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 /* We need definitions from the SSE header files*/
  31 #include <xmmintrin.h>
  32
  33 #ifndef __SSE2__
  34 #pragma GCC push_options
  35 #pragma GCC target("sse2")
  36 #define __DISABLE_SSE2__
  37 #endif /* __SSE2__ */
  38
  39 /* SSE2 */
  40 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  41 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  42 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  43 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  44 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  45
  46 /* The Intel API is flexible enough that we must allow aliasing with other
  47    vector types, and their scalar components.  */
  48 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  49 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  50
  51 /* Create a selector for use with the SHUFPD instruction.  */
  52 #define _MM_SHUFFLE2(fp1,fp0) \
  53  (((fp1) << 1) | (fp0))
  54
  55 /* Create a vector with element 0 as F and the rest zero.  */
  56 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  57 _mm_set_sd (double __F)
  58 {
  59   return __extension__ (__m128d){ __F, 0.0 };
  60 }
  61
  62 /* Create a vector with both elements equal to F.  */
  63 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  64 _mm_set1_pd (double __F)
  65 {
  66   return __extension__ (__m128d){ __F, __F };
  67 }
  68
  69 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  70 _mm_set_pd1 (double __F)
  71 {
  72   return _mm_set1_pd (__F);
  73 }
  74
  75 /* Create a vector with the lower value X and upper value W.  */
  76 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  77 _mm_set_pd (double __W, double __X)
  78 {
  79   return __extension__ (__m128d){ __X, __W };
  80 }
  81
  82 /* Create a vector with the lower value W and upper value X.  */
  83 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  84 _mm_setr_pd (double __W, double __X)
  85 {
  86   return __extension__ (__m128d){ __W, __X };
  87 }
  88
  89 /* Create a vector of zeros.  */
  90 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  91 _mm_setzero_pd (void)
  92 {
  93   return __extension__ (__m128d){ 0.0, 0.0 };
  94 }
  95
  96 /* Sets the low DPFP value of A from the low value of B.  */
  97 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  98 _mm_move_sd (__m128d __A, __m128d __B)
  99 {
 100   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 101 }
 102
 103 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 104 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 105 _mm_load_pd (double const *__P)
 106 {
 107   return *(__m128d *)__P;
 108 }
 109
 110 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 111 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 112 _mm_loadu_pd (double const *__P)
 113 {
 114   return __builtin_ia32_loadupd (__P);
 115 }
 116
 117 /* Create a vector with all two elements equal to *P.  */
 118 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 119 _mm_load1_pd (double const *__P)
 120 {
 121   return _mm_set1_pd (*__P);
 122 }
 123
 124 /* Create a vector with element 0 as *P and the rest zero.  */
 125 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 126 _mm_load_sd (double const *__P)
 127 {
 128   return _mm_set_sd (*__P);
 129 }
 130
 131 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 132 _mm_load_pd1 (double const *__P)
 133 {
 134   return _mm_load1_pd (__P);
 135 }
 136
 137 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 138 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 139 _mm_loadr_pd (double const *__P)
 140 {
 141   __m128d __tmp = _mm_load_pd (__P);
 142   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 143 }
 144
 145 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 146 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 147 _mm_store_pd (double *__P, __m128d __A)
 148 {
 149   *(__m128d *)__P = __A;
 150 }
 151
 152 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 153 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 154 _mm_storeu_pd (double *__P, __m128d __A)
 155 {
 156   __builtin_ia32_storeupd (__P, __A);
 157 }
 158
 159 /* Stores the lower DPFP value.  */
 160 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 161 _mm_store_sd (double *__P, __m128d __A)
 162 {
 163   *__P = __builtin_ia32_vec_ext_v2df (__A, 0);
 164 }
 165
 166 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 167 _mm_cvtsd_f64 (__m128d __A)
 168 {
 169   return __builtin_ia32_vec_ext_v2df (__A, 0);
 170 }
 171
 172 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 173 _mm_storel_pd (double *__P, __m128d __A)
 174 {
 175   _mm_store_sd (__P, __A);
 176 }
 177
 178 /* Stores the upper DPFP value.  */
 179 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 180 _mm_storeh_pd (double *__P, __m128d __A)
 181 {
 182   *__P = __builtin_ia32_vec_ext_v2df (__A, 1);
 183 }
 184
 185 /* Store the lower DPFP value across two words.
 186    The address must be 16-byte aligned.  */
 187 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 188 _mm_store1_pd (double *__P, __m128d __A)
 189 {
 190   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 191 }
 192
 193 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 194 _mm_store_pd1 (double *__P, __m128d __A)
 195 {
 196   _mm_store1_pd (__P, __A);
 197 }
 198
 199 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 200 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 201 _mm_storer_pd (double *__P, __m128d __A)
 202 {
 203   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 204 }
 205
 206 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 207 _mm_cvtsi128_si32 (__m128i __A)
 208 {
 209   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 210 }
 211
 212 #ifdef __x86_64__
 213 /* Intel intrinsic.  */
 214 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 215 _mm_cvtsi128_si64 (__m128i __A)
 216 {
 217   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 218 }
 219
 220 /* Microsoft intrinsic.  */
 221 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 222 _mm_cvtsi128_si64x (__m128i __A)
 223 {
 224   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 225 }
 226 #endif
 227
 228 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 229 _mm_add_pd (__m128d __A, __m128d __B)
 230 {
 231   return (__m128d)__builtin_ia32_addpd ((__v2df)__A, (__v2df)__B);
 232 }
 233
 234 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 235 _mm_add_sd (__m128d __A, __m128d __B)
 236 {
 237   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 238 }
 239
 240 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 241 _mm_sub_pd (__m128d __A, __m128d __B)
 242 {
 243   return (__m128d)__builtin_ia32_subpd ((__v2df)__A, (__v2df)__B);
 244 }
 245
 246 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 247 _mm_sub_sd (__m128d __A, __m128d __B)
 248 {
 249   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 250 }
 251
 252 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 253 _mm_mul_pd (__m128d __A, __m128d __B)
 254 {
 255   return (__m128d)__builtin_ia32_mulpd ((__v2df)__A, (__v2df)__B);
 256 }
 257
 258 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 259 _mm_mul_sd (__m128d __A, __m128d __B)
 260 {
 261   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 262 }
 263
 264 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 265 _mm_div_pd (__m128d __A, __m128d __B)
 266 {
 267   return (__m128d)__builtin_ia32_divpd ((__v2df)__A, (__v2df)__B);
 268 }
 269
 270 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 271 _mm_div_sd (__m128d __A, __m128d __B)
 272 {
 273   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 274 }
 275
 276 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 277 _mm_sqrt_pd (__m128d __A)
 278 {
 279   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 280 }
 281
 282 /* Return pair {sqrt (A[0), B[1]}.  */
 283 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 284 _mm_sqrt_sd (__m128d __A, __m128d __B)
 285 {
 286   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 287   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 288 }
 289
 290 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 291 _mm_min_pd (__m128d __A, __m128d __B)
 292 {
 293   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 294 }
 295
 296 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 297 _mm_min_sd (__m128d __A, __m128d __B)
 298 {
 299   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 300 }
 301
 302 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 303 _mm_max_pd (__m128d __A, __m128d __B)
 304 {
 305   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 306 }
 307
 308 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 309 _mm_max_sd (__m128d __A, __m128d __B)
 310 {
 311   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 312 }
 313
 314 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 315 _mm_and_pd (__m128d __A, __m128d __B)
 316 {
 317   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 318 }
 319
 320 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 321 _mm_andnot_pd (__m128d __A, __m128d __B)
 322 {
 323   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 324 }
 325
 326 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 327 _mm_or_pd (__m128d __A, __m128d __B)
 328 {
 329   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 330 }
 331
 332 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 333 _mm_xor_pd (__m128d __A, __m128d __B)
 334 {
 335   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 336 }
 337
 338 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 339 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 340 {
 341   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 342 }
 343
 344 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 345 _mm_cmplt_pd (__m128d __A, __m128d __B)
 346 {
 347   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 348 }
 349
 350 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 351 _mm_cmple_pd (__m128d __A, __m128d __B)
 352 {
 353   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 354 }
 355
 356 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 357 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 358 {
 359   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 360 }
 361
 362 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 363 _mm_cmpge_pd (__m128d __A, __m128d __B)
 364 {
 365   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 366 }
 367
 368 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 369 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 370 {
 371   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 372 }
 373
 374 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 375 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 376 {
 377   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 378 }
 379
 380 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 381 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 382 {
 383   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 384 }
 385
 386 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 387 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 388 {
 389   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 390 }
 391
 392 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 393 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 394 {
 395   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 396 }
 397
 398 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 399 _mm_cmpord_pd (__m128d __A, __m128d __B)
 400 {
 401   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 402 }
 403
 404 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 405 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 406 {
 407   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 408 }
 409
 410 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 411 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 412 {
 413   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 414 }
 415
 416 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 417 _mm_cmplt_sd (__m128d __A, __m128d __B)
 418 {
 419   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 420 }
 421
 422 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 423 _mm_cmple_sd (__m128d __A, __m128d __B)
 424 {
 425   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 426 }
 427
 428 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 429 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 430 {
 431   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 432                                          (__v2df)
 433                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 434                                                                  (__v2df)
 435                                                                  __A));
 436 }
 437
 438 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 439 _mm_cmpge_sd (__m128d __A, __m128d __B)
 440 {
 441   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 442                                          (__v2df)
 443                                          __builtin_ia32_cmplesd ((__v2df) __B,
 444                                                                  (__v2df)
 445                                                                  __A));
 446 }
 447
 448 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 449 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 450 {
 451   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 452 }
 453
 454 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 455 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 456 {
 457   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 458 }
 459
 460 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 461 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 462 {
 463   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 464 }
 465
 466 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 467 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 468 {
 469   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 470                                          (__v2df)
 471                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 472                                                                   (__v2df)
 473                                                                   __A));
 474 }
 475
 476 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 477 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 478 {
 479   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 480                                          (__v2df)
 481                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 482                                                                   (__v2df)
 483                                                                   __A));
 484 }
 485
 486 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 487 _mm_cmpord_sd (__m128d __A, __m128d __B)
 488 {
 489   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 490 }
 491
 492 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 493 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 494 {
 495   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 496 }
 497
 498 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 499 _mm_comieq_sd (__m128d __A, __m128d __B)
 500 {
 501   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 502 }
 503
 504 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 505 _mm_comilt_sd (__m128d __A, __m128d __B)
 506 {
 507   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 508 }
 509
 510 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 511 _mm_comile_sd (__m128d __A, __m128d __B)
 512 {
 513   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 514 }
 515
 516 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 517 _mm_comigt_sd (__m128d __A, __m128d __B)
 518 {
 519   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 520 }
 521
 522 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 523 _mm_comige_sd (__m128d __A, __m128d __B)
 524 {
 525   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 526 }
 527
 528 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 529 _mm_comineq_sd (__m128d __A, __m128d __B)
 530 {
 531   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 532 }
 533
 534 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 535 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 536 {
 537   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 538 }
 539
 540 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 541 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 542 {
 543   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 544 }
 545
 546 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 547 _mm_ucomile_sd (__m128d __A, __m128d __B)
 548 {
 549   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 550 }
 551
 552 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 553 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 554 {
 555   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 556 }
 557
 558 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 559 _mm_ucomige_sd (__m128d __A, __m128d __B)
 560 {
 561   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 562 }
 563
 564 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 565 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 566 {
 567   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 568 }
 569
 570 /* Create a vector of Qi, where i is the element number.  */
 571
 572 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 573 _mm_set_epi64x (long long __q1, long long __q0)
 574 {
 575   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 576 }
 577
 578 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 579 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 580 {
 581   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 582 }
 583
 584 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 585 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 586 {
 587   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 588 }
 589
 590 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 591 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 592                short __q3, short __q2, short __q1, short __q0)
 593 {
 594   return __extension__ (__m128i)(__v8hi){
 595     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 596 }
 597
 598 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 599 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 600               char __q11, char __q10, char __q09, char __q08,
 601               char __q07, char __q06, char __q05, char __q04,
 602               char __q03, char __q02, char __q01, char __q00)
 603 {
 604   return __extension__ (__m128i)(__v16qi){
 605     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 606     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 607   };
 608 }
 609
 610 /* Set all of the elements of the vector to A.  */
 611
 612 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 613 _mm_set1_epi64x (long long __A)
 614 {
 615   return _mm_set_epi64x (__A, __A);
 616 }
 617
 618 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 619 _mm_set1_epi64 (__m64 __A)
 620 {
 621   return _mm_set_epi64 (__A, __A);
 622 }
 623
 624 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 625 _mm_set1_epi32 (int __A)
 626 {
 627   return _mm_set_epi32 (__A, __A, __A, __A);
 628 }
 629
 630 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 631 _mm_set1_epi16 (short __A)
 632 {
 633   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 634 }
 635
 636 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 637 _mm_set1_epi8 (char __A)
 638 {
 639   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 640                        __A, __A, __A, __A, __A, __A, __A, __A);
 641 }
 642
 643 /* Create a vector of Qi, where i is the element number.
 644    The parameter order is reversed from the _mm_set_epi* functions.  */
 645
 646 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 647 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 648 {
 649   return _mm_set_epi64 (__q1, __q0);
 650 }
 651
 652 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 653 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 654 {
 655   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 656 }
 657
 658 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 659 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 660                 short __q4, short __q5, short __q6, short __q7)
 661 {
 662   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 663 }
 664
 665 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 666 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 667                char __q04, char __q05, char __q06, char __q07,
 668                char __q08, char __q09, char __q10, char __q11,
 669                char __q12, char __q13, char __q14, char __q15)
 670 {
 671   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 672                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 673 }
 674
 675 /* Create a vector with element 0 as *P and the rest zero.  */
 676
 677 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 678 _mm_load_si128 (__m128i const *__P)
 679 {
 680   return *__P;
 681 }
 682
 683 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 684 _mm_loadu_si128 (__m128i const *__P)
 685 {
 686   return (__m128i) __builtin_ia32_loaddqu ((char const *)__P);
 687 }
 688
 689 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 690 _mm_loadl_epi64 (__m128i const *__P)
 691 {
 692   return _mm_set_epi64 ((__m64)0LL, *(__m64 *)__P);
 693 }
 694
 695 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 696 _mm_store_si128 (__m128i *__P, __m128i __B)
 697 {
 698   *__P = __B;
 699 }
 700
 701 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 702 _mm_storeu_si128 (__m128i *__P, __m128i __B)
 703 {
 704   __builtin_ia32_storedqu ((char *)__P, (__v16qi)__B);
 705 }
 706
 707 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 708 _mm_storel_epi64 (__m128i *__P, __m128i __B)
 709 {
 710   *(long long *)__P = __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 711 }
 712
 713 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 714 _mm_movepi64_pi64 (__m128i __B)
 715 {
 716   return (__m64) __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 717 }
 718
 719 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 720 _mm_movpi64_epi64 (__m64 __A)
 721 {
 722   return _mm_set_epi64 ((__m64)0LL, __A);
 723 }
 724
 725 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 726 _mm_move_epi64 (__m128i __A)
 727 {
 728   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 729 }
 730
 731 /* Create a vector of zeros.  */
 732 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 733 _mm_setzero_si128 (void)
 734 {
 735   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 736 }
 737
 738 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 739 _mm_cvtepi32_pd (__m128i __A)
 740 {
 741   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 742 }
 743
 744 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 745 _mm_cvtepi32_ps (__m128i __A)
 746 {
 747   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 748 }
 749
 750 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 751 _mm_cvtpd_epi32 (__m128d __A)
 752 {
 753   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 754 }
 755
 756 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 757 _mm_cvtpd_pi32 (__m128d __A)
 758 {
 759   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 760 }
 761
 762 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 763 _mm_cvtpd_ps (__m128d __A)
 764 {
 765   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 766 }
 767
 768 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 769 _mm_cvttpd_epi32 (__m128d __A)
 770 {
 771   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 772 }
 773
 774 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 775 _mm_cvttpd_pi32 (__m128d __A)
 776 {
 777   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 778 }
 779
 780 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 781 _mm_cvtpi32_pd (__m64 __A)
 782 {
 783   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 784 }
 785
 786 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 787 _mm_cvtps_epi32 (__m128 __A)
 788 {
 789   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 790 }
 791
 792 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 793 _mm_cvttps_epi32 (__m128 __A)
 794 {
 795   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 796 }
 797
 798 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 799 _mm_cvtps_pd (__m128 __A)
 800 {
 801   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 802 }
 803
 804 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 805 _mm_cvtsd_si32 (__m128d __A)
 806 {
 807   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 808 }
 809
 810 #ifdef __x86_64__
 811 /* Intel intrinsic.  */
 812 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 813 _mm_cvtsd_si64 (__m128d __A)
 814 {
 815   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 816 }
 817
 818 /* Microsoft intrinsic.  */
 819 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 820 _mm_cvtsd_si64x (__m128d __A)
 821 {
 822   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 823 }
 824 #endif
 825
 826 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 827 _mm_cvttsd_si32 (__m128d __A)
 828 {
 829   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 830 }
 831
 832 #ifdef __x86_64__
 833 /* Intel intrinsic.  */
 834 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 835 _mm_cvttsd_si64 (__m128d __A)
 836 {
 837   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 838 }
 839
 840 /* Microsoft intrinsic.  */
 841 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 842 _mm_cvttsd_si64x (__m128d __A)
 843 {
 844   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 845 }
 846 #endif
 847
 848 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 849 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 850 {
 851   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 852 }
 853
 854 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 855 _mm_cvtsi32_sd (__m128d __A, int __B)
 856 {
 857   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 858 }
 859
 860 #ifdef __x86_64__
 861 /* Intel intrinsic.  */
 862 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 863 _mm_cvtsi64_sd (__m128d __A, long long __B)
 864 {
 865   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 866 }
 867
 868 /* Microsoft intrinsic.  */
 869 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 870 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 871 {
 872   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 873 }
 874 #endif
 875
 876 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 877 _mm_cvtss_sd (__m128d __A, __m128 __B)
 878 {
 879   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 880 }
 881
 882 #ifdef __OPTIMIZE__
 883 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 884 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 885 {
 886   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 887 }
 888 #else
 889 #define _mm_shuffle_pd(A, B, N)                                         \
 890   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 891                                    (__v2df)(__m128d)(B), (int)(N)))
 892 #endif
 893
 894 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 895 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 896 {
 897   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 898 }
 899
 900 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 901 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 902 {
 903   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 904 }
 905
 906 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 907 _mm_loadh_pd (__m128d __A, double const *__B)
 908 {
 909   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 910 }
 911
 912 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 913 _mm_loadl_pd (__m128d __A, double const *__B)
 914 {
 915   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 916 }
 917
 918 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 919 _mm_movemask_pd (__m128d __A)
 920 {
 921   return __builtin_ia32_movmskpd ((__v2df)__A);
 922 }
 923
 924 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 925 _mm_packs_epi16 (__m128i __A, __m128i __B)
 926 {
 927   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 928 }
 929
 930 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 931 _mm_packs_epi32 (__m128i __A, __m128i __B)
 932 {
 933   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 934 }
 935
 936 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 937 _mm_packus_epi16 (__m128i __A, __m128i __B)
 938 {
 939   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 940 }
 941
 942 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 943 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 944 {
 945   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 946 }
 947
 948 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 949 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 950 {
 951   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 952 }
 953
 954 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 955 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 956 {
 957   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 958 }
 959
 960 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 961 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 962 {
 963   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 964 }
 965
 966 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 967 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 968 {
 969   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 970 }
 971
 972 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 973 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 974 {
 975   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
 976 }
 977
 978 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 979 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
 980 {
 981   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
 982 }
 983
 984 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 985 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
 986 {
 987   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
 988 }
 989
 990 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 991 _mm_add_epi8 (__m128i __A, __m128i __B)
 992 {
 993   return (__m128i)__builtin_ia32_paddb128 ((__v16qi)__A, (__v16qi)__B);
 994 }
 995
 996 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 997 _mm_add_epi16 (__m128i __A, __m128i __B)
 998 {
 999   return (__m128i)__builtin_ia32_paddw128 ((__v8hi)__A, (__v8hi)__B);
1000 }
1001
1002 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1003 _mm_add_epi32 (__m128i __A, __m128i __B)
1004 {
1005   return (__m128i)__builtin_ia32_paddd128 ((__v4si)__A, (__v4si)__B);
1006 }
1007
1008 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1009 _mm_add_epi64 (__m128i __A, __m128i __B)
1010 {
1011   return (__m128i)__builtin_ia32_paddq128 ((__v2di)__A, (__v2di)__B);
1012 }
1013
1014 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1015 _mm_adds_epi8 (__m128i __A, __m128i __B)
1016 {
1017   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1018 }
1019
1020 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1021 _mm_adds_epi16 (__m128i __A, __m128i __B)
1022 {
1023   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1024 }
1025
1026 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1027 _mm_adds_epu8 (__m128i __A, __m128i __B)
1028 {
1029   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1030 }
1031
1032 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1033 _mm_adds_epu16 (__m128i __A, __m128i __B)
1034 {
1035   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1036 }
1037
1038 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1039 _mm_sub_epi8 (__m128i __A, __m128i __B)
1040 {
1041   return (__m128i)__builtin_ia32_psubb128 ((__v16qi)__A, (__v16qi)__B);
1042 }
1043
1044 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1045 _mm_sub_epi16 (__m128i __A, __m128i __B)
1046 {
1047   return (__m128i)__builtin_ia32_psubw128 ((__v8hi)__A, (__v8hi)__B);
1048 }
1049
1050 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1051 _mm_sub_epi32 (__m128i __A, __m128i __B)
1052 {
1053   return (__m128i)__builtin_ia32_psubd128 ((__v4si)__A, (__v4si)__B);
1054 }
1055
1056 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1057 _mm_sub_epi64 (__m128i __A, __m128i __B)
1058 {
1059   return (__m128i)__builtin_ia32_psubq128 ((__v2di)__A, (__v2di)__B);
1060 }
1061
1062 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1063 _mm_subs_epi8 (__m128i __A, __m128i __B)
1064 {
1065   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1066 }
1067
1068 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1069 _mm_subs_epi16 (__m128i __A, __m128i __B)
1070 {
1071   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1072 }
1073
1074 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1075 _mm_subs_epu8 (__m128i __A, __m128i __B)
1076 {
1077   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1078 }
1079
1080 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1081 _mm_subs_epu16 (__m128i __A, __m128i __B)
1082 {
1083   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1084 }
1085
1086 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1087 _mm_madd_epi16 (__m128i __A, __m128i __B)
1088 {
1089   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1090 }
1091
1092 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1093 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1094 {
1095   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1096 }
1097
1098 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1099 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1100 {
1101   return (__m128i)__builtin_ia32_pmullw128 ((__v8hi)__A, (__v8hi)__B);
1102 }
1103
1104 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1105 _mm_mul_su32 (__m64 __A, __m64 __B)
1106 {
1107   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1108 }
1109
1110 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1111 _mm_mul_epu32 (__m128i __A, __m128i __B)
1112 {
1113   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1114 }
1115
1116 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1117 _mm_slli_epi16 (__m128i __A, int __B)
1118 {
1119   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1120 }
1121
1122 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1123 _mm_slli_epi32 (__m128i __A, int __B)
1124 {
1125   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1126 }
1127
1128 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1129 _mm_slli_epi64 (__m128i __A, int __B)
1130 {
1131   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1132 }
1133
1134 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1135 _mm_srai_epi16 (__m128i __A, int __B)
1136 {
1137   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1138 }
1139
1140 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1141 _mm_srai_epi32 (__m128i __A, int __B)
1142 {
1143   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1144 }
1145
1146 #ifdef __OPTIMIZE__
1147 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1148 _mm_srli_si128 (__m128i __A, const int __N)
1149 {
1150   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1151 }
1152
1153 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1154 _mm_slli_si128 (__m128i __A, const int __N)
1155 {
1156   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1157 }
1158 #else
1159 #define _mm_srli_si128(A, N) \
1160   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1161 #define _mm_slli_si128(A, N) \
1162   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1163 #endif
1164
1165 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1166 _mm_srli_epi16 (__m128i __A, int __B)
1167 {
1168   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1169 }
1170
1171 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1172 _mm_srli_epi32 (__m128i __A, int __B)
1173 {
1174   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1175 }
1176
1177 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1178 _mm_srli_epi64 (__m128i __A, int __B)
1179 {
1180   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1181 }
1182
1183 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1184 _mm_sll_epi16 (__m128i __A, __m128i __B)
1185 {
1186   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1187 }
1188
1189 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1190 _mm_sll_epi32 (__m128i __A, __m128i __B)
1191 {
1192   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1193 }
1194
1195 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1196 _mm_sll_epi64 (__m128i __A, __m128i __B)
1197 {
1198   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1199 }
1200
1201 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1202 _mm_sra_epi16 (__m128i __A, __m128i __B)
1203 {
1204   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1205 }
1206
1207 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1208 _mm_sra_epi32 (__m128i __A, __m128i __B)
1209 {
1210   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1211 }
1212
1213 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1214 _mm_srl_epi16 (__m128i __A, __m128i __B)
1215 {
1216   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1217 }
1218
1219 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1220 _mm_srl_epi32 (__m128i __A, __m128i __B)
1221 {
1222   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1223 }
1224
1225 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1226 _mm_srl_epi64 (__m128i __A, __m128i __B)
1227 {
1228   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1229 }
1230
1231 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1232 _mm_and_si128 (__m128i __A, __m128i __B)
1233 {
1234   return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
1235 }
1236
1237 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1238 _mm_andnot_si128 (__m128i __A, __m128i __B)
1239 {
1240   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1241 }
1242
1243 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1244 _mm_or_si128 (__m128i __A, __m128i __B)
1245 {
1246   return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
1247 }
1248
1249 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1250 _mm_xor_si128 (__m128i __A, __m128i __B)
1251 {
1252   return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);
1253 }
1254
1255 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1256 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1257 {
1258   return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);
1259 }
1260
1261 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1262 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1263 {
1264   return (__m128i)__builtin_ia32_pcmpeqw128 ((__v8hi)__A, (__v8hi)__B);
1265 }
1266
1267 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1268 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1269 {
1270   return (__m128i)__builtin_ia32_pcmpeqd128 ((__v4si)__A, (__v4si)__B);
1271 }
1272
1273 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1274 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1275 {
1276   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__B, (__v16qi)__A);
1277 }
1278
1279 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1280 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1281 {
1282   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__B, (__v8hi)__A);
1283 }
1284
1285 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1286 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1287 {
1288   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__B, (__v4si)__A);
1289 }
1290
1291 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1292 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1293 {
1294   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__A, (__v16qi)__B);
1295 }
1296
1297 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1298 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1299 {
1300   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__A, (__v8hi)__B);
1301 }
1302
1303 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1304 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1305 {
1306   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__A, (__v4si)__B);
1307 }
1308
1309 #ifdef __OPTIMIZE__
1310 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1311 _mm_extract_epi16 (__m128i const __A, int const __N)
1312 {
1313   return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1314 }
1315
1316 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1317 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1318 {
1319   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1320 }
1321 #else
1322 #define _mm_extract_epi16(A, N) \
1323   ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1324 #define _mm_insert_epi16(A, D, N)                               \
1325   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1326                                           (int)(D), (int)(N)))
1327 #endif
1328
1329 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1330 _mm_max_epi16 (__m128i __A, __m128i __B)
1331 {
1332   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1333 }
1334
1335 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1336 _mm_max_epu8 (__m128i __A, __m128i __B)
1337 {
1338   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1339 }
1340
1341 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1342 _mm_min_epi16 (__m128i __A, __m128i __B)
1343 {
1344   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1345 }
1346
1347 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1348 _mm_min_epu8 (__m128i __A, __m128i __B)
1349 {
1350   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1351 }
1352
1353 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1354 _mm_movemask_epi8 (__m128i __A)
1355 {
1356   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1357 }
1358
1359 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1360 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1361 {
1362   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1363 }
1364
1365 #ifdef __OPTIMIZE__
1366 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1367 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1368 {
1369   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1370 }
1371
1372 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1373 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1374 {
1375   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1376 }
1377
1378 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1379 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1380 {
1381   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1382 }
1383 #else
1384 #define _mm_shufflehi_epi16(A, N) \
1385   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1386 #define _mm_shufflelo_epi16(A, N) \
1387   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1388 #define _mm_shuffle_epi32(A, N) \
1389   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1390 #endif
1391
1392 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1393 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1394 {
1395   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1396 }
1397
1398 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1399 _mm_avg_epu8 (__m128i __A, __m128i __B)
1400 {
1401   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1402 }
1403
1404 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1405 _mm_avg_epu16 (__m128i __A, __m128i __B)
1406 {
1407   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1408 }
1409
1410 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1411 _mm_sad_epu8 (__m128i __A, __m128i __B)
1412 {
1413   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1414 }
1415
1416 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1417 _mm_stream_si32 (int *__A, int __B)
1418 {
1419   __builtin_ia32_movnti (__A, __B);
1420 }
1421
1422 #ifdef __x86_64__
1423 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1424 _mm_stream_si64 (long long int *__A, long long int __B)
1425 {
1426   __builtin_ia32_movnti64 (__A, __B);
1427 }
1428 #endif
1429
1430 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1431 _mm_stream_si128 (__m128i *__A, __m128i __B)
1432 {
1433   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1434 }
1435
1436 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1437 _mm_stream_pd (double *__A, __m128d __B)
1438 {
1439   __builtin_ia32_movntpd (__A, (__v2df)__B);
1440 }
1441
1442 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1443 _mm_clflush (void const *__A)
1444 {
1445   __builtin_ia32_clflush (__A);
1446 }
1447
1448 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1449 _mm_lfence (void)
1450 {
1451   __builtin_ia32_lfence ();
1452 }
1453
1454 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1455 _mm_mfence (void)
1456 {
1457   __builtin_ia32_mfence ();
1458 }
1459
1460 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1461 _mm_cvtsi32_si128 (int __A)
1462 {
1463   return _mm_set_epi32 (0, 0, 0, __A);
1464 }
1465
1466 #ifdef __x86_64__
1467 /* Intel intrinsic.  */
1468 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1469 _mm_cvtsi64_si128 (long long __A)
1470 {
1471   return _mm_set_epi64x (0, __A);
1472 }
1473
1474 /* Microsoft intrinsic.  */
1475 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1476 _mm_cvtsi64x_si128 (long long __A)
1477 {
1478   return _mm_set_epi64x (0, __A);
1479 }
1480 #endif
1481
1482 /* Casts between various SP, DP, INT vector types.  Note that these do no
1483    conversion of values, they just change the type.  */
1484 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1485 _mm_castpd_ps(__m128d __A)
1486 {
1487   return (__m128) __A;
1488 }
1489
1490 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1491 _mm_castpd_si128(__m128d __A)
1492 {
1493   return (__m128i) __A;
1494 }
1495
1496 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1497 _mm_castps_pd(__m128 __A)
1498 {
1499   return (__m128d) __A;
1500 }
1501
1502 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1503 _mm_castps_si128(__m128 __A)
1504 {
1505   return (__m128i) __A;
1506 }
1507
1508 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1509 _mm_castsi128_ps(__m128i __A)
1510 {
1511   return (__m128) __A;
1512 }
1513
1514 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1515 _mm_castsi128_pd(__m128i __A)
1516 {
1517   return (__m128d) __A;
1518 }
1519
1520 #ifdef __DISABLE_SSE2__
1521 #undef __DISABLE_SSE2__
1522 #pragma GCC pop_options
1523 #endif /* __DISABLE_SSE2__ */
1524
1525 #endif /* _EMMINTRIN_H_INCLUDED */