gcc/config/i386/emmintrin.h

   1 /* Copyright (C) 2003-2014 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 unsigned long long __v2du __attribute__ ((__vector_size__ (16)));
  43 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  44 typedef unsigned int __v4su __attribute__ ((__vector_size__ (16)));
  45 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  46 typedef unsigned short __v8hu __attribute__ ((__vector_size__ (16)));
  47 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  48 typedef unsigned char __v16qu __attribute__ ((__vector_size__ (16)));
  49
  50 /* The Intel API is flexible enough that we must allow aliasing with other
  51    vector types, and their scalar components.  */
  52 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  53 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  54
  55 /* Create a selector for use with the SHUFPD instruction.  */
  56 #define _MM_SHUFFLE2(fp1,fp0) \
  57  (((fp1) << 1) | (fp0))
  58
  59 /* Create a vector with element 0 as F and the rest zero.  */
  60 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  61 _mm_set_sd (double __F)
  62 {
  63   return __extension__ (__m128d){ __F, 0.0 };
  64 }
  65
  66 /* Create a vector with both elements equal to F.  */
  67 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  68 _mm_set1_pd (double __F)
  69 {
  70   return __extension__ (__m128d){ __F, __F };
  71 }
  72
  73 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  74 _mm_set_pd1 (double __F)
  75 {
  76   return _mm_set1_pd (__F);
  77 }
  78
  79 /* Create a vector with the lower value X and upper value W.  */
  80 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  81 _mm_set_pd (double __W, double __X)
  82 {
  83   return __extension__ (__m128d){ __X, __W };
  84 }
  85
  86 /* Create a vector with the lower value W and upper value X.  */
  87 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  88 _mm_setr_pd (double __W, double __X)
  89 {
  90   return __extension__ (__m128d){ __W, __X };
  91 }
  92
  93 /* Create an undefined vector.  */
  94 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  95 _mm_undefined_pd (void)
  96 {
  97   __m128d __Y = __Y;
  98   return __Y;
  99 }
 100
 101 /* Create a vector of zeros.  */
 102 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 103 _mm_setzero_pd (void)
 104 {
 105   return __extension__ (__m128d){ 0.0, 0.0 };
 106 }
 107
 108 /* Sets the low DPFP value of A from the low value of B.  */
 109 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 110 _mm_move_sd (__m128d __A, __m128d __B)
 111 {
 112   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 113 }
 114
 115 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 116 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 117 _mm_load_pd (double const *__P)
 118 {
 119   return *(__m128d *)__P;
 120 }
 121
 122 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 123 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 124 _mm_loadu_pd (double const *__P)
 125 {
 126   return __builtin_ia32_loadupd (__P);
 127 }
 128
 129 /* Create a vector with all two elements equal to *P.  */
 130 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 131 _mm_load1_pd (double const *__P)
 132 {
 133   return _mm_set1_pd (*__P);
 134 }
 135
 136 /* Create a vector with element 0 as *P and the rest zero.  */
 137 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 138 _mm_load_sd (double const *__P)
 139 {
 140   return _mm_set_sd (*__P);
 141 }
 142
 143 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 144 _mm_load_pd1 (double const *__P)
 145 {
 146   return _mm_load1_pd (__P);
 147 }
 148
 149 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 150 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 151 _mm_loadr_pd (double const *__P)
 152 {
 153   __m128d __tmp = _mm_load_pd (__P);
 154   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 155 }
 156
 157 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 158 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 159 _mm_store_pd (double *__P, __m128d __A)
 160 {
 161   *(__m128d *)__P = __A;
 162 }
 163
 164 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 165 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 166 _mm_storeu_pd (double *__P, __m128d __A)
 167 {
 168   __builtin_ia32_storeupd (__P, __A);
 169 }
 170
 171 /* Stores the lower DPFP value.  */
 172 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 173 _mm_store_sd (double *__P, __m128d __A)
 174 {
 175   *__P = ((__v2df)__A)[0];
 176 }
 177
 178 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 179 _mm_cvtsd_f64 (__m128d __A)
 180 {
 181   return ((__v2df)__A)[0];
 182 }
 183
 184 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 185 _mm_storel_pd (double *__P, __m128d __A)
 186 {
 187   _mm_store_sd (__P, __A);
 188 }
 189
 190 /* Stores the upper DPFP value.  */
 191 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 192 _mm_storeh_pd (double *__P, __m128d __A)
 193 {
 194   *__P = ((__v2df)__A)[1];
 195 }
 196
 197 /* Store the lower DPFP value across two words.
 198    The address must be 16-byte aligned.  */
 199 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 200 _mm_store1_pd (double *__P, __m128d __A)
 201 {
 202   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 203 }
 204
 205 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 206 _mm_store_pd1 (double *__P, __m128d __A)
 207 {
 208   _mm_store1_pd (__P, __A);
 209 }
 210
 211 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 212 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 213 _mm_storer_pd (double *__P, __m128d __A)
 214 {
 215   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 216 }
 217
 218 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 219 _mm_cvtsi128_si32 (__m128i __A)
 220 {
 221   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 222 }
 223
 224 #ifdef __x86_64__
 225 /* Intel intrinsic.  */
 226 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 227 _mm_cvtsi128_si64 (__m128i __A)
 228 {
 229   return ((__v2di)__A)[0];
 230 }
 231
 232 /* Microsoft intrinsic.  */
 233 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 234 _mm_cvtsi128_si64x (__m128i __A)
 235 {
 236   return ((__v2di)__A)[0];
 237 }
 238 #endif
 239
 240 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 241 _mm_add_pd (__m128d __A, __m128d __B)
 242 {
 243   return (__m128d) ((__v2df)__A + (__v2df)__B);
 244 }
 245
 246 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 247 _mm_add_sd (__m128d __A, __m128d __B)
 248 {
 249   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 250 }
 251
 252 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 253 _mm_sub_pd (__m128d __A, __m128d __B)
 254 {
 255   return (__m128d) ((__v2df)__A - (__v2df)__B);
 256 }
 257
 258 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 259 _mm_sub_sd (__m128d __A, __m128d __B)
 260 {
 261   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 262 }
 263
 264 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 265 _mm_mul_pd (__m128d __A, __m128d __B)
 266 {
 267   return (__m128d) ((__v2df)__A * (__v2df)__B);
 268 }
 269
 270 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 271 _mm_mul_sd (__m128d __A, __m128d __B)
 272 {
 273   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 274 }
 275
 276 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 277 _mm_div_pd (__m128d __A, __m128d __B)
 278 {
 279   return (__m128d) ((__v2df)__A / (__v2df)__B);
 280 }
 281
 282 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 283 _mm_div_sd (__m128d __A, __m128d __B)
 284 {
 285   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 286 }
 287
 288 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 289 _mm_sqrt_pd (__m128d __A)
 290 {
 291   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 292 }
 293
 294 /* Return pair {sqrt (B[0]), A[1]}.  */
 295 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 296 _mm_sqrt_sd (__m128d __A, __m128d __B)
 297 {
 298   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 299   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 300 }
 301
 302 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 303 _mm_min_pd (__m128d __A, __m128d __B)
 304 {
 305   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 306 }
 307
 308 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 309 _mm_min_sd (__m128d __A, __m128d __B)
 310 {
 311   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 312 }
 313
 314 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 315 _mm_max_pd (__m128d __A, __m128d __B)
 316 {
 317   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 318 }
 319
 320 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 321 _mm_max_sd (__m128d __A, __m128d __B)
 322 {
 323   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 324 }
 325
 326 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 327 _mm_and_pd (__m128d __A, __m128d __B)
 328 {
 329   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 330 }
 331
 332 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 333 _mm_andnot_pd (__m128d __A, __m128d __B)
 334 {
 335   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 336 }
 337
 338 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 339 _mm_or_pd (__m128d __A, __m128d __B)
 340 {
 341   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 342 }
 343
 344 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 345 _mm_xor_pd (__m128d __A, __m128d __B)
 346 {
 347   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 348 }
 349
 350 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 351 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 352 {
 353   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 354 }
 355
 356 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 357 _mm_cmplt_pd (__m128d __A, __m128d __B)
 358 {
 359   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 360 }
 361
 362 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 363 _mm_cmple_pd (__m128d __A, __m128d __B)
 364 {
 365   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 366 }
 367
 368 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 369 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 370 {
 371   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 372 }
 373
 374 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 375 _mm_cmpge_pd (__m128d __A, __m128d __B)
 376 {
 377   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 378 }
 379
 380 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 381 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 382 {
 383   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 384 }
 385
 386 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 387 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 388 {
 389   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 390 }
 391
 392 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 393 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 394 {
 395   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 396 }
 397
 398 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 399 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 400 {
 401   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 402 }
 403
 404 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 405 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 406 {
 407   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 408 }
 409
 410 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 411 _mm_cmpord_pd (__m128d __A, __m128d __B)
 412 {
 413   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 414 }
 415
 416 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 417 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 418 {
 419   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 420 }
 421
 422 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 423 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 424 {
 425   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 426 }
 427
 428 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 429 _mm_cmplt_sd (__m128d __A, __m128d __B)
 430 {
 431   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 432 }
 433
 434 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 435 _mm_cmple_sd (__m128d __A, __m128d __B)
 436 {
 437   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 438 }
 439
 440 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 441 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 442 {
 443   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 444                                          (__v2df)
 445                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 446                                                                  (__v2df)
 447                                                                  __A));
 448 }
 449
 450 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 451 _mm_cmpge_sd (__m128d __A, __m128d __B)
 452 {
 453   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 454                                          (__v2df)
 455                                          __builtin_ia32_cmplesd ((__v2df) __B,
 456                                                                  (__v2df)
 457                                                                  __A));
 458 }
 459
 460 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 461 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 462 {
 463   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 464 }
 465
 466 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 467 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 468 {
 469   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 470 }
 471
 472 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 473 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 474 {
 475   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 476 }
 477
 478 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 479 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 480 {
 481   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 482                                          (__v2df)
 483                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 484                                                                   (__v2df)
 485                                                                   __A));
 486 }
 487
 488 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 489 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 490 {
 491   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 492                                          (__v2df)
 493                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 494                                                                   (__v2df)
 495                                                                   __A));
 496 }
 497
 498 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 499 _mm_cmpord_sd (__m128d __A, __m128d __B)
 500 {
 501   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 502 }
 503
 504 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 505 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 506 {
 507   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 508 }
 509
 510 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 511 _mm_comieq_sd (__m128d __A, __m128d __B)
 512 {
 513   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 514 }
 515
 516 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 517 _mm_comilt_sd (__m128d __A, __m128d __B)
 518 {
 519   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 520 }
 521
 522 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 523 _mm_comile_sd (__m128d __A, __m128d __B)
 524 {
 525   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 526 }
 527
 528 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 529 _mm_comigt_sd (__m128d __A, __m128d __B)
 530 {
 531   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 532 }
 533
 534 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 535 _mm_comige_sd (__m128d __A, __m128d __B)
 536 {
 537   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 538 }
 539
 540 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 541 _mm_comineq_sd (__m128d __A, __m128d __B)
 542 {
 543   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 544 }
 545
 546 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 547 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 548 {
 549   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 550 }
 551
 552 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 553 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 554 {
 555   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 556 }
 557
 558 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 559 _mm_ucomile_sd (__m128d __A, __m128d __B)
 560 {
 561   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 562 }
 563
 564 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 565 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 566 {
 567   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 568 }
 569
 570 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 571 _mm_ucomige_sd (__m128d __A, __m128d __B)
 572 {
 573   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 574 }
 575
 576 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 577 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 578 {
 579   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 580 }
 581
 582 /* Create a vector of Qi, where i is the element number.  */
 583
 584 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 585 _mm_set_epi64x (long long __q1, long long __q0)
 586 {
 587   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 588 }
 589
 590 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 591 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 592 {
 593   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 594 }
 595
 596 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 597 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 598 {
 599   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 600 }
 601
 602 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 603 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 604                short __q3, short __q2, short __q1, short __q0)
 605 {
 606   return __extension__ (__m128i)(__v8hi){
 607     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 608 }
 609
 610 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 611 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 612               char __q11, char __q10, char __q09, char __q08,
 613               char __q07, char __q06, char __q05, char __q04,
 614               char __q03, char __q02, char __q01, char __q00)
 615 {
 616   return __extension__ (__m128i)(__v16qi){
 617     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 618     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 619   };
 620 }
 621
 622 /* Set all of the elements of the vector to A.  */
 623
 624 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 625 _mm_set1_epi64x (long long __A)
 626 {
 627   return _mm_set_epi64x (__A, __A);
 628 }
 629
 630 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 631 _mm_set1_epi64 (__m64 __A)
 632 {
 633   return _mm_set_epi64 (__A, __A);
 634 }
 635
 636 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 637 _mm_set1_epi32 (int __A)
 638 {
 639   return _mm_set_epi32 (__A, __A, __A, __A);
 640 }
 641
 642 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 643 _mm_set1_epi16 (short __A)
 644 {
 645   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 646 }
 647
 648 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 649 _mm_set1_epi8 (char __A)
 650 {
 651   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 652                        __A, __A, __A, __A, __A, __A, __A, __A);
 653 }
 654
 655 /* Create a vector of Qi, where i is the element number.
 656    The parameter order is reversed from the _mm_set_epi* functions.  */
 657
 658 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 659 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 660 {
 661   return _mm_set_epi64 (__q1, __q0);
 662 }
 663
 664 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 665 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 666 {
 667   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 668 }
 669
 670 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 671 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 672                 short __q4, short __q5, short __q6, short __q7)
 673 {
 674   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 675 }
 676
 677 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 678 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 679                char __q04, char __q05, char __q06, char __q07,
 680                char __q08, char __q09, char __q10, char __q11,
 681                char __q12, char __q13, char __q14, char __q15)
 682 {
 683   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 684                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 685 }
 686
 687 /* Create a vector with element 0 as *P and the rest zero.  */
 688
 689 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 690 _mm_load_si128 (__m128i const *__P)
 691 {
 692   return *__P;
 693 }
 694
 695 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 696 _mm_loadu_si128 (__m128i const *__P)
 697 {
 698   return (__m128i) __builtin_ia32_loaddqu ((char const *)__P);
 699 }
 700
 701 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 702 _mm_loadl_epi64 (__m128i const *__P)
 703 {
 704   return _mm_set_epi64 ((__m64)0LL, *(__m64 *)__P);
 705 }
 706
 707 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 708 _mm_store_si128 (__m128i *__P, __m128i __B)
 709 {
 710   *__P = __B;
 711 }
 712
 713 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 714 _mm_storeu_si128 (__m128i *__P, __m128i __B)
 715 {
 716   __builtin_ia32_storedqu ((char *)__P, (__v16qi)__B);
 717 }
 718
 719 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 720 _mm_storel_epi64 (__m128i *__P, __m128i __B)
 721 {
 722   *(long long *)__P = ((__v2di)__B)[0];
 723 }
 724
 725 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 726 _mm_movepi64_pi64 (__m128i __B)
 727 {
 728   return (__m64) ((__v2di)__B)[0];
 729 }
 730
 731 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 732 _mm_movpi64_epi64 (__m64 __A)
 733 {
 734   return _mm_set_epi64 ((__m64)0LL, __A);
 735 }
 736
 737 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 738 _mm_move_epi64 (__m128i __A)
 739 {
 740   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 741 }
 742
 743 /* Create an undefined vector.  */
 744 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 745 _mm_undefined_si128 (void)
 746 {
 747   __m128i __Y = __Y;
 748   return __Y;
 749 }
 750
 751 /* Create a vector of zeros.  */
 752 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 753 _mm_setzero_si128 (void)
 754 {
 755   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 756 }
 757
 758 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 759 _mm_cvtepi32_pd (__m128i __A)
 760 {
 761   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 762 }
 763
 764 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 765 _mm_cvtepi32_ps (__m128i __A)
 766 {
 767   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 768 }
 769
 770 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 771 _mm_cvtpd_epi32 (__m128d __A)
 772 {
 773   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 774 }
 775
 776 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 777 _mm_cvtpd_pi32 (__m128d __A)
 778 {
 779   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 780 }
 781
 782 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 783 _mm_cvtpd_ps (__m128d __A)
 784 {
 785   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 786 }
 787
 788 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 789 _mm_cvttpd_epi32 (__m128d __A)
 790 {
 791   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 792 }
 793
 794 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 795 _mm_cvttpd_pi32 (__m128d __A)
 796 {
 797   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 798 }
 799
 800 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 801 _mm_cvtpi32_pd (__m64 __A)
 802 {
 803   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 804 }
 805
 806 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 807 _mm_cvtps_epi32 (__m128 __A)
 808 {
 809   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 810 }
 811
 812 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 813 _mm_cvttps_epi32 (__m128 __A)
 814 {
 815   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 816 }
 817
 818 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 819 _mm_cvtps_pd (__m128 __A)
 820 {
 821   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 822 }
 823
 824 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 825 _mm_cvtsd_si32 (__m128d __A)
 826 {
 827   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 828 }
 829
 830 #ifdef __x86_64__
 831 /* Intel intrinsic.  */
 832 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 833 _mm_cvtsd_si64 (__m128d __A)
 834 {
 835   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 836 }
 837
 838 /* Microsoft intrinsic.  */
 839 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 840 _mm_cvtsd_si64x (__m128d __A)
 841 {
 842   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 843 }
 844 #endif
 845
 846 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 847 _mm_cvttsd_si32 (__m128d __A)
 848 {
 849   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 850 }
 851
 852 #ifdef __x86_64__
 853 /* Intel intrinsic.  */
 854 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 855 _mm_cvttsd_si64 (__m128d __A)
 856 {
 857   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 858 }
 859
 860 /* Microsoft intrinsic.  */
 861 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 862 _mm_cvttsd_si64x (__m128d __A)
 863 {
 864   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 865 }
 866 #endif
 867
 868 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 869 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 870 {
 871   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 872 }
 873
 874 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 875 _mm_cvtsi32_sd (__m128d __A, int __B)
 876 {
 877   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 878 }
 879
 880 #ifdef __x86_64__
 881 /* Intel intrinsic.  */
 882 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 883 _mm_cvtsi64_sd (__m128d __A, long long __B)
 884 {
 885   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 886 }
 887
 888 /* Microsoft intrinsic.  */
 889 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 890 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 891 {
 892   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 893 }
 894 #endif
 895
 896 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 897 _mm_cvtss_sd (__m128d __A, __m128 __B)
 898 {
 899   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 900 }
 901
 902 #ifdef __OPTIMIZE__
 903 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 904 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 905 {
 906   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 907 }
 908 #else
 909 #define _mm_shuffle_pd(A, B, N)                                         \
 910   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 911                                    (__v2df)(__m128d)(B), (int)(N)))
 912 #endif
 913
 914 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 915 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 916 {
 917   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 918 }
 919
 920 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 921 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 922 {
 923   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 924 }
 925
 926 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 927 _mm_loadh_pd (__m128d __A, double const *__B)
 928 {
 929   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 930 }
 931
 932 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 933 _mm_loadl_pd (__m128d __A, double const *__B)
 934 {
 935   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 936 }
 937
 938 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 939 _mm_movemask_pd (__m128d __A)
 940 {
 941   return __builtin_ia32_movmskpd ((__v2df)__A);
 942 }
 943
 944 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 945 _mm_packs_epi16 (__m128i __A, __m128i __B)
 946 {
 947   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 948 }
 949
 950 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 951 _mm_packs_epi32 (__m128i __A, __m128i __B)
 952 {
 953   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 954 }
 955
 956 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 957 _mm_packus_epi16 (__m128i __A, __m128i __B)
 958 {
 959   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 960 }
 961
 962 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 963 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 964 {
 965   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 966 }
 967
 968 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 969 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 970 {
 971   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 972 }
 973
 974 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 975 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 976 {
 977   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 978 }
 979
 980 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 981 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 982 {
 983   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 984 }
 985
 986 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 987 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 988 {
 989   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 990 }
 991
 992 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 993 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 994 {
 995   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
 996 }
 997
 998 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 999 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
1000 {
1001   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
1002 }
1003
1004 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1005 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
1006 {
1007   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
1008 }
1009
1010 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1011 _mm_add_epi8 (__m128i __A, __m128i __B)
1012 {
1013   return (__m128i) ((__v16qu)__A + (__v16qu)__B);
1014 }
1015
1016 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1017 _mm_add_epi16 (__m128i __A, __m128i __B)
1018 {
1019   return (__m128i) ((__v8hu)__A + (__v8hu)__B);
1020 }
1021
1022 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1023 _mm_add_epi32 (__m128i __A, __m128i __B)
1024 {
1025   return (__m128i) ((__v4su)__A + (__v4su)__B);
1026 }
1027
1028 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1029 _mm_add_epi64 (__m128i __A, __m128i __B)
1030 {
1031   return (__m128i) ((__v2du)__A + (__v2du)__B);
1032 }
1033
1034 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1035 _mm_adds_epi8 (__m128i __A, __m128i __B)
1036 {
1037   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1038 }
1039
1040 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1041 _mm_adds_epi16 (__m128i __A, __m128i __B)
1042 {
1043   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1044 }
1045
1046 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1047 _mm_adds_epu8 (__m128i __A, __m128i __B)
1048 {
1049   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1050 }
1051
1052 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1053 _mm_adds_epu16 (__m128i __A, __m128i __B)
1054 {
1055   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1056 }
1057
1058 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1059 _mm_sub_epi8 (__m128i __A, __m128i __B)
1060 {
1061   return (__m128i) ((__v16qu)__A - (__v16qu)__B);
1062 }
1063
1064 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1065 _mm_sub_epi16 (__m128i __A, __m128i __B)
1066 {
1067   return (__m128i) ((__v8hu)__A - (__v8hu)__B);
1068 }
1069
1070 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1071 _mm_sub_epi32 (__m128i __A, __m128i __B)
1072 {
1073   return (__m128i) ((__v4su)__A - (__v4su)__B);
1074 }
1075
1076 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1077 _mm_sub_epi64 (__m128i __A, __m128i __B)
1078 {
1079   return (__m128i) ((__v2du)__A - (__v2du)__B);
1080 }
1081
1082 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1083 _mm_subs_epi8 (__m128i __A, __m128i __B)
1084 {
1085   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1086 }
1087
1088 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1089 _mm_subs_epi16 (__m128i __A, __m128i __B)
1090 {
1091   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1092 }
1093
1094 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1095 _mm_subs_epu8 (__m128i __A, __m128i __B)
1096 {
1097   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1098 }
1099
1100 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1101 _mm_subs_epu16 (__m128i __A, __m128i __B)
1102 {
1103   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1104 }
1105
1106 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1107 _mm_madd_epi16 (__m128i __A, __m128i __B)
1108 {
1109   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1110 }
1111
1112 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1113 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1114 {
1115   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1116 }
1117
1118 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1119 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1120 {
1121   return (__m128i) ((__v8hu)__A * (__v8hu)__B);
1122 }
1123
1124 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1125 _mm_mul_su32 (__m64 __A, __m64 __B)
1126 {
1127   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1128 }
1129
1130 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1131 _mm_mul_epu32 (__m128i __A, __m128i __B)
1132 {
1133   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1134 }
1135
1136 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1137 _mm_slli_epi16 (__m128i __A, int __B)
1138 {
1139   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1140 }
1141
1142 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1143 _mm_slli_epi32 (__m128i __A, int __B)
1144 {
1145   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1146 }
1147
1148 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1149 _mm_slli_epi64 (__m128i __A, int __B)
1150 {
1151   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1152 }
1153
1154 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1155 _mm_srai_epi16 (__m128i __A, int __B)
1156 {
1157   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1158 }
1159
1160 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1161 _mm_srai_epi32 (__m128i __A, int __B)
1162 {
1163   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1164 }
1165
1166 #ifdef __OPTIMIZE__
1167 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1168 _mm_srli_si128 (__m128i __A, const int __N)
1169 {
1170   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1171 }
1172
1173 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1174 _mm_slli_si128 (__m128i __A, const int __N)
1175 {
1176   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1177 }
1178 #else
1179 #define _mm_srli_si128(A, N) \
1180   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1181 #define _mm_slli_si128(A, N) \
1182   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1183 #endif
1184
1185 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1186 _mm_srli_epi16 (__m128i __A, int __B)
1187 {
1188   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1189 }
1190
1191 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1192 _mm_srli_epi32 (__m128i __A, int __B)
1193 {
1194   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1195 }
1196
1197 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1198 _mm_srli_epi64 (__m128i __A, int __B)
1199 {
1200   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1201 }
1202
1203 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1204 _mm_sll_epi16 (__m128i __A, __m128i __B)
1205 {
1206   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1207 }
1208
1209 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1210 _mm_sll_epi32 (__m128i __A, __m128i __B)
1211 {
1212   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1213 }
1214
1215 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1216 _mm_sll_epi64 (__m128i __A, __m128i __B)
1217 {
1218   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1219 }
1220
1221 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1222 _mm_sra_epi16 (__m128i __A, __m128i __B)
1223 {
1224   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1225 }
1226
1227 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1228 _mm_sra_epi32 (__m128i __A, __m128i __B)
1229 {
1230   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1231 }
1232
1233 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1234 _mm_srl_epi16 (__m128i __A, __m128i __B)
1235 {
1236   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1237 }
1238
1239 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1240 _mm_srl_epi32 (__m128i __A, __m128i __B)
1241 {
1242   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1243 }
1244
1245 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1246 _mm_srl_epi64 (__m128i __A, __m128i __B)
1247 {
1248   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1249 }
1250
1251 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1252 _mm_and_si128 (__m128i __A, __m128i __B)
1253 {
1254   return (__m128i) ((__v2du)__A & (__v2du)__B);
1255 }
1256
1257 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1258 _mm_andnot_si128 (__m128i __A, __m128i __B)
1259 {
1260   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1261 }
1262
1263 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1264 _mm_or_si128 (__m128i __A, __m128i __B)
1265 {
1266   return (__m128i) ((__v2du)__A | (__v2du)__B);
1267 }
1268
1269 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1270 _mm_xor_si128 (__m128i __A, __m128i __B)
1271 {
1272   return (__m128i) ((__v2du)__A ^ (__v2du)__B);
1273 }
1274
1275 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1276 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1277 {
1278   return (__m128i) ((__v16qi)__A == (__v16qi)__B);
1279 }
1280
1281 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1282 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1283 {
1284   return (__m128i) ((__v8hi)__A == (__v8hi)__B);
1285 }
1286
1287 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1288 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1289 {
1290   return (__m128i) ((__v4si)__A == (__v4si)__B);
1291 }
1292
1293 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1294 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1295 {
1296   return (__m128i) ((__v16qi)__A < (__v16qi)__B);
1297 }
1298
1299 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1300 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1301 {
1302   return (__m128i) ((__v8hi)__A < (__v8hi)__B);
1303 }
1304
1305 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1306 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1307 {
1308   return (__m128i) ((__v4si)__A < (__v4si)__B);
1309 }
1310
1311 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1312 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1313 {
1314   return (__m128i) ((__v16qi)__A > (__v16qi)__B);
1315 }
1316
1317 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1318 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1319 {
1320   return (__m128i) ((__v8hi)__A > (__v8hi)__B);
1321 }
1322
1323 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1324 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1325 {
1326   return (__m128i) ((__v4si)__A > (__v4si)__B);
1327 }
1328
1329 #ifdef __OPTIMIZE__
1330 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1331 _mm_extract_epi16 (__m128i const __A, int const __N)
1332 {
1333   return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1334 }
1335
1336 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1337 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1338 {
1339   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1340 }
1341 #else
1342 #define _mm_extract_epi16(A, N) \
1343   ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1344 #define _mm_insert_epi16(A, D, N)                               \
1345   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1346                                           (int)(D), (int)(N)))
1347 #endif
1348
1349 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1350 _mm_max_epi16 (__m128i __A, __m128i __B)
1351 {
1352   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1353 }
1354
1355 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1356 _mm_max_epu8 (__m128i __A, __m128i __B)
1357 {
1358   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1359 }
1360
1361 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1362 _mm_min_epi16 (__m128i __A, __m128i __B)
1363 {
1364   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1365 }
1366
1367 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1368 _mm_min_epu8 (__m128i __A, __m128i __B)
1369 {
1370   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1371 }
1372
1373 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1374 _mm_movemask_epi8 (__m128i __A)
1375 {
1376   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1377 }
1378
1379 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1380 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1381 {
1382   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1383 }
1384
1385 #ifdef __OPTIMIZE__
1386 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1387 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1388 {
1389   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1390 }
1391
1392 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1393 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1394 {
1395   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1396 }
1397
1398 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1399 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1400 {
1401   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1402 }
1403 #else
1404 #define _mm_shufflehi_epi16(A, N) \
1405   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1406 #define _mm_shufflelo_epi16(A, N) \
1407   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1408 #define _mm_shuffle_epi32(A, N) \
1409   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1410 #endif
1411
1412 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1413 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1414 {
1415   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1416 }
1417
1418 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1419 _mm_avg_epu8 (__m128i __A, __m128i __B)
1420 {
1421   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1422 }
1423
1424 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1425 _mm_avg_epu16 (__m128i __A, __m128i __B)
1426 {
1427   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1428 }
1429
1430 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1431 _mm_sad_epu8 (__m128i __A, __m128i __B)
1432 {
1433   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1434 }
1435
1436 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1437 _mm_stream_si32 (int *__A, int __B)
1438 {
1439   __builtin_ia32_movnti (__A, __B);
1440 }
1441
1442 #ifdef __x86_64__
1443 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1444 _mm_stream_si64 (long long int *__A, long long int __B)
1445 {
1446   __builtin_ia32_movnti64 (__A, __B);
1447 }
1448 #endif
1449
1450 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1451 _mm_stream_si128 (__m128i *__A, __m128i __B)
1452 {
1453   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1454 }
1455
1456 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1457 _mm_stream_pd (double *__A, __m128d __B)
1458 {
1459   __builtin_ia32_movntpd (__A, (__v2df)__B);
1460 }
1461
1462 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1463 _mm_clflush (void const *__A)
1464 {
1465   __builtin_ia32_clflush (__A);
1466 }
1467
1468 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1469 _mm_lfence (void)
1470 {
1471   __builtin_ia32_lfence ();
1472 }
1473
1474 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1475 _mm_mfence (void)
1476 {
1477   __builtin_ia32_mfence ();
1478 }
1479
1480 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1481 _mm_cvtsi32_si128 (int __A)
1482 {
1483   return _mm_set_epi32 (0, 0, 0, __A);
1484 }
1485
1486 #ifdef __x86_64__
1487 /* Intel intrinsic.  */
1488 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1489 _mm_cvtsi64_si128 (long long __A)
1490 {
1491   return _mm_set_epi64x (0, __A);
1492 }
1493
1494 /* Microsoft intrinsic.  */
1495 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1496 _mm_cvtsi64x_si128 (long long __A)
1497 {
1498   return _mm_set_epi64x (0, __A);
1499 }
1500 #endif
1501
1502 /* Casts between various SP, DP, INT vector types.  Note that these do no
1503    conversion of values, they just change the type.  */
1504 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1505 _mm_castpd_ps(__m128d __A)
1506 {
1507   return (__m128) __A;
1508 }
1509
1510 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1511 _mm_castpd_si128(__m128d __A)
1512 {
1513   return (__m128i) __A;
1514 }
1515
1516 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1517 _mm_castps_pd(__m128 __A)
1518 {
1519   return (__m128d) __A;
1520 }
1521
1522 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1523 _mm_castps_si128(__m128 __A)
1524 {
1525   return (__m128i) __A;
1526 }
1527
1528 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1529 _mm_castsi128_ps(__m128i __A)
1530 {
1531   return (__m128) __A;
1532 }
1533
1534 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1535 _mm_castsi128_pd(__m128i __A)
1536 {
1537   return (__m128d) __A;
1538 }
1539
1540 #ifdef __DISABLE_SSE2__
1541 #undef __DISABLE_SSE2__
1542 #pragma GCC pop_options
1543 #endif /* __DISABLE_SSE2__ */
1544
1545 #endif /* _EMMINTRIN_H_INCLUDED */