gcc/config/i386/emmintrin.h

   1 /* Copyright (C) 2003-2013 Free Software Foundation, Inc.
   2
   3    This file is part of GCC.
   4
   5    GCC is free software; you can redistribute it and/or modify
   6    it under the terms of the GNU General Public License as published by
   7    the Free Software Foundation; either version 3, or (at your option)
   8    any later version.
   9
  10    GCC is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13    GNU General Public License for more details.
  14
  15    Under Section 7 of GPL version 3, you are granted additional
  16    permissions described in the GCC Runtime Library Exception, version
  17    3.1, as published by the Free Software Foundation.
  18
  19    You should have received a copy of the GNU General Public License and
  20    a copy of the GCC Runtime Library Exception along with this program;
  21    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  22    <http://www.gnu.org/licenses/>.  */
  23
  24 /* Implemented from the specification included in the Intel C++ Compiler
  25    User Guide and Reference, version 9.0.  */
  26
  27 #ifndef _EMMINTRIN_H_INCLUDED
  28 #define _EMMINTRIN_H_INCLUDED
  29
  30 #ifndef __SSE2__
  31 # error "SSE2 instruction set not enabled"
  32 #else
  33
  34 /* We need definitions from the SSE header files*/
  35 #include <xmmintrin.h>
  36
  37 /* SSE2 */
  38 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  39 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  40 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  41 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  42 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  43
  44 /* The Intel API is flexible enough that we must allow aliasing with other
  45    vector types, and their scalar components.  */
  46 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  47 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  48
  49 /* Create a selector for use with the SHUFPD instruction.  */
  50 #define _MM_SHUFFLE2(fp1,fp0) \
  51  (((fp1) << 1) | (fp0))
  52
  53 /* Create a vector with element 0 as F and the rest zero.  */
  54 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  55 _mm_set_sd (double __F)
  56 {
  57   return __extension__ (__m128d){ __F, 0.0 };
  58 }
  59
  60 /* Create a vector with both elements equal to F.  */
  61 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  62 _mm_set1_pd (double __F)
  63 {
  64   return __extension__ (__m128d){ __F, __F };
  65 }
  66
  67 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  68 _mm_set_pd1 (double __F)
  69 {
  70   return _mm_set1_pd (__F);
  71 }
  72
  73 /* Create a vector with the lower value X and upper value W.  */
  74 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  75 _mm_set_pd (double __W, double __X)
  76 {
  77   return __extension__ (__m128d){ __X, __W };
  78 }
  79
  80 /* Create a vector with the lower value W and upper value X.  */
  81 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  82 _mm_setr_pd (double __W, double __X)
  83 {
  84   return __extension__ (__m128d){ __W, __X };
  85 }
  86
  87 /* Create a vector of zeros.  */
  88 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  89 _mm_setzero_pd (void)
  90 {
  91   return __extension__ (__m128d){ 0.0, 0.0 };
  92 }
  93
  94 /* Sets the low DPFP value of A from the low value of B.  */
  95 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  96 _mm_move_sd (__m128d __A, __m128d __B)
  97 {
  98   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
  99 }
 100
 101 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 102 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 103 _mm_load_pd (double const *__P)
 104 {
 105   return *(__m128d *)__P;
 106 }
 107
 108 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 109 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 110 _mm_loadu_pd (double const *__P)
 111 {
 112   return __builtin_ia32_loadupd (__P);
 113 }
 114
 115 /* Create a vector with all two elements equal to *P.  */
 116 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 117 _mm_load1_pd (double const *__P)
 118 {
 119   return _mm_set1_pd (*__P);
 120 }
 121
 122 /* Create a vector with element 0 as *P and the rest zero.  */
 123 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 124 _mm_load_sd (double const *__P)
 125 {
 126   return _mm_set_sd (*__P);
 127 }
 128
 129 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 130 _mm_load_pd1 (double const *__P)
 131 {
 132   return _mm_load1_pd (__P);
 133 }
 134
 135 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 136 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 137 _mm_loadr_pd (double const *__P)
 138 {
 139   __m128d __tmp = _mm_load_pd (__P);
 140   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 141 }
 142
 143 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 144 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 145 _mm_store_pd (double *__P, __m128d __A)
 146 {
 147   *(__m128d *)__P = __A;
 148 }
 149
 150 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 151 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 152 _mm_storeu_pd (double *__P, __m128d __A)
 153 {
 154   __builtin_ia32_storeupd (__P, __A);
 155 }
 156
 157 /* Stores the lower DPFP value.  */
 158 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 159 _mm_store_sd (double *__P, __m128d __A)
 160 {
 161   *__P = __builtin_ia32_vec_ext_v2df (__A, 0);
 162 }
 163
 164 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 165 _mm_cvtsd_f64 (__m128d __A)
 166 {
 167   return __builtin_ia32_vec_ext_v2df (__A, 0);
 168 }
 169
 170 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 171 _mm_storel_pd (double *__P, __m128d __A)
 172 {
 173   _mm_store_sd (__P, __A);
 174 }
 175
 176 /* Stores the upper DPFP value.  */
 177 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 178 _mm_storeh_pd (double *__P, __m128d __A)
 179 {
 180   *__P = __builtin_ia32_vec_ext_v2df (__A, 1);
 181 }
 182
 183 /* Store the lower DPFP value across two words.
 184    The address must be 16-byte aligned.  */
 185 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 186 _mm_store1_pd (double *__P, __m128d __A)
 187 {
 188   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 189 }
 190
 191 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 192 _mm_store_pd1 (double *__P, __m128d __A)
 193 {
 194   _mm_store1_pd (__P, __A);
 195 }
 196
 197 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 198 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 199 _mm_storer_pd (double *__P, __m128d __A)
 200 {
 201   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 202 }
 203
 204 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 205 _mm_cvtsi128_si32 (__m128i __A)
 206 {
 207   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 208 }
 209
 210 #ifdef __x86_64__
 211 /* Intel intrinsic.  */
 212 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 213 _mm_cvtsi128_si64 (__m128i __A)
 214 {
 215   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 216 }
 217
 218 /* Microsoft intrinsic.  */
 219 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 220 _mm_cvtsi128_si64x (__m128i __A)
 221 {
 222   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 223 }
 224 #endif
 225
 226 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 227 _mm_add_pd (__m128d __A, __m128d __B)
 228 {
 229   return (__m128d)__builtin_ia32_addpd ((__v2df)__A, (__v2df)__B);
 230 }
 231
 232 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 233 _mm_add_sd (__m128d __A, __m128d __B)
 234 {
 235   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 236 }
 237
 238 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 239 _mm_sub_pd (__m128d __A, __m128d __B)
 240 {
 241   return (__m128d)__builtin_ia32_subpd ((__v2df)__A, (__v2df)__B);
 242 }
 243
 244 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 245 _mm_sub_sd (__m128d __A, __m128d __B)
 246 {
 247   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 248 }
 249
 250 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 251 _mm_mul_pd (__m128d __A, __m128d __B)
 252 {
 253   return (__m128d)__builtin_ia32_mulpd ((__v2df)__A, (__v2df)__B);
 254 }
 255
 256 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 257 _mm_mul_sd (__m128d __A, __m128d __B)
 258 {
 259   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 260 }
 261
 262 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 263 _mm_div_pd (__m128d __A, __m128d __B)
 264 {
 265   return (__m128d)__builtin_ia32_divpd ((__v2df)__A, (__v2df)__B);
 266 }
 267
 268 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 269 _mm_div_sd (__m128d __A, __m128d __B)
 270 {
 271   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 272 }
 273
 274 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 275 _mm_sqrt_pd (__m128d __A)
 276 {
 277   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 278 }
 279
 280 /* Return pair {sqrt (A[0), B[1]}.  */
 281 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 282 _mm_sqrt_sd (__m128d __A, __m128d __B)
 283 {
 284   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 285   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 286 }
 287
 288 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 289 _mm_min_pd (__m128d __A, __m128d __B)
 290 {
 291   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 292 }
 293
 294 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 295 _mm_min_sd (__m128d __A, __m128d __B)
 296 {
 297   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 298 }
 299
 300 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 301 _mm_max_pd (__m128d __A, __m128d __B)
 302 {
 303   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 304 }
 305
 306 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 307 _mm_max_sd (__m128d __A, __m128d __B)
 308 {
 309   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 310 }
 311
 312 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 313 _mm_and_pd (__m128d __A, __m128d __B)
 314 {
 315   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 316 }
 317
 318 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 319 _mm_andnot_pd (__m128d __A, __m128d __B)
 320 {
 321   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 322 }
 323
 324 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 325 _mm_or_pd (__m128d __A, __m128d __B)
 326 {
 327   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 328 }
 329
 330 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 331 _mm_xor_pd (__m128d __A, __m128d __B)
 332 {
 333   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 334 }
 335
 336 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 337 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 338 {
 339   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 340 }
 341
 342 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 343 _mm_cmplt_pd (__m128d __A, __m128d __B)
 344 {
 345   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 346 }
 347
 348 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 349 _mm_cmple_pd (__m128d __A, __m128d __B)
 350 {
 351   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 352 }
 353
 354 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 355 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 356 {
 357   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 358 }
 359
 360 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 361 _mm_cmpge_pd (__m128d __A, __m128d __B)
 362 {
 363   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 364 }
 365
 366 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 367 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 368 {
 369   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 370 }
 371
 372 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 373 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 374 {
 375   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 376 }
 377
 378 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 379 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 380 {
 381   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 382 }
 383
 384 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 385 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 386 {
 387   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 388 }
 389
 390 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 391 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 392 {
 393   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 394 }
 395
 396 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 397 _mm_cmpord_pd (__m128d __A, __m128d __B)
 398 {
 399   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 400 }
 401
 402 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 403 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 404 {
 405   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 406 }
 407
 408 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 409 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 410 {
 411   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 412 }
 413
 414 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 415 _mm_cmplt_sd (__m128d __A, __m128d __B)
 416 {
 417   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 418 }
 419
 420 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 421 _mm_cmple_sd (__m128d __A, __m128d __B)
 422 {
 423   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 424 }
 425
 426 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 427 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 428 {
 429   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 430                                          (__v2df)
 431                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 432                                                                  (__v2df)
 433                                                                  __A));
 434 }
 435
 436 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 437 _mm_cmpge_sd (__m128d __A, __m128d __B)
 438 {
 439   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 440                                          (__v2df)
 441                                          __builtin_ia32_cmplesd ((__v2df) __B,
 442                                                                  (__v2df)
 443                                                                  __A));
 444 }
 445
 446 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 447 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 448 {
 449   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 450 }
 451
 452 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 453 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 454 {
 455   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 456 }
 457
 458 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 459 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 460 {
 461   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 462 }
 463
 464 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 465 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 466 {
 467   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 468                                          (__v2df)
 469                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 470                                                                   (__v2df)
 471                                                                   __A));
 472 }
 473
 474 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 475 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 476 {
 477   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 478                                          (__v2df)
 479                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 480                                                                   (__v2df)
 481                                                                   __A));
 482 }
 483
 484 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 485 _mm_cmpord_sd (__m128d __A, __m128d __B)
 486 {
 487   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 488 }
 489
 490 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 491 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 492 {
 493   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 494 }
 495
 496 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 497 _mm_comieq_sd (__m128d __A, __m128d __B)
 498 {
 499   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 500 }
 501
 502 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 503 _mm_comilt_sd (__m128d __A, __m128d __B)
 504 {
 505   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 506 }
 507
 508 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 509 _mm_comile_sd (__m128d __A, __m128d __B)
 510 {
 511   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 512 }
 513
 514 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 515 _mm_comigt_sd (__m128d __A, __m128d __B)
 516 {
 517   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 518 }
 519
 520 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 521 _mm_comige_sd (__m128d __A, __m128d __B)
 522 {
 523   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 524 }
 525
 526 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 527 _mm_comineq_sd (__m128d __A, __m128d __B)
 528 {
 529   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 530 }
 531
 532 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 533 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 534 {
 535   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 536 }
 537
 538 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 539 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 540 {
 541   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 542 }
 543
 544 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 545 _mm_ucomile_sd (__m128d __A, __m128d __B)
 546 {
 547   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 548 }
 549
 550 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 551 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 552 {
 553   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 554 }
 555
 556 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 557 _mm_ucomige_sd (__m128d __A, __m128d __B)
 558 {
 559   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 560 }
 561
 562 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 563 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 564 {
 565   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 566 }
 567
 568 /* Create a vector of Qi, where i is the element number.  */
 569
 570 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 571 _mm_set_epi64x (long long __q1, long long __q0)
 572 {
 573   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 574 }
 575
 576 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 577 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 578 {
 579   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 580 }
 581
 582 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 583 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 584 {
 585   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 586 }
 587
 588 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 589 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 590                short __q3, short __q2, short __q1, short __q0)
 591 {
 592   return __extension__ (__m128i)(__v8hi){
 593     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 594 }
 595
 596 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 597 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 598               char __q11, char __q10, char __q09, char __q08,
 599               char __q07, char __q06, char __q05, char __q04,
 600               char __q03, char __q02, char __q01, char __q00)
 601 {
 602   return __extension__ (__m128i)(__v16qi){
 603     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 604     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 605   };
 606 }
 607
 608 /* Set all of the elements of the vector to A.  */
 609
 610 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 611 _mm_set1_epi64x (long long __A)
 612 {
 613   return _mm_set_epi64x (__A, __A);
 614 }
 615
 616 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 617 _mm_set1_epi64 (__m64 __A)
 618 {
 619   return _mm_set_epi64 (__A, __A);
 620 }
 621
 622 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 623 _mm_set1_epi32 (int __A)
 624 {
 625   return _mm_set_epi32 (__A, __A, __A, __A);
 626 }
 627
 628 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 629 _mm_set1_epi16 (short __A)
 630 {
 631   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 632 }
 633
 634 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 635 _mm_set1_epi8 (char __A)
 636 {
 637   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 638                        __A, __A, __A, __A, __A, __A, __A, __A);
 639 }
 640
 641 /* Create a vector of Qi, where i is the element number.
 642    The parameter order is reversed from the _mm_set_epi* functions.  */
 643
 644 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 645 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 646 {
 647   return _mm_set_epi64 (__q1, __q0);
 648 }
 649
 650 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 651 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 652 {
 653   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 654 }
 655
 656 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 657 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 658                 short __q4, short __q5, short __q6, short __q7)
 659 {
 660   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 661 }
 662
 663 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 664 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 665                char __q04, char __q05, char __q06, char __q07,
 666                char __q08, char __q09, char __q10, char __q11,
 667                char __q12, char __q13, char __q14, char __q15)
 668 {
 669   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 670                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 671 }
 672
 673 /* Create a vector with element 0 as *P and the rest zero.  */
 674
 675 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 676 _mm_load_si128 (__m128i const *__P)
 677 {
 678   return *__P;
 679 }
 680
 681 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 682 _mm_loadu_si128 (__m128i const *__P)
 683 {
 684   return (__m128i) __builtin_ia32_loaddqu ((char const *)__P);
 685 }
 686
 687 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 688 _mm_loadl_epi64 (__m128i const *__P)
 689 {
 690   return _mm_set_epi64 ((__m64)0LL, *(__m64 *)__P);
 691 }
 692
 693 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 694 _mm_store_si128 (__m128i *__P, __m128i __B)
 695 {
 696   *__P = __B;
 697 }
 698
 699 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 700 _mm_storeu_si128 (__m128i *__P, __m128i __B)
 701 {
 702   __builtin_ia32_storedqu ((char *)__P, (__v16qi)__B);
 703 }
 704
 705 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 706 _mm_storel_epi64 (__m128i *__P, __m128i __B)
 707 {
 708   *(long long *)__P = __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 709 }
 710
 711 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 712 _mm_movepi64_pi64 (__m128i __B)
 713 {
 714   return (__m64) __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 715 }
 716
 717 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 718 _mm_movpi64_epi64 (__m64 __A)
 719 {
 720   return _mm_set_epi64 ((__m64)0LL, __A);
 721 }
 722
 723 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 724 _mm_move_epi64 (__m128i __A)
 725 {
 726   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 727 }
 728
 729 /* Create a vector of zeros.  */
 730 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 731 _mm_setzero_si128 (void)
 732 {
 733   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 734 }
 735
 736 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 737 _mm_cvtepi32_pd (__m128i __A)
 738 {
 739   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 740 }
 741
 742 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 743 _mm_cvtepi32_ps (__m128i __A)
 744 {
 745   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 746 }
 747
 748 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 749 _mm_cvtpd_epi32 (__m128d __A)
 750 {
 751   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 752 }
 753
 754 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 755 _mm_cvtpd_pi32 (__m128d __A)
 756 {
 757   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 758 }
 759
 760 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 761 _mm_cvtpd_ps (__m128d __A)
 762 {
 763   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 764 }
 765
 766 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 767 _mm_cvttpd_epi32 (__m128d __A)
 768 {
 769   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 770 }
 771
 772 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 773 _mm_cvttpd_pi32 (__m128d __A)
 774 {
 775   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 776 }
 777
 778 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 779 _mm_cvtpi32_pd (__m64 __A)
 780 {
 781   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 782 }
 783
 784 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 785 _mm_cvtps_epi32 (__m128 __A)
 786 {
 787   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 788 }
 789
 790 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 791 _mm_cvttps_epi32 (__m128 __A)
 792 {
 793   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 794 }
 795
 796 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 797 _mm_cvtps_pd (__m128 __A)
 798 {
 799   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 800 }
 801
 802 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 803 _mm_cvtsd_si32 (__m128d __A)
 804 {
 805   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 806 }
 807
 808 #ifdef __x86_64__
 809 /* Intel intrinsic.  */
 810 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 811 _mm_cvtsd_si64 (__m128d __A)
 812 {
 813   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 814 }
 815
 816 /* Microsoft intrinsic.  */
 817 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 818 _mm_cvtsd_si64x (__m128d __A)
 819 {
 820   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 821 }
 822 #endif
 823
 824 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 825 _mm_cvttsd_si32 (__m128d __A)
 826 {
 827   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 828 }
 829
 830 #ifdef __x86_64__
 831 /* Intel intrinsic.  */
 832 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 833 _mm_cvttsd_si64 (__m128d __A)
 834 {
 835   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 836 }
 837
 838 /* Microsoft intrinsic.  */
 839 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 840 _mm_cvttsd_si64x (__m128d __A)
 841 {
 842   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 843 }
 844 #endif
 845
 846 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 847 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 848 {
 849   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 850 }
 851
 852 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 853 _mm_cvtsi32_sd (__m128d __A, int __B)
 854 {
 855   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 856 }
 857
 858 #ifdef __x86_64__
 859 /* Intel intrinsic.  */
 860 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 861 _mm_cvtsi64_sd (__m128d __A, long long __B)
 862 {
 863   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 864 }
 865
 866 /* Microsoft intrinsic.  */
 867 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 868 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 869 {
 870   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 871 }
 872 #endif
 873
 874 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 875 _mm_cvtss_sd (__m128d __A, __m128 __B)
 876 {
 877   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 878 }
 879
 880 #ifdef __OPTIMIZE__
 881 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 882 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 883 {
 884   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 885 }
 886 #else
 887 #define _mm_shuffle_pd(A, B, N)                                         \
 888   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 889                                    (__v2df)(__m128d)(B), (int)(N)))
 890 #endif
 891
 892 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 893 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 894 {
 895   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 896 }
 897
 898 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 899 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 900 {
 901   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 902 }
 903
 904 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 905 _mm_loadh_pd (__m128d __A, double const *__B)
 906 {
 907   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 908 }
 909
 910 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 911 _mm_loadl_pd (__m128d __A, double const *__B)
 912 {
 913   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 914 }
 915
 916 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 917 _mm_movemask_pd (__m128d __A)
 918 {
 919   return __builtin_ia32_movmskpd ((__v2df)__A);
 920 }
 921
 922 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 923 _mm_packs_epi16 (__m128i __A, __m128i __B)
 924 {
 925   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 926 }
 927
 928 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 929 _mm_packs_epi32 (__m128i __A, __m128i __B)
 930 {
 931   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 932 }
 933
 934 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 935 _mm_packus_epi16 (__m128i __A, __m128i __B)
 936 {
 937   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 938 }
 939
 940 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 941 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 942 {
 943   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 944 }
 945
 946 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 947 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 948 {
 949   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 950 }
 951
 952 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 953 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 954 {
 955   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 956 }
 957
 958 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 959 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 960 {
 961   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 962 }
 963
 964 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 965 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 966 {
 967   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 968 }
 969
 970 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 971 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 972 {
 973   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
 974 }
 975
 976 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 977 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
 978 {
 979   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
 980 }
 981
 982 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 983 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
 984 {
 985   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
 986 }
 987
 988 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 989 _mm_add_epi8 (__m128i __A, __m128i __B)
 990 {
 991   return (__m128i)__builtin_ia32_paddb128 ((__v16qi)__A, (__v16qi)__B);
 992 }
 993
 994 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 995 _mm_add_epi16 (__m128i __A, __m128i __B)
 996 {
 997   return (__m128i)__builtin_ia32_paddw128 ((__v8hi)__A, (__v8hi)__B);
 998 }
 999
1000 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1001 _mm_add_epi32 (__m128i __A, __m128i __B)
1002 {
1003   return (__m128i)__builtin_ia32_paddd128 ((__v4si)__A, (__v4si)__B);
1004 }
1005
1006 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1007 _mm_add_epi64 (__m128i __A, __m128i __B)
1008 {
1009   return (__m128i)__builtin_ia32_paddq128 ((__v2di)__A, (__v2di)__B);
1010 }
1011
1012 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1013 _mm_adds_epi8 (__m128i __A, __m128i __B)
1014 {
1015   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1016 }
1017
1018 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1019 _mm_adds_epi16 (__m128i __A, __m128i __B)
1020 {
1021   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1022 }
1023
1024 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1025 _mm_adds_epu8 (__m128i __A, __m128i __B)
1026 {
1027   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1028 }
1029
1030 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1031 _mm_adds_epu16 (__m128i __A, __m128i __B)
1032 {
1033   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1034 }
1035
1036 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1037 _mm_sub_epi8 (__m128i __A, __m128i __B)
1038 {
1039   return (__m128i)__builtin_ia32_psubb128 ((__v16qi)__A, (__v16qi)__B);
1040 }
1041
1042 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1043 _mm_sub_epi16 (__m128i __A, __m128i __B)
1044 {
1045   return (__m128i)__builtin_ia32_psubw128 ((__v8hi)__A, (__v8hi)__B);
1046 }
1047
1048 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1049 _mm_sub_epi32 (__m128i __A, __m128i __B)
1050 {
1051   return (__m128i)__builtin_ia32_psubd128 ((__v4si)__A, (__v4si)__B);
1052 }
1053
1054 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1055 _mm_sub_epi64 (__m128i __A, __m128i __B)
1056 {
1057   return (__m128i)__builtin_ia32_psubq128 ((__v2di)__A, (__v2di)__B);
1058 }
1059
1060 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1061 _mm_subs_epi8 (__m128i __A, __m128i __B)
1062 {
1063   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1064 }
1065
1066 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1067 _mm_subs_epi16 (__m128i __A, __m128i __B)
1068 {
1069   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1070 }
1071
1072 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1073 _mm_subs_epu8 (__m128i __A, __m128i __B)
1074 {
1075   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1076 }
1077
1078 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1079 _mm_subs_epu16 (__m128i __A, __m128i __B)
1080 {
1081   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1082 }
1083
1084 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1085 _mm_madd_epi16 (__m128i __A, __m128i __B)
1086 {
1087   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1088 }
1089
1090 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1091 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1092 {
1093   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1094 }
1095
1096 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1097 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1098 {
1099   return (__m128i)__builtin_ia32_pmullw128 ((__v8hi)__A, (__v8hi)__B);
1100 }
1101
1102 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1103 _mm_mul_su32 (__m64 __A, __m64 __B)
1104 {
1105   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1106 }
1107
1108 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1109 _mm_mul_epu32 (__m128i __A, __m128i __B)
1110 {
1111   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1112 }
1113
1114 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1115 _mm_slli_epi16 (__m128i __A, int __B)
1116 {
1117   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1118 }
1119
1120 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1121 _mm_slli_epi32 (__m128i __A, int __B)
1122 {
1123   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1124 }
1125
1126 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1127 _mm_slli_epi64 (__m128i __A, int __B)
1128 {
1129   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1130 }
1131
1132 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1133 _mm_srai_epi16 (__m128i __A, int __B)
1134 {
1135   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1136 }
1137
1138 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1139 _mm_srai_epi32 (__m128i __A, int __B)
1140 {
1141   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1142 }
1143
1144 #ifdef __OPTIMIZE__
1145 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1146 _mm_srli_si128 (__m128i __A, const int __N)
1147 {
1148   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1149 }
1150
1151 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1152 _mm_slli_si128 (__m128i __A, const int __N)
1153 {
1154   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1155 }
1156 #else
1157 #define _mm_srli_si128(A, N) \
1158   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1159 #define _mm_slli_si128(A, N) \
1160   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1161 #endif
1162
1163 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1164 _mm_srli_epi16 (__m128i __A, int __B)
1165 {
1166   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1167 }
1168
1169 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1170 _mm_srli_epi32 (__m128i __A, int __B)
1171 {
1172   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1173 }
1174
1175 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1176 _mm_srli_epi64 (__m128i __A, int __B)
1177 {
1178   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1179 }
1180
1181 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1182 _mm_sll_epi16 (__m128i __A, __m128i __B)
1183 {
1184   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1185 }
1186
1187 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1188 _mm_sll_epi32 (__m128i __A, __m128i __B)
1189 {
1190   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1191 }
1192
1193 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1194 _mm_sll_epi64 (__m128i __A, __m128i __B)
1195 {
1196   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1197 }
1198
1199 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1200 _mm_sra_epi16 (__m128i __A, __m128i __B)
1201 {
1202   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1203 }
1204
1205 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1206 _mm_sra_epi32 (__m128i __A, __m128i __B)
1207 {
1208   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1209 }
1210
1211 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1212 _mm_srl_epi16 (__m128i __A, __m128i __B)
1213 {
1214   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1215 }
1216
1217 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1218 _mm_srl_epi32 (__m128i __A, __m128i __B)
1219 {
1220   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1221 }
1222
1223 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1224 _mm_srl_epi64 (__m128i __A, __m128i __B)
1225 {
1226   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1227 }
1228
1229 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1230 _mm_and_si128 (__m128i __A, __m128i __B)
1231 {
1232   return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
1233 }
1234
1235 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1236 _mm_andnot_si128 (__m128i __A, __m128i __B)
1237 {
1238   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1239 }
1240
1241 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1242 _mm_or_si128 (__m128i __A, __m128i __B)
1243 {
1244   return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
1245 }
1246
1247 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1248 _mm_xor_si128 (__m128i __A, __m128i __B)
1249 {
1250   return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);
1251 }
1252
1253 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1254 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1255 {
1256   return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);
1257 }
1258
1259 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1260 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1261 {
1262   return (__m128i)__builtin_ia32_pcmpeqw128 ((__v8hi)__A, (__v8hi)__B);
1263 }
1264
1265 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1266 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1267 {
1268   return (__m128i)__builtin_ia32_pcmpeqd128 ((__v4si)__A, (__v4si)__B);
1269 }
1270
1271 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1272 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1273 {
1274   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__B, (__v16qi)__A);
1275 }
1276
1277 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1278 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1279 {
1280   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__B, (__v8hi)__A);
1281 }
1282
1283 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1284 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1285 {
1286   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__B, (__v4si)__A);
1287 }
1288
1289 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1290 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1291 {
1292   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__A, (__v16qi)__B);
1293 }
1294
1295 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1296 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1297 {
1298   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__A, (__v8hi)__B);
1299 }
1300
1301 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1302 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1303 {
1304   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__A, (__v4si)__B);
1305 }
1306
1307 #ifdef __OPTIMIZE__
1308 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1309 _mm_extract_epi16 (__m128i const __A, int const __N)
1310 {
1311   return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1312 }
1313
1314 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1315 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1316 {
1317   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1318 }
1319 #else
1320 #define _mm_extract_epi16(A, N) \
1321   ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1322 #define _mm_insert_epi16(A, D, N)                               \
1323   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1324                                           (int)(D), (int)(N)))
1325 #endif
1326
1327 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1328 _mm_max_epi16 (__m128i __A, __m128i __B)
1329 {
1330   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1331 }
1332
1333 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1334 _mm_max_epu8 (__m128i __A, __m128i __B)
1335 {
1336   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1337 }
1338
1339 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1340 _mm_min_epi16 (__m128i __A, __m128i __B)
1341 {
1342   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1343 }
1344
1345 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1346 _mm_min_epu8 (__m128i __A, __m128i __B)
1347 {
1348   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1349 }
1350
1351 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1352 _mm_movemask_epi8 (__m128i __A)
1353 {
1354   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1355 }
1356
1357 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1358 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1359 {
1360   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1361 }
1362
1363 #ifdef __OPTIMIZE__
1364 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1365 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1366 {
1367   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1368 }
1369
1370 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1371 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1372 {
1373   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1374 }
1375
1376 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1377 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1378 {
1379   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1380 }
1381 #else
1382 #define _mm_shufflehi_epi16(A, N) \
1383   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1384 #define _mm_shufflelo_epi16(A, N) \
1385   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1386 #define _mm_shuffle_epi32(A, N) \
1387   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1388 #endif
1389
1390 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1391 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1392 {
1393   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1394 }
1395
1396 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1397 _mm_avg_epu8 (__m128i __A, __m128i __B)
1398 {
1399   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1400 }
1401
1402 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1403 _mm_avg_epu16 (__m128i __A, __m128i __B)
1404 {
1405   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1406 }
1407
1408 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1409 _mm_sad_epu8 (__m128i __A, __m128i __B)
1410 {
1411   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1412 }
1413
1414 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1415 _mm_stream_si32 (int *__A, int __B)
1416 {
1417   __builtin_ia32_movnti (__A, __B);
1418 }
1419
1420 #ifdef __x86_64__
1421 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1422 _mm_stream_si64 (long long int *__A, long long int __B)
1423 {
1424   __builtin_ia32_movnti64 (__A, __B);
1425 }
1426 #endif
1427
1428 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1429 _mm_stream_si128 (__m128i *__A, __m128i __B)
1430 {
1431   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1432 }
1433
1434 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1435 _mm_stream_pd (double *__A, __m128d __B)
1436 {
1437   __builtin_ia32_movntpd (__A, (__v2df)__B);
1438 }
1439
1440 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1441 _mm_clflush (void const *__A)
1442 {
1443   __builtin_ia32_clflush (__A);
1444 }
1445
1446 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1447 _mm_lfence (void)
1448 {
1449   __builtin_ia32_lfence ();
1450 }
1451
1452 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1453 _mm_mfence (void)
1454 {
1455   __builtin_ia32_mfence ();
1456 }
1457
1458 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1459 _mm_cvtsi32_si128 (int __A)
1460 {
1461   return _mm_set_epi32 (0, 0, 0, __A);
1462 }
1463
1464 #ifdef __x86_64__
1465 /* Intel intrinsic.  */
1466 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1467 _mm_cvtsi64_si128 (long long __A)
1468 {
1469   return _mm_set_epi64x (0, __A);
1470 }
1471
1472 /* Microsoft intrinsic.  */
1473 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1474 _mm_cvtsi64x_si128 (long long __A)
1475 {
1476   return _mm_set_epi64x (0, __A);
1477 }
1478 #endif
1479
1480 /* Casts between various SP, DP, INT vector types.  Note that these do no
1481    conversion of values, they just change the type.  */
1482 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1483 _mm_castpd_ps(__m128d __A)
1484 {
1485   return (__m128) __A;
1486 }
1487
1488 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1489 _mm_castpd_si128(__m128d __A)
1490 {
1491   return (__m128i) __A;
1492 }
1493
1494 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1495 _mm_castps_pd(__m128 __A)
1496 {
1497   return (__m128d) __A;
1498 }
1499
1500 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1501 _mm_castps_si128(__m128 __A)
1502 {
1503   return (__m128i) __A;
1504 }
1505
1506 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1507 _mm_castsi128_ps(__m128i __A)
1508 {
1509   return (__m128) __A;
1510 }
1511
1512 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1513 _mm_castsi128_pd(__m128i __A)
1514 {
1515   return (__m128d) __A;
1516 }
1517
1518 #endif /* __SSE2__  */
1519
1520 #endif /* _EMMINTRIN_H_INCLUDED */