __A, __m128i __B)
1069 {
1070   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1071 }
1072
1073 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1074 _mm_subs_epu8 (__m128i __A, __m128i __B)
1075 {
1076   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1077 }
1078
1079 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1080 _mm_subs_epu16 (__m128i __A, __m128i __B)
1081 {
1082   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1083 }
1084
1085 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1086 _mm_madd_epi16 (__m128i __A, __m128i __B)
1087 {
1088   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1089 }
1090
1091 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1092 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1093 {
1094   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1095 }
1096
1097 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1098 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1099 {
1100   return (__m128i)__builtin_ia32_pmullw128 ((__v8hi)__A, (__v8hi)__B);
1101 }
1102
1103 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1104 _mm_mul_su32 (__m64 __A, __m64 __B)
1105 {
1106   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1107 }
1108
1109 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1110 _mm_mul_epu32 (__m128i __A, __m128i __B)
1111 {
1112   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1113 }
1114
1115 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1116 _mm_slli_epi16 (__m128i __A, int __B)
1117 {
1118   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1119 }
1120
1121 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1122 _mm_slli_epi32 (__m128i __A, int __B)
1123 {
1124   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1125 }
1126
1127 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1128 _mm_slli_epi64 (__m128i __A, int __B)
1129 {
1130   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1131 }
1132
1133 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1134 _mm_srai_epi16 (__m128i __A, int __B)
1135 {
1136   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1137 }
1138
1139 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1140 _mm_srai_epi32 (__m128i __A, int __B)
1141 {
1142   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1143 }
1144
1145 #ifdef __OPTIMIZE__
1146 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1147 _mm_srli_si128 (__m128i __A, const int __N)
1148 {
1149   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1150 }
1151
1152 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1153 _mm_slli_si128 (__m128i __A, const int __N)
1154 {
1155   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1156 }
1157 #else
1158 #define _mm_srli_si128(A, N) \
1159   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1160 #define _mm_slli_si128(A, N) \
1161   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1162 #endif
1163
1164 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1165 _mm_srli_epi16 (__m128i __A, int __B)
1166 {
1167   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1168 }
1169
1170 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1171 _mm_srli_epi32 (__m128i __A, int __B)
1172 {
1173   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1174 }
1175
1176 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1177 _mm_srli_epi64 (__m128i __A, int __B)
1178 {
1179   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1180 }
1181
1182 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1183 _mm_sll_epi16 (__m128i __A, __m128i __B)
1184 {
1185   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1186 }
1187
1188 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1189 _mm_sll_epi32 (__m128i __A, __m128i __B)
1190 {
1191   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1192 }
1193
1194 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1195 _mm_sll_epi64 (__m128i __A, __m128i __B)
1196 {
1197   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1198 }
1199
1200 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1201 _mm_sra_epi16 (__m128i __A, __m128i __B)
1202 {
1203   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1204 }
1205
1206 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1207 _mm_sra_epi32 (__m128i __A, __m128i __B)
1208 {
1209   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1210 }
1211
1212 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1213 _mm_srl_epi16 (__m128i __A, __m128i __B)
1214 {
1215   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1216 }
1217
1218 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1219 _mm_srl_epi32 (__m128i __A, __m128i __B)
1220 {
1221   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1222 }
1223
1224 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1225 _mm_srl_epi64 (__m128i __A, __m128i __B)
1226 {
1227   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1228 }
1229
1230 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1231 _mm_and_si128 (__m128i __A, __m128i __B)
1232 {
1233   return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
1234 }
1235
1236 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1237 _mm_andnot_si128 (__m128i __A, __m128i __B)
1238 {
1239   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1240 }
1241
1242 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1243 _mm_or_si128 (__m128i __A, __m128i __B)
1244 {
1245   return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
1246 }
1247
1248 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1249 _mm_xor_si128 (__m128i __A, __m128i __B)
1250 {
1251   return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);
1252 }
1253
1254 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1255 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1256 {
1257   return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);
1258 }
1259
1260 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1261 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1262 {
1263   return (__m128i)__builtin_ia32_pcmpeqw128 ((__v8hi)__A, (__v8hi)__B);
1264 }
1265
1266 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1267 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1268 {
1269   return (__m128i)__builtin_ia32_pcmpeqd128 ((__v4si)__A, (__v4si)__B);
1270 }
1271
1272 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1273 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1274 {
1275   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__B, (__v16qi)__A);
1276 }
1277
1278 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1279 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1280 {
1281   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__B, (__v8hi)__A);
1282 }
1283
1284 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1285 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1286 {
1287   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__B, (__v4si)__A);
1288 }
1289
1290 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1291 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1292 {
1293   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__A, (__v16qi)__B);
1294 }
1295
1296 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1297 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1298 {
1299   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__A, (__v8hi)__B);
1300 }
1301
1302 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1303 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1304 {
1305   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__A, (__v4si)__B);
1306 }
1307
1308 #ifdef __OPTIMIZE__
1309 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1310 _mm_extract_epi16 (__m128i const __A, int const __N)
1311 {
1312   return __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1313 }
1314
1315 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1316 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1317 {
1318   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1319 }
1320 #else
1321 #define _mm_extract_epi16(A, N) \
1322   ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1323 #define _mm_insert_epi16(A, D, N)                               \
1324   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1325                                           (int)(D), (int)(N)))
1326 #endif
1327
1328 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1329 _mm_max_epi16 (__m128i __A, __m128i __B)
1330 {
1331   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1332 }
1333
1334 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1335 _mm_max_epu8 (__m128i __A, __m128i __B)
1336 {
1337   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1338 }
1339
1340 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1341 _mm_min_epi16 (__m128i __A, __m128i __B)
1342 {
1343   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1344 }
1345
1346 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1347 _mm_min_epu8 (__m128i __A, __m128i __B)
1348 {
1349   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1350 }
1351
1352 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1353 _mm_movemask_epi8 (__m128i __A)
1354 {
1355   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1356 }
1357
1358 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1359 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1360 {
1361   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1362 }
1363
1364 #ifdef __OPTIMIZE__
1365 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1366 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1367 {
1368   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1369 }
1370
1371 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1372 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1373 {
1374   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1375 }
1376
1377 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1378 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1379 {
1380   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1381 }
1382 #else
1383 #define _mm_shufflehi_epi16(A, N) \
1384   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1385 #define _mm_shufflelo_epi16(A, N) \
1386   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1387 #define _mm_shuffle_epi32(A, N) \
1388   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1389 #endif
1390
1391 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1392 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1393 {
1394   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1395 }
1396
1397 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1398 _mm_avg_epu8 (__m128i __A, __m128i __B)
1399 {
1400   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1401 }
1402
1403 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1404 _mm_avg_epu16 (__m128i __A, __m128i __B)
1405 {
1406   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1407 }
1408
1409 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1410 _mm_sad_epu8 (__m128i __A, __m128i __B)
1411 {
1412   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1413 }
1414
1415 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1416 _mm_stream_si32 (int *__A, int __B)
1417 {
1418   __builtin_ia32_movnti (__A, __B);
1419 }
1420
1421 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1422 _mm_stream_si128 (__m128i *__A, __m128i __B)
1423 {
1424   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1425 }
1426
1427 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1428 _mm_stream_pd (double *__A, __m128d __B)
1429 {
1430   __builtin_ia32_movntpd (__A, (__v2df)__B);
1431 }
1432
1433 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1434 _mm_clflush (void const *__A)
1435 {
1436   __builtin_ia32_clflush (__A);
1437 }
1438
1439 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1440 _mm_lfence (void)
1441 {
1442   __builtin_ia32_lfence ();
1443 }
1444
1445 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1446 _mm_mfence (void)
1447 {
1448   __builtin_ia32_mfence ();
1449 }
1450
1451 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1452 _mm_cvtsi32_si128 (int __A)
1453 {
1454   return _mm_set_epi32 (0, 0, 0, __A);
1455 }
1456
1457 #ifdef __x86_64__
1458 /* Intel intrinsic.  */
1459 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1460 _mm_cvtsi64_si128 (long long __A)
1461 {
1462   return _mm_set_epi64x (0, __A);
1463 }
1464
1465 /* Microsoft intrinsic.  */
1466 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1467 _mm_cvtsi64x_si128 (long long __A)
1468 {
1469   return _mm_set_epi64x (0, __A);
1470 }
1471 #endif
1472
1473 /* Casts between various SP, DP, INT vector types.  Note that these do no
1474    conversion of values, they just change the type.  */
1475 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1476 _mm_castpd_ps(__m128d __A)
1477 {
1478   return (__m128) __A;
1479 }
1480
1481 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1482 _mm_castpd_si128(__m128d __A)
1483 {
1484   return (__m128i) __A;
1485 }
1486
1487 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1488 _mm_castps_pd(__m128 __A)
1489 {
1490   return (__m128d) __A;
1491 }
1492
1493 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1494 _mm_castps_si128(__m128 __A)
1495 {
1496   return (__m128i) __A;
1497 }
1498
1499 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1500 _mm_castsi128_ps(__m128i __A)
1501 {
1502   return (__m128) __A;
1503 }
1504
1505 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1506 _mm_castsi128_pd(__m128i __A)
1507 {
1508   return (__m128d) __A;
1509 }
1510
1511 #endif /* __SSE2__  */
1512
1513 #endif /* _EMMINTRIN_H_INCLUDED */