gcc/config/i386/avx512vlintrin.h

   1 /* Copyright (C) 2014-2018 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 #ifndef _IMMINTRIN_H_INCLUDED
  25 #error "Never use <avx512vlintrin.h> directly; include <immintrin.h> instead."
  26 #endif
  27
  28 #ifndef _AVX512VLINTRIN_H_INCLUDED
  29 #define _AVX512VLINTRIN_H_INCLUDED
  30
  31 #ifndef __AVX512VL__
  32 #pragma GCC push_options
  33 #pragma GCC target("avx512vl")
  34 #define __DISABLE_AVX512VL__
  35 #endif /* __AVX512VL__ */
  36
  37 /* Internal data types for implementing the intrinsics.  */
  38 typedef unsigned int __mmask32;
  39
  40 extern __inline __m256d
  41 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  42 _mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A)
  43 {
  44   return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A,
  45                                                   (__v4df) __W,
  46                                                   (__mmask8) __U);
  47 }
  48
  49 extern __inline __m256d
  50 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  51 _mm256_maskz_mov_pd (__mmask8 __U, __m256d __A)
  52 {
  53   return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A,
  54                                                   (__v4df)
  55                                                   _mm256_setzero_pd (),
  56                                                   (__mmask8) __U);
  57 }
  58
  59 extern __inline __m128d
  60 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  61 _mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A)
  62 {
  63   return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A,
  64                                                   (__v2df) __W,
  65                                                   (__mmask8) __U);
  66 }
  67
  68 extern __inline __m128d
  69 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  70 _mm_maskz_mov_pd (__mmask8 __U, __m128d __A)
  71 {
  72   return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A,
  73                                                   (__v2df)
  74                                                   _mm_setzero_pd (),
  75                                                   (__mmask8) __U);
  76 }
  77
  78 extern __inline __m256d
  79 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  80 _mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P)
  81 {
  82   return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
  83                                                    (__v4df) __W,
  84                                                    (__mmask8) __U);
  85 }
  86
  87 extern __inline __m256d
  88 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  89 _mm256_maskz_load_pd (__mmask8 __U, void const *__P)
  90 {
  91   return (__m256d) __builtin_ia32_loadapd256_mask ((__v4df *) __P,
  92                                                    (__v4df)
  93                                                    _mm256_setzero_pd (),
  94                                                    (__mmask8) __U);
  95 }
  96
  97 extern __inline __m128d
  98 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  99 _mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P)
 100 {
 101   return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
 102                                                    (__v2df) __W,
 103                                                    (__mmask8) __U);
 104 }
 105
 106 extern __inline __m128d
 107 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 108 _mm_maskz_load_pd (__mmask8 __U, void const *__P)
 109 {
 110   return (__m128d) __builtin_ia32_loadapd128_mask ((__v2df *) __P,
 111                                                    (__v2df)
 112                                                    _mm_setzero_pd (),
 113                                                    (__mmask8) __U);
 114 }
 115
 116 extern __inline void
 117 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 118 _mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A)
 119 {
 120   __builtin_ia32_storeapd256_mask ((__v4df *) __P,
 121                                    (__v4df) __A,
 122                                    (__mmask8) __U);
 123 }
 124
 125 extern __inline void
 126 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 127 _mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A)
 128 {
 129   __builtin_ia32_storeapd128_mask ((__v2df *) __P,
 130                                    (__v2df) __A,
 131                                    (__mmask8) __U);
 132 }
 133
 134 extern __inline __m256
 135 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 136 _mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A)
 137 {
 138   return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A,
 139                                                  (__v8sf) __W,
 140                                                  (__mmask8) __U);
 141 }
 142
 143 extern __inline __m256
 144 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 145 _mm256_maskz_mov_ps (__mmask8 __U, __m256 __A)
 146 {
 147   return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A,
 148                                                  (__v8sf)
 149                                                  _mm256_setzero_ps (),
 150                                                  (__mmask8) __U);
 151 }
 152
 153 extern __inline __m128
 154 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 155 _mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A)
 156 {
 157   return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A,
 158                                                  (__v4sf) __W,
 159                                                  (__mmask8) __U);
 160 }
 161
 162 extern __inline __m128
 163 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 164 _mm_maskz_mov_ps (__mmask8 __U, __m128 __A)
 165 {
 166   return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A,
 167                                                  (__v4sf)
 168                                                  _mm_setzero_ps (),
 169                                                  (__mmask8) __U);
 170 }
 171
 172 extern __inline __m256
 173 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 174 _mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P)
 175 {
 176   return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
 177                                                   (__v8sf) __W,
 178                                                   (__mmask8) __U);
 179 }
 180
 181 extern __inline __m256
 182 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 183 _mm256_maskz_load_ps (__mmask8 __U, void const *__P)
 184 {
 185   return (__m256) __builtin_ia32_loadaps256_mask ((__v8sf *) __P,
 186                                                   (__v8sf)
 187                                                   _mm256_setzero_ps (),
 188                                                   (__mmask8) __U);
 189 }
 190
 191 extern __inline __m128
 192 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 193 _mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P)
 194 {
 195   return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
 196                                                   (__v4sf) __W,
 197                                                   (__mmask8) __U);
 198 }
 199
 200 extern __inline __m128
 201 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 202 _mm_maskz_load_ps (__mmask8 __U, void const *__P)
 203 {
 204   return (__m128) __builtin_ia32_loadaps128_mask ((__v4sf *) __P,
 205                                                   (__v4sf)
 206                                                   _mm_setzero_ps (),
 207                                                   (__mmask8) __U);
 208 }
 209
 210 extern __inline void
 211 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 212 _mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A)
 213 {
 214   __builtin_ia32_storeaps256_mask ((__v8sf *) __P,
 215                                    (__v8sf) __A,
 216                                    (__mmask8) __U);
 217 }
 218
 219 extern __inline void
 220 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 221 _mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A)
 222 {
 223   __builtin_ia32_storeaps128_mask ((__v4sf *) __P,
 224                                    (__v4sf) __A,
 225                                    (__mmask8) __U);
 226 }
 227
 228 extern __inline __m256i
 229 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 230 _mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
 231 {
 232   return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
 233                                                      (__v4di) __W,
 234                                                      (__mmask8) __U);
 235 }
 236
 237 extern __inline __m256i
 238 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 239 _mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A)
 240 {
 241   return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A,
 242                                                      (__v4di)
 243                                                      _mm256_setzero_si256 (),
 244                                                      (__mmask8) __U);
 245 }
 246
 247 extern __inline __m128i
 248 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 249 _mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
 250 {
 251   return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
 252                                                      (__v2di) __W,
 253                                                      (__mmask8) __U);
 254 }
 255
 256 extern __inline __m128i
 257 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 258 _mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A)
 259 {
 260   return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A,
 261                                                      (__v2di)
 262                                                      _mm_setzero_si128 (),
 263                                                      (__mmask8) __U);
 264 }
 265
 266 extern __inline __m256i
 267 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 268 _mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P)
 269 {
 270   return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
 271                                                         (__v4di) __W,
 272                                                         (__mmask8)
 273                                                         __U);
 274 }
 275
 276 extern __inline __m256i
 277 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 278 _mm256_maskz_load_epi64 (__mmask8 __U, void const *__P)
 279 {
 280   return (__m256i) __builtin_ia32_movdqa64load256_mask ((__v4di *) __P,
 281                                                         (__v4di)
 282                                                         _mm256_setzero_si256 (),
 283                                                         (__mmask8)
 284                                                         __U);
 285 }
 286
 287 extern __inline __m128i
 288 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 289 _mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P)
 290 {
 291   return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
 292                                                         (__v2di) __W,
 293                                                         (__mmask8)
 294                                                         __U);
 295 }
 296
 297 extern __inline __m128i
 298 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 299 _mm_maskz_load_epi64 (__mmask8 __U, void const *__P)
 300 {
 301   return (__m128i) __builtin_ia32_movdqa64load128_mask ((__v2di *) __P,
 302                                                         (__v2di)
 303                                                         _mm_setzero_si128 (),
 304                                                         (__mmask8)
 305                                                         __U);
 306 }
 307
 308 extern __inline void
 309 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 310 _mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A)
 311 {
 312   __builtin_ia32_movdqa64store256_mask ((__v4di *) __P,
 313                                         (__v4di) __A,
 314                                         (__mmask8) __U);
 315 }
 316
 317 extern __inline void
 318 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 319 _mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A)
 320 {
 321   __builtin_ia32_movdqa64store128_mask ((__v2di *) __P,
 322                                         (__v2di) __A,
 323                                         (__mmask8) __U);
 324 }
 325
 326 extern __inline __m256i
 327 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 328 _mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
 329 {
 330   return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A,
 331                                                      (__v8si) __W,
 332                                                      (__mmask8) __U);
 333 }
 334
 335 extern __inline __m256i
 336 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 337 _mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A)
 338 {
 339   return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A,
 340                                                      (__v8si)
 341                                                      _mm256_setzero_si256 (),
 342                                                      (__mmask8) __U);
 343 }
 344
 345 extern __inline __m128i
 346 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 347 _mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
 348 {
 349   return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A,
 350                                                      (__v4si) __W,
 351                                                      (__mmask8) __U);
 352 }
 353
 354 extern __inline __m128i
 355 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 356 _mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A)
 357 {
 358   return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A,
 359                                                      (__v4si)
 360                                                      _mm_setzero_si128 (),
 361                                                      (__mmask8) __U);
 362 }
 363
 364 extern __inline __m256i
 365 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 366 _mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P)
 367 {
 368   return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
 369                                                         (__v8si) __W,
 370                                                         (__mmask8)
 371                                                         __U);
 372 }
 373
 374 extern __inline __m256i
 375 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 376 _mm256_maskz_load_epi32 (__mmask8 __U, void const *__P)
 377 {
 378   return (__m256i) __builtin_ia32_movdqa32load256_mask ((__v8si *) __P,
 379                                                         (__v8si)
 380                                                         _mm256_setzero_si256 (),
 381                                                         (__mmask8)
 382                                                         __U);
 383 }
 384
 385 extern __inline __m128i
 386 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 387 _mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P)
 388 {
 389   return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
 390                                                         (__v4si) __W,
 391                                                         (__mmask8)
 392                                                         __U);
 393 }
 394
 395 extern __inline __m128i
 396 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 397 _mm_maskz_load_epi32 (__mmask8 __U, void const *__P)
 398 {
 399   return (__m128i) __builtin_ia32_movdqa32load128_mask ((__v4si *) __P,
 400                                                         (__v4si)
 401                                                         _mm_setzero_si128 (),
 402                                                         (__mmask8)
 403                                                         __U);
 404 }
 405
 406 extern __inline void
 407 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 408 _mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A)
 409 {
 410   __builtin_ia32_movdqa32store256_mask ((__v8si *) __P,
 411                                         (__v8si) __A,
 412                                         (__mmask8) __U);
 413 }
 414
 415 extern __inline void
 416 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 417 _mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A)
 418 {
 419   __builtin_ia32_movdqa32store128_mask ((__v4si *) __P,
 420                                         (__v4si) __A,
 421                                         (__mmask8) __U);
 422 }
 423
 424 extern __inline __m128d
 425 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 426 _mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
 427 {
 428   return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
 429                                                  (__v2df) __B,
 430                                                  (__v2df) __W,
 431                                                  (__mmask8) __U);
 432 }
 433
 434 extern __inline __m128d
 435 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 436 _mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B)
 437 {
 438   return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A,
 439                                                  (__v2df) __B,
 440                                                  (__v2df)
 441                                                  _mm_setzero_pd (),
 442                                                  (__mmask8) __U);
 443 }
 444
 445 extern __inline __m256d
 446 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 447 _mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A,
 448                     __m256d __B)
 449 {
 450   return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
 451                                                  (__v4df) __B,
 452                                                  (__v4df) __W,
 453                                                  (__mmask8) __U);
 454 }
 455
 456 extern __inline __m256d
 457 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 458 _mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B)
 459 {
 460   return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A,
 461                                                  (__v4df) __B,
 462                                                  (__v4df)
 463                                                  _mm256_setzero_pd (),
 464                                                  (__mmask8) __U);
 465 }
 466
 467 extern __inline __m128
 468 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 469 _mm_mask_add_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
 470 {
 471   return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
 472                                                 (__v4sf) __B,
 473                                                 (__v4sf) __W,
 474                                                 (__mmask8) __U);
 475 }
 476
 477 extern __inline __m128
 478 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 479 _mm_maskz_add_ps (__mmask8 __U, __m128 __A, __m128 __B)
 480 {
 481   return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A,
 482                                                 (__v4sf) __B,
 483                                                 (__v4sf)
 484                                                 _mm_setzero_ps (),
 485                                                 (__mmask8) __U);
 486 }
 487
 488 extern __inline __m256
 489 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 490 _mm256_mask_add_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
 491 {
 492   return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
 493                                                 (__v8sf) __B,
 494                                                 (__v8sf) __W,
 495                                                 (__mmask8) __U);
 496 }
 497
 498 extern __inline __m256
 499 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 500 _mm256_maskz_add_ps (__mmask8 __U, __m256 __A, __m256 __B)
 501 {
 502   return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A,
 503                                                 (__v8sf) __B,
 504                                                 (__v8sf)
 505                                                 _mm256_setzero_ps (),
 506                                                 (__mmask8) __U);
 507 }
 508
 509 extern __inline __m128d
 510 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 511 _mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
 512 {
 513   return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
 514                                                  (__v2df) __B,
 515                                                  (__v2df) __W,
 516                                                  (__mmask8) __U);
 517 }
 518
 519 extern __inline __m128d
 520 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 521 _mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B)
 522 {
 523   return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A,
 524                                                  (__v2df) __B,
 525                                                  (__v2df)
 526                                                  _mm_setzero_pd (),
 527                                                  (__mmask8) __U);
 528 }
 529
 530 extern __inline __m256d
 531 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 532 _mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A,
 533                     __m256d __B)
 534 {
 535   return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
 536                                                  (__v4df) __B,
 537                                                  (__v4df) __W,
 538                                                  (__mmask8) __U);
 539 }
 540
 541 extern __inline __m256d
 542 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 543 _mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B)
 544 {
 545   return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A,
 546                                                  (__v4df) __B,
 547                                                  (__v4df)
 548                                                  _mm256_setzero_pd (),
 549                                                  (__mmask8) __U);
 550 }
 551
 552 extern __inline __m128
 553 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 554 _mm_mask_sub_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
 555 {
 556   return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
 557                                                 (__v4sf) __B,
 558                                                 (__v4sf) __W,
 559                                                 (__mmask8) __U);
 560 }
 561
 562 extern __inline __m128
 563 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 564 _mm_maskz_sub_ps (__mmask8 __U, __m128 __A, __m128 __B)
 565 {
 566   return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A,
 567                                                 (__v4sf) __B,
 568                                                 (__v4sf)
 569                                                 _mm_setzero_ps (),
 570                                                 (__mmask8) __U);
 571 }
 572
 573 extern __inline __m256
 574 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 575 _mm256_mask_sub_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
 576 {
 577   return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
 578                                                 (__v8sf) __B,
 579                                                 (__v8sf) __W,
 580                                                 (__mmask8) __U);
 581 }
 582
 583 extern __inline __m256
 584 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 585 _mm256_maskz_sub_ps (__mmask8 __U, __m256 __A, __m256 __B)
 586 {
 587   return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A,
 588                                                 (__v8sf) __B,
 589                                                 (__v8sf)
 590                                                 _mm256_setzero_ps (),
 591                                                 (__mmask8) __U);
 592 }
 593
 594 extern __inline void
 595 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 596 _mm256_store_epi64 (void *__P, __m256i __A)
 597 {
 598   *(__m256i *) __P = __A;
 599 }
 600
 601 extern __inline void
 602 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 603 _mm_store_epi64 (void *__P, __m128i __A)
 604 {
 605   *(__m128i *) __P = __A;
 606 }
 607
 608 extern __inline __m256d
 609 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 610 _mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P)
 611 {
 612   return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P,
 613                                                    (__v4df) __W,
 614                                                    (__mmask8) __U);
 615 }
 616
 617 extern __inline __m256d
 618 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 619 _mm256_maskz_loadu_pd (__mmask8 __U, void const *__P)
 620 {
 621   return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P,
 622                                                    (__v4df)
 623                                                    _mm256_setzero_pd (),
 624                                                    (__mmask8) __U);
 625 }
 626
 627 extern __inline __m128d
 628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 629 _mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P)
 630 {
 631   return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P,
 632                                                    (__v2df) __W,
 633                                                    (__mmask8) __U);
 634 }
 635
 636 extern __inline __m128d
 637 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 638 _mm_maskz_loadu_pd (__mmask8 __U, void const *__P)
 639 {
 640   return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P,
 641                                                    (__v2df)
 642                                                    _mm_setzero_pd (),
 643                                                    (__mmask8) __U);
 644 }
 645
 646 extern __inline void
 647 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 648 _mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A)
 649 {
 650   __builtin_ia32_storeupd256_mask ((double *) __P,
 651                                    (__v4df) __A,
 652                                    (__mmask8) __U);
 653 }
 654
 655 extern __inline void
 656 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 657 _mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A)
 658 {
 659   __builtin_ia32_storeupd128_mask ((double *) __P,
 660                                    (__v2df) __A,
 661                                    (__mmask8) __U);
 662 }
 663
 664 extern __inline __m256
 665 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 666 _mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P)
 667 {
 668   return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P,
 669                                                   (__v8sf) __W,
 670                                                   (__mmask8) __U);
 671 }
 672
 673 extern __inline __m256
 674 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 675 _mm256_maskz_loadu_ps (__mmask8 __U, void const *__P)
 676 {
 677   return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P,
 678                                                   (__v8sf)
 679                                                   _mm256_setzero_ps (),
 680                                                   (__mmask8) __U);
 681 }
 682
 683 extern __inline __m128
 684 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 685 _mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P)
 686 {
 687   return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P,
 688                                                   (__v4sf) __W,
 689                                                   (__mmask8) __U);
 690 }
 691
 692 extern __inline __m128
 693 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 694 _mm_maskz_loadu_ps (__mmask8 __U, void const *__P)
 695 {
 696   return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P,
 697                                                   (__v4sf)
 698                                                   _mm_setzero_ps (),
 699                                                   (__mmask8) __U);
 700 }
 701
 702 extern __inline void
 703 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 704 _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
 705 {
 706   __builtin_ia32_storeups256_mask ((float *) __P,
 707                                    (__v8sf) __A,
 708                                    (__mmask8) __U);
 709 }
 710
 711 extern __inline void
 712 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 713 _mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A)
 714 {
 715   __builtin_ia32_storeups128_mask ((float *) __P,
 716                                    (__v4sf) __A,
 717                                    (__mmask8) __U);
 718 }
 719
 720 extern __inline __m256i
 721 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 722 _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P)
 723 {
 724   return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P,
 725                                                      (__v4di) __W,
 726                                                      (__mmask8) __U);
 727 }
 728
 729 extern __inline __m256i
 730 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 731 _mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
 732 {
 733   return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P,
 734                                                      (__v4di)
 735                                                      _mm256_setzero_si256 (),
 736                                                      (__mmask8) __U);
 737 }
 738
 739 extern __inline __m128i
 740 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 741 _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
 742 {
 743   return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P,
 744                                                      (__v2di) __W,
 745                                                      (__mmask8) __U);
 746 }
 747
 748 extern __inline __m128i
 749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 750 _mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P)
 751 {
 752   return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P,
 753                                                      (__v2di)
 754                                                      _mm_setzero_si128 (),
 755                                                      (__mmask8) __U);
 756 }
 757
 758 extern __inline void
 759 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 760 _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A)
 761 {
 762   __builtin_ia32_storedqudi256_mask ((long long *) __P,
 763                                      (__v4di) __A,
 764                                      (__mmask8) __U);
 765 }
 766
 767 extern __inline void
 768 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 769 _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A)
 770 {
 771   __builtin_ia32_storedqudi128_mask ((long long *) __P,
 772                                      (__v2di) __A,
 773                                      (__mmask8) __U);
 774 }
 775
 776 extern __inline __m256i
 777 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 778 _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P)
 779 {
 780   return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P,
 781                                                      (__v8si) __W,
 782                                                      (__mmask8) __U);
 783 }
 784
 785 extern __inline __m256i
 786 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 787 _mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
 788 {
 789   return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P,
 790                                                      (__v8si)
 791                                                      _mm256_setzero_si256 (),
 792                                                      (__mmask8) __U);
 793 }
 794
 795 extern __inline __m128i
 796 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 797 _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
 798 {
 799   return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P,
 800                                                      (__v4si) __W,
 801                                                      (__mmask8) __U);
 802 }
 803
 804 extern __inline __m128i
 805 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 806 _mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P)
 807 {
 808   return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P,
 809                                                      (__v4si)
 810                                                      _mm_setzero_si128 (),
 811                                                      (__mmask8) __U);
 812 }
 813
 814 extern __inline void
 815 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 816 _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A)
 817 {
 818   __builtin_ia32_storedqusi256_mask ((int *) __P,
 819                                      (__v8si) __A,
 820                                      (__mmask8) __U);
 821 }
 822
 823 extern __inline void
 824 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 825 _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A)
 826 {
 827   __builtin_ia32_storedqusi128_mask ((int *) __P,
 828                                      (__v4si) __A,
 829                                      (__mmask8) __U);
 830 }
 831
 832 extern __inline __m256i
 833 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 834 _mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
 835 {
 836   return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
 837                                                  (__v8si) __W,
 838                                                  (__mmask8) __U);
 839 }
 840
 841 extern __inline __m256i
 842 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 843 _mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A)
 844 {
 845   return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A,
 846                                                  (__v8si)
 847                                                  _mm256_setzero_si256 (),
 848                                                  (__mmask8) __U);
 849 }
 850
 851 extern __inline __m128i
 852 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 853 _mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
 854 {
 855   return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
 856                                                  (__v4si) __W,
 857                                                  (__mmask8) __U);
 858 }
 859
 860 extern __inline __m128i
 861 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 862 _mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A)
 863 {
 864   return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A,
 865                                                  (__v4si)
 866                                                  _mm_setzero_si128 (),
 867                                                  (__mmask8) __U);
 868 }
 869
 870 extern __inline __m256i
 871 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 872 _mm256_abs_epi64 (__m256i __A)
 873 {
 874   return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
 875                                                  (__v4di)
 876                                                  _mm256_setzero_si256 (),
 877                                                  (__mmask8) -1);
 878 }
 879
 880 extern __inline __m256i
 881 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 882 _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
 883 {
 884   return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
 885                                                  (__v4di) __W,
 886                                                  (__mmask8) __U);
 887 }
 888
 889 extern __inline __m256i
 890 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 891 _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A)
 892 {
 893   return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A,
 894                                                  (__v4di)
 895                                                  _mm256_setzero_si256 (),
 896                                                  (__mmask8) __U);
 897 }
 898
 899 extern __inline __m128i
 900 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 901 _mm_abs_epi64 (__m128i __A)
 902 {
 903   return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
 904                                                  (__v2di)
 905                                                  _mm_setzero_si128 (),
 906                                                  (__mmask8) -1);
 907 }
 908
 909 extern __inline __m128i
 910 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 911 _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
 912 {
 913   return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
 914                                                  (__v2di) __W,
 915                                                  (__mmask8) __U);
 916 }
 917
 918 extern __inline __m128i
 919 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 920 _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A)
 921 {
 922   return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A,
 923                                                  (__v2di)
 924                                                  _mm_setzero_si128 (),
 925                                                  (__mmask8) __U);
 926 }
 927
 928 extern __inline __m128i
 929 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 930 _mm256_cvtpd_epu32 (__m256d __A)
 931 {
 932   return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
 933                                                      (__v4si)
 934                                                      _mm_setzero_si128 (),
 935                                                      (__mmask8) -1);
 936 }
 937
 938 extern __inline __m128i
 939 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 940 _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A)
 941 {
 942   return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
 943                                                      (__v4si) __W,
 944                                                      (__mmask8) __U);
 945 }
 946
 947 extern __inline __m128i
 948 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 949 _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A)
 950 {
 951   return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A,
 952                                                      (__v4si)
 953                                                      _mm_setzero_si128 (),
 954                                                      (__mmask8) __U);
 955 }
 956
 957 extern __inline __m128i
 958 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 959 _mm_cvtpd_epu32 (__m128d __A)
 960 {
 961   return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
 962                                                      (__v4si)
 963                                                      _mm_setzero_si128 (),
 964                                                      (__mmask8) -1);
 965 }
 966
 967 extern __inline __m128i
 968 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 969 _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A)
 970 {
 971   return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
 972                                                      (__v4si) __W,
 973                                                      (__mmask8) __U);
 974 }
 975
 976 extern __inline __m128i
 977 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 978 _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A)
 979 {
 980   return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A,
 981                                                      (__v4si)
 982                                                      _mm_setzero_si128 (),
 983                                                      (__mmask8) __U);
 984 }
 985
 986 extern __inline __m256i
 987 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 988 _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A)
 989 {
 990   return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
 991                                                      (__v8si) __W,
 992                                                      (__mmask8) __U);
 993 }
 994
 995 extern __inline __m256i
 996 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 997 _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A)
 998 {
 999   return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A,
1000                                                      (__v8si)
1001                                                      _mm256_setzero_si256 (),
1002                                                      (__mmask8) __U);
1003 }
1004
1005 extern __inline __m128i
1006 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1007 _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A)
1008 {
1009   return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
1010                                                      (__v4si) __W,
1011                                                      (__mmask8) __U);
1012 }
1013
1014 extern __inline __m128i
1015 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1016 _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A)
1017 {
1018   return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A,
1019                                                      (__v4si)
1020                                                      _mm_setzero_si128 (),
1021                                                      (__mmask8) __U);
1022 }
1023
1024 extern __inline __m256i
1025 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1026 _mm256_cvttps_epu32 (__m256 __A)
1027 {
1028   return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
1029                                                       (__v8si)
1030                                                       _mm256_setzero_si256 (),
1031                                                       (__mmask8) -1);
1032 }
1033
1034 extern __inline __m256i
1035 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1036 _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A)
1037 {
1038   return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
1039                                                       (__v8si) __W,
1040                                                       (__mmask8) __U);
1041 }
1042
1043 extern __inline __m256i
1044 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1045 _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A)
1046 {
1047   return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A,
1048                                                       (__v8si)
1049                                                       _mm256_setzero_si256 (),
1050                                                       (__mmask8) __U);
1051 }
1052
1053 extern __inline __m128i
1054 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1055 _mm_cvttps_epu32 (__m128 __A)
1056 {
1057   return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
1058                                                       (__v4si)
1059                                                       _mm_setzero_si128 (),
1060                                                       (__mmask8) -1);
1061 }
1062
1063 extern __inline __m128i
1064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1065 _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A)
1066 {
1067   return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
1068                                                       (__v4si) __W,
1069                                                       (__mmask8) __U);
1070 }
1071
1072 extern __inline __m128i
1073 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1074 _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A)
1075 {
1076   return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A,
1077                                                       (__v4si)
1078                                                       _mm_setzero_si128 (),
1079                                                       (__mmask8) __U);
1080 }
1081
1082 extern __inline __m128i
1083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1084 _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A)
1085 {
1086   return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
1087                                                      (__v4si) __W,
1088                                                      (__mmask8) __U);
1089 }
1090
1091 extern __inline __m128i
1092 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1093 _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A)
1094 {
1095   return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A,
1096                                                      (__v4si)
1097                                                      _mm_setzero_si128 (),
1098                                                      (__mmask8) __U);
1099 }
1100
1101 extern __inline __m128i
1102 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1103 _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A)
1104 {
1105   return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
1106                                                      (__v4si) __W,
1107                                                      (__mmask8) __U);
1108 }
1109
1110 extern __inline __m128i
1111 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1112 _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A)
1113 {
1114   return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A,
1115                                                      (__v4si)
1116                                                      _mm_setzero_si128 (),
1117                                                      (__mmask8) __U);
1118 }
1119
1120 extern __inline __m128i
1121 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1122 _mm256_cvttpd_epu32 (__m256d __A)
1123 {
1124   return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
1125                                                       (__v4si)
1126                                                       _mm_setzero_si128 (),
1127                                                       (__mmask8) -1);
1128 }
1129
1130 extern __inline __m128i
1131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1132 _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A)
1133 {
1134   return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
1135                                                       (__v4si) __W,
1136                                                       (__mmask8) __U);
1137 }
1138
1139 extern __inline __m128i
1140 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1141 _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A)
1142 {
1143   return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A,
1144                                                       (__v4si)
1145                                                       _mm_setzero_si128 (),
1146                                                       (__mmask8) __U);
1147 }
1148
1149 extern __inline __m128i
1150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1151 _mm_cvttpd_epu32 (__m128d __A)
1152 {
1153   return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
1154                                                       (__v4si)
1155                                                       _mm_setzero_si128 (),
1156                                                       (__mmask8) -1);
1157 }
1158
1159 extern __inline __m128i
1160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1161 _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A)
1162 {
1163   return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
1164                                                       (__v4si) __W,
1165                                                       (__mmask8) __U);
1166 }
1167
1168 extern __inline __m128i
1169 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1170 _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A)
1171 {
1172   return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A,
1173                                                       (__v4si)
1174                                                       _mm_setzero_si128 (),
1175                                                       (__mmask8) __U);
1176 }
1177
1178 extern __inline __m128i
1179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1180 _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A)
1181 {
1182   return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
1183                                                     (__v4si) __W,
1184                                                     (__mmask8) __U);
1185 }
1186
1187 extern __inline __m128i
1188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1189 _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A)
1190 {
1191   return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A,
1192                                                     (__v4si)
1193                                                     _mm_setzero_si128 (),
1194                                                     (__mmask8) __U);
1195 }
1196
1197 extern __inline __m128i
1198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1199 _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A)
1200 {
1201   return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
1202                                                     (__v4si) __W,
1203                                                     (__mmask8) __U);
1204 }
1205
1206 extern __inline __m128i
1207 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1208 _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A)
1209 {
1210   return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A,
1211                                                     (__v4si)
1212                                                     _mm_setzero_si128 (),
1213                                                     (__mmask8) __U);
1214 }
1215
1216 extern __inline __m256d
1217 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1218 _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A)
1219 {
1220   return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
1221                                                     (__v4df) __W,
1222                                                     (__mmask8) __U);
1223 }
1224
1225 extern __inline __m256d
1226 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1227 _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A)
1228 {
1229   return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A,
1230                                                     (__v4df)
1231                                                     _mm256_setzero_pd (),
1232                                                     (__mmask8) __U);
1233 }
1234
1235 extern __inline __m128d
1236 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1237 _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A)
1238 {
1239   return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
1240                                                     (__v2df) __W,
1241                                                     (__mmask8) __U);
1242 }
1243
1244 extern __inline __m128d
1245 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1246 _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A)
1247 {
1248   return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A,
1249                                                     (__v2df)
1250                                                     _mm_setzero_pd (),
1251                                                     (__mmask8) __U);
1252 }
1253
1254 extern __inline __m256d
1255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1256 _mm256_cvtepu32_pd (__m128i __A)
1257 {
1258   return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
1259                                                      (__v4df)
1260                                                      _mm256_setzero_pd (),
1261                                                      (__mmask8) -1);
1262 }
1263
1264 extern __inline __m256d
1265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1266 _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A)
1267 {
1268   return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
1269                                                      (__v4df) __W,
1270                                                      (__mmask8) __U);
1271 }
1272
1273 extern __inline __m256d
1274 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1275 _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A)
1276 {
1277   return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A,
1278                                                      (__v4df)
1279                                                      _mm256_setzero_pd (),
1280                                                      (__mmask8) __U);
1281 }
1282
1283 extern __inline __m128d
1284 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1285 _mm_cvtepu32_pd (__m128i __A)
1286 {
1287   return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
1288                                                      (__v2df)
1289                                                      _mm_setzero_pd (),
1290                                                      (__mmask8) -1);
1291 }
1292
1293 extern __inline __m128d
1294 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1295 _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A)
1296 {
1297   return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
1298                                                      (__v2df) __W,
1299                                                      (__mmask8) __U);
1300 }
1301
1302 extern __inline __m128d
1303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1304 _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A)
1305 {
1306   return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A,
1307                                                      (__v2df)
1308                                                      _mm_setzero_pd (),
1309                                                      (__mmask8) __U);
1310 }
1311
1312 extern __inline __m256
1313 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1314 _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A)
1315 {
1316   return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
1317                                                    (__v8sf) __W,
1318                                                    (__mmask8) __U);
1319 }
1320
1321 extern __inline __m256
1322 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1323 _mm256_maskz_cvtepi32_ps (__mmask8 __U, __m256i __A)
1324 {
1325   return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A,
1326                                                    (__v8sf)
1327                                                    _mm256_setzero_ps (),
1328                                                    (__mmask8) __U);
1329 }
1330
1331 extern __inline __m128
1332 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1333 _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A)
1334 {
1335   return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
1336                                                    (__v4sf) __W,
1337                                                    (__mmask8) __U);
1338 }
1339
1340 extern __inline __m128
1341 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1342 _mm_maskz_cvtepi32_ps (__mmask8 __U, __m128i __A)
1343 {
1344   return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A,
1345                                                    (__v4sf)
1346                                                    _mm_setzero_ps (),
1347                                                    (__mmask8) __U);
1348 }
1349
1350 extern __inline __m256
1351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1352 _mm256_cvtepu32_ps (__m256i __A)
1353 {
1354   return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
1355                                                     (__v8sf)
1356                                                     _mm256_setzero_ps (),
1357                                                     (__mmask8) -1);
1358 }
1359
1360 extern __inline __m256
1361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1362 _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A)
1363 {
1364   return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
1365                                                     (__v8sf) __W,
1366                                                     (__mmask8) __U);
1367 }
1368
1369 extern __inline __m256
1370 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1371 _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A)
1372 {
1373   return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A,
1374                                                     (__v8sf)
1375                                                     _mm256_setzero_ps (),
1376                                                     (__mmask8) __U);
1377 }
1378
1379 extern __inline __m128
1380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1381 _mm_cvtepu32_ps (__m128i __A)
1382 {
1383   return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
1384                                                     (__v4sf)
1385                                                     _mm_setzero_ps (),
1386                                                     (__mmask8) -1);
1387 }
1388
1389 extern __inline __m128
1390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1391 _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A)
1392 {
1393   return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
1394                                                     (__v4sf) __W,
1395                                                     (__mmask8) __U);
1396 }
1397
1398 extern __inline __m128
1399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1400 _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A)
1401 {
1402   return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A,
1403                                                     (__v4sf)
1404                                                     _mm_setzero_ps (),
1405                                                     (__mmask8) __U);
1406 }
1407
1408 extern __inline __m256d
1409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1410 _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A)
1411 {
1412   return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
1413                                                     (__v4df) __W,
1414                                                     (__mmask8) __U);
1415 }
1416
1417 extern __inline __m256d
1418 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1419 _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A)
1420 {
1421   return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A,
1422                                                     (__v4df)
1423                                                     _mm256_setzero_pd (),
1424                                                     (__mmask8) __U);
1425 }
1426
1427 extern __inline __m128d
1428 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1429 _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A)
1430 {
1431   return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
1432                                                     (__v2df) __W,
1433                                                     (__mmask8) __U);
1434 }
1435
1436 extern __inline __m128d
1437 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1438 _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A)
1439 {
1440   return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A,
1441                                                     (__v2df)
1442                                                     _mm_setzero_pd (),
1443                                                     (__mmask8) __U);
1444 }
1445
1446 extern __inline __m128i
1447 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1448 _mm_cvtepi32_epi8 (__m128i __A)
1449 {
1450   return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
1451                                                   (__v16qi)
1452                                                   _mm_undefined_si128 (),
1453                                                   (__mmask8) -1);
1454 }
1455
1456 extern __inline void
1457 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1458 _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1459 {
1460   __builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
1461 }
1462
1463 extern __inline __m128i
1464 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1465 _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
1466 {
1467   return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
1468                                                   (__v16qi) __O, __M);
1469 }
1470
1471 extern __inline __m128i
1472 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1473 _mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A)
1474 {
1475   return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A,
1476                                                   (__v16qi)
1477                                                   _mm_setzero_si128 (),
1478                                                   __M);
1479 }
1480
1481 extern __inline __m128i
1482 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1483 _mm256_cvtepi32_epi8 (__m256i __A)
1484 {
1485   return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
1486                                                   (__v16qi)
1487                                                   _mm_undefined_si128 (),
1488                                                   (__mmask8) -1);
1489 }
1490
1491 extern __inline __m128i
1492 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1493 _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
1494 {
1495   return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
1496                                                   (__v16qi) __O, __M);
1497 }
1498
1499 extern __inline void
1500 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1501 _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
1502 {
1503   __builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
1504 }
1505
1506 extern __inline __m128i
1507 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1508 _mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A)
1509 {
1510   return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
1511                                                   (__v16qi)
1512                                                   _mm_setzero_si128 (),
1513                                                   __M);
1514 }
1515
1516 extern __inline __m128i
1517 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1518 _mm_cvtsepi32_epi8 (__m128i __A)
1519 {
1520   return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
1521                                                    (__v16qi)
1522                                                    _mm_undefined_si128 (),
1523                                                    (__mmask8) -1);
1524 }
1525
1526 extern __inline void
1527 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1528 _mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1529 {
1530   __builtin_ia32_pmovsdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
1531 }
1532
1533 extern __inline __m128i
1534 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1535 _mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
1536 {
1537   return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
1538                                                    (__v16qi) __O, __M);
1539 }
1540
1541 extern __inline __m128i
1542 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1543 _mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A)
1544 {
1545   return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A,
1546                                                    (__v16qi)
1547                                                    _mm_setzero_si128 (),
1548                                                    __M);
1549 }
1550
1551 extern __inline __m128i
1552 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1553 _mm256_cvtsepi32_epi8 (__m256i __A)
1554 {
1555   return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
1556                                                    (__v16qi)
1557                                                    _mm_undefined_si128 (),
1558                                                    (__mmask8) -1);
1559 }
1560
1561 extern __inline void
1562 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1563 _mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
1564 {
1565   __builtin_ia32_pmovsdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
1566 }
1567
1568 extern __inline __m128i
1569 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1570 _mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
1571 {
1572   return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
1573                                                    (__v16qi) __O, __M);
1574 }
1575
1576 extern __inline __m128i
1577 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1578 _mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A)
1579 {
1580   return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A,
1581                                                    (__v16qi)
1582                                                    _mm_setzero_si128 (),
1583                                                    __M);
1584 }
1585
1586 extern __inline __m128i
1587 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1588 _mm_cvtusepi32_epi8 (__m128i __A)
1589 {
1590   return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
1591                                                     (__v16qi)
1592                                                     _mm_undefined_si128 (),
1593                                                     (__mmask8) -1);
1594 }
1595
1596 extern __inline void
1597 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1598 _mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1599 {
1600   __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
1601 }
1602
1603 extern __inline __m128i
1604 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1605 _mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
1606 {
1607   return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
1608                                                     (__v16qi) __O,
1609                                                     __M);
1610 }
1611
1612 extern __inline __m128i
1613 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1614 _mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A)
1615 {
1616   return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A,
1617                                                     (__v16qi)
1618                                                     _mm_setzero_si128 (),
1619                                                     __M);
1620 }
1621
1622 extern __inline __m128i
1623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1624 _mm256_cvtusepi32_epi8 (__m256i __A)
1625 {
1626   return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
1627                                                     (__v16qi)
1628                                                     _mm_undefined_si128 (),
1629                                                     (__mmask8) -1);
1630 }
1631
1632 extern __inline void
1633 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1634 _mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
1635 {
1636   __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M);
1637 }
1638
1639 extern __inline __m128i
1640 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1641 _mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
1642 {
1643   return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
1644                                                     (__v16qi) __O,
1645                                                     __M);
1646 }
1647
1648 extern __inline __m128i
1649 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1650 _mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A)
1651 {
1652   return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A,
1653                                                     (__v16qi)
1654                                                     _mm_setzero_si128 (),
1655                                                     __M);
1656 }
1657
1658 extern __inline __m128i
1659 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1660 _mm_cvtepi32_epi16 (__m128i __A)
1661 {
1662   return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
1663                                                   (__v8hi)
1664                                                   _mm_setzero_si128 (),
1665                                                   (__mmask8) -1);
1666 }
1667
1668 extern __inline void
1669 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1670 _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
1671 {
1672   __builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
1673 }
1674
1675 extern __inline __m128i
1676 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1677 _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
1678 {
1679   return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
1680                                                   (__v8hi) __O, __M);
1681 }
1682
1683 extern __inline __m128i
1684 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1685 _mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A)
1686 {
1687   return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A,
1688                                                   (__v8hi)
1689                                                   _mm_setzero_si128 (),
1690                                                   __M);
1691 }
1692
1693 extern __inline __m128i
1694 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1695 _mm256_cvtepi32_epi16 (__m256i __A)
1696 {
1697   return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
1698                                                   (__v8hi)
1699                                                   _mm_setzero_si128 (),
1700                                                   (__mmask8) -1);
1701 }
1702
1703 extern __inline void
1704 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1705 _mm256_mask_cvtepi32_storeu_epi16 (void *  __P, __mmask8 __M, __m256i __A)
1706 {
1707   __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
1708 }
1709
1710 extern __inline __m128i
1711 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1712 _mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
1713 {
1714   return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
1715                                                   (__v8hi) __O, __M);
1716 }
1717
1718 extern __inline __m128i
1719 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1720 _mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A)
1721 {
1722   return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A,
1723                                                   (__v8hi)
1724                                                   _mm_setzero_si128 (),
1725                                                   __M);
1726 }
1727
1728 extern __inline __m128i
1729 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1730 _mm_cvtsepi32_epi16 (__m128i __A)
1731 {
1732   return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
1733                                                    (__v8hi)
1734                                                    _mm_setzero_si128 (),
1735                                                    (__mmask8) -1);
1736 }
1737
1738 extern __inline void
1739 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1740 _mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
1741 {
1742   __builtin_ia32_pmovsdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
1743 }
1744
1745 extern __inline __m128i
1746 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1747 _mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
1748 {
1749   return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
1750                                                    (__v8hi)__O,
1751                                                    __M);
1752 }
1753
1754 extern __inline __m128i
1755 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1756 _mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A)
1757 {
1758   return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A,
1759                                                    (__v8hi)
1760                                                    _mm_setzero_si128 (),
1761                                                    __M);
1762 }
1763
1764 extern __inline __m128i
1765 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1766 _mm256_cvtsepi32_epi16 (__m256i __A)
1767 {
1768   return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
1769                                                    (__v8hi)
1770                                                    _mm_undefined_si128 (),
1771                                                    (__mmask8) -1);
1772 }
1773
1774 extern __inline void
1775 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1776 _mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
1777 {
1778   __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
1779 }
1780
1781 extern __inline __m128i
1782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1783 _mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
1784 {
1785   return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
1786                                                    (__v8hi) __O, __M);
1787 }
1788
1789 extern __inline __m128i
1790 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1791 _mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A)
1792 {
1793   return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A,
1794                                                    (__v8hi)
1795                                                    _mm_setzero_si128 (),
1796                                                    __M);
1797 }
1798
1799 extern __inline __m128i
1800 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1801 _mm_cvtusepi32_epi16 (__m128i __A)
1802 {
1803   return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
1804                                                     (__v8hi)
1805                                                     _mm_undefined_si128 (),
1806                                                     (__mmask8) -1);
1807 }
1808
1809 extern __inline void
1810 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1811 _mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
1812 {
1813   __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
1814 }
1815
1816 extern __inline __m128i
1817 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1818 _mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
1819 {
1820   return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
1821                                                     (__v8hi) __O, __M);
1822 }
1823
1824 extern __inline __m128i
1825 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1826 _mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A)
1827 {
1828   return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A,
1829                                                     (__v8hi)
1830                                                     _mm_setzero_si128 (),
1831                                                     __M);
1832 }
1833
1834 extern __inline __m128i
1835 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1836 _mm256_cvtusepi32_epi16 (__m256i __A)
1837 {
1838   return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
1839                                                     (__v8hi)
1840                                                     _mm_undefined_si128 (),
1841                                                     (__mmask8) -1);
1842 }
1843
1844 extern __inline void
1845 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1846 _mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
1847 {
1848   __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
1849 }
1850
1851 extern __inline __m128i
1852 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1853 _mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
1854 {
1855   return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
1856                                                     (__v8hi) __O, __M);
1857 }
1858
1859 extern __inline __m128i
1860 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1861 _mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A)
1862 {
1863   return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A,
1864                                                     (__v8hi)
1865                                                     _mm_setzero_si128 (),
1866                                                     __M);
1867 }
1868
1869 extern __inline __m128i
1870 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1871 _mm_cvtepi64_epi8 (__m128i __A)
1872 {
1873   return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
1874                                                   (__v16qi)
1875                                                   _mm_undefined_si128 (),
1876                                                   (__mmask8) -1);
1877 }
1878
1879 extern __inline void
1880 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1881 _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1882 {
1883   __builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
1884 }
1885
1886 extern __inline __m128i
1887 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1888 _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
1889 {
1890   return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
1891                                                   (__v16qi) __O, __M);
1892 }
1893
1894 extern __inline __m128i
1895 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1896 _mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A)
1897 {
1898   return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A,
1899                                                   (__v16qi)
1900                                                   _mm_setzero_si128 (),
1901                                                   __M);
1902 }
1903
1904 extern __inline __m128i
1905 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1906 _mm256_cvtepi64_epi8 (__m256i __A)
1907 {
1908   return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
1909                                                   (__v16qi)
1910                                                   _mm_undefined_si128 (),
1911                                                   (__mmask8) -1);
1912 }
1913
1914 extern __inline void
1915 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1916 _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
1917 {
1918   __builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
1919 }
1920
1921 extern __inline __m128i
1922 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1923 _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
1924 {
1925   return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
1926                                                   (__v16qi) __O, __M);
1927 }
1928
1929 extern __inline __m128i
1930 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1931 _mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A)
1932 {
1933   return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A,
1934                                                   (__v16qi)
1935                                                   _mm_setzero_si128 (),
1936                                                   __M);
1937 }
1938
1939 extern __inline __m128i
1940 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1941 _mm_cvtsepi64_epi8 (__m128i __A)
1942 {
1943   return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
1944                                                    (__v16qi)
1945                                                    _mm_undefined_si128 (),
1946                                                    (__mmask8) -1);
1947 }
1948
1949 extern __inline void
1950 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1951 _mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
1952 {
1953   __builtin_ia32_pmovsqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
1954 }
1955
1956 extern __inline __m128i
1957 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1958 _mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
1959 {
1960   return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
1961                                                    (__v16qi) __O, __M);
1962 }
1963
1964 extern __inline __m128i
1965 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1966 _mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A)
1967 {
1968   return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A,
1969                                                    (__v16qi)
1970                                                    _mm_setzero_si128 (),
1971                                                    __M);
1972 }
1973
1974 extern __inline __m128i
1975 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1976 _mm256_cvtsepi64_epi8 (__m256i __A)
1977 {
1978   return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
1979                                                    (__v16qi)
1980                                                    _mm_undefined_si128 (),
1981                                                    (__mmask8) -1);
1982 }
1983
1984 extern __inline void
1985 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1986 _mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
1987 {
1988   __builtin_ia32_pmovsqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
1989 }
1990
1991 extern __inline __m128i
1992 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
1993 _mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
1994 {
1995   return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
1996                                                    (__v16qi) __O, __M);
1997 }
1998
1999 extern __inline __m128i
2000 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2001 _mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A)
2002 {
2003   return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A,
2004                                                    (__v16qi)
2005                                                    _mm_setzero_si128 (),
2006                                                    __M);
2007 }
2008
2009 extern __inline __m128i
2010 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2011 _mm_cvtusepi64_epi8 (__m128i __A)
2012 {
2013   return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
2014                                                     (__v16qi)
2015                                                     _mm_undefined_si128 (),
2016                                                     (__mmask8) -1);
2017 }
2018
2019 extern __inline void
2020 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2021 _mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
2022 {
2023   __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
2024 }
2025
2026 extern __inline __m128i
2027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2028 _mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A)
2029 {
2030   return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
2031                                                     (__v16qi) __O,
2032                                                     __M);
2033 }
2034
2035 extern __inline __m128i
2036 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2037 _mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A)
2038 {
2039   return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A,
2040                                                     (__v16qi)
2041                                                     _mm_setzero_si128 (),
2042                                                     __M);
2043 }
2044
2045 extern __inline __m128i
2046 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2047 _mm256_cvtusepi64_epi8 (__m256i __A)
2048 {
2049   return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
2050                                                     (__v16qi)
2051                                                     _mm_undefined_si128 (),
2052                                                     (__mmask8) -1);
2053 }
2054
2055 extern __inline void
2056 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2057 _mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
2058 {
2059   __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
2060 }
2061
2062 extern __inline __m128i
2063 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2064 _mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A)
2065 {
2066   return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
2067                                                     (__v16qi) __O,
2068                                                     __M);
2069 }
2070
2071 extern __inline __m128i
2072 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2073 _mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A)
2074 {
2075   return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A,
2076                                                     (__v16qi)
2077                                                     _mm_setzero_si128 (),
2078                                                     __M);
2079 }
2080
2081 extern __inline __m128i
2082 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2083 _mm_cvtepi64_epi16 (__m128i __A)
2084 {
2085   return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
2086                                                   (__v8hi)
2087                                                   _mm_undefined_si128 (),
2088                                                   (__mmask8) -1);
2089 }
2090
2091 extern __inline void
2092 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2093 _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
2094 {
2095   __builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
2096 }
2097
2098 extern __inline __m128i
2099 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2100 _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
2101 {
2102   return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
2103                                                   (__v8hi)__O,
2104                                                   __M);
2105 }
2106
2107 extern __inline __m128i
2108 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2109 _mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A)
2110 {
2111   return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A,
2112                                                   (__v8hi)
2113                                                   _mm_setzero_si128 (),
2114                                                   __M);
2115 }
2116
2117 extern __inline __m128i
2118 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2119 _mm256_cvtepi64_epi16 (__m256i __A)
2120 {
2121   return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
2122                                                   (__v8hi)
2123                                                   _mm_undefined_si128 (),
2124                                                   (__mmask8) -1);
2125 }
2126
2127 extern __inline void
2128 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2129 _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
2130 {
2131   __builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
2132 }
2133
2134 extern __inline __m128i
2135 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2136 _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
2137 {
2138   return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
2139                                                   (__v8hi) __O, __M);
2140 }
2141
2142 extern __inline __m128i
2143 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2144 _mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A)
2145 {
2146   return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A,
2147                                                   (__v8hi)
2148                                                   _mm_setzero_si128 (),
2149                                                   __M);
2150 }
2151
2152 extern __inline __m128i
2153 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2154 _mm_cvtsepi64_epi16 (__m128i __A)
2155 {
2156   return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
2157                                                    (__v8hi)
2158                                                    _mm_undefined_si128 (),
2159                                                    (__mmask8) -1);
2160 }
2161
2162 extern __inline void
2163 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2164 _mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
2165 {
2166   __builtin_ia32_pmovsqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
2167 }
2168
2169 extern __inline __m128i
2170 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2171 _mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
2172 {
2173   return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
2174                                                    (__v8hi) __O, __M);
2175 }
2176
2177 extern __inline __m128i
2178 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2179 _mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A)
2180 {
2181   return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A,
2182                                                    (__v8hi)
2183                                                    _mm_setzero_si128 (),
2184                                                    __M);
2185 }
2186
2187 extern __inline __m128i
2188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2189 _mm256_cvtsepi64_epi16 (__m256i __A)
2190 {
2191   return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
2192                                                    (__v8hi)
2193                                                    _mm_undefined_si128 (),
2194                                                    (__mmask8) -1);
2195 }
2196
2197 extern __inline void
2198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2199 _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
2200 {
2201   __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
2202 }
2203
2204 extern __inline __m128i
2205 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2206 _mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
2207 {
2208   return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
2209                                                    (__v8hi) __O, __M);
2210 }
2211
2212 extern __inline __m128i
2213 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2214 _mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A)
2215 {
2216   return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A,
2217                                                    (__v8hi)
2218                                                    _mm_setzero_si128 (),
2219                                                    __M);
2220 }
2221
2222 extern __inline __m128i
2223 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2224 _mm_cvtusepi64_epi16 (__m128i __A)
2225 {
2226   return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
2227                                                     (__v8hi)
2228                                                     _mm_undefined_si128 (),
2229                                                     (__mmask8) -1);
2230 }
2231
2232 extern __inline void
2233 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2234 _mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
2235 {
2236   __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
2237 }
2238
2239 extern __inline __m128i
2240 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2241 _mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A)
2242 {
2243   return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
2244                                                     (__v8hi) __O, __M);
2245 }
2246
2247 extern __inline __m128i
2248 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2249 _mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A)
2250 {
2251   return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A,
2252                                                     (__v8hi)
2253                                                     _mm_setzero_si128 (),
2254                                                     __M);
2255 }
2256
2257 extern __inline __m128i
2258 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2259 _mm256_cvtusepi64_epi16 (__m256i __A)
2260 {
2261   return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
2262                                                     (__v8hi)
2263                                                     _mm_undefined_si128 (),
2264                                                     (__mmask8) -1);
2265 }
2266
2267 extern __inline void
2268 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2269 _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
2270 {
2271   __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
2272 }
2273
2274 extern __inline __m128i
2275 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2276 _mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A)
2277 {
2278   return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
2279                                                     (__v8hi) __O, __M);
2280 }
2281
2282 extern __inline __m128i
2283 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2284 _mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A)
2285 {
2286   return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A,
2287                                                     (__v8hi)
2288                                                     _mm_setzero_si128 (),
2289                                                     __M);
2290 }
2291
2292 extern __inline __m128i
2293 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2294 _mm_cvtepi64_epi32 (__m128i __A)
2295 {
2296   return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
2297                                                   (__v4si)
2298                                                   _mm_undefined_si128 (),
2299                                                   (__mmask8) -1);
2300 }
2301
2302 extern __inline void
2303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2304 _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
2305 {
2306   __builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
2307 }
2308
2309 extern __inline __m128i
2310 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2311 _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
2312 {
2313   return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
2314                                                   (__v4si) __O, __M);
2315 }
2316
2317 extern __inline __m128i
2318 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2319 _mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A)
2320 {
2321   return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A,
2322                                                   (__v4si)
2323                                                   _mm_setzero_si128 (),
2324                                                   __M);
2325 }
2326
2327 extern __inline __m128i
2328 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2329 _mm256_cvtepi64_epi32 (__m256i __A)
2330 {
2331   return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
2332                                                   (__v4si)
2333                                                   _mm_undefined_si128 (),
2334                                                   (__mmask8) -1);
2335 }
2336
2337 extern __inline void
2338 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2339 _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
2340 {
2341   __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
2342 }
2343
2344 extern __inline __m128i
2345 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2346 _mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
2347 {
2348   return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
2349                                                   (__v4si) __O, __M);
2350 }
2351
2352 extern __inline __m128i
2353 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2354 _mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A)
2355 {
2356   return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A,
2357                                                   (__v4si)
2358                                                   _mm_setzero_si128 (),
2359                                                   __M);
2360 }
2361
2362 extern __inline __m128i
2363 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2364 _mm_cvtsepi64_epi32 (__m128i __A)
2365 {
2366   return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
2367                                                    (__v4si)
2368                                                    _mm_undefined_si128 (),
2369                                                    (__mmask8) -1);
2370 }
2371
2372 extern __inline void
2373 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2374 _mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
2375 {
2376   __builtin_ia32_pmovsqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
2377 }
2378
2379 extern __inline __m128i
2380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2381 _mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
2382 {
2383   return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
2384                                                    (__v4si) __O, __M);
2385 }
2386
2387 extern __inline __m128i
2388 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2389 _mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A)
2390 {
2391   return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A,
2392                                                    (__v4si)
2393                                                    _mm_setzero_si128 (),
2394                                                    __M);
2395 }
2396
2397 extern __inline __m128i
2398 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2399 _mm256_cvtsepi64_epi32 (__m256i __A)
2400 {
2401   return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
2402                                                    (__v4si)
2403                                                    _mm_undefined_si128 (),
2404                                                    (__mmask8) -1);
2405 }
2406
2407 extern __inline void
2408 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2409 _mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
2410 {
2411   __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
2412 }
2413
2414 extern __inline __m128i
2415 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2416 _mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
2417 {
2418   return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
2419                                                    (__v4si)__O,
2420                                                    __M);
2421 }
2422
2423 extern __inline __m128i
2424 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2425 _mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A)
2426 {
2427   return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A,
2428                                                    (__v4si)
2429                                                    _mm_setzero_si128 (),
2430                                                    __M);
2431 }
2432
2433 extern __inline __m128i
2434 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2435 _mm_cvtusepi64_epi32 (__m128i __A)
2436 {
2437   return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
2438                                                     (__v4si)
2439                                                     _mm_undefined_si128 (),
2440                                                     (__mmask8) -1);
2441 }
2442
2443 extern __inline void
2444 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2445 _mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
2446 {
2447   __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
2448 }
2449
2450 extern __inline __m128i
2451 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2452 _mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
2453 {
2454   return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
2455                                                     (__v4si) __O, __M);
2456 }
2457
2458 extern __inline __m128i
2459 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2460 _mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A)
2461 {
2462   return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A,
2463                                                     (__v4si)
2464                                                     _mm_setzero_si128 (),
2465                                                     __M);
2466 }
2467
2468 extern __inline __m128i
2469 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2470 _mm256_cvtusepi64_epi32 (__m256i __A)
2471 {
2472   return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
2473                                                     (__v4si)
2474                                                     _mm_undefined_si128 (),
2475                                                     (__mmask8) -1);
2476 }
2477
2478 extern __inline void
2479 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2480 _mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
2481 {
2482   __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
2483 }
2484
2485 extern __inline __m128i
2486 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2487 _mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A)
2488 {
2489   return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
2490                                                     (__v4si) __O, __M);
2491 }
2492
2493 extern __inline __m128i
2494 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2495 _mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A)
2496 {
2497   return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A,
2498                                                     (__v4si)
2499                                                     _mm_setzero_si128 (),
2500                                                     __M);
2501 }
2502
2503 extern __inline __m256
2504 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2505 _mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A)
2506 {
2507   return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A,
2508                                                       (__v8sf) __O,
2509                                                       __M);
2510 }
2511
2512 extern __inline __m256
2513 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2514 _mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
2515 {
2516   return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A,
2517                                                       (__v8sf)
2518                                                       _mm256_setzero_ps (),
2519                                                       __M);
2520 }
2521
2522 extern __inline __m128
2523 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2524 _mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A)
2525 {
2526   return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A,
2527                                                       (__v4sf) __O,
2528                                                       __M);
2529 }
2530
2531 extern __inline __m128
2532 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2533 _mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A)
2534 {
2535   return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A,
2536                                                       (__v4sf)
2537                                                       _mm_setzero_ps (),
2538                                                       __M);
2539 }
2540
2541 extern __inline __m256d
2542 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2543 _mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A)
2544 {
2545   return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A,
2546                                                        (__v4df) __O,
2547                                                        __M);
2548 }
2549
2550 extern __inline __m256d
2551 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2552 _mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A)
2553 {
2554   return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A,
2555                                                        (__v4df)
2556                                                        _mm256_setzero_pd (),
2557                                                        __M);
2558 }
2559
2560 extern __inline __m256i
2561 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2562 _mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A)
2563 {
2564   return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A,
2565                                                        (__v8si) __O,
2566                                                        __M);
2567 }
2568
2569 extern __inline __m256i
2570 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2571 _mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
2572 {
2573   return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A,
2574                                                        (__v8si)
2575                                                        _mm256_setzero_si256 (),
2576                                                        __M);
2577 }
2578
2579 extern __inline __m256i
2580 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2581 _mm256_mask_set1_epi32 (__m256i __O, __mmask8 __M, int __A)
2582 {
2583   return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A, (__v8si) __O,
2584                                                            __M);
2585 }
2586
2587 extern __inline __m256i
2588 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2589 _mm256_maskz_set1_epi32 (__mmask8 __M, int __A)
2590 {
2591   return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A,
2592                                                            (__v8si)
2593                                                            _mm256_setzero_si256 (),
2594                                                            __M);
2595 }
2596
2597 extern __inline __m128i
2598 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2599 _mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A)
2600 {
2601   return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A,
2602                                                        (__v4si) __O,
2603                                                        __M);
2604 }
2605
2606 extern __inline __m128i
2607 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2608 _mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A)
2609 {
2610   return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A,
2611                                                        (__v4si)
2612                                                        _mm_setzero_si128 (),
2613                                                        __M);
2614 }
2615
2616 extern __inline __m128i
2617 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2618 _mm_mask_set1_epi32 (__m128i __O, __mmask8 __M, int __A)
2619 {
2620   return (__m128i) __builtin_ia32_pbroadcastd128_gpr_mask (__A, (__v4si) __O,
2621                                                            __M);
2622 }
2623
2624 extern __inline __m128i
2625 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2626 _mm_maskz_set1_epi32 (__mmask8 __M, int __A)
2627 {
2628   return (__m128i)
2629          __builtin_ia32_pbroadcastd128_gpr_mask (__A,
2630                                                  (__v4si) _mm_setzero_si128 (),
2631                                                  __M);
2632 }
2633
2634 extern __inline __m256i
2635 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2636 _mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A)
2637 {
2638   return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A,
2639                                                        (__v4di) __O,
2640                                                        __M);
2641 }
2642
2643 extern __inline __m256i
2644 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2645 _mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
2646 {
2647   return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A,
2648                                                        (__v4di)
2649                                                        _mm256_setzero_si256 (),
2650                                                        __M);
2651 }
2652
2653 extern __inline __m256i
2654 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2655 _mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A)
2656 {
2657   return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O,
2658                                                            __M);
2659 }
2660
2661 extern __inline __m256i
2662 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2663 _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
2664 {
2665   return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
2666                                                            (__v4di)
2667                                                            _mm256_setzero_si256 (),
2668                                                            __M);
2669 }
2670
2671 extern __inline __m128i
2672 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2673 _mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A)
2674 {
2675   return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A,
2676                                                        (__v2di) __O,
2677                                                        __M);
2678 }
2679
2680 extern __inline __m128i
2681 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2682 _mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
2683 {
2684   return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A,
2685                                                        (__v2di)
2686                                                        _mm_setzero_si128 (),
2687                                                        __M);
2688 }
2689
2690 extern __inline __m128i
2691 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2692 _mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
2693 {
2694   return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O,
2695                                                            __M);
2696 }
2697
2698 extern __inline __m128i
2699 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2700 _mm_maskz_set1_epi64 (__mmask8 __M, long long __A)
2701 {
2702   return (__m128i)
2703          __builtin_ia32_pbroadcastq128_gpr_mask (__A,
2704                                                  (__v2di) _mm_setzero_si128 (),
2705                                                  __M);
2706 }
2707
2708 extern __inline __m256
2709 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2710 _mm256_broadcast_f32x4 (__m128 __A)
2711 {
2712   return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
2713                                                           (__v8sf)_mm256_undefined_pd (),
2714                                                           (__mmask8) -1);
2715 }
2716
2717 extern __inline __m256
2718 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2719 _mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A)
2720 {
2721   return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
2722                                                           (__v8sf) __O,
2723                                                           __M);
2724 }
2725
2726 extern __inline __m256
2727 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2728 _mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A)
2729 {
2730   return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A,
2731                                                           (__v8sf)
2732                                                           _mm256_setzero_ps (),
2733                                                           __M);
2734 }
2735
2736 extern __inline __m256i
2737 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2738 _mm256_broadcast_i32x4 (__m128i __A)
2739 {
2740   return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
2741                                                            __A,
2742                                                            (__v8si)_mm256_undefined_si256 (),
2743                                                            (__mmask8) -1);
2744 }
2745
2746 extern __inline __m256i
2747 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2748 _mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A)
2749 {
2750   return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
2751                                                            __A,
2752                                                            (__v8si)
2753                                                            __O, __M);
2754 }
2755
2756 extern __inline __m256i
2757 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2758 _mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A)
2759 {
2760   return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si)
2761                                                            __A,
2762                                                            (__v8si)
2763                                                            _mm256_setzero_si256 (),
2764                                                            __M);
2765 }
2766
2767 extern __inline __m256i
2768 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2769 _mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
2770 {
2771   return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
2772                                                     (__v8si) __W,
2773                                                     (__mmask8) __U);
2774 }
2775
2776 extern __inline __m256i
2777 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2778 _mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
2779 {
2780   return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A,
2781                                                     (__v8si)
2782                                                     _mm256_setzero_si256 (),
2783                                                     (__mmask8) __U);
2784 }
2785
2786 extern __inline __m128i
2787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2788 _mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
2789 {
2790   return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
2791                                                     (__v4si) __W,
2792                                                     (__mmask8) __U);
2793 }
2794
2795 extern __inline __m128i
2796 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2797 _mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A)
2798 {
2799   return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A,
2800                                                     (__v4si)
2801                                                     _mm_setzero_si128 (),
2802                                                     (__mmask8) __U);
2803 }
2804
2805 extern __inline __m256i
2806 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2807 _mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
2808 {
2809   return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
2810                                                     (__v4di) __W,
2811                                                     (__mmask8) __U);
2812 }
2813
2814 extern __inline __m256i
2815 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2816 _mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
2817 {
2818   return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A,
2819                                                     (__v4di)
2820                                                     _mm256_setzero_si256 (),
2821                                                     (__mmask8) __U);
2822 }
2823
2824 extern __inline __m128i
2825 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2826 _mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
2827 {
2828   return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
2829                                                     (__v2di) __W,
2830                                                     (__mmask8) __U);
2831 }
2832
2833 extern __inline __m128i
2834 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2835 _mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A)
2836 {
2837   return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A,
2838                                                     (__v2di)
2839                                                     _mm_setzero_si128 (),
2840                                                     (__mmask8) __U);
2841 }
2842
2843 extern __inline __m256i
2844 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2845 _mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
2846 {
2847   return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
2848                                                     (__v8si) __W,
2849                                                     (__mmask8) __U);
2850 }
2851
2852 extern __inline __m256i
2853 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2854 _mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
2855 {
2856   return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A,
2857                                                     (__v8si)
2858                                                     _mm256_setzero_si256 (),
2859                                                     (__mmask8) __U);
2860 }
2861
2862 extern __inline __m128i
2863 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2864 _mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
2865 {
2866   return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
2867                                                     (__v4si) __W,
2868                                                     (__mmask8) __U);
2869 }
2870
2871 extern __inline __m128i
2872 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2873 _mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A)
2874 {
2875   return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A,
2876                                                     (__v4si)
2877                                                     _mm_setzero_si128 (),
2878                                                     (__mmask8) __U);
2879 }
2880
2881 extern __inline __m256i
2882 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2883 _mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
2884 {
2885   return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
2886                                                     (__v4di) __W,
2887                                                     (__mmask8) __U);
2888 }
2889
2890 extern __inline __m256i
2891 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2892 _mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
2893 {
2894   return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A,
2895                                                     (__v4di)
2896                                                     _mm256_setzero_si256 (),
2897                                                     (__mmask8) __U);
2898 }
2899
2900 extern __inline __m128i
2901 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2902 _mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
2903 {
2904   return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
2905                                                     (__v2di) __W,
2906                                                     (__mmask8) __U);
2907 }
2908
2909 extern __inline __m128i
2910 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2911 _mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A)
2912 {
2913   return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A,
2914                                                     (__v2di)
2915                                                     _mm_setzero_si128 (),
2916                                                     (__mmask8) __U);
2917 }
2918
2919 extern __inline __m256i
2920 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2921 _mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
2922 {
2923   return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
2924                                                     (__v4di) __W,
2925                                                     (__mmask8) __U);
2926 }
2927
2928 extern __inline __m256i
2929 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2930 _mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
2931 {
2932   return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X,
2933                                                     (__v4di)
2934                                                     _mm256_setzero_si256 (),
2935                                                     (__mmask8) __U);
2936 }
2937
2938 extern __inline __m128i
2939 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2940 _mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
2941 {
2942   return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
2943                                                     (__v2di) __W,
2944                                                     (__mmask8) __U);
2945 }
2946
2947 extern __inline __m128i
2948 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2949 _mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X)
2950 {
2951   return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X,
2952                                                     (__v2di)
2953                                                     _mm_setzero_si128 (),
2954                                                     (__mmask8) __U);
2955 }
2956
2957 extern __inline __m256i
2958 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2959 _mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
2960 {
2961   return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
2962                                                     (__v8si) __W,
2963                                                     (__mmask8) __U);
2964 }
2965
2966 extern __inline __m256i
2967 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2968 _mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
2969 {
2970   return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A,
2971                                                     (__v8si)
2972                                                     _mm256_setzero_si256 (),
2973                                                     (__mmask8) __U);
2974 }
2975
2976 extern __inline __m128i
2977 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2978 _mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
2979 {
2980   return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
2981                                                     (__v4si) __W,
2982                                                     (__mmask8) __U);
2983 }
2984
2985 extern __inline __m128i
2986 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2987 _mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A)
2988 {
2989   return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A,
2990                                                     (__v4si)
2991                                                     _mm_setzero_si128 (),
2992                                                     (__mmask8) __U);
2993 }
2994
2995 extern __inline __m256i
2996 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
2997 _mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
2998 {
2999   return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
3000                                                     (__v4di) __W,
3001                                                     (__mmask8) __U);
3002 }
3003
3004 extern __inline __m256i
3005 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3006 _mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
3007 {
3008   return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A,
3009                                                     (__v4di)
3010                                                     _mm256_setzero_si256 (),
3011                                                     (__mmask8) __U);
3012 }
3013
3014 extern __inline __m128i
3015 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3016 _mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
3017 {
3018   return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
3019                                                     (__v2di) __W,
3020                                                     (__mmask8) __U);
3021 }
3022
3023 extern __inline __m128i
3024 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3025 _mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A)
3026 {
3027   return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A,
3028                                                     (__v2di)
3029                                                     _mm_setzero_si128 (),
3030                                                     (__mmask8) __U);
3031 }
3032
3033 extern __inline __m256i
3034 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3035 _mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A)
3036 {
3037   return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
3038                                                     (__v8si) __W,
3039                                                     (__mmask8) __U);
3040 }
3041
3042 extern __inline __m256i
3043 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3044 _mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
3045 {
3046   return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A,
3047                                                     (__v8si)
3048                                                     _mm256_setzero_si256 (),
3049                                                     (__mmask8) __U);
3050 }
3051
3052 extern __inline __m128i
3053 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3054 _mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
3055 {
3056   return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
3057                                                     (__v4si) __W,
3058                                                     (__mmask8) __U);
3059 }
3060
3061 extern __inline __m128i
3062 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3063 _mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A)
3064 {
3065   return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A,
3066                                                     (__v4si)
3067                                                     _mm_setzero_si128 (),
3068                                                     (__mmask8) __U);
3069 }
3070
3071 extern __inline __m256i
3072 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3073 _mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A)
3074 {
3075   return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
3076                                                     (__v4di) __W,
3077                                                     (__mmask8) __U);
3078 }
3079
3080 extern __inline __m256i
3081 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3082 _mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
3083 {
3084   return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A,
3085                                                     (__v4di)
3086                                                     _mm256_setzero_si256 (),
3087                                                     (__mmask8) __U);
3088 }
3089
3090 extern __inline __m128i
3091 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3092 _mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
3093 {
3094   return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
3095                                                     (__v2di) __W,
3096                                                     (__mmask8) __U);
3097 }
3098
3099 extern __inline __m128i
3100 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3101 _mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A)
3102 {
3103   return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A,
3104                                                     (__v2di)
3105                                                     _mm_setzero_si128 (),
3106                                                     (__mmask8) __U);
3107 }
3108
3109 extern __inline __m256i
3110 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3111 _mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X)
3112 {
3113   return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
3114                                                     (__v4di) __W,
3115                                                     (__mmask8) __U);
3116 }
3117
3118 extern __inline __m256i
3119 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3120 _mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
3121 {
3122   return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X,
3123                                                     (__v4di)
3124                                                     _mm256_setzero_si256 (),
3125                                                     (__mmask8) __U);
3126 }
3127
3128 extern __inline __m128i
3129 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3130 _mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X)
3131 {
3132   return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
3133                                                     (__v2di) __W,
3134                                                     (__mmask8) __U);
3135 }
3136
3137 extern __inline __m128i
3138 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3139 _mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X)
3140 {
3141   return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X,
3142                                                     (__v2di)
3143                                                     _mm_setzero_si128 (),
3144                                                     (__mmask8) __U);
3145 }
3146
3147 extern __inline __m256d
3148 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3149 _mm256_rcp14_pd (__m256d __A)
3150 {
3151   return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
3152                                               (__v4df)
3153                                               _mm256_setzero_pd (),
3154                                               (__mmask8) -1);
3155 }
3156
3157 extern __inline __m256d
3158 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3159 _mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A)
3160 {
3161   return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
3162                                               (__v4df) __W,
3163                                               (__mmask8) __U);
3164 }
3165
3166 extern __inline __m256d
3167 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3168 _mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A)
3169 {
3170   return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A,
3171                                               (__v4df)
3172                                               _mm256_setzero_pd (),
3173                                               (__mmask8) __U);
3174 }
3175
3176 extern __inline __m128d
3177 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3178 _mm_rcp14_pd (__m128d __A)
3179 {
3180   return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
3181                                               (__v2df)
3182                                               _mm_setzero_pd (),
3183                                               (__mmask8) -1);
3184 }
3185
3186 extern __inline __m128d
3187 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3188 _mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A)
3189 {
3190   return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
3191                                               (__v2df) __W,
3192                                               (__mmask8) __U);
3193 }
3194
3195 extern __inline __m128d
3196 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3197 _mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A)
3198 {
3199   return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A,
3200                                               (__v2df)
3201                                               _mm_setzero_pd (),
3202                                               (__mmask8) __U);
3203 }
3204
3205 extern __inline __m256
3206 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3207 _mm256_rcp14_ps (__m256 __A)
3208 {
3209   return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
3210                                              (__v8sf)
3211                                              _mm256_setzero_ps (),
3212                                              (__mmask8) -1);
3213 }
3214
3215 extern __inline __m256
3216 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3217 _mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A)
3218 {
3219   return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
3220                                              (__v8sf) __W,
3221                                              (__mmask8) __U);
3222 }
3223
3224 extern __inline __m256
3225 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3226 _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A)
3227 {
3228   return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A,
3229                                              (__v8sf)
3230                                              _mm256_setzero_ps (),
3231                                              (__mmask8) __U);
3232 }
3233
3234 extern __inline __m128
3235 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3236 _mm_rcp14_ps (__m128 __A)
3237 {
3238   return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
3239                                              (__v4sf)
3240                                              _mm_setzero_ps (),
3241                                              (__mmask8) -1);
3242 }
3243
3244 extern __inline __m128
3245 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3246 _mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A)
3247 {
3248   return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
3249                                              (__v4sf) __W,
3250                                              (__mmask8) __U);
3251 }
3252
3253 extern __inline __m128
3254 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3255 _mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A)
3256 {
3257   return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A,
3258                                              (__v4sf)
3259                                              _mm_setzero_ps (),
3260                                              (__mmask8) __U);
3261 }
3262
3263 extern __inline __m256d
3264 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3265 _mm256_rsqrt14_pd (__m256d __A)
3266 {
3267   return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
3268                                                      (__v4df)
3269                                                      _mm256_setzero_pd (),
3270                                                      (__mmask8) -1);
3271 }
3272
3273 extern __inline __m256d
3274 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3275 _mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A)
3276 {
3277   return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
3278                                                      (__v4df) __W,
3279                                                      (__mmask8) __U);
3280 }
3281
3282 extern __inline __m256d
3283 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3284 _mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A)
3285 {
3286   return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A,
3287                                                      (__v4df)
3288                                                      _mm256_setzero_pd (),
3289                                                      (__mmask8) __U);
3290 }
3291
3292 extern __inline __m128d
3293 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3294 _mm_rsqrt14_pd (__m128d __A)
3295 {
3296   return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
3297                                                      (__v2df)
3298                                                      _mm_setzero_pd (),
3299                                                      (__mmask8) -1);
3300 }
3301
3302 extern __inline __m128d
3303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3304 _mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A)
3305 {
3306   return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
3307                                                      (__v2df) __W,
3308                                                      (__mmask8) __U);
3309 }
3310
3311 extern __inline __m128d
3312 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3313 _mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A)
3314 {
3315   return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A,
3316                                                      (__v2df)
3317                                                      _mm_setzero_pd (),
3318                                                      (__mmask8) __U);
3319 }
3320
3321 extern __inline __m256
3322 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3323 _mm256_rsqrt14_ps (__m256 __A)
3324 {
3325   return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
3326                                                     (__v8sf)
3327                                                     _mm256_setzero_ps (),
3328                                                     (__mmask8) -1);
3329 }
3330
3331 extern __inline __m256
3332 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3333 _mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A)
3334 {
3335   return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
3336                                                     (__v8sf) __W,
3337                                                     (__mmask8) __U);
3338 }
3339
3340 extern __inline __m256
3341 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3342 _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
3343 {
3344   return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A,
3345                                                     (__v8sf)
3346                                                     _mm256_setzero_ps (),
3347                                                     (__mmask8) __U);
3348 }
3349
3350 extern __inline __m128
3351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3352 _mm_rsqrt14_ps (__m128 __A)
3353 {
3354   return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
3355                                                     (__v4sf)
3356                                                     _mm_setzero_ps (),
3357                                                     (__mmask8) -1);
3358 }
3359
3360 extern __inline __m128
3361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3362 _mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A)
3363 {
3364   return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
3365                                                     (__v4sf) __W,
3366                                                     (__mmask8) __U);
3367 }
3368
3369 extern __inline __m128
3370 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3371 _mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A)
3372 {
3373   return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A,
3374                                                     (__v4sf)
3375                                                     _mm_setzero_ps (),
3376                                                     (__mmask8) __U);
3377 }
3378
3379 extern __inline __m256d
3380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3381 _mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A)
3382 {
3383   return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A,
3384                                                   (__v4df) __W,
3385                                                   (__mmask8) __U);
3386 }
3387
3388 extern __inline __m256d
3389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3390 _mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A)
3391 {
3392   return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A,
3393                                                   (__v4df)
3394                                                   _mm256_setzero_pd (),
3395                                                   (__mmask8) __U);
3396 }
3397
3398 extern __inline __m128d
3399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3400 _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A)
3401 {
3402   return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A,
3403                                                   (__v2df) __W,
3404                                                   (__mmask8) __U);
3405 }
3406
3407 extern __inline __m128d
3408 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3409 _mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A)
3410 {
3411   return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A,
3412                                                   (__v2df)
3413                                                   _mm_setzero_pd (),
3414                                                   (__mmask8) __U);
3415 }
3416
3417 extern __inline __m256
3418 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3419 _mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A)
3420 {
3421   return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A,
3422                                                  (__v8sf) __W,
3423                                                  (__mmask8) __U);
3424 }
3425
3426 extern __inline __m256
3427 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3428 _mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A)
3429 {
3430   return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A,
3431                                                  (__v8sf)
3432                                                  _mm256_setzero_ps (),
3433                                                  (__mmask8) __U);
3434 }
3435
3436 extern __inline __m128
3437 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3438 _mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A)
3439 {
3440   return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A,
3441                                                  (__v4sf) __W,
3442                                                  (__mmask8) __U);
3443 }
3444
3445 extern __inline __m128
3446 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3447 _mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A)
3448 {
3449   return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A,
3450                                                  (__v4sf)
3451                                                  _mm_setzero_ps (),
3452                                                  (__mmask8) __U);
3453 }
3454
3455 extern __inline __m256i
3456 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3457 _mm256_mask_add_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
3458                        __m256i __B)
3459 {
3460   return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A,
3461                                                  (__v8si) __B,
3462                                                  (__v8si) __W,
3463                                                  (__mmask8) __U);
3464 }
3465
3466 extern __inline __m256i
3467 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3468 _mm256_maskz_add_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
3469 {
3470   return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A,
3471                                                  (__v8si) __B,
3472                                                  (__v8si)
3473                                                  _mm256_setzero_si256 (),
3474                                                  (__mmask8) __U);
3475 }
3476
3477 extern __inline __m256i
3478 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3479 _mm256_mask_add_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
3480                        __m256i __B)
3481 {
3482   return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A,
3483                                                  (__v4di) __B,
3484                                                  (__v4di) __W,
3485                                                  (__mmask8) __U);
3486 }
3487
3488 extern __inline __m256i
3489 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3490 _mm256_maskz_add_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
3491 {
3492   return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A,
3493                                                  (__v4di) __B,
3494                                                  (__v4di)
3495                                                  _mm256_setzero_si256 (),
3496                                                  (__mmask8) __U);
3497 }
3498
3499 extern __inline __m256i
3500 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3501 _mm256_mask_sub_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
3502                        __m256i __B)
3503 {
3504   return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A,
3505                                                  (__v8si) __B,
3506                                                  (__v8si) __W,
3507                                                  (__mmask8) __U);
3508 }
3509
3510 extern __inline __m256i
3511 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3512 _mm256_maskz_sub_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
3513 {
3514   return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A,
3515                                                  (__v8si) __B,
3516                                                  (__v8si)
3517                                                  _mm256_setzero_si256 (),
3518                                                  (__mmask8) __U);
3519 }
3520
3521 extern __inline __m256i
3522 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3523 _mm256_mask_sub_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
3524                        __m256i __B)
3525 {
3526   return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A,
3527                                                  (__v4di) __B,
3528                                                  (__v4di) __W,
3529                                                  (__mmask8) __U);
3530 }
3531
3532 extern __inline __m256i
3533 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3534 _mm256_maskz_sub_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
3535 {
3536   return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A,
3537                                                  (__v4di) __B,
3538                                                  (__v4di)
3539                                                  _mm256_setzero_si256 (),
3540                                                  (__mmask8) __U);
3541 }
3542
3543 extern __inline __m128i
3544 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3545 _mm_mask_add_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
3546                     __m128i __B)
3547 {
3548   return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A,
3549                                                  (__v4si) __B,
3550                                                  (__v4si) __W,
3551                                                  (__mmask8) __U);
3552 }
3553
3554 extern __inline __m128i
3555 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3556 _mm_maskz_add_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
3557 {
3558   return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A,
3559                                                  (__v4si) __B,
3560                                                  (__v4si)
3561                                                  _mm_setzero_si128 (),
3562                                                  (__mmask8) __U);
3563 }
3564
3565 extern __inline __m128i
3566 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3567 _mm_mask_add_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
3568                     __m128i __B)
3569 {
3570   return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A,
3571                                                  (__v2di) __B,
3572                                                  (__v2di) __W,
3573                                                  (__mmask8) __U);
3574 }
3575
3576 extern __inline __m128i
3577 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3578 _mm_maskz_add_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
3579 {
3580   return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A,
3581                                                  (__v2di) __B,
3582                                                  (__v2di)
3583                                                  _mm_setzero_si128 (),
3584                                                  (__mmask8) __U);
3585 }
3586
3587 extern __inline __m128i
3588 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3589 _mm_mask_sub_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
3590                     __m128i __B)
3591 {
3592   return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A,
3593                                                  (__v4si) __B,
3594                                                  (__v4si) __W,
3595                                                  (__mmask8) __U);
3596 }
3597
3598 extern __inline __m128i
3599 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3600 _mm_maskz_sub_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
3601 {
3602   return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A,
3603                                                  (__v4si) __B,
3604                                                  (__v4si)
3605                                                  _mm_setzero_si128 (),
3606                                                  (__mmask8) __U);
3607 }
3608
3609 extern __inline __m128i
3610 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3611 _mm_mask_sub_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
3612                     __m128i __B)
3613 {
3614   return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A,
3615                                                  (__v2di) __B,
3616                                                  (__v2di) __W,
3617                                                  (__mmask8) __U);
3618 }
3619
3620 extern __inline __m128i
3621 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3622 _mm_maskz_sub_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
3623 {
3624   return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A,
3625                                                  (__v2di) __B,
3626                                                  (__v2di)
3627                                                  _mm_setzero_si128 (),
3628                                                  (__mmask8) __U);
3629 }
3630
3631 extern __inline __m256
3632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3633 _mm256_getexp_ps (__m256 __A)
3634 {
3635   return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3636                                                    (__v8sf)
3637                                                    _mm256_setzero_ps (),
3638                                                    (__mmask8) -1);
3639 }
3640
3641 extern __inline __m256
3642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3643 _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A)
3644 {
3645   return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3646                                                    (__v8sf) __W,
3647                                                    (__mmask8) __U);
3648 }
3649
3650 extern __inline __m256
3651 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3652 _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A)
3653 {
3654   return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A,
3655                                                    (__v8sf)
3656                                                    _mm256_setzero_ps (),
3657                                                    (__mmask8) __U);
3658 }
3659
3660 extern __inline __m256d
3661 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3662 _mm256_getexp_pd (__m256d __A)
3663 {
3664   return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3665                                                     (__v4df)
3666                                                     _mm256_setzero_pd (),
3667                                                     (__mmask8) -1);
3668 }
3669
3670 extern __inline __m256d
3671 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3672 _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A)
3673 {
3674   return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3675                                                     (__v4df) __W,
3676                                                     (__mmask8) __U);
3677 }
3678
3679 extern __inline __m256d
3680 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3681 _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A)
3682 {
3683   return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A,
3684                                                     (__v4df)
3685                                                     _mm256_setzero_pd (),
3686                                                     (__mmask8) __U);
3687 }
3688
3689 extern __inline __m128
3690 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3691 _mm_getexp_ps (__m128 __A)
3692 {
3693   return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3694                                                    (__v4sf)
3695                                                    _mm_setzero_ps (),
3696                                                    (__mmask8) -1);
3697 }
3698
3699 extern __inline __m128
3700 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3701 _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A)
3702 {
3703   return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3704                                                    (__v4sf) __W,
3705                                                    (__mmask8) __U);
3706 }
3707
3708 extern __inline __m128
3709 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3710 _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A)
3711 {
3712   return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A,
3713                                                    (__v4sf)
3714                                                    _mm_setzero_ps (),
3715                                                    (__mmask8) __U);
3716 }
3717
3718 extern __inline __m128d
3719 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3720 _mm_getexp_pd (__m128d __A)
3721 {
3722   return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3723                                                     (__v2df)
3724                                                     _mm_setzero_pd (),
3725                                                     (__mmask8) -1);
3726 }
3727
3728 extern __inline __m128d
3729 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3730 _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A)
3731 {
3732   return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3733                                                     (__v2df) __W,
3734                                                     (__mmask8) __U);
3735 }
3736
3737 extern __inline __m128d
3738 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3739 _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A)
3740 {
3741   return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A,
3742                                                     (__v2df)
3743                                                     _mm_setzero_pd (),
3744                                                     (__mmask8) __U);
3745 }
3746
3747 extern __inline __m256i
3748 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3749 _mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
3750                        __m128i __B)
3751 {
3752   return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
3753                                                  (__v4si) __B,
3754                                                  (__v8si) __W,
3755                                                  (__mmask8) __U);
3756 }
3757
3758 extern __inline __m256i
3759 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3760 _mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
3761 {
3762   return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A,
3763                                                  (__v4si) __B,
3764                                                  (__v8si)
3765                                                  _mm256_setzero_si256 (),
3766                                                  (__mmask8) __U);
3767 }
3768
3769 extern __inline __m128i
3770 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3771 _mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
3772                     __m128i __B)
3773 {
3774   return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
3775                                                  (__v4si) __B,
3776                                                  (__v4si) __W,
3777                                                  (__mmask8) __U);
3778 }
3779
3780 extern __inline __m128i
3781 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3782 _mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
3783 {
3784   return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A,
3785                                                  (__v4si) __B,
3786                                                  (__v4si)
3787                                                  _mm_setzero_si128 (),
3788                                                  (__mmask8) __U);
3789 }
3790
3791 extern __inline __m256i
3792 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3793 _mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
3794                        __m128i __B)
3795 {
3796   return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
3797                                                  (__v2di) __B,
3798                                                  (__v4di) __W,
3799                                                  (__mmask8) __U);
3800 }
3801
3802 extern __inline __m256i
3803 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3804 _mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
3805 {
3806   return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A,
3807                                                  (__v2di) __B,
3808                                                  (__v4di)
3809                                                  _mm256_setzero_si256 (),
3810                                                  (__mmask8) __U);
3811 }
3812
3813 extern __inline __m128i
3814 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3815 _mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
3816                     __m128i __B)
3817 {
3818   return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
3819                                                  (__v2di) __B,
3820                                                  (__v2di) __W,
3821                                                  (__mmask8) __U);
3822 }
3823
3824 extern __inline __m128i
3825 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3826 _mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
3827 {
3828   return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A,
3829                                                  (__v2di) __B,
3830                                                  (__v2di)
3831                                                  _mm_setzero_si128 (),
3832                                                  (__mmask8) __U);
3833 }
3834
3835 extern __inline __m256i
3836 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3837 _mm256_mask_and_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
3838                        __m256i __B)
3839 {
3840   return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
3841                                                  (__v8si) __B,
3842                                                  (__v8si) __W,
3843                                                  (__mmask8) __U);
3844 }
3845
3846 extern __inline __m256i
3847 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3848 _mm256_maskz_and_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
3849 {
3850   return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A,
3851                                                  (__v8si) __B,
3852                                                  (__v8si)
3853                                                  _mm256_setzero_si256 (),
3854                                                  (__mmask8) __U);
3855 }
3856
3857 extern __inline __m256d
3858 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3859 _mm256_scalef_pd (__m256d __A, __m256d __B)
3860 {
3861   return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3862                                                     (__v4df) __B,
3863                                                     (__v4df)
3864                                                     _mm256_setzero_pd (),
3865                                                     (__mmask8) -1);
3866 }
3867
3868 extern __inline __m256d
3869 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3870 _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A,
3871                        __m256d __B)
3872 {
3873   return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3874                                                     (__v4df) __B,
3875                                                     (__v4df) __W,
3876                                                     (__mmask8) __U);
3877 }
3878
3879 extern __inline __m256d
3880 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3881 _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B)
3882 {
3883   return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A,
3884                                                     (__v4df) __B,
3885                                                     (__v4df)
3886                                                     _mm256_setzero_pd (),
3887                                                     (__mmask8) __U);
3888 }
3889
3890 extern __inline __m256
3891 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3892 _mm256_scalef_ps (__m256 __A, __m256 __B)
3893 {
3894   return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3895                                                    (__v8sf) __B,
3896                                                    (__v8sf)
3897                                                    _mm256_setzero_ps (),
3898                                                    (__mmask8) -1);
3899 }
3900
3901 extern __inline __m256
3902 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3903 _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A,
3904                        __m256 __B)
3905 {
3906   return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3907                                                    (__v8sf) __B,
3908                                                    (__v8sf) __W,
3909                                                    (__mmask8) __U);
3910 }
3911
3912 extern __inline __m256
3913 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3914 _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B)
3915 {
3916   return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A,
3917                                                    (__v8sf) __B,
3918                                                    (__v8sf)
3919                                                    _mm256_setzero_ps (),
3920                                                    (__mmask8) __U);
3921 }
3922
3923 extern __inline __m128d
3924 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3925 _mm_scalef_pd (__m128d __A, __m128d __B)
3926 {
3927   return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3928                                                     (__v2df) __B,
3929                                                     (__v2df)
3930                                                     _mm_setzero_pd (),
3931                                                     (__mmask8) -1);
3932 }
3933
3934 extern __inline __m128d
3935 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3936 _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A,
3937                     __m128d __B)
3938 {
3939   return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3940                                                     (__v2df) __B,
3941                                                     (__v2df) __W,
3942                                                     (__mmask8) __U);
3943 }
3944
3945 extern __inline __m128d
3946 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3947 _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B)
3948 {
3949   return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A,
3950                                                     (__v2df) __B,
3951                                                     (__v2df)
3952                                                     _mm_setzero_pd (),
3953                                                     (__mmask8) __U);
3954 }
3955
3956 extern __inline __m128
3957 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3958 _mm_scalef_ps (__m128 __A, __m128 __B)
3959 {
3960   return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3961                                                    (__v4sf) __B,
3962                                                    (__v4sf)
3963                                                    _mm_setzero_ps (),
3964                                                    (__mmask8) -1);
3965 }
3966
3967 extern __inline __m128
3968 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3969 _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
3970 {
3971   return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3972                                                    (__v4sf) __B,
3973                                                    (__v4sf) __W,
3974                                                    (__mmask8) __U);
3975 }
3976
3977 extern __inline __m128
3978 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3979 _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B)
3980 {
3981   return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A,
3982                                                    (__v4sf) __B,
3983                                                    (__v4sf)
3984                                                    _mm_setzero_ps (),
3985                                                    (__mmask8) __U);
3986 }
3987
3988 extern __inline __m256d
3989 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
3990 _mm256_mask_fmadd_pd (__m256d __A, __mmask8 __U, __m256d __B,
3991                       __m256d __C)
3992 {
3993   return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A,
3994                                                     (__v4df) __B,
3995                                                     (__v4df) __C,
3996                                                     (__mmask8) __U);
3997 }
3998
3999 extern __inline __m256d
4000 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4001 _mm256_mask3_fmadd_pd (__m256d __A, __m256d __B, __m256d __C,
4002                        __mmask8 __U)
4003 {
4004   return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A,
4005                                                      (__v4df) __B,
4006                                                      (__v4df) __C,
4007                                                      (__mmask8) __U);
4008 }
4009
4010 extern __inline __m256d
4011 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4012 _mm256_maskz_fmadd_pd (__mmask8 __U, __m256d __A, __m256d __B,
4013                        __m256d __C)
4014 {
4015   return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A,
4016                                                      (__v4df) __B,
4017                                                      (__v4df) __C,
4018                                                      (__mmask8) __U);
4019 }
4020
4021 extern __inline __m128d
4022 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4023 _mm_mask_fmadd_pd (__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
4024 {
4025   return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A,
4026                                                     (__v2df) __B,
4027                                                     (__v2df) __C,
4028                                                     (__mmask8) __U);
4029 }
4030
4031 extern __inline __m128d
4032 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4033 _mm_mask3_fmadd_pd (__m128d __A, __m128d __B, __m128d __C,
4034                     __mmask8 __U)
4035 {
4036   return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A,
4037                                                      (__v2df) __B,
4038                                                      (__v2df) __C,
4039                                                      (__mmask8) __U);
4040 }
4041
4042 extern __inline __m128d
4043 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4044 _mm_maskz_fmadd_pd (__mmask8 __U, __m128d __A, __m128d __B,
4045                     __m128d __C)
4046 {
4047   return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A,
4048                                                      (__v2df) __B,
4049                                                      (__v2df) __C,
4050                                                      (__mmask8) __U);
4051 }
4052
4053 extern __inline __m256
4054 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4055 _mm256_mask_fmadd_ps (__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
4056 {
4057   return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A,
4058                                                    (__v8sf) __B,
4059                                                    (__v8sf) __C,
4060                                                    (__mmask8) __U);
4061 }
4062
4063 extern __inline __m256
4064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4065 _mm256_mask3_fmadd_ps (__m256 __A, __m256 __B, __m256 __C,
4066                        __mmask8 __U)
4067 {
4068   return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A,
4069                                                     (__v8sf) __B,
4070                                                     (__v8sf) __C,
4071                                                     (__mmask8) __U);
4072 }
4073
4074 extern __inline __m256
4075 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4076 _mm256_maskz_fmadd_ps (__mmask8 __U, __m256 __A, __m256 __B,
4077                        __m256 __C)
4078 {
4079   return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A,
4080                                                     (__v8sf) __B,
4081                                                     (__v8sf) __C,
4082                                                     (__mmask8) __U);
4083 }
4084
4085 extern __inline __m128
4086 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4087 _mm_mask_fmadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4088 {
4089   return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A,
4090                                                    (__v4sf) __B,
4091                                                    (__v4sf) __C,
4092                                                    (__mmask8) __U);
4093 }
4094
4095 extern __inline __m128
4096 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4097 _mm_mask3_fmadd_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
4098 {
4099   return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A,
4100                                                     (__v4sf) __B,
4101                                                     (__v4sf) __C,
4102                                                     (__mmask8) __U);
4103 }
4104
4105 extern __inline __m128
4106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4107 _mm_maskz_fmadd_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
4108 {
4109   return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A,
4110                                                     (__v4sf) __B,
4111                                                     (__v4sf) __C,
4112                                                     (__mmask8) __U);
4113 }
4114
4115 extern __inline __m256d
4116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4117 _mm256_mask_fmsub_pd (__m256d __A, __mmask8 __U, __m256d __B,
4118                       __m256d __C)
4119 {
4120   return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A,
4121                                                     (__v4df) __B,
4122                                                     -(__v4df) __C,
4123                                                     (__mmask8) __U);
4124 }
4125
4126 extern __inline __m256d
4127 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4128 _mm256_mask3_fmsub_pd (__m256d __A, __m256d __B, __m256d __C,
4129                        __mmask8 __U)
4130 {
4131   return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A,
4132                                                      (__v4df) __B,
4133                                                      (__v4df) __C,
4134                                                      (__mmask8) __U);
4135 }
4136
4137 extern __inline __m256d
4138 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4139 _mm256_maskz_fmsub_pd (__mmask8 __U, __m256d __A, __m256d __B,
4140                        __m256d __C)
4141 {
4142   return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A,
4143                                                      (__v4df) __B,
4144                                                      -(__v4df) __C,
4145                                                      (__mmask8) __U);
4146 }
4147
4148 extern __inline __m128d
4149 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4150 _mm_mask_fmsub_pd (__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
4151 {
4152   return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A,
4153                                                     (__v2df) __B,
4154                                                     -(__v2df) __C,
4155                                                     (__mmask8) __U);
4156 }
4157
4158 extern __inline __m128d
4159 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4160 _mm_mask3_fmsub_pd (__m128d __A, __m128d __B, __m128d __C,
4161                     __mmask8 __U)
4162 {
4163   return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A,
4164                                                      (__v2df) __B,
4165                                                      (__v2df) __C,
4166                                                      (__mmask8) __U);
4167 }
4168
4169 extern __inline __m128d
4170 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4171 _mm_maskz_fmsub_pd (__mmask8 __U, __m128d __A, __m128d __B,
4172                     __m128d __C)
4173 {
4174   return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A,
4175                                                      (__v2df) __B,
4176                                                      -(__v2df) __C,
4177                                                      (__mmask8) __U);
4178 }
4179
4180 extern __inline __m256
4181 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4182 _mm256_mask_fmsub_ps (__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
4183 {
4184   return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A,
4185                                                    (__v8sf) __B,
4186                                                    -(__v8sf) __C,
4187                                                    (__mmask8) __U);
4188 }
4189
4190 extern __inline __m256
4191 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4192 _mm256_mask3_fmsub_ps (__m256 __A, __m256 __B, __m256 __C,
4193                        __mmask8 __U)
4194 {
4195   return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A,
4196                                                     (__v8sf) __B,
4197                                                     (__v8sf) __C,
4198                                                     (__mmask8) __U);
4199 }
4200
4201 extern __inline __m256
4202 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4203 _mm256_maskz_fmsub_ps (__mmask8 __U, __m256 __A, __m256 __B,
4204                        __m256 __C)
4205 {
4206   return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A,
4207                                                     (__v8sf) __B,
4208                                                     -(__v8sf) __C,
4209                                                     (__mmask8) __U);
4210 }
4211
4212 extern __inline __m128
4213 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4214 _mm_mask_fmsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4215 {
4216   return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A,
4217                                                    (__v4sf) __B,
4218                                                    -(__v4sf) __C,
4219                                                    (__mmask8) __U);
4220 }
4221
4222 extern __inline __m128
4223 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4224 _mm_mask3_fmsub_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
4225 {
4226   return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A,
4227                                                     (__v4sf) __B,
4228                                                     (__v4sf) __C,
4229                                                     (__mmask8) __U);
4230 }
4231
4232 extern __inline __m128
4233 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4234 _mm_maskz_fmsub_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
4235 {
4236   return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A,
4237                                                     (__v4sf) __B,
4238                                                     -(__v4sf) __C,
4239                                                     (__mmask8) __U);
4240 }
4241
4242 extern __inline __m256d
4243 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4244 _mm256_mask_fmaddsub_pd (__m256d __A, __mmask8 __U, __m256d __B,
4245                          __m256d __C)
4246 {
4247   return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A,
4248                                                        (__v4df) __B,
4249                                                        (__v4df) __C,
4250                                                        (__mmask8) __U);
4251 }
4252
4253 extern __inline __m256d
4254 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4255 _mm256_mask3_fmaddsub_pd (__m256d __A, __m256d __B, __m256d __C,
4256                           __mmask8 __U)
4257 {
4258   return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A,
4259                                                         (__v4df) __B,
4260                                                         (__v4df) __C,
4261                                                         (__mmask8)
4262                                                         __U);
4263 }
4264
4265 extern __inline __m256d
4266 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4267 _mm256_maskz_fmaddsub_pd (__mmask8 __U, __m256d __A, __m256d __B,
4268                           __m256d __C)
4269 {
4270   return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A,
4271                                                         (__v4df) __B,
4272                                                         (__v4df) __C,
4273                                                         (__mmask8)
4274                                                         __U);
4275 }
4276
4277 extern __inline __m128d
4278 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4279 _mm_mask_fmaddsub_pd (__m128d __A, __mmask8 __U, __m128d __B,
4280                       __m128d __C)
4281 {
4282   return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A,
4283                                                        (__v2df) __B,
4284                                                        (__v2df) __C,
4285                                                        (__mmask8) __U);
4286 }
4287
4288 extern __inline __m128d
4289 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4290 _mm_mask3_fmaddsub_pd (__m128d __A, __m128d __B, __m128d __C,
4291                        __mmask8 __U)
4292 {
4293   return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A,
4294                                                         (__v2df) __B,
4295                                                         (__v2df) __C,
4296                                                         (__mmask8)
4297                                                         __U);
4298 }
4299
4300 extern __inline __m128d
4301 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4302 _mm_maskz_fmaddsub_pd (__mmask8 __U, __m128d __A, __m128d __B,
4303                        __m128d __C)
4304 {
4305   return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A,
4306                                                         (__v2df) __B,
4307                                                         (__v2df) __C,
4308                                                         (__mmask8)
4309                                                         __U);
4310 }
4311
4312 extern __inline __m256
4313 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4314 _mm256_mask_fmaddsub_ps (__m256 __A, __mmask8 __U, __m256 __B,
4315                          __m256 __C)
4316 {
4317   return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A,
4318                                                       (__v8sf) __B,
4319                                                       (__v8sf) __C,
4320                                                       (__mmask8) __U);
4321 }
4322
4323 extern __inline __m256
4324 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4325 _mm256_mask3_fmaddsub_ps (__m256 __A, __m256 __B, __m256 __C,
4326                           __mmask8 __U)
4327 {
4328   return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A,
4329                                                        (__v8sf) __B,
4330                                                        (__v8sf) __C,
4331                                                        (__mmask8) __U);
4332 }
4333
4334 extern __inline __m256
4335 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4336 _mm256_maskz_fmaddsub_ps (__mmask8 __U, __m256 __A, __m256 __B,
4337                           __m256 __C)
4338 {
4339   return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A,
4340                                                        (__v8sf) __B,
4341                                                        (__v8sf) __C,
4342                                                        (__mmask8) __U);
4343 }
4344
4345 extern __inline __m128
4346 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4347 _mm_mask_fmaddsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4348 {
4349   return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A,
4350                                                       (__v4sf) __B,
4351                                                       (__v4sf) __C,
4352                                                       (__mmask8) __U);
4353 }
4354
4355 extern __inline __m128
4356 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4357 _mm_mask3_fmaddsub_ps (__m128 __A, __m128 __B, __m128 __C,
4358                        __mmask8 __U)
4359 {
4360   return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A,
4361                                                        (__v4sf) __B,
4362                                                        (__v4sf) __C,
4363                                                        (__mmask8) __U);
4364 }
4365
4366 extern __inline __m128
4367 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4368 _mm_maskz_fmaddsub_ps (__mmask8 __U, __m128 __A, __m128 __B,
4369                        __m128 __C)
4370 {
4371   return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A,
4372                                                        (__v4sf) __B,
4373                                                        (__v4sf) __C,
4374                                                        (__mmask8) __U);
4375 }
4376
4377 extern __inline __m256d
4378 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4379 _mm256_mask_fmsubadd_pd (__m256d __A, __mmask8 __U, __m256d __B,
4380                          __m256d __C)
4381 {
4382   return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A,
4383                                                        (__v4df) __B,
4384                                                        -(__v4df) __C,
4385                                                        (__mmask8) __U);
4386 }
4387
4388 extern __inline __m256d
4389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4390 _mm256_mask3_fmsubadd_pd (__m256d __A, __m256d __B, __m256d __C,
4391                           __mmask8 __U)
4392 {
4393   return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A,
4394                                                         (__v4df) __B,
4395                                                         (__v4df) __C,
4396                                                         (__mmask8)
4397                                                         __U);
4398 }
4399
4400 extern __inline __m256d
4401 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4402 _mm256_maskz_fmsubadd_pd (__mmask8 __U, __m256d __A, __m256d __B,
4403                           __m256d __C)
4404 {
4405   return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A,
4406                                                         (__v4df) __B,
4407                                                         -(__v4df) __C,
4408                                                         (__mmask8)
4409                                                         __U);
4410 }
4411
4412 extern __inline __m128d
4413 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4414 _mm_mask_fmsubadd_pd (__m128d __A, __mmask8 __U, __m128d __B,
4415                       __m128d __C)
4416 {
4417   return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A,
4418                                                        (__v2df) __B,
4419                                                        -(__v2df) __C,
4420                                                        (__mmask8) __U);
4421 }
4422
4423 extern __inline __m128d
4424 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4425 _mm_mask3_fmsubadd_pd (__m128d __A, __m128d __B, __m128d __C,
4426                        __mmask8 __U)
4427 {
4428   return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A,
4429                                                         (__v2df) __B,
4430                                                         (__v2df) __C,
4431                                                         (__mmask8)
4432                                                         __U);
4433 }
4434
4435 extern __inline __m128d
4436 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4437 _mm_maskz_fmsubadd_pd (__mmask8 __U, __m128d __A, __m128d __B,
4438                        __m128d __C)
4439 {
4440   return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A,
4441                                                         (__v2df) __B,
4442                                                         -(__v2df) __C,
4443                                                         (__mmask8)
4444                                                         __U);
4445 }
4446
4447 extern __inline __m256
4448 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4449 _mm256_mask_fmsubadd_ps (__m256 __A, __mmask8 __U, __m256 __B,
4450                          __m256 __C)
4451 {
4452   return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A,
4453                                                       (__v8sf) __B,
4454                                                       -(__v8sf) __C,
4455                                                       (__mmask8) __U);
4456 }
4457
4458 extern __inline __m256
4459 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4460 _mm256_mask3_fmsubadd_ps (__m256 __A, __m256 __B, __m256 __C,
4461                           __mmask8 __U)
4462 {
4463   return (__m256) __builtin_ia32_vfmsubaddps256_mask3 ((__v8sf) __A,
4464                                                        (__v8sf) __B,
4465                                                        (__v8sf) __C,
4466                                                        (__mmask8) __U);
4467 }
4468
4469 extern __inline __m256
4470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4471 _mm256_maskz_fmsubadd_ps (__mmask8 __U, __m256 __A, __m256 __B,
4472                           __m256 __C)
4473 {
4474   return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A,
4475                                                        (__v8sf) __B,
4476                                                        -(__v8sf) __C,
4477                                                        (__mmask8) __U);
4478 }
4479
4480 extern __inline __m128
4481 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4482 _mm_mask_fmsubadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4483 {
4484   return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A,
4485                                                       (__v4sf) __B,
4486                                                       -(__v4sf) __C,
4487                                                       (__mmask8) __U);
4488 }
4489
4490 extern __inline __m128
4491 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4492 _mm_mask3_fmsubadd_ps (__m128 __A, __m128 __B, __m128 __C,
4493                        __mmask8 __U)
4494 {
4495   return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A,
4496                                                        (__v4sf) __B,
4497                                                        (__v4sf) __C,
4498                                                        (__mmask8) __U);
4499 }
4500
4501 extern __inline __m128
4502 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4503 _mm_maskz_fmsubadd_ps (__mmask8 __U, __m128 __A, __m128 __B,
4504                        __m128 __C)
4505 {
4506   return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A,
4507                                                        (__v4sf) __B,
4508                                                        -(__v4sf) __C,
4509                                                        (__mmask8) __U);
4510 }
4511
4512 extern __inline __m256d
4513 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4514 _mm256_mask_fnmadd_pd (__m256d __A, __mmask8 __U, __m256d __B,
4515                        __m256d __C)
4516 {
4517   return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A,
4518                                                      (__v4df) __B,
4519                                                      (__v4df) __C,
4520                                                      (__mmask8) __U);
4521 }
4522
4523 extern __inline __m256d
4524 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4525 _mm256_mask3_fnmadd_pd (__m256d __A, __m256d __B, __m256d __C,
4526                         __mmask8 __U)
4527 {
4528   return (__m256d) __builtin_ia32_vfmaddpd256_mask3 (-(__v4df) __A,
4529                                                      (__v4df) __B,
4530                                                      (__v4df) __C,
4531                                                      (__mmask8) __U);
4532 }
4533
4534 extern __inline __m256d
4535 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4536 _mm256_maskz_fnmadd_pd (__mmask8 __U, __m256d __A, __m256d __B,
4537                         __m256d __C)
4538 {
4539   return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A,
4540                                                      (__v4df) __B,
4541                                                      (__v4df) __C,
4542                                                      (__mmask8) __U);
4543 }
4544
4545 extern __inline __m128d
4546 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4547 _mm_mask_fnmadd_pd (__m128d __A, __mmask8 __U, __m128d __B,
4548                     __m128d __C)
4549 {
4550   return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A,
4551                                                      (__v2df) __B,
4552                                                      (__v2df) __C,
4553                                                      (__mmask8) __U);
4554 }
4555
4556 extern __inline __m128d
4557 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4558 _mm_mask3_fnmadd_pd (__m128d __A, __m128d __B, __m128d __C,
4559                      __mmask8 __U)
4560 {
4561   return (__m128d) __builtin_ia32_vfmaddpd128_mask3 (-(__v2df) __A,
4562                                                      (__v2df) __B,
4563                                                      (__v2df) __C,
4564                                                      (__mmask8) __U);
4565 }
4566
4567 extern __inline __m128d
4568 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4569 _mm_maskz_fnmadd_pd (__mmask8 __U, __m128d __A, __m128d __B,
4570                      __m128d __C)
4571 {
4572   return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A,
4573                                                      (__v2df) __B,
4574                                                      (__v2df) __C,
4575                                                      (__mmask8) __U);
4576 }
4577
4578 extern __inline __m256
4579 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4580 _mm256_mask_fnmadd_ps (__m256 __A, __mmask8 __U, __m256 __B,
4581                        __m256 __C)
4582 {
4583   return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A,
4584                                                     (__v8sf) __B,
4585                                                     (__v8sf) __C,
4586                                                     (__mmask8) __U);
4587 }
4588
4589 extern __inline __m256
4590 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4591 _mm256_mask3_fnmadd_ps (__m256 __A, __m256 __B, __m256 __C,
4592                         __mmask8 __U)
4593 {
4594   return (__m256) __builtin_ia32_vfmaddps256_mask3 (-(__v8sf) __A,
4595                                                     (__v8sf) __B,
4596                                                     (__v8sf) __C,
4597                                                     (__mmask8) __U);
4598 }
4599
4600 extern __inline __m256
4601 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4602 _mm256_maskz_fnmadd_ps (__mmask8 __U, __m256 __A, __m256 __B,
4603                         __m256 __C)
4604 {
4605   return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A,
4606                                                     (__v8sf) __B,
4607                                                     (__v8sf) __C,
4608                                                     (__mmask8) __U);
4609 }
4610
4611 extern __inline __m128
4612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4613 _mm_mask_fnmadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4614 {
4615   return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A,
4616                                                     (__v4sf) __B,
4617                                                     (__v4sf) __C,
4618                                                     (__mmask8) __U);
4619 }
4620
4621 extern __inline __m128
4622 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4623 _mm_mask3_fnmadd_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
4624 {
4625   return (__m128) __builtin_ia32_vfmaddps128_mask3 (-(__v4sf) __A,
4626                                                     (__v4sf) __B,
4627                                                     (__v4sf) __C,
4628                                                     (__mmask8) __U);
4629 }
4630
4631 extern __inline __m128
4632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4633 _mm_maskz_fnmadd_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
4634 {
4635   return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A,
4636                                                     (__v4sf) __B,
4637                                                     (__v4sf) __C,
4638                                                     (__mmask8) __U);
4639 }
4640
4641 extern __inline __m256d
4642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4643 _mm256_mask_fnmsub_pd (__m256d __A, __mmask8 __U, __m256d __B,
4644                        __m256d __C)
4645 {
4646   return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A,
4647                                                      (__v4df) __B,
4648                                                      (__v4df) __C,
4649                                                      (__mmask8) __U);
4650 }
4651
4652 extern __inline __m256d
4653 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4654 _mm256_mask3_fnmsub_pd (__m256d __A, __m256d __B, __m256d __C,
4655                         __mmask8 __U)
4656 {
4657   return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A,
4658                                                       (__v4df) __B,
4659                                                       (__v4df) __C,
4660                                                       (__mmask8) __U);
4661 }
4662
4663 extern __inline __m256d
4664 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4665 _mm256_maskz_fnmsub_pd (__mmask8 __U, __m256d __A, __m256d __B,
4666                         __m256d __C)
4667 {
4668   return (__m256d) __builtin_ia32_vfmaddpd256_maskz (-(__v4df) __A,
4669                                                      (__v4df) __B,
4670                                                      -(__v4df) __C,
4671                                                      (__mmask8) __U);
4672 }
4673
4674 extern __inline __m128d
4675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4676 _mm_mask_fnmsub_pd (__m128d __A, __mmask8 __U, __m128d __B,
4677                     __m128d __C)
4678 {
4679   return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A,
4680                                                      (__v2df) __B,
4681                                                      (__v2df) __C,
4682                                                      (__mmask8) __U);
4683 }
4684
4685 extern __inline __m128d
4686 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4687 _mm_mask3_fnmsub_pd (__m128d __A, __m128d __B, __m128d __C,
4688                      __mmask8 __U)
4689 {
4690   return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A,
4691                                                       (__v2df) __B,
4692                                                       (__v2df) __C,
4693                                                       (__mmask8) __U);
4694 }
4695
4696 extern __inline __m128d
4697 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4698 _mm_maskz_fnmsub_pd (__mmask8 __U, __m128d __A, __m128d __B,
4699                      __m128d __C)
4700 {
4701   return (__m128d) __builtin_ia32_vfmaddpd128_maskz (-(__v2df) __A,
4702                                                      (__v2df) __B,
4703                                                      -(__v2df) __C,
4704                                                      (__mmask8) __U);
4705 }
4706
4707 extern __inline __m256
4708 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4709 _mm256_mask_fnmsub_ps (__m256 __A, __mmask8 __U, __m256 __B,
4710                        __m256 __C)
4711 {
4712   return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A,
4713                                                     (__v8sf) __B,
4714                                                     (__v8sf) __C,
4715                                                     (__mmask8) __U);
4716 }
4717
4718 extern __inline __m256
4719 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4720 _mm256_mask3_fnmsub_ps (__m256 __A, __m256 __B, __m256 __C,
4721                         __mmask8 __U)
4722 {
4723   return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A,
4724                                                      (__v8sf) __B,
4725                                                      (__v8sf) __C,
4726                                                      (__mmask8) __U);
4727 }
4728
4729 extern __inline __m256
4730 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4731 _mm256_maskz_fnmsub_ps (__mmask8 __U, __m256 __A, __m256 __B,
4732                         __m256 __C)
4733 {
4734   return (__m256) __builtin_ia32_vfmaddps256_maskz (-(__v8sf) __A,
4735                                                     (__v8sf) __B,
4736                                                     -(__v8sf) __C,
4737                                                     (__mmask8) __U);
4738 }
4739
4740 extern __inline __m128
4741 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4742 _mm_mask_fnmsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
4743 {
4744   return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A,
4745                                                     (__v4sf) __B,
4746                                                     (__v4sf) __C,
4747                                                     (__mmask8) __U);
4748 }
4749
4750 extern __inline __m128
4751 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4752 _mm_mask3_fnmsub_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
4753 {
4754   return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A,
4755                                                      (__v4sf) __B,
4756                                                      (__v4sf) __C,
4757                                                      (__mmask8) __U);
4758 }
4759
4760 extern __inline __m128
4761 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4762 _mm_maskz_fnmsub_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C)
4763 {
4764   return (__m128) __builtin_ia32_vfmaddps128_maskz (-(__v4sf) __A,
4765                                                     (__v4sf) __B,
4766                                                     -(__v4sf) __C,
4767                                                     (__mmask8) __U);
4768 }
4769
4770 extern __inline __m128i
4771 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4772 _mm_mask_and_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
4773                     __m128i __B)
4774 {
4775   return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
4776                                                  (__v4si) __B,
4777                                                  (__v4si) __W,
4778                                                  (__mmask8) __U);
4779 }
4780
4781 extern __inline __m128i
4782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4783 _mm_maskz_and_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
4784 {
4785   return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A,
4786                                                  (__v4si) __B,
4787                                                  (__v4si)
4788                                                  _mm_setzero_si128 (),
4789                                                  (__mmask8) __U);
4790 }
4791
4792 extern __inline __m256i
4793 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4794 _mm256_mask_andnot_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
4795                           __m256i __B)
4796 {
4797   return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
4798                                                   (__v8si) __B,
4799                                                   (__v8si) __W,
4800                                                   (__mmask8) __U);
4801 }
4802
4803 extern __inline __m256i
4804 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4805 _mm256_maskz_andnot_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
4806 {
4807   return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A,
4808                                                   (__v8si) __B,
4809                                                   (__v8si)
4810                                                   _mm256_setzero_si256 (),
4811                                                   (__mmask8) __U);
4812 }
4813
4814 extern __inline __m128i
4815 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4816 _mm_mask_andnot_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
4817                        __m128i __B)
4818 {
4819   return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
4820                                                   (__v4si) __B,
4821                                                   (__v4si) __W,
4822                                                   (__mmask8) __U);
4823 }
4824
4825 extern __inline __m128i
4826 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4827 _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
4828 {
4829   return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A,
4830                                                   (__v4si) __B,
4831                                                   (__v4si)
4832                                                   _mm_setzero_si128 (),
4833                                                   (__mmask8) __U);
4834 }
4835
4836 extern __inline __m256i
4837 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4838 _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
4839                       __m256i __B)
4840 {
4841   return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
4842                                                 (__v8si) __B,
4843                                                 (__v8si) __W,
4844                                                 (__mmask8) __U);
4845 }
4846
4847 extern __inline __m256i
4848 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4849 _mm256_maskz_or_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
4850 {
4851   return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A,
4852                                                 (__v8si) __B,
4853                                                 (__v8si)
4854                                                 _mm256_setzero_si256 (),
4855                                                 (__mmask8) __U);
4856 }
4857
4858 extern __inline __m128i
4859 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4860 _mm_mask_or_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
4861 {
4862   return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
4863                                                 (__v4si) __B,
4864                                                 (__v4si) __W,
4865                                                 (__mmask8) __U);
4866 }
4867
4868 extern __inline __m128i
4869 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4870 _mm_maskz_or_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
4871 {
4872   return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A,
4873                                                 (__v4si) __B,
4874                                                 (__v4si)
4875                                                 _mm_setzero_si128 (),
4876                                                 (__mmask8) __U);
4877 }
4878
4879 extern __inline __m256i
4880 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4881 _mm256_mask_xor_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
4882                        __m256i __B)
4883 {
4884   return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
4885                                                  (__v8si) __B,
4886                                                  (__v8si) __W,
4887                                                  (__mmask8) __U);
4888 }
4889
4890 extern __inline __m256i
4891 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4892 _mm256_maskz_xor_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
4893 {
4894   return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A,
4895                                                  (__v8si) __B,
4896                                                  (__v8si)
4897                                                  _mm256_setzero_si256 (),
4898                                                  (__mmask8) __U);
4899 }
4900
4901 extern __inline __m128i
4902 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4903 _mm_mask_xor_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
4904                     __m128i __B)
4905 {
4906   return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
4907                                                  (__v4si) __B,
4908                                                  (__v4si) __W,
4909                                                  (__mmask8) __U);
4910 }
4911
4912 extern __inline __m128i
4913 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4914 _mm_maskz_xor_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
4915 {
4916   return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A,
4917                                                  (__v4si) __B,
4918                                                  (__v4si)
4919                                                  _mm_setzero_si128 (),
4920                                                  (__mmask8) __U);
4921 }
4922
4923 extern __inline __m128
4924 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4925 _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A)
4926 {
4927   return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
4928                                                 (__v4sf) __W,
4929                                                 (__mmask8) __U);
4930 }
4931
4932 extern __inline __m128
4933 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4934 _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A)
4935 {
4936   return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
4937                                                 (__v4sf)
4938                                                 _mm_setzero_ps (),
4939                                                 (__mmask8) __U);
4940 }
4941
4942 extern __inline __m128
4943 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4944 _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A)
4945 {
4946   return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A,
4947                                                    (__v4sf) __W,
4948                                                    (__mmask8) __U);
4949 }
4950
4951 extern __inline __m128
4952 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4953 _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A)
4954 {
4955   return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A,
4956                                                    (__v4sf)
4957                                                    _mm_setzero_ps (),
4958                                                    (__mmask8) __U);
4959 }
4960
4961 extern __inline __m256i
4962 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4963 _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A)
4964 {
4965   return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A,
4966                                                     (__v8si) __W,
4967                                                     (__mmask8) __U);
4968 }
4969
4970 extern __inline __m256i
4971 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4972 _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A)
4973 {
4974   return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A,
4975                                                     (__v8si)
4976                                                     _mm256_setzero_si256 (),
4977                                                     (__mmask8) __U);
4978 }
4979
4980 extern __inline __m128i
4981 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4982 _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A)
4983 {
4984   return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A,
4985                                                     (__v4si) __W,
4986                                                     (__mmask8) __U);
4987 }
4988
4989 extern __inline __m128i
4990 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
4991 _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A)
4992 {
4993   return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A,
4994                                                     (__v4si)
4995                                                     _mm_setzero_si128 (),
4996                                                     (__mmask8) __U);
4997 }
4998
4999 extern __inline __m256i
5000 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5001 _mm256_cvtps_epu32 (__m256 __A)
5002 {
5003   return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
5004                                                      (__v8si)
5005                                                      _mm256_setzero_si256 (),
5006                                                      (__mmask8) -1);
5007 }
5008
5009 extern __inline __m256i
5010 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5011 _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A)
5012 {
5013   return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
5014                                                      (__v8si) __W,
5015                                                      (__mmask8) __U);
5016 }
5017
5018 extern __inline __m256i
5019 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5020 _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A)
5021 {
5022   return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A,
5023                                                      (__v8si)
5024                                                      _mm256_setzero_si256 (),
5025                                                      (__mmask8) __U);
5026 }
5027
5028 extern __inline __m128i
5029 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5030 _mm_cvtps_epu32 (__m128 __A)
5031 {
5032   return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
5033                                                      (__v4si)
5034                                                      _mm_setzero_si128 (),
5035                                                      (__mmask8) -1);
5036 }
5037
5038 extern __inline __m128i
5039 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5040 _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A)
5041 {
5042   return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
5043                                                      (__v4si) __W,
5044                                                      (__mmask8) __U);
5045 }
5046
5047 extern __inline __m128i
5048 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5049 _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A)
5050 {
5051   return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A,
5052                                                      (__v4si)
5053                                                      _mm_setzero_si128 (),
5054                                                      (__mmask8) __U);
5055 }
5056
5057 extern __inline __m256d
5058 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5059 _mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A)
5060 {
5061   return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A,
5062                                                    (__v4df) __W,
5063                                                    (__mmask8) __U);
5064 }
5065
5066 extern __inline __m256d
5067 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5068 _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A)
5069 {
5070   return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A,
5071                                                    (__v4df)
5072                                                    _mm256_setzero_pd (),
5073                                                    (__mmask8) __U);
5074 }
5075
5076 extern __inline __m128d
5077 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5078 _mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A)
5079 {
5080   return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A,
5081                                                    (__v2df) __W,
5082                                                    (__mmask8) __U);
5083 }
5084
5085 extern __inline __m128d
5086 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5087 _mm_maskz_movedup_pd (__mmask8 __U, __m128d __A)
5088 {
5089   return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A,
5090                                                    (__v2df)
5091                                                    _mm_setzero_pd (),
5092                                                    (__mmask8) __U);
5093 }
5094
5095 extern __inline __m256
5096 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5097 _mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A)
5098 {
5099   return (__m256) __builtin_ia32_movshdup256_mask ((__v8sf) __A,
5100                                                    (__v8sf) __W,
5101                                                    (__mmask8) __U);
5102 }
5103
5104 extern __inline __m256
5105 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5106 _mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A)
5107 {
5108   return (__m256) __builtin_ia32_movshdup256_mask ((__v8sf) __A,
5109                                                    (__v8sf)
5110                                                    _mm256_setzero_ps (),
5111                                                    (__mmask8) __U);
5112 }
5113
5114 extern __inline __m128
5115 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5116 _mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A)
5117 {
5118   return (__m128) __builtin_ia32_movshdup128_mask ((__v4sf) __A,
5119                                                    (__v4sf) __W,
5120                                                    (__mmask8) __U);
5121 }
5122
5123 extern __inline __m128
5124 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5125 _mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A)
5126 {
5127   return (__m128) __builtin_ia32_movshdup128_mask ((__v4sf) __A,
5128                                                    (__v4sf)
5129                                                    _mm_setzero_ps (),
5130                                                    (__mmask8) __U);
5131 }
5132
5133 extern __inline __m256
5134 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5135 _mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A)
5136 {
5137   return (__m256) __builtin_ia32_movsldup256_mask ((__v8sf) __A,
5138                                                    (__v8sf) __W,
5139                                                    (__mmask8) __U);
5140 }
5141
5142 extern __inline __m256
5143 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5144 _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A)
5145 {
5146   return (__m256) __builtin_ia32_movsldup256_mask ((__v8sf) __A,
5147                                                    (__v8sf)
5148                                                    _mm256_setzero_ps (),
5149                                                    (__mmask8) __U);
5150 }
5151
5152 extern __inline __m128
5153 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5154 _mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A)
5155 {
5156   return (__m128) __builtin_ia32_movsldup128_mask ((__v4sf) __A,
5157                                                    (__v4sf) __W,
5158                                                    (__mmask8) __U);
5159 }
5160
5161 extern __inline __m128
5162 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5163 _mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A)
5164 {
5165   return (__m128) __builtin_ia32_movsldup128_mask ((__v4sf) __A,
5166                                                    (__v4sf)
5167                                                    _mm_setzero_ps (),
5168                                                    (__mmask8) __U);
5169 }
5170
5171 extern __inline __m128i
5172 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5173 _mm_mask_unpackhi_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5174                          __m128i __B)
5175 {
5176   return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A,
5177                                                      (__v4si) __B,
5178                                                      (__v4si) __W,
5179                                                      (__mmask8) __U);
5180 }
5181
5182 extern __inline __m128i
5183 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5184 _mm_maskz_unpackhi_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5185 {
5186   return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A,
5187                                                      (__v4si) __B,
5188                                                      (__v4si)
5189                                                      _mm_setzero_si128 (),
5190                                                      (__mmask8) __U);
5191 }
5192
5193 extern __inline __m256i
5194 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5195 _mm256_mask_unpackhi_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5196                             __m256i __B)
5197 {
5198   return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A,
5199                                                      (__v8si) __B,
5200                                                      (__v8si) __W,
5201                                                      (__mmask8) __U);
5202 }
5203
5204 extern __inline __m256i
5205 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5206 _mm256_maskz_unpackhi_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
5207 {
5208   return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A,
5209                                                      (__v8si) __B,
5210                                                      (__v8si)
5211                                                      _mm256_setzero_si256 (),
5212                                                      (__mmask8) __U);
5213 }
5214
5215 extern __inline __m128i
5216 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5217 _mm_mask_unpackhi_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5218                          __m128i __B)
5219 {
5220   return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A,
5221                                                       (__v2di) __B,
5222                                                       (__v2di) __W,
5223                                                       (__mmask8) __U);
5224 }
5225
5226 extern __inline __m128i
5227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5228 _mm_maskz_unpackhi_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5229 {
5230   return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A,
5231                                                       (__v2di) __B,
5232                                                       (__v2di)
5233                                                       _mm_setzero_si128 (),
5234                                                       (__mmask8) __U);
5235 }
5236
5237 extern __inline __m256i
5238 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5239 _mm256_mask_unpackhi_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5240                             __m256i __B)
5241 {
5242   return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A,
5243                                                       (__v4di) __B,
5244                                                       (__v4di) __W,
5245                                                       (__mmask8) __U);
5246 }
5247
5248 extern __inline __m256i
5249 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5250 _mm256_maskz_unpackhi_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
5251 {
5252   return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A,
5253                                                       (__v4di) __B,
5254                                                       (__v4di)
5255                                                       _mm256_setzero_si256 (),
5256                                                       (__mmask8) __U);
5257 }
5258
5259 extern __inline __m128i
5260 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5261 _mm_mask_unpacklo_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
5262                          __m128i __B)
5263 {
5264   return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A,
5265                                                      (__v4si) __B,
5266                                                      (__v4si) __W,
5267                                                      (__mmask8) __U);
5268 }
5269
5270 extern __inline __m128i
5271 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5272 _mm_maskz_unpacklo_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
5273 {
5274   return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A,
5275                                                      (__v4si) __B,
5276                                                      (__v4si)
5277                                                      _mm_setzero_si128 (),
5278                                                      (__mmask8) __U);
5279 }
5280
5281 extern __inline __m256i
5282 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5283 _mm256_mask_unpacklo_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
5284                             __m256i __B)
5285 {
5286   return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A,
5287                                                      (__v8si) __B,
5288                                                      (__v8si) __W,
5289                                                      (__mmask8) __U);
5290 }
5291
5292 extern __inline __m256i
5293 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5294 _mm256_maskz_unpacklo_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
5295 {
5296   return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A,
5297                                                      (__v8si) __B,
5298                                                      (__v8si)
5299                                                      _mm256_setzero_si256 (),
5300                                                      (__mmask8) __U);
5301 }
5302
5303 extern __inline __m128i
5304 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5305 _mm_mask_unpacklo_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
5306                          __m128i __B)
5307 {
5308   return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A,
5309                                                       (__v2di) __B,
5310                                                       (__v2di) __W,
5311                                                       (__mmask8) __U);
5312 }
5313
5314 extern __inline __m128i
5315 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5316 _mm_maskz_unpacklo_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
5317 {
5318   return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A,
5319                                                       (__v2di) __B,
5320                                                       (__v2di)
5321                                                       _mm_setzero_si128 (),
5322                                                       (__mmask8) __U);
5323 }
5324
5325 extern __inline __m256i
5326 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5327 _mm256_mask_unpacklo_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
5328                             __m256i __B)
5329 {
5330   return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A,
5331                                                       (__v4di) __B,
5332                                                       (__v4di) __W,
5333                                                       (__mmask8) __U);
5334 }
5335
5336 extern __inline __m256i
5337 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5338 _mm256_maskz_unpacklo_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
5339 {
5340   return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A,
5341                                                       (__v4di) __B,
5342                                                       (__v4di)
5343                                                       _mm256_setzero_si256 (),
5344                                                       (__mmask8) __U);
5345 }
5346
5347 extern __inline __mmask8
5348 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5349 _mm_cmpeq_epu32_mask (__m128i __A, __m128i __B)
5350 {
5351   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A,
5352                                                    (__v4si) __B, 0,
5353                                                    (__mmask8) -1);
5354 }
5355
5356 extern __inline __mmask8
5357 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5358 _mm_cmpeq_epi32_mask (__m128i __A, __m128i __B)
5359 {
5360   return (__mmask8) __builtin_ia32_pcmpeqd128_mask ((__v4si) __A,
5361                                                     (__v4si) __B,
5362                                                     (__mmask8) -1);
5363 }
5364
5365 extern __inline __mmask8
5366 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5367 _mm_mask_cmpeq_epu32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5368 {
5369   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A,
5370                                                    (__v4si) __B, 0, __U);
5371 }
5372
5373 extern __inline __mmask8
5374 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5375 _mm_mask_cmpeq_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5376 {
5377   return (__mmask8) __builtin_ia32_pcmpeqd128_mask ((__v4si) __A,
5378                                                     (__v4si) __B, __U);
5379 }
5380
5381 extern __inline __mmask8
5382 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5383 _mm256_cmpeq_epu32_mask (__m256i __A, __m256i __B)
5384 {
5385   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A,
5386                                                    (__v8si) __B, 0,
5387                                                    (__mmask8) -1);
5388 }
5389
5390 extern __inline __mmask8
5391 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5392 _mm256_cmpeq_epi32_mask (__m256i __A, __m256i __B)
5393 {
5394   return (__mmask8) __builtin_ia32_pcmpeqd256_mask ((__v8si) __A,
5395                                                     (__v8si) __B,
5396                                                     (__mmask8) -1);
5397 }
5398
5399 extern __inline __mmask8
5400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5401 _mm256_mask_cmpeq_epu32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5402 {
5403   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A,
5404                                                    (__v8si) __B, 0, __U);
5405 }
5406
5407 extern __inline __mmask8
5408 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5409 _mm256_mask_cmpeq_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5410 {
5411   return (__mmask8) __builtin_ia32_pcmpeqd256_mask ((__v8si) __A,
5412                                                     (__v8si) __B, __U);
5413 }
5414
5415 extern __inline __mmask8
5416 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5417 _mm_cmpeq_epu64_mask (__m128i __A, __m128i __B)
5418 {
5419   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A,
5420                                                    (__v2di) __B, 0,
5421                                                    (__mmask8) -1);
5422 }
5423
5424 extern __inline __mmask8
5425 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5426 _mm_cmpeq_epi64_mask (__m128i __A, __m128i __B)
5427 {
5428   return (__mmask8) __builtin_ia32_pcmpeqq128_mask ((__v2di) __A,
5429                                                     (__v2di) __B,
5430                                                     (__mmask8) -1);
5431 }
5432
5433 extern __inline __mmask8
5434 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5435 _mm_mask_cmpeq_epu64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5436 {
5437   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A,
5438                                                    (__v2di) __B, 0, __U);
5439 }
5440
5441 extern __inline __mmask8
5442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5443 _mm_mask_cmpeq_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5444 {
5445   return (__mmask8) __builtin_ia32_pcmpeqq128_mask ((__v2di) __A,
5446                                                     (__v2di) __B, __U);
5447 }
5448
5449 extern __inline __mmask8
5450 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5451 _mm256_cmpeq_epu64_mask (__m256i __A, __m256i __B)
5452 {
5453   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A,
5454                                                    (__v4di) __B, 0,
5455                                                    (__mmask8) -1);
5456 }
5457
5458 extern __inline __mmask8
5459 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5460 _mm256_cmpeq_epi64_mask (__m256i __A, __m256i __B)
5461 {
5462   return (__mmask8) __builtin_ia32_pcmpeqq256_mask ((__v4di) __A,
5463                                                     (__v4di) __B,
5464                                                     (__mmask8) -1);
5465 }
5466
5467 extern __inline __mmask8
5468 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5469 _mm256_mask_cmpeq_epu64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5470 {
5471   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A,
5472                                                    (__v4di) __B, 0, __U);
5473 }
5474
5475 extern __inline __mmask8
5476 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5477 _mm256_mask_cmpeq_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5478 {
5479   return (__mmask8) __builtin_ia32_pcmpeqq256_mask ((__v4di) __A,
5480                                                     (__v4di) __B, __U);
5481 }
5482
5483 extern __inline __mmask8
5484 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5485 _mm_cmpgt_epu32_mask (__m128i __A, __m128i __B)
5486 {
5487   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A,
5488                                                    (__v4si) __B, 6,
5489                                                    (__mmask8) -1);
5490 }
5491
5492 extern __inline __mmask8
5493 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5494 _mm_cmpgt_epi32_mask (__m128i __A, __m128i __B)
5495 {
5496   return (__mmask8) __builtin_ia32_pcmpgtd128_mask ((__v4si) __A,
5497                                                     (__v4si) __B,
5498                                                     (__mmask8) -1);
5499 }
5500
5501 extern __inline __mmask8
5502 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5503 _mm_mask_cmpgt_epu32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5504 {
5505   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A,
5506                                                    (__v4si) __B, 6, __U);
5507 }
5508
5509 extern __inline __mmask8
5510 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5511 _mm_mask_cmpgt_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5512 {
5513   return (__mmask8) __builtin_ia32_pcmpgtd128_mask ((__v4si) __A,
5514                                                     (__v4si) __B, __U);
5515 }
5516
5517 extern __inline __mmask8
5518 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5519 _mm256_cmpgt_epu32_mask (__m256i __A, __m256i __B)
5520 {
5521   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A,
5522                                                    (__v8si) __B, 6,
5523                                                    (__mmask8) -1);
5524 }
5525
5526 extern __inline __mmask8
5527 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5528 _mm256_cmpgt_epi32_mask (__m256i __A, __m256i __B)
5529 {
5530   return (__mmask8) __builtin_ia32_pcmpgtd256_mask ((__v8si) __A,
5531                                                     (__v8si) __B,
5532                                                     (__mmask8) -1);
5533 }
5534
5535 extern __inline __mmask8
5536 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5537 _mm256_mask_cmpgt_epu32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5538 {
5539   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A,
5540                                                    (__v8si) __B, 6, __U);
5541 }
5542
5543 extern __inline __mmask8
5544 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5545 _mm256_mask_cmpgt_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5546 {
5547   return (__mmask8) __builtin_ia32_pcmpgtd256_mask ((__v8si) __A,
5548                                                     (__v8si) __B, __U);
5549 }
5550
5551 extern __inline __mmask8
5552 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5553 _mm_cmpgt_epu64_mask (__m128i __A, __m128i __B)
5554 {
5555   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A,
5556                                                    (__v2di) __B, 6,
5557                                                    (__mmask8) -1);
5558 }
5559
5560 extern __inline __mmask8
5561 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5562 _mm_cmpgt_epi64_mask (__m128i __A, __m128i __B)
5563 {
5564   return (__mmask8) __builtin_ia32_pcmpgtq128_mask ((__v2di) __A,
5565                                                     (__v2di) __B,
5566                                                     (__mmask8) -1);
5567 }
5568
5569 extern __inline __mmask8
5570 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5571 _mm_mask_cmpgt_epu64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5572 {
5573   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A,
5574                                                    (__v2di) __B, 6, __U);
5575 }
5576
5577 extern __inline __mmask8
5578 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5579 _mm_mask_cmpgt_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5580 {
5581   return (__mmask8) __builtin_ia32_pcmpgtq128_mask ((__v2di) __A,
5582                                                     (__v2di) __B, __U);
5583 }
5584
5585 extern __inline __mmask8
5586 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5587 _mm256_cmpgt_epu64_mask (__m256i __A, __m256i __B)
5588 {
5589   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A,
5590                                                    (__v4di) __B, 6,
5591                                                    (__mmask8) -1);
5592 }
5593
5594 extern __inline __mmask8
5595 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5596 _mm256_cmpgt_epi64_mask (__m256i __A, __m256i __B)
5597 {
5598   return (__mmask8) __builtin_ia32_pcmpgtq256_mask ((__v4di) __A,
5599                                                     (__v4di) __B,
5600                                                     (__mmask8) -1);
5601 }
5602
5603 extern __inline __mmask8
5604 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5605 _mm256_mask_cmpgt_epu64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5606 {
5607   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A,
5608                                                    (__v4di) __B, 6, __U);
5609 }
5610
5611 extern __inline __mmask8
5612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5613 _mm256_mask_cmpgt_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5614 {
5615   return (__mmask8) __builtin_ia32_pcmpgtq256_mask ((__v4di) __A,
5616                                                     (__v4di) __B, __U);
5617 }
5618
5619 extern __inline __mmask8
5620 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5621 _mm_test_epi32_mask (__m128i __A, __m128i __B)
5622 {
5623   return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
5624                                                (__v4si) __B,
5625                                                (__mmask8) -1);
5626 }
5627
5628 extern __inline __mmask8
5629 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5630 _mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5631 {
5632   return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A,
5633                                                (__v4si) __B, __U);
5634 }
5635
5636 extern __inline __mmask8
5637 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5638 _mm256_test_epi32_mask (__m256i __A, __m256i __B)
5639 {
5640   return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
5641                                                (__v8si) __B,
5642                                                (__mmask8) -1);
5643 }
5644
5645 extern __inline __mmask8
5646 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5647 _mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5648 {
5649   return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A,
5650                                                (__v8si) __B, __U);
5651 }
5652
5653 extern __inline __mmask8
5654 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5655 _mm_test_epi64_mask (__m128i __A, __m128i __B)
5656 {
5657   return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
5658                                                (__v2di) __B,
5659                                                (__mmask8) -1);
5660 }
5661
5662 extern __inline __mmask8
5663 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5664 _mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5665 {
5666   return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A,
5667                                                (__v2di) __B, __U);
5668 }
5669
5670 extern __inline __mmask8
5671 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5672 _mm256_test_epi64_mask (__m256i __A, __m256i __B)
5673 {
5674   return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
5675                                                (__v4di) __B,
5676                                                (__mmask8) -1);
5677 }
5678
5679 extern __inline __mmask8
5680 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5681 _mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5682 {
5683   return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A,
5684                                                (__v4di) __B, __U);
5685 }
5686
5687 extern __inline __mmask8
5688 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5689 _mm_testn_epi32_mask (__m128i __A, __m128i __B)
5690 {
5691   return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
5692                                                 (__v4si) __B,
5693                                                 (__mmask8) -1);
5694 }
5695
5696 extern __inline __mmask8
5697 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5698 _mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B)
5699 {
5700   return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A,
5701                                                 (__v4si) __B, __U);
5702 }
5703
5704 extern __inline __mmask8
5705 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5706 _mm256_testn_epi32_mask (__m256i __A, __m256i __B)
5707 {
5708   return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
5709                                                 (__v8si) __B,
5710                                                 (__mmask8) -1);
5711 }
5712
5713 extern __inline __mmask8
5714 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5715 _mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B)
5716 {
5717   return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A,
5718                                                 (__v8si) __B, __U);
5719 }
5720
5721 extern __inline __mmask8
5722 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5723 _mm_testn_epi64_mask (__m128i __A, __m128i __B)
5724 {
5725   return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
5726                                                 (__v2di) __B,
5727                                                 (__mmask8) -1);
5728 }
5729
5730 extern __inline __mmask8
5731 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5732 _mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B)
5733 {
5734   return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A,
5735                                                 (__v2di) __B, __U);
5736 }
5737
5738 extern __inline __mmask8
5739 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5740 _mm256_testn_epi64_mask (__m256i __A, __m256i __B)
5741 {
5742   return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
5743                                                 (__v4di) __B,
5744                                                 (__mmask8) -1);
5745 }
5746
5747 extern __inline __mmask8
5748 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5749 _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
5750 {
5751   return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A,
5752                                                 (__v4di) __B, __U);
5753 }
5754
5755 extern __inline __m256d
5756 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5757 _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A)
5758 {
5759   return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A,
5760                                                       (__v4df) __W,
5761                                                       (__mmask8) __U);
5762 }
5763
5764 extern __inline __m256d
5765 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5766 _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A)
5767 {
5768   return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A,
5769                                                       (__v4df)
5770                                                       _mm256_setzero_pd (),
5771                                                       (__mmask8) __U);
5772 }
5773
5774 extern __inline void
5775 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5776 _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A)
5777 {
5778   __builtin_ia32_compressstoredf256_mask ((__v4df *) __P,
5779                                           (__v4df) __A,
5780                                           (__mmask8) __U);
5781 }
5782
5783 extern __inline __m128d
5784 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5785 _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A)
5786 {
5787   return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A,
5788                                                       (__v2df) __W,
5789                                                       (__mmask8) __U);
5790 }
5791
5792 extern __inline __m128d
5793 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5794 _mm_maskz_compress_pd (__mmask8 __U, __m128d __A)
5795 {
5796   return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A,
5797                                                       (__v2df)
5798                                                       _mm_setzero_pd (),
5799                                                       (__mmask8) __U);
5800 }
5801
5802 extern __inline void
5803 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5804 _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A)
5805 {
5806   __builtin_ia32_compressstoredf128_mask ((__v2df *) __P,
5807                                           (__v2df) __A,
5808                                           (__mmask8) __U);
5809 }
5810
5811 extern __inline __m256
5812 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5813 _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A)
5814 {
5815   return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A,
5816                                                      (__v8sf) __W,
5817                                                      (__mmask8) __U);
5818 }
5819
5820 extern __inline __m256
5821 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5822 _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A)
5823 {
5824   return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A,
5825                                                      (__v8sf)
5826                                                      _mm256_setzero_ps (),
5827                                                      (__mmask8) __U);
5828 }
5829
5830 extern __inline void
5831 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5832 _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A)
5833 {
5834   __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P,
5835                                           (__v8sf) __A,
5836                                           (__mmask8) __U);
5837 }
5838
5839 extern __inline __m128
5840 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5841 _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A)
5842 {
5843   return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A,
5844                                                      (__v4sf) __W,
5845                                                      (__mmask8) __U);
5846 }
5847
5848 extern __inline __m128
5849 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5850 _mm_maskz_compress_ps (__mmask8 __U, __m128 __A)
5851 {
5852   return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A,
5853                                                      (__v4sf)
5854                                                      _mm_setzero_ps (),
5855                                                      (__mmask8) __U);
5856 }
5857
5858 extern __inline void
5859 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5860 _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A)
5861 {
5862   __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P,
5863                                           (__v4sf) __A,
5864                                           (__mmask8) __U);
5865 }
5866
5867 extern __inline __m256i
5868 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5869 _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
5870 {
5871   return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A,
5872                                                       (__v4di) __W,
5873                                                       (__mmask8) __U);
5874 }
5875
5876 extern __inline __m256i
5877 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5878 _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A)
5879 {
5880   return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A,
5881                                                       (__v4di)
5882                                                       _mm256_setzero_si256 (),
5883                                                       (__mmask8) __U);
5884 }
5885
5886 extern __inline void
5887 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5888 _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A)
5889 {
5890   __builtin_ia32_compressstoredi256_mask ((__v4di *) __P,
5891                                           (__v4di) __A,
5892                                           (__mmask8) __U);
5893 }
5894
5895 extern __inline __m128i
5896 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5897 _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
5898 {
5899   return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A,
5900                                                       (__v2di) __W,
5901                                                       (__mmask8) __U);
5902 }
5903
5904 extern __inline __m128i
5905 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5906 _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A)
5907 {
5908   return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A,
5909                                                       (__v2di)
5910                                                       _mm_setzero_si128 (),
5911                                                       (__mmask8) __U);
5912 }
5913
5914 extern __inline void
5915 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5916 _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A)
5917 {
5918   __builtin_ia32_compressstoredi128_mask ((__v2di *) __P,
5919                                           (__v2di) __A,
5920                                           (__mmask8) __U);
5921 }
5922
5923 extern __inline __m256i
5924 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5925 _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
5926 {
5927   return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A,
5928                                                       (__v8si) __W,
5929                                                       (__mmask8) __U);
5930 }
5931
5932 extern __inline __m256i
5933 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5934 _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A)
5935 {
5936   return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A,
5937                                                       (__v8si)
5938                                                       _mm256_setzero_si256 (),
5939                                                       (__mmask8) __U);
5940 }
5941
5942 extern __inline void
5943 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5944 _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A)
5945 {
5946   __builtin_ia32_compressstoresi256_mask ((__v8si *) __P,
5947                                           (__v8si) __A,
5948                                           (__mmask8) __U);
5949 }
5950
5951 extern __inline __m128i
5952 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5953 _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
5954 {
5955   return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A,
5956                                                       (__v4si) __W,
5957                                                       (__mmask8) __U);
5958 }
5959
5960 extern __inline __m128i
5961 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5962 _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A)
5963 {
5964   return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A,
5965                                                       (__v4si)
5966                                                       _mm_setzero_si128 (),
5967                                                       (__mmask8) __U);
5968 }
5969
5970 extern __inline void
5971 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5972 _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A)
5973 {
5974   __builtin_ia32_compressstoresi128_mask ((__v4si *) __P,
5975                                           (__v4si) __A,
5976                                           (__mmask8) __U);
5977 }
5978
5979 extern __inline __m256d
5980 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5981 _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A)
5982 {
5983   return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A,
5984                                                     (__v4df) __W,
5985                                                     (__mmask8) __U);
5986 }
5987
5988 extern __inline __m256d
5989 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
5990 _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A)
5991 {
5992   return (__m256d) __builtin_ia32_expanddf256_maskz ((__v4df) __A,
5993                                                      (__v4df)
5994                                                      _mm256_setzero_pd (),
5995                                                      (__mmask8) __U);
5996 }
5997
5998 extern __inline __m256d
5999 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6000 _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P)
6001 {
6002   return (__m256d) __builtin_ia32_expandloaddf256_mask ((__v4df *) __P,
6003                                                         (__v4df) __W,
6004                                                         (__mmask8)
6005                                                         __U);
6006 }
6007
6008 extern __inline __m256d
6009 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6010 _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P)
6011 {
6012   return (__m256d) __builtin_ia32_expandloaddf256_maskz ((__v4df *) __P,
6013                                                          (__v4df)
6014                                                          _mm256_setzero_pd (),
6015                                                          (__mmask8)
6016                                                          __U);
6017 }
6018
6019 extern __inline __m128d
6020 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6021 _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A)
6022 {
6023   return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A,
6024                                                     (__v2df) __W,
6025                                                     (__mmask8) __U);
6026 }
6027
6028 extern __inline __m128d
6029 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6030 _mm_maskz_expand_pd (__mmask8 __U, __m128d __A)
6031 {
6032   return (__m128d) __builtin_ia32_expanddf128_maskz ((__v2df) __A,
6033                                                      (__v2df)
6034                                                      _mm_setzero_pd (),
6035                                                      (__mmask8) __U);
6036 }
6037
6038 extern __inline __m128d
6039 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6040 _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P)
6041 {
6042   return (__m128d) __builtin_ia32_expandloaddf128_mask ((__v2df *) __P,
6043                                                         (__v2df) __W,
6044                                                         (__mmask8)
6045                                                         __U);
6046 }
6047
6048 extern __inline __m128d
6049 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6050 _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P)
6051 {
6052   return (__m128d) __builtin_ia32_expandloaddf128_maskz ((__v2df *) __P,
6053                                                          (__v2df)
6054                                                          _mm_setzero_pd (),
6055                                                          (__mmask8)
6056                                                          __U);
6057 }
6058
6059 extern __inline __m256
6060 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6061 _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A)
6062 {
6063   return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A,
6064                                                    (__v8sf) __W,
6065                                                    (__mmask8) __U);
6066 }
6067
6068 extern __inline __m256
6069 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6070 _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A)
6071 {
6072   return (__m256) __builtin_ia32_expandsf256_maskz ((__v8sf) __A,
6073                                                     (__v8sf)
6074                                                     _mm256_setzero_ps (),
6075                                                     (__mmask8) __U);
6076 }
6077
6078 extern __inline __m256
6079 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6080 _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P)
6081 {
6082   return (__m256) __builtin_ia32_expandloadsf256_mask ((__v8sf *) __P,
6083                                                        (__v8sf) __W,
6084                                                        (__mmask8) __U);
6085 }
6086
6087 extern __inline __m256
6088 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6089 _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P)
6090 {
6091   return (__m256) __builtin_ia32_expandloadsf256_maskz ((__v8sf *) __P,
6092                                                         (__v8sf)
6093                                                         _mm256_setzero_ps (),
6094                                                         (__mmask8)
6095                                                         __U);
6096 }
6097
6098 extern __inline __m128
6099 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6100 _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A)
6101 {
6102   return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A,
6103                                                    (__v4sf) __W,
6104                                                    (__mmask8) __U);
6105 }
6106
6107 extern __inline __m128
6108 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6109 _mm_maskz_expand_ps (__mmask8 __U, __m128 __A)
6110 {
6111   return (__m128) __builtin_ia32_expandsf128_maskz ((__v4sf) __A,
6112                                                     (__v4sf)
6113                                                     _mm_setzero_ps (),
6114                                                     (__mmask8) __U);
6115 }
6116
6117 extern __inline __m128
6118 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6119 _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P)
6120 {
6121   return (__m128) __builtin_ia32_expandloadsf128_mask ((__v4sf *) __P,
6122                                                        (__v4sf) __W,
6123                                                        (__mmask8) __U);
6124 }
6125
6126 extern __inline __m128
6127 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6128 _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P)
6129 {
6130   return (__m128) __builtin_ia32_expandloadsf128_maskz ((__v4sf *) __P,
6131                                                         (__v4sf)
6132                                                         _mm_setzero_ps (),
6133                                                         (__mmask8)
6134                                                         __U);
6135 }
6136
6137 extern __inline __m256i
6138 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6139 _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
6140 {
6141   return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A,
6142                                                     (__v4di) __W,
6143                                                     (__mmask8) __U);
6144 }
6145
6146 extern __inline __m256i
6147 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6148 _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A)
6149 {
6150   return (__m256i) __builtin_ia32_expanddi256_maskz ((__v4di) __A,
6151                                                      (__v4di)
6152                                                      _mm256_setzero_si256 (),
6153                                                      (__mmask8) __U);
6154 }
6155
6156 extern __inline __m256i
6157 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6158 _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U,
6159                                void const *__P)
6160 {
6161   return (__m256i) __builtin_ia32_expandloaddi256_mask ((__v4di *) __P,
6162                                                         (__v4di) __W,
6163                                                         (__mmask8)
6164                                                         __U);
6165 }
6166
6167 extern __inline __m256i
6168 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6169 _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P)
6170 {
6171   return (__m256i) __builtin_ia32_expandloaddi256_maskz ((__v4di *) __P,
6172                                                          (__v4di)
6173                                                          _mm256_setzero_si256 (),
6174                                                          (__mmask8)
6175                                                          __U);
6176 }
6177
6178 extern __inline __m128i
6179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6180 _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
6181 {
6182   return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A,
6183                                                     (__v2di) __W,
6184                                                     (__mmask8) __U);
6185 }
6186
6187 extern __inline __m128i
6188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6189 _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A)
6190 {
6191   return (__m128i) __builtin_ia32_expanddi128_maskz ((__v2di) __A,
6192                                                      (__v2di)
6193                                                      _mm_setzero_si128 (),
6194                                                      (__mmask8) __U);
6195 }
6196
6197 extern __inline __m128i
6198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6199 _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P)
6200 {
6201   return (__m128i) __builtin_ia32_expandloaddi128_mask ((__v2di *) __P,
6202                                                         (__v2di) __W,
6203                                                         (__mmask8)
6204                                                         __U);
6205 }
6206
6207 extern __inline __m128i
6208 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6209 _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P)
6210 {
6211   return (__m128i) __builtin_ia32_expandloaddi128_maskz ((__v2di *) __P,
6212                                                          (__v2di)
6213                                                          _mm_setzero_si128 (),
6214                                                          (__mmask8)
6215                                                          __U);
6216 }
6217
6218 extern __inline __m256i
6219 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6220 _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
6221 {
6222   return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A,
6223                                                     (__v8si) __W,
6224                                                     (__mmask8) __U);
6225 }
6226
6227 extern __inline __m256i
6228 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6229 _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A)
6230 {
6231   return (__m256i) __builtin_ia32_expandsi256_maskz ((__v8si) __A,
6232                                                      (__v8si)
6233                                                      _mm256_setzero_si256 (),
6234                                                      (__mmask8) __U);
6235 }
6236
6237 extern __inline __m256i
6238 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6239 _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U,
6240                                void const *__P)
6241 {
6242   return (__m256i) __builtin_ia32_expandloadsi256_mask ((__v8si *) __P,
6243                                                         (__v8si) __W,
6244                                                         (__mmask8)
6245                                                         __U);
6246 }
6247
6248 extern __inline __m256i
6249 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6250 _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P)
6251 {
6252   return (__m256i) __builtin_ia32_expandloadsi256_maskz ((__v8si *) __P,
6253                                                          (__v8si)
6254                                                          _mm256_setzero_si256 (),
6255                                                          (__mmask8)
6256                                                          __U);
6257 }
6258
6259 extern __inline __m128i
6260 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6261 _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
6262 {
6263   return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A,
6264                                                     (__v4si) __W,
6265                                                     (__mmask8) __U);
6266 }
6267
6268 extern __inline __m128i
6269 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6270 _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A)
6271 {
6272   return (__m128i) __builtin_ia32_expandsi128_maskz ((__v4si) __A,
6273                                                      (__v4si)
6274                                                      _mm_setzero_si128 (),
6275                                                      (__mmask8) __U);
6276 }
6277
6278 extern __inline __m128i
6279 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6280 _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P)
6281 {
6282   return (__m128i) __builtin_ia32_expandloadsi128_mask ((__v4si *) __P,
6283                                                         (__v4si) __W,
6284                                                         (__mmask8)
6285                                                         __U);
6286 }
6287
6288 extern __inline __m128i
6289 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6290 _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P)
6291 {
6292   return (__m128i) __builtin_ia32_expandloadsi128_maskz ((__v4si *) __P,
6293                                                          (__v4si)
6294                                                          _mm_setzero_si128 (),
6295                                                          (__mmask8)
6296                                                          __U);
6297 }
6298
6299 extern __inline __m256d
6300 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6301 _mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B)
6302 {
6303   return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
6304                                                         /* idx */ ,
6305                                                         (__v4df) __A,
6306                                                         (__v4df) __B,
6307                                                         (__mmask8) -1);
6308 }
6309
6310 extern __inline __m256d
6311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6312 _mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I,
6313                              __m256d __B)
6314 {
6315   return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I
6316                                                         /* idx */ ,
6317                                                         (__v4df) __A,
6318                                                         (__v4df) __B,
6319                                                         (__mmask8)
6320                                                         __U);
6321 }
6322
6323 extern __inline __m256d
6324 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6325 _mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U,
6326                               __m256d __B)
6327 {
6328   return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A,
6329                                                         (__v4di) __I
6330                                                         /* idx */ ,
6331                                                         (__v4df) __B,
6332                                                         (__mmask8)
6333                                                         __U);
6334 }
6335
6336 extern __inline __m256d
6337 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6338 _mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I,
6339                               __m256d __B)
6340 {
6341   return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I
6342                                                          /* idx */ ,
6343                                                          (__v4df) __A,
6344                                                          (__v4df) __B,
6345                                                          (__mmask8)
6346                                                          __U);
6347 }
6348
6349 extern __inline __m256
6350 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6351 _mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B)
6352 {
6353   return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
6354                                                        /* idx */ ,
6355                                                        (__v8sf) __A,
6356                                                        (__v8sf) __B,
6357                                                        (__mmask8) -1);
6358 }
6359
6360 extern __inline __m256
6361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6362 _mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I,
6363                              __m256 __B)
6364 {
6365   return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I
6366                                                        /* idx */ ,
6367                                                        (__v8sf) __A,
6368                                                        (__v8sf) __B,
6369                                                        (__mmask8) __U);
6370 }
6371
6372 extern __inline __m256
6373 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6374 _mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U,
6375                               __m256 __B)
6376 {
6377   return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A,
6378                                                        (__v8si) __I
6379                                                        /* idx */ ,
6380                                                        (__v8sf) __B,
6381                                                        (__mmask8) __U);
6382 }
6383
6384 extern __inline __m256
6385 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6386 _mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I,
6387                               __m256 __B)
6388 {
6389   return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I
6390                                                         /* idx */ ,
6391                                                         (__v8sf) __A,
6392                                                         (__v8sf) __B,
6393                                                         (__mmask8)
6394                                                         __U);
6395 }
6396
6397 extern __inline __m128i
6398 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6399 _mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B)
6400 {
6401   return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
6402                                                        /* idx */ ,
6403                                                        (__v2di) __A,
6404                                                        (__v2di) __B,
6405                                                        (__mmask8) -1);
6406 }
6407
6408 extern __inline __m128i
6409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6410 _mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I,
6411                              __m128i __B)
6412 {
6413   return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I
6414                                                        /* idx */ ,
6415                                                        (__v2di) __A,
6416                                                        (__v2di) __B,
6417                                                        (__mmask8) __U);
6418 }
6419
6420 extern __inline __m128i
6421 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6422 _mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U,
6423                               __m128i __B)
6424 {
6425   return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A,
6426                                                        (__v2di) __I
6427                                                        /* idx */ ,
6428                                                        (__v2di) __B,
6429                                                        (__mmask8) __U);
6430 }
6431
6432 extern __inline __m128i
6433 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6434 _mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I,
6435                               __m128i __B)
6436 {
6437   return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I
6438                                                         /* idx */ ,
6439                                                         (__v2di) __A,
6440                                                         (__v2di) __B,
6441                                                         (__mmask8)
6442                                                         __U);
6443 }
6444
6445 extern __inline __m128i
6446 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6447 _mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B)
6448 {
6449   return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
6450                                                        /* idx */ ,
6451                                                        (__v4si) __A,
6452                                                        (__v4si) __B,
6453                                                        (__mmask8) -1);
6454 }
6455
6456 extern __inline __m128i
6457 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6458 _mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I,
6459                              __m128i __B)
6460 {
6461   return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I
6462                                                        /* idx */ ,
6463                                                        (__v4si) __A,
6464                                                        (__v4si) __B,
6465                                                        (__mmask8) __U);
6466 }
6467
6468 extern __inline __m128i
6469 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6470 _mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U,
6471                               __m128i __B)
6472 {
6473   return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A,
6474                                                        (__v4si) __I
6475                                                        /* idx */ ,
6476                                                        (__v4si) __B,
6477                                                        (__mmask8) __U);
6478 }
6479
6480 extern __inline __m128i
6481 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6482 _mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I,
6483                               __m128i __B)
6484 {
6485   return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I
6486                                                         /* idx */ ,
6487                                                         (__v4si) __A,
6488                                                         (__v4si) __B,
6489                                                         (__mmask8)
6490                                                         __U);
6491 }
6492
6493 extern __inline __m256i
6494 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6495 _mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B)
6496 {
6497   return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
6498                                                        /* idx */ ,
6499                                                        (__v4di) __A,
6500                                                        (__v4di) __B,
6501                                                        (__mmask8) -1);
6502 }
6503
6504 extern __inline __m256i
6505 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6506 _mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I,
6507                                 __m256i __B)
6508 {
6509   return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I
6510                                                        /* idx */ ,
6511                                                        (__v4di) __A,
6512                                                        (__v4di) __B,
6513                                                        (__mmask8) __U);
6514 }
6515
6516 extern __inline __m256i
6517 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6518 _mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I,
6519                                  __mmask8 __U, __m256i __B)
6520 {
6521   return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A,
6522                                                        (__v4di) __I
6523                                                        /* idx */ ,
6524                                                        (__v4di) __B,
6525                                                        (__mmask8) __U);
6526 }
6527
6528 extern __inline __m256i
6529 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6530 _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A,
6531                                  __m256i __I, __m256i __B)
6532 {
6533   return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I
6534                                                         /* idx */ ,
6535                                                         (__v4di) __A,
6536                                                         (__v4di) __B,
6537                                                         (__mmask8)
6538                                                         __U);
6539 }
6540
6541 extern __inline __m256i
6542 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6543 _mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B)
6544 {
6545   return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
6546                                                        /* idx */ ,
6547                                                        (__v8si) __A,
6548                                                        (__v8si) __B,
6549                                                        (__mmask8) -1);
6550 }
6551
6552 extern __inline __m256i
6553 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6554 _mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I,
6555                                 __m256i __B)
6556 {
6557   return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I
6558                                                        /* idx */ ,
6559                                                        (__v8si) __A,
6560                                                        (__v8si) __B,
6561                                                        (__mmask8) __U);
6562 }
6563
6564 extern __inline __m256i
6565 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6566 _mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I,
6567                                  __mmask8 __U, __m256i __B)
6568 {
6569   return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A,
6570                                                        (__v8si) __I
6571                                                        /* idx */ ,
6572                                                        (__v8si) __B,
6573                                                        (__mmask8) __U);
6574 }
6575
6576 extern __inline __m256i
6577 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6578 _mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A,
6579                                  __m256i __I, __m256i __B)
6580 {
6581   return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I
6582                                                         /* idx */ ,
6583                                                         (__v8si) __A,
6584                                                         (__v8si) __B,
6585                                                         (__mmask8)
6586                                                         __U);
6587 }
6588
6589 extern __inline __m128d
6590 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6591 _mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B)
6592 {
6593   return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
6594                                                         /* idx */ ,
6595                                                         (__v2df) __A,
6596                                                         (__v2df) __B,
6597                                                         (__mmask8) -1);
6598 }
6599
6600 extern __inline __m128d
6601 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6602 _mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I,
6603                           __m128d __B)
6604 {
6605   return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I
6606                                                         /* idx */ ,
6607                                                         (__v2df) __A,
6608                                                         (__v2df) __B,
6609                                                         (__mmask8)
6610                                                         __U);
6611 }
6612
6613 extern __inline __m128d
6614 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6615 _mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U,
6616                            __m128d __B)
6617 {
6618   return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A,
6619                                                         (__v2di) __I
6620                                                         /* idx */ ,
6621                                                         (__v2df) __B,
6622                                                         (__mmask8)
6623                                                         __U);
6624 }
6625
6626 extern __inline __m128d
6627 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6628 _mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I,
6629                            __m128d __B)
6630 {
6631   return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I
6632                                                          /* idx */ ,
6633                                                          (__v2df) __A,
6634                                                          (__v2df) __B,
6635                                                          (__mmask8)
6636                                                          __U);
6637 }
6638
6639 extern __inline __m128
6640 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6641 _mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B)
6642 {
6643   return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
6644                                                        /* idx */ ,
6645                                                        (__v4sf) __A,
6646                                                        (__v4sf) __B,
6647                                                        (__mmask8) -1);
6648 }
6649
6650 extern __inline __m128
6651 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6652 _mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I,
6653                           __m128 __B)
6654 {
6655   return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I
6656                                                        /* idx */ ,
6657                                                        (__v4sf) __A,
6658                                                        (__v4sf) __B,
6659                                                        (__mmask8) __U);
6660 }
6661
6662 extern __inline __m128
6663 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6664 _mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U,
6665                            __m128 __B)
6666 {
6667   return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A,
6668                                                        (__v4si) __I
6669                                                        /* idx */ ,
6670                                                        (__v4sf) __B,
6671                                                        (__mmask8) __U);
6672 }
6673
6674 extern __inline __m128
6675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6676 _mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I,
6677                            __m128 __B)
6678 {
6679   return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I
6680                                                         /* idx */ ,
6681                                                         (__v4sf) __A,
6682                                                         (__v4sf) __B,
6683                                                         (__mmask8)
6684                                                         __U);
6685 }
6686
6687 extern __inline __m128i
6688 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6689 _mm_srav_epi64 (__m128i __X, __m128i __Y)
6690 {
6691   return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
6692                                                   (__v2di) __Y,
6693                                                   (__v2di)
6694                                                   _mm_setzero_si128 (),
6695                                                   (__mmask8) -1);
6696 }
6697
6698 extern __inline __m128i
6699 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6700 _mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
6701                      __m128i __Y)
6702 {
6703   return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
6704                                                   (__v2di) __Y,
6705                                                   (__v2di) __W,
6706                                                   (__mmask8) __U);
6707 }
6708
6709 extern __inline __m128i
6710 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6711 _mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
6712 {
6713   return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X,
6714                                                   (__v2di) __Y,
6715                                                   (__v2di)
6716                                                   _mm_setzero_si128 (),
6717                                                   (__mmask8) __U);
6718 }
6719
6720 extern __inline __m256i
6721 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6722 _mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
6723                         __m256i __Y)
6724 {
6725   return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
6726                                                  (__v8si) __Y,
6727                                                  (__v8si) __W,
6728                                                  (__mmask8) __U);
6729 }
6730
6731 extern __inline __m256i
6732 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6733 _mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
6734 {
6735   return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X,
6736                                                  (__v8si) __Y,
6737                                                  (__v8si)
6738                                                  _mm256_setzero_si256 (),
6739                                                  (__mmask8) __U);
6740 }
6741
6742 extern __inline __m128i
6743 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6744 _mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
6745                      __m128i __Y)
6746 {
6747   return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
6748                                                  (__v4si) __Y,
6749                                                  (__v4si) __W,
6750                                                  (__mmask8) __U);
6751 }
6752
6753 extern __inline __m128i
6754 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6755 _mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
6756 {
6757   return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X,
6758                                                  (__v4si) __Y,
6759                                                  (__v4si)
6760                                                  _mm_setzero_si128 (),
6761                                                  (__mmask8) __U);
6762 }
6763
6764 extern __inline __m256i
6765 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6766 _mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
6767                         __m256i __Y)
6768 {
6769   return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
6770                                                  (__v4di) __Y,
6771                                                  (__v4di) __W,
6772                                                  (__mmask8) __U);
6773 }
6774
6775 extern __inline __m256i
6776 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6777 _mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
6778 {
6779   return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X,
6780                                                  (__v4di) __Y,
6781                                                  (__v4di)
6782                                                  _mm256_setzero_si256 (),
6783                                                  (__mmask8) __U);
6784 }
6785
6786 extern __inline __m128i
6787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6788 _mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
6789                      __m128i __Y)
6790 {
6791   return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
6792                                                  (__v2di) __Y,
6793                                                  (__v2di) __W,
6794                                                  (__mmask8) __U);
6795 }
6796
6797 extern __inline __m128i
6798 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6799 _mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
6800 {
6801   return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X,
6802                                                  (__v2di) __Y,
6803                                                  (__v2di)
6804                                                  _mm_setzero_si128 (),
6805                                                  (__mmask8) __U);
6806 }
6807
6808 extern __inline __m256i
6809 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6810 _mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
6811                         __m256i __Y)
6812 {
6813   return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
6814                                                  (__v8si) __Y,
6815                                                  (__v8si) __W,
6816                                                  (__mmask8) __U);
6817 }
6818
6819 extern __inline __m256i
6820 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6821 _mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
6822 {
6823   return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X,
6824                                                  (__v8si) __Y,
6825                                                  (__v8si)
6826                                                  _mm256_setzero_si256 (),
6827                                                  (__mmask8) __U);
6828 }
6829
6830 extern __inline __m128i
6831 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6832 _mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
6833                      __m128i __Y)
6834 {
6835   return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
6836                                                  (__v4si) __Y,
6837                                                  (__v4si) __W,
6838                                                  (__mmask8) __U);
6839 }
6840
6841 extern __inline __m128i
6842 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6843 _mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
6844 {
6845   return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X,
6846                                                  (__v4si) __Y,
6847                                                  (__v4si)
6848                                                  _mm_setzero_si128 (),
6849                                                  (__mmask8) __U);
6850 }
6851
6852 extern __inline __m256i
6853 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6854 _mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X,
6855                         __m256i __Y)
6856 {
6857   return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
6858                                                  (__v8si) __Y,
6859                                                  (__v8si) __W,
6860                                                  (__mmask8) __U);
6861 }
6862
6863 extern __inline __m256i
6864 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6865 _mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y)
6866 {
6867   return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X,
6868                                                  (__v8si) __Y,
6869                                                  (__v8si)
6870                                                  _mm256_setzero_si256 (),
6871                                                  (__mmask8) __U);
6872 }
6873
6874 extern __inline __m128i
6875 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6876 _mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X,
6877                      __m128i __Y)
6878 {
6879   return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
6880                                                  (__v4si) __Y,
6881                                                  (__v4si) __W,
6882                                                  (__mmask8) __U);
6883 }
6884
6885 extern __inline __m128i
6886 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6887 _mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y)
6888 {
6889   return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X,
6890                                                  (__v4si) __Y,
6891                                                  (__v4si)
6892                                                  _mm_setzero_si128 (),
6893                                                  (__mmask8) __U);
6894 }
6895
6896 extern __inline __m256i
6897 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6898 _mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
6899                         __m256i __Y)
6900 {
6901   return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
6902                                                  (__v4di) __Y,
6903                                                  (__v4di) __W,
6904                                                  (__mmask8) __U);
6905 }
6906
6907 extern __inline __m256i
6908 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6909 _mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
6910 {
6911   return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X,
6912                                                  (__v4di) __Y,
6913                                                  (__v4di)
6914                                                  _mm256_setzero_si256 (),
6915                                                  (__mmask8) __U);
6916 }
6917
6918 extern __inline __m128i
6919 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6920 _mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X,
6921                      __m128i __Y)
6922 {
6923   return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
6924                                                  (__v2di) __Y,
6925                                                  (__v2di) __W,
6926                                                  (__mmask8) __U);
6927 }
6928
6929 extern __inline __m128i
6930 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6931 _mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y)
6932 {
6933   return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X,
6934                                                  (__v2di) __Y,
6935                                                  (__v2di)
6936                                                  _mm_setzero_si128 (),
6937                                                  (__mmask8) __U);
6938 }
6939
6940 extern __inline __m256i
6941 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6942 _mm256_rolv_epi32 (__m256i __A, __m256i __B)
6943 {
6944   return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
6945                                                   (__v8si) __B,
6946                                                   (__v8si)
6947                                                   _mm256_setzero_si256 (),
6948                                                   (__mmask8) -1);
6949 }
6950
6951 extern __inline __m256i
6952 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6953 _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
6954                         __m256i __B)
6955 {
6956   return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
6957                                                   (__v8si) __B,
6958                                                   (__v8si) __W,
6959                                                   (__mmask8) __U);
6960 }
6961
6962 extern __inline __m256i
6963 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6964 _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
6965 {
6966   return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A,
6967                                                   (__v8si) __B,
6968                                                   (__v8si)
6969                                                   _mm256_setzero_si256 (),
6970                                                   (__mmask8) __U);
6971 }
6972
6973 extern __inline __m128i
6974 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6975 _mm_rolv_epi32 (__m128i __A, __m128i __B)
6976 {
6977   return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
6978                                                   (__v4si) __B,
6979                                                   (__v4si)
6980                                                   _mm_setzero_si128 (),
6981                                                   (__mmask8) -1);
6982 }
6983
6984 extern __inline __m128i
6985 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6986 _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
6987                      __m128i __B)
6988 {
6989   return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
6990                                                   (__v4si) __B,
6991                                                   (__v4si) __W,
6992                                                   (__mmask8) __U);
6993 }
6994
6995 extern __inline __m128i
6996 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
6997 _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
6998 {
6999   return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A,
7000                                                   (__v4si) __B,
7001                                                   (__v4si)
7002                                                   _mm_setzero_si128 (),
7003                                                   (__mmask8) __U);
7004 }
7005
7006 extern __inline __m256i
7007 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7008 _mm256_rorv_epi32 (__m256i __A, __m256i __B)
7009 {
7010   return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
7011                                                   (__v8si) __B,
7012                                                   (__v8si)
7013                                                   _mm256_setzero_si256 (),
7014                                                   (__mmask8) -1);
7015 }
7016
7017 extern __inline __m256i
7018 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7019 _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
7020                         __m256i __B)
7021 {
7022   return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
7023                                                   (__v8si) __B,
7024                                                   (__v8si) __W,
7025                                                   (__mmask8) __U);
7026 }
7027
7028 extern __inline __m256i
7029 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7030 _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
7031 {
7032   return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A,
7033                                                   (__v8si) __B,
7034                                                   (__v8si)
7035                                                   _mm256_setzero_si256 (),
7036                                                   (__mmask8) __U);
7037 }
7038
7039 extern __inline __m128i
7040 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7041 _mm_rorv_epi32 (__m128i __A, __m128i __B)
7042 {
7043   return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
7044                                                   (__v4si) __B,
7045                                                   (__v4si)
7046                                                   _mm_setzero_si128 (),
7047                                                   (__mmask8) -1);
7048 }
7049
7050 extern __inline __m128i
7051 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7052 _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
7053                      __m128i __B)
7054 {
7055   return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
7056                                                   (__v4si) __B,
7057                                                   (__v4si) __W,
7058                                                   (__mmask8) __U);
7059 }
7060
7061 extern __inline __m128i
7062 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7063 _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
7064 {
7065   return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A,
7066                                                   (__v4si) __B,
7067                                                   (__v4si)
7068                                                   _mm_setzero_si128 (),
7069                                                   (__mmask8) __U);
7070 }
7071
7072 extern __inline __m256i
7073 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7074 _mm256_rolv_epi64 (__m256i __A, __m256i __B)
7075 {
7076   return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
7077                                                   (__v4di) __B,
7078                                                   (__v4di)
7079                                                   _mm256_setzero_si256 (),
7080                                                   (__mmask8) -1);
7081 }
7082
7083 extern __inline __m256i
7084 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7085 _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7086                         __m256i __B)
7087 {
7088   return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
7089                                                   (__v4di) __B,
7090                                                   (__v4di) __W,
7091                                                   (__mmask8) __U);
7092 }
7093
7094 extern __inline __m256i
7095 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7096 _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7097 {
7098   return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A,
7099                                                   (__v4di) __B,
7100                                                   (__v4di)
7101                                                   _mm256_setzero_si256 (),
7102                                                   (__mmask8) __U);
7103 }
7104
7105 extern __inline __m128i
7106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7107 _mm_rolv_epi64 (__m128i __A, __m128i __B)
7108 {
7109   return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
7110                                                   (__v2di) __B,
7111                                                   (__v2di)
7112                                                   _mm_setzero_si128 (),
7113                                                   (__mmask8) -1);
7114 }
7115
7116 extern __inline __m128i
7117 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7118 _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
7119                      __m128i __B)
7120 {
7121   return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
7122                                                   (__v2di) __B,
7123                                                   (__v2di) __W,
7124                                                   (__mmask8) __U);
7125 }
7126
7127 extern __inline __m128i
7128 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7129 _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7130 {
7131   return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A,
7132                                                   (__v2di) __B,
7133                                                   (__v2di)
7134                                                   _mm_setzero_si128 (),
7135                                                   (__mmask8) __U);
7136 }
7137
7138 extern __inline __m256i
7139 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7140 _mm256_rorv_epi64 (__m256i __A, __m256i __B)
7141 {
7142   return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
7143                                                   (__v4di) __B,
7144                                                   (__v4di)
7145                                                   _mm256_setzero_si256 (),
7146                                                   (__mmask8) -1);
7147 }
7148
7149 extern __inline __m256i
7150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7151 _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7152                         __m256i __B)
7153 {
7154   return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
7155                                                   (__v4di) __B,
7156                                                   (__v4di) __W,
7157                                                   (__mmask8) __U);
7158 }
7159
7160 extern __inline __m256i
7161 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7162 _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7163 {
7164   return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A,
7165                                                   (__v4di) __B,
7166                                                   (__v4di)
7167                                                   _mm256_setzero_si256 (),
7168                                                   (__mmask8) __U);
7169 }
7170
7171 extern __inline __m128i
7172 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7173 _mm_rorv_epi64 (__m128i __A, __m128i __B)
7174 {
7175   return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
7176                                                   (__v2di) __B,
7177                                                   (__v2di)
7178                                                   _mm_setzero_si128 (),
7179                                                   (__mmask8) -1);
7180 }
7181
7182 extern __inline __m128i
7183 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7184 _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
7185                      __m128i __B)
7186 {
7187   return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
7188                                                   (__v2di) __B,
7189                                                   (__v2di) __W,
7190                                                   (__mmask8) __U);
7191 }
7192
7193 extern __inline __m128i
7194 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7195 _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7196 {
7197   return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A,
7198                                                   (__v2di) __B,
7199                                                   (__v2di)
7200                                                   _mm_setzero_si128 (),
7201                                                   (__mmask8) __U);
7202 }
7203
7204 extern __inline __m256i
7205 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7206 _mm256_srav_epi64 (__m256i __X, __m256i __Y)
7207 {
7208   return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
7209                                                   (__v4di) __Y,
7210                                                   (__v4di)
7211                                                   _mm256_setzero_si256 (),
7212                                                   (__mmask8) -1);
7213 }
7214
7215 extern __inline __m256i
7216 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7217 _mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X,
7218                         __m256i __Y)
7219 {
7220   return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
7221                                                   (__v4di) __Y,
7222                                                   (__v4di) __W,
7223                                                   (__mmask8) __U);
7224 }
7225
7226 extern __inline __m256i
7227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7228 _mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
7229 {
7230   return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X,
7231                                                   (__v4di) __Y,
7232                                                   (__v4di)
7233                                                   _mm256_setzero_si256 (),
7234                                                   (__mmask8) __U);
7235 }
7236
7237 extern __inline __m256i
7238 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7239 _mm256_mask_and_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7240                        __m256i __B)
7241 {
7242   return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
7243                                                  (__v4di) __B,
7244                                                  (__v4di) __W, __U);
7245 }
7246
7247 extern __inline __m256i
7248 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7249 _mm256_maskz_and_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7250 {
7251   return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A,
7252                                                  (__v4di) __B,
7253                                                  (__v4di)
7254                                                  _mm256_setzero_pd (),
7255                                                  __U);
7256 }
7257
7258 extern __inline __m128i
7259 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7260 _mm_mask_and_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
7261                     __m128i __B)
7262 {
7263   return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
7264                                                  (__v2di) __B,
7265                                                  (__v2di) __W, __U);
7266 }
7267
7268 extern __inline __m128i
7269 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7270 _mm_maskz_and_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7271 {
7272   return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A,
7273                                                  (__v2di) __B,
7274                                                  (__v2di)
7275                                                  _mm_setzero_pd (),
7276                                                  __U);
7277 }
7278
7279 extern __inline __m256i
7280 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7281 _mm256_mask_andnot_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7282                           __m256i __B)
7283 {
7284   return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
7285                                                   (__v4di) __B,
7286                                                   (__v4di) __W, __U);
7287 }
7288
7289 extern __inline __m256i
7290 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7291 _mm256_maskz_andnot_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7292 {
7293   return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A,
7294                                                   (__v4di) __B,
7295                                                   (__v4di)
7296                                                   _mm256_setzero_pd (),
7297                                                   __U);
7298 }
7299
7300 extern __inline __m128i
7301 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7302 _mm_mask_andnot_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
7303                        __m128i __B)
7304 {
7305   return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
7306                                                   (__v2di) __B,
7307                                                   (__v2di) __W, __U);
7308 }
7309
7310 extern __inline __m128i
7311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7312 _mm_maskz_andnot_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7313 {
7314   return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A,
7315                                                   (__v2di) __B,
7316                                                   (__v2di)
7317                                                   _mm_setzero_pd (),
7318                                                   __U);
7319 }
7320
7321 extern __inline __m256i
7322 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7323 _mm256_mask_or_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7324                       __m256i __B)
7325 {
7326   return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
7327                                                 (__v4di) __B,
7328                                                 (__v4di) __W,
7329                                                 (__mmask8) __U);
7330 }
7331
7332 extern __inline __m256i
7333 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7334 _mm256_maskz_or_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7335 {
7336   return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A,
7337                                                 (__v4di) __B,
7338                                                 (__v4di)
7339                                                 _mm256_setzero_si256 (),
7340                                                 (__mmask8) __U);
7341 }
7342
7343 extern __inline __m128i
7344 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7345 _mm_mask_or_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
7346 {
7347   return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
7348                                                 (__v2di) __B,
7349                                                 (__v2di) __W,
7350                                                 (__mmask8) __U);
7351 }
7352
7353 extern __inline __m128i
7354 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7355 _mm_maskz_or_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7356 {
7357   return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A,
7358                                                 (__v2di) __B,
7359                                                 (__v2di)
7360                                                 _mm_setzero_si128 (),
7361                                                 (__mmask8) __U);
7362 }
7363
7364 extern __inline __m256i
7365 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7366 _mm256_mask_xor_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
7367                        __m256i __B)
7368 {
7369   return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
7370                                                  (__v4di) __B,
7371                                                  (__v4di) __W,
7372                                                  (__mmask8) __U);
7373 }
7374
7375 extern __inline __m256i
7376 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7377 _mm256_maskz_xor_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
7378 {
7379   return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A,
7380                                                  (__v4di) __B,
7381                                                  (__v4di)
7382                                                  _mm256_setzero_si256 (),
7383                                                  (__mmask8) __U);
7384 }
7385
7386 extern __inline __m128i
7387 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7388 _mm_mask_xor_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
7389                     __m128i __B)
7390 {
7391   return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
7392                                                  (__v2di) __B,
7393                                                  (__v2di) __W,
7394                                                  (__mmask8) __U);
7395 }
7396
7397 extern __inline __m128i
7398 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7399 _mm_maskz_xor_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
7400 {
7401   return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A,
7402                                                  (__v2di) __B,
7403                                                  (__v2di)
7404                                                  _mm_setzero_si128 (),
7405                                                  (__mmask8) __U);
7406 }
7407
7408 extern __inline __m256d
7409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7410 _mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A,
7411                     __m256d __B)
7412 {
7413   return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A,
7414                                                  (__v4df) __B,
7415                                                  (__v4df) __W,
7416                                                  (__mmask8) __U);
7417 }
7418
7419 extern __inline __m256d
7420 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7421 _mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B)
7422 {
7423   return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A,
7424                                                  (__v4df) __B,
7425                                                  (__v4df)
7426                                                  _mm256_setzero_pd (),
7427                                                  (__mmask8) __U);
7428 }
7429
7430 extern __inline __m256
7431 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7432 _mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
7433 {
7434   return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A,
7435                                                 (__v8sf) __B,
7436                                                 (__v8sf) __W,
7437                                                 (__mmask8) __U);
7438 }
7439
7440 extern __inline __m256
7441 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7442 _mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B)
7443 {
7444   return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A,
7445                                                 (__v8sf) __B,
7446                                                 (__v8sf)
7447                                                 _mm256_setzero_ps (),
7448                                                 (__mmask8) __U);
7449 }
7450
7451 extern __inline __m128
7452 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7453 _mm_mask_div_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
7454 {
7455   return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A,
7456                                              (__v4sf) __B,
7457                                              (__v4sf) __W,
7458                                              (__mmask8) __U);
7459 }
7460
7461 extern __inline __m128
7462 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7463 _mm_maskz_div_ps (__mmask8 __U, __m128 __A, __m128 __B)
7464 {
7465   return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A,
7466                                              (__v4sf) __B,
7467                                              (__v4sf)
7468                                              _mm_setzero_ps (),
7469                                              (__mmask8) __U);
7470 }
7471
7472 extern __inline __m128d
7473 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7474 _mm_mask_div_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
7475 {
7476   return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A,
7477                                               (__v2df) __B,
7478                                               (__v2df) __W,
7479                                               (__mmask8) __U);
7480 }
7481
7482 extern __inline __m128d
7483 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7484 _mm_maskz_div_pd (__mmask8 __U, __m128d __A, __m128d __B)
7485 {
7486   return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A,
7487                                               (__v2df) __B,
7488                                               (__v2df)
7489                                               _mm_setzero_pd (),
7490                                               (__mmask8) __U);
7491 }
7492
7493 extern __inline __m256d
7494 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7495 _mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A,
7496                     __m256d __B)
7497 {
7498   return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A,
7499                                                  (__v4df) __B,
7500                                                  (__v4df) __W,
7501                                                  (__mmask8) __U);
7502 }
7503
7504 extern __inline __m256d
7505 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7506 _mm256_mask_div_pd (__m256d __W, __mmask8 __U, __m256d __A,
7507                     __m256d __B)
7508 {
7509   return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A,
7510                                                  (__v4df) __B,
7511                                                  (__v4df) __W,
7512                                                  (__mmask8) __U);
7513 }
7514
7515 extern __inline __m256d
7516 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7517 _mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B)
7518 {
7519   return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A,
7520                                                  (__v4df) __B,
7521                                                  (__v4df)
7522                                                  _mm256_setzero_pd (),
7523                                                  (__mmask8) __U);
7524 }
7525
7526 extern __inline __m256
7527 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7528 _mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
7529 {
7530   return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A,
7531                                                 (__v8sf) __B,
7532                                                 (__v8sf) __W,
7533                                                 (__mmask8) __U);
7534 }
7535
7536 extern __inline __m256d
7537 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7538 _mm256_maskz_div_pd (__mmask8 __U, __m256d __A, __m256d __B)
7539 {
7540   return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A,
7541                                                  (__v4df) __B,
7542                                                  (__v4df)
7543                                                  _mm256_setzero_pd (),
7544                                                  (__mmask8) __U);
7545 }
7546
7547 extern __inline __m256
7548 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7549 _mm256_mask_div_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
7550 {
7551   return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A,
7552                                                 (__v8sf) __B,
7553                                                 (__v8sf) __W,
7554                                                 (__mmask8) __U);
7555 }
7556
7557 extern __inline __m256
7558 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7559 _mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B)
7560 {
7561   return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A,
7562                                                 (__v8sf) __B,
7563                                                 (__v8sf)
7564                                                 _mm256_setzero_ps (),
7565                                                 (__mmask8) __U);
7566 }
7567
7568 extern __inline __m256
7569 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7570 _mm256_maskz_div_ps (__mmask8 __U, __m256 __A, __m256 __B)
7571 {
7572   return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A,
7573                                                 (__v8sf) __B,
7574                                                 (__v8sf)
7575                                                 _mm256_setzero_ps (),
7576                                                 (__mmask8) __U);
7577 }
7578
7579 extern __inline __m128
7580 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7581 _mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
7582 {
7583   return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A,
7584                                              (__v4sf) __B,
7585                                              (__v4sf) __W,
7586                                              (__mmask8) __U);
7587 }
7588
7589 extern __inline __m128
7590 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7591 _mm_mask_mul_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
7592 {
7593   return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A,
7594                                              (__v4sf) __B,
7595                                              (__v4sf) __W,
7596                                              (__mmask8) __U);
7597 }
7598
7599 extern __inline __m128
7600 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7601 _mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B)
7602 {
7603   return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A,
7604                                              (__v4sf) __B,
7605                                              (__v4sf)
7606                                              _mm_setzero_ps (),
7607                                              (__mmask8) __U);
7608 }
7609
7610 extern __inline __m128
7611 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7612 _mm_maskz_mul_ps (__mmask8 __U, __m128 __A, __m128 __B)
7613 {
7614   return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A,
7615                                              (__v4sf) __B,
7616                                              (__v4sf)
7617                                              _mm_setzero_ps (),
7618                                              (__mmask8) __U);
7619 }
7620
7621 extern __inline __m128
7622 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7623 _mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
7624 {
7625   return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A,
7626                                              (__v4sf) __B,
7627                                              (__v4sf) __W,
7628                                              (__mmask8) __U);
7629 }
7630
7631 extern __inline __m128
7632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7633 _mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B)
7634 {
7635   return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A,
7636                                              (__v4sf) __B,
7637                                              (__v4sf)
7638                                              _mm_setzero_ps (),
7639                                              (__mmask8) __U);
7640 }
7641
7642 extern __inline __m128d
7643 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7644 _mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
7645 {
7646   return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A,
7647                                               (__v2df) __B,
7648                                               (__v2df) __W,
7649                                               (__mmask8) __U);
7650 }
7651
7652 extern __inline __m128d
7653 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7654 _mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B)
7655 {
7656   return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A,
7657                                               (__v2df) __B,
7658                                               (__v2df)
7659                                               _mm_setzero_pd (),
7660                                               (__mmask8) __U);
7661 }
7662
7663 extern __inline __m128d
7664 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7665 _mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
7666 {
7667   return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A,
7668                                               (__v2df) __B,
7669                                               (__v2df) __W,
7670                                               (__mmask8) __U);
7671 }
7672
7673 extern __inline __m128d
7674 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7675 _mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B)
7676 {
7677   return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A,
7678                                               (__v2df) __B,
7679                                               (__v2df)
7680                                               _mm_setzero_pd (),
7681                                               (__mmask8) __U);
7682 }
7683
7684 extern __inline __m128d
7685 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7686 _mm_mask_mul_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
7687 {
7688   return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A,
7689                                               (__v2df) __B,
7690                                               (__v2df) __W,
7691                                               (__mmask8) __U);
7692 }
7693
7694 extern __inline __m128d
7695 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7696 _mm_maskz_mul_pd (__mmask8 __U, __m128d __A, __m128d __B)
7697 {
7698   return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A,
7699                                               (__v2df) __B,
7700                                               (__v2df)
7701                                               _mm_setzero_pd (),
7702                                               (__mmask8) __U);
7703 }
7704
7705 extern __inline __m256
7706 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7707 _mm256_mask_mul_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
7708 {
7709   return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A,
7710                                                 (__v8sf) __B,
7711                                                 (__v8sf) __W,
7712                                                 (__mmask8) __U);
7713 }
7714
7715 extern __inline __m256
7716 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7717 _mm256_maskz_mul_ps (__mmask8 __U, __m256 __A, __m256 __B)
7718 {
7719   return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A,
7720                                                 (__v8sf) __B,
7721                                                 (__v8sf)
7722                                                 _mm256_setzero_ps (),
7723                                                 (__mmask8) __U);
7724 }
7725
7726 extern __inline __m256d
7727 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7728 _mm256_mask_mul_pd (__m256d __W, __mmask8 __U, __m256d __A,
7729                     __m256d __B)
7730 {
7731   return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A,
7732                                                  (__v4df) __B,
7733                                                  (__v4df) __W,
7734                                                  (__mmask8) __U);
7735 }
7736
7737 extern __inline __m256d
7738 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7739 _mm256_maskz_mul_pd (__mmask8 __U, __m256d __A, __m256d __B)
7740 {
7741   return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A,
7742                                                  (__v4df) __B,
7743                                                  (__v4df)
7744                                                  _mm256_setzero_pd (),
7745                                                  (__mmask8) __U);
7746 }
7747
7748 extern __inline __m256i
7749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7750 _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B)
7751 {
7752   return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
7753                                                   (__v4di) __B,
7754                                                   (__v4di)
7755                                                   _mm256_setzero_si256 (),
7756                                                   __M);
7757 }
7758
7759 extern __inline __m256i
7760 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7761 _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A,
7762                        __m256i __B)
7763 {
7764   return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
7765                                                   (__v4di) __B,
7766                                                   (__v4di) __W, __M);
7767 }
7768
7769 extern __inline __m256i
7770 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7771 _mm256_min_epi64 (__m256i __A, __m256i __B)
7772 {
7773   return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
7774                                                   (__v4di) __B,
7775                                                   (__v4di)
7776                                                   _mm256_setzero_si256 (),
7777                                                   (__mmask8) -1);
7778 }
7779
7780 extern __inline __m256i
7781 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7782 _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A,
7783                        __m256i __B)
7784 {
7785   return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
7786                                                   (__v4di) __B,
7787                                                   (__v4di) __W, __M);
7788 }
7789
7790 extern __inline __m256i
7791 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7792 _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B)
7793 {
7794   return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A,
7795                                                   (__v4di) __B,
7796                                                   (__v4di)
7797                                                   _mm256_setzero_si256 (),
7798                                                   __M);
7799 }
7800
7801 extern __inline __m256i
7802 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7803 _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B)
7804 {
7805   return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
7806                                                   (__v4di) __B,
7807                                                   (__v4di)
7808                                                   _mm256_setzero_si256 (),
7809                                                   __M);
7810 }
7811
7812 extern __inline __m256i
7813 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7814 _mm256_max_epi64 (__m256i __A, __m256i __B)
7815 {
7816   return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A,
7817                                                   (__v4di) __B,
7818                                                   (__v4di)
7819                                                   _mm256_setzero_si256 (),
7820                                                   (__mmask8) -1);
7821 }
7822
7823 extern __inline __m256i
7824 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7825 _mm256_max_epu64 (__m256i __A, __m256i __B)
7826 {
7827   return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
7828                                                   (__v4di) __B,
7829                                                   (__v4di)
7830                                                   _mm256_setzero_si256 (),
7831                                                   (__mmask8) -1);
7832 }
7833
7834 extern __inline __m256i
7835 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7836 _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A,
7837                        __m256i __B)
7838 {
7839   return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A,
7840                                                   (__v4di) __B,
7841                                                   (__v4di) __W, __M);
7842 }
7843
7844 extern __inline __m256i
7845 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7846 _mm256_min_epu64 (__m256i __A, __m256i __B)
7847 {
7848   return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
7849                                                   (__v4di) __B,
7850                                                   (__v4di)
7851                                                   _mm256_setzero_si256 (),
7852                                                   (__mmask8) -1);
7853 }
7854
7855 extern __inline __m256i
7856 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7857 _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A,
7858                        __m256i __B)
7859 {
7860   return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
7861                                                   (__v4di) __B,
7862                                                   (__v4di) __W, __M);
7863 }
7864
7865 extern __inline __m256i
7866 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7867 _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B)
7868 {
7869   return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A,
7870                                                   (__v4di) __B,
7871                                                   (__v4di)
7872                                                   _mm256_setzero_si256 (),
7873                                                   __M);
7874 }
7875
7876 extern __inline __m256i
7877 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7878 _mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B)
7879 {
7880   return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A,
7881                                                   (__v8si) __B,
7882                                                   (__v8si)
7883                                                   _mm256_setzero_si256 (),
7884                                                   __M);
7885 }
7886
7887 extern __inline __m256i
7888 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7889 _mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
7890                        __m256i __B)
7891 {
7892   return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A,
7893                                                   (__v8si) __B,
7894                                                   (__v8si) __W, __M);
7895 }
7896
7897 extern __inline __m256i
7898 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7899 _mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B)
7900 {
7901   return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A,
7902                                                   (__v8si) __B,
7903                                                   (__v8si)
7904                                                   _mm256_setzero_si256 (),
7905                                                   __M);
7906 }
7907
7908 extern __inline __m256i
7909 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7910 _mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
7911                        __m256i __B)
7912 {
7913   return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A,
7914                                                   (__v8si) __B,
7915                                                   (__v8si) __W, __M);
7916 }
7917
7918 extern __inline __m256i
7919 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7920 _mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B)
7921 {
7922   return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A,
7923                                                   (__v8si) __B,
7924                                                   (__v8si)
7925                                                   _mm256_setzero_si256 (),
7926                                                   __M);
7927 }
7928
7929 extern __inline __m256i
7930 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7931 _mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A,
7932                        __m256i __B)
7933 {
7934   return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A,
7935                                                   (__v8si) __B,
7936                                                   (__v8si) __W, __M);
7937 }
7938
7939 extern __inline __m256i
7940 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7941 _mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B)
7942 {
7943   return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A,
7944                                                   (__v8si) __B,
7945                                                   (__v8si)
7946                                                   _mm256_setzero_si256 (),
7947                                                   __M);
7948 }
7949
7950 extern __inline __m256i
7951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7952 _mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A,
7953                        __m256i __B)
7954 {
7955   return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A,
7956                                                   (__v8si) __B,
7957                                                   (__v8si) __W, __M);
7958 }
7959
7960 extern __inline __m128i
7961 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7962 _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B)
7963 {
7964   return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
7965                                                   (__v2di) __B,
7966                                                   (__v2di)
7967                                                   _mm_setzero_si128 (),
7968                                                   __M);
7969 }
7970
7971 extern __inline __m128i
7972 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7973 _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A,
7974                     __m128i __B)
7975 {
7976   return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
7977                                                   (__v2di) __B,
7978                                                   (__v2di) __W, __M);
7979 }
7980
7981 extern __inline __m128i
7982 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7983 _mm_min_epi64 (__m128i __A, __m128i __B)
7984 {
7985   return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
7986                                                   (__v2di) __B,
7987                                                   (__v2di)
7988                                                   _mm_setzero_si128 (),
7989                                                   (__mmask8) -1);
7990 }
7991
7992 extern __inline __m128i
7993 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
7994 _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A,
7995                     __m128i __B)
7996 {
7997   return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
7998                                                   (__v2di) __B,
7999                                                   (__v2di) __W, __M);
8000 }
8001
8002 extern __inline __m128i
8003 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8004 _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B)
8005 {
8006   return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A,
8007                                                   (__v2di) __B,
8008                                                   (__v2di)
8009                                                   _mm_setzero_si128 (),
8010                                                   __M);
8011 }
8012
8013 extern __inline __m128i
8014 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8015 _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B)
8016 {
8017   return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
8018                                                   (__v2di) __B,
8019                                                   (__v2di)
8020                                                   _mm_setzero_si128 (),
8021                                                   __M);
8022 }
8023
8024 extern __inline __m128i
8025 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8026 _mm_max_epi64 (__m128i __A, __m128i __B)
8027 {
8028   return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A,
8029                                                   (__v2di) __B,
8030                                                   (__v2di)
8031                                                   _mm_setzero_si128 (),
8032                                                   (__mmask8) -1);
8033 }
8034
8035 extern __inline __m128i
8036 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8037 _mm_max_epu64 (__m128i __A, __m128i __B)
8038 {
8039   return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
8040                                                   (__v2di) __B,
8041                                                   (__v2di)
8042                                                   _mm_setzero_si128 (),
8043                                                   (__mmask8) -1);
8044 }
8045
8046 extern __inline __m128i
8047 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8048 _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A,
8049                     __m128i __B)
8050 {
8051   return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A,
8052                                                   (__v2di) __B,
8053                                                   (__v2di) __W, __M);
8054 }
8055
8056 extern __inline __m128i
8057 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8058 _mm_min_epu64 (__m128i __A, __m128i __B)
8059 {
8060   return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
8061                                                   (__v2di) __B,
8062                                                   (__v2di)
8063                                                   _mm_setzero_si128 (),
8064                                                   (__mmask8) -1);
8065 }
8066
8067 extern __inline __m128i
8068 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8069 _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A,
8070                     __m128i __B)
8071 {
8072   return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
8073                                                   (__v2di) __B,
8074                                                   (__v2di) __W, __M);
8075 }
8076
8077 extern __inline __m128i
8078 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8079 _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B)
8080 {
8081   return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A,
8082                                                   (__v2di) __B,
8083                                                   (__v2di)
8084                                                   _mm_setzero_si128 (),
8085                                                   __M);
8086 }
8087
8088 extern __inline __m128i
8089 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8090 _mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B)
8091 {
8092   return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A,
8093                                                   (__v4si) __B,
8094                                                   (__v4si)
8095                                                   _mm_setzero_si128 (),
8096                                                   __M);
8097 }
8098
8099 extern __inline __m128i
8100 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8101 _mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A,
8102                     __m128i __B)
8103 {
8104   return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A,
8105                                                   (__v4si) __B,
8106                                                   (__v4si) __W, __M);
8107 }
8108
8109 extern __inline __m128i
8110 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8111 _mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B)
8112 {
8113   return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A,
8114                                                   (__v4si) __B,
8115                                                   (__v4si)
8116                                                   _mm_setzero_si128 (),
8117                                                   __M);
8118 }
8119
8120 extern __inline __m128i
8121 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8122 _mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A,
8123                     __m128i __B)
8124 {
8125   return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A,
8126                                                   (__v4si) __B,
8127                                                   (__v4si) __W, __M);
8128 }
8129
8130 extern __inline __m128i
8131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8132 _mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B)
8133 {
8134   return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A,
8135                                                   (__v4si) __B,
8136                                                   (__v4si)
8137                                                   _mm_setzero_si128 (),
8138                                                   __M);
8139 }
8140
8141 extern __inline __m128i
8142 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8143 _mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A,
8144                     __m128i __B)
8145 {
8146   return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A,
8147                                                   (__v4si) __B,
8148                                                   (__v4si) __W, __M);
8149 }
8150
8151 extern __inline __m128i
8152 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8153 _mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B)
8154 {
8155   return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A,
8156                                                   (__v4si) __B,
8157                                                   (__v4si)
8158                                                   _mm_setzero_si128 (),
8159                                                   __M);
8160 }
8161
8162 extern __inline __m128i
8163 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8164 _mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A,
8165                     __m128i __B)
8166 {
8167   return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A,
8168                                                   (__v4si) __B,
8169                                                   (__v4si) __W, __M);
8170 }
8171
8172 #ifndef __AVX512CD__
8173 #pragma GCC push_options
8174 #pragma GCC target("avx512vl,avx512cd")
8175 #define __DISABLE_AVX512VLCD__
8176 #endif
8177
8178 extern __inline __m128i
8179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8180 _mm_broadcastmb_epi64 (__mmask8 __A)
8181 {
8182   return (__m128i) __builtin_ia32_broadcastmb128 (__A);
8183 }
8184
8185 extern __inline __m256i
8186 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8187 _mm256_broadcastmb_epi64 (__mmask8 __A)
8188 {
8189   return (__m256i) __builtin_ia32_broadcastmb256 (__A);
8190 }
8191
8192 extern __inline __m128i
8193 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8194 _mm_broadcastmw_epi32 (__mmask16 __A)
8195 {
8196   return (__m128i) __builtin_ia32_broadcastmw128 (__A);
8197 }
8198
8199 extern __inline __m256i
8200 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8201 _mm256_broadcastmw_epi32 (__mmask16 __A)
8202 {
8203   return (__m256i) __builtin_ia32_broadcastmw256 (__A);
8204 }
8205
8206 extern __inline __m256i
8207 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8208 _mm256_lzcnt_epi32 (__m256i __A)
8209 {
8210   return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
8211                                                      (__v8si)
8212                                                      _mm256_setzero_si256 (),
8213                                                      (__mmask8) -1);
8214 }
8215
8216 extern __inline __m256i
8217 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8218 _mm256_mask_lzcnt_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
8219 {
8220   return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
8221                                                      (__v8si) __W,
8222                                                      (__mmask8) __U);
8223 }
8224
8225 extern __inline __m256i
8226 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8227 _mm256_maskz_lzcnt_epi32 (__mmask8 __U, __m256i __A)
8228 {
8229   return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A,
8230                                                      (__v8si)
8231                                                      _mm256_setzero_si256 (),
8232                                                      (__mmask8) __U);
8233 }
8234
8235 extern __inline __m256i
8236 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8237 _mm256_lzcnt_epi64 (__m256i __A)
8238 {
8239   return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
8240                                                      (__v4di)
8241                                                      _mm256_setzero_si256 (),
8242                                                      (__mmask8) -1);
8243 }
8244
8245 extern __inline __m256i
8246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8247 _mm256_mask_lzcnt_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
8248 {
8249   return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
8250                                                      (__v4di) __W,
8251                                                      (__mmask8) __U);
8252 }
8253
8254 extern __inline __m256i
8255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8256 _mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A)
8257 {
8258   return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A,
8259                                                      (__v4di)
8260                                                      _mm256_setzero_si256 (),
8261                                                      (__mmask8) __U);
8262 }
8263
8264 extern __inline __m256i
8265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8266 _mm256_conflict_epi64 (__m256i __A)
8267 {
8268   return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A,
8269                                                          (__v4di)
8270                                                          _mm256_setzero_si256 (),
8271                                                          (__mmask8) -1);
8272 }
8273
8274 extern __inline __m256i
8275 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8276 _mm256_mask_conflict_epi64 (__m256i __W, __mmask8 __U, __m256i __A)
8277 {
8278   return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A,
8279                                                          (__v4di) __W,
8280                                                          (__mmask8)
8281                                                          __U);
8282 }
8283
8284 extern __inline __m256i
8285 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8286 _mm256_maskz_conflict_epi64 (__mmask8 __U, __m256i __A)
8287 {
8288   return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A,
8289                                                          (__v4di)
8290                                                          _mm256_setzero_si256 (),
8291                                                          (__mmask8)
8292                                                          __U);
8293 }
8294
8295 extern __inline __m256i
8296 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8297 _mm256_conflict_epi32 (__m256i __A)
8298 {
8299   return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A,
8300                                                          (__v8si)
8301                                                          _mm256_setzero_si256 (),
8302                                                          (__mmask8) -1);
8303 }
8304
8305 extern __inline __m256i
8306 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8307 _mm256_mask_conflict_epi32 (__m256i __W, __mmask8 __U, __m256i __A)
8308 {
8309   return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A,
8310                                                          (__v8si) __W,
8311                                                          (__mmask8)
8312                                                          __U);
8313 }
8314
8315 extern __inline __m256i
8316 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8317 _mm256_maskz_conflict_epi32 (__mmask8 __U, __m256i __A)
8318 {
8319   return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A,
8320                                                          (__v8si)
8321                                                          _mm256_setzero_si256 (),
8322                                                          (__mmask8)
8323                                                          __U);
8324 }
8325
8326 extern __inline __m128i
8327 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8328 _mm_lzcnt_epi32 (__m128i __A)
8329 {
8330   return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
8331                                                      (__v4si)
8332                                                      _mm_setzero_si128 (),
8333                                                      (__mmask8) -1);
8334 }
8335
8336 extern __inline __m128i
8337 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8338 _mm_mask_lzcnt_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
8339 {
8340   return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
8341                                                      (__v4si) __W,
8342                                                      (__mmask8) __U);
8343 }
8344
8345 extern __inline __m128i
8346 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8347 _mm_maskz_lzcnt_epi32 (__mmask8 __U, __m128i __A)
8348 {
8349   return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A,
8350                                                      (__v4si)
8351                                                      _mm_setzero_si128 (),
8352                                                      (__mmask8) __U);
8353 }
8354
8355 extern __inline __m128i
8356 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8357 _mm_lzcnt_epi64 (__m128i __A)
8358 {
8359   return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
8360                                                      (__v2di)
8361                                                      _mm_setzero_si128 (),
8362                                                      (__mmask8) -1);
8363 }
8364
8365 extern __inline __m128i
8366 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8367 _mm_mask_lzcnt_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
8368 {
8369   return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
8370                                                      (__v2di) __W,
8371                                                      (__mmask8) __U);
8372 }
8373
8374 extern __inline __m128i
8375 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8376 _mm_maskz_lzcnt_epi64 (__mmask8 __U, __m128i __A)
8377 {
8378   return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A,
8379                                                      (__v2di)
8380                                                      _mm_setzero_si128 (),
8381                                                      (__mmask8) __U);
8382 }
8383
8384 extern __inline __m128i
8385 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8386 _mm_conflict_epi64 (__m128i __A)
8387 {
8388   return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A,
8389                                                          (__v2di)
8390                                                          _mm_setzero_si128 (),
8391                                                          (__mmask8) -1);
8392 }
8393
8394 extern __inline __m128i
8395 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8396 _mm_mask_conflict_epi64 (__m128i __W, __mmask8 __U, __m128i __A)
8397 {
8398   return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A,
8399                                                          (__v2di) __W,
8400                                                          (__mmask8)
8401                                                          __U);
8402 }
8403
8404 extern __inline __m128i
8405 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8406 _mm_maskz_conflict_epi64 (__mmask8 __U, __m128i __A)
8407 {
8408   return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A,
8409                                                          (__v2di)
8410                                                          _mm_setzero_si128 (),
8411                                                          (__mmask8)
8412                                                          __U);
8413 }
8414
8415 extern __inline __m128i
8416 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8417 _mm_conflict_epi32 (__m128i __A)
8418 {
8419   return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A,
8420                                                          (__v4si)
8421                                                          _mm_setzero_si128 (),
8422                                                          (__mmask8) -1);
8423 }
8424
8425 extern __inline __m128i
8426 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8427 _mm_mask_conflict_epi32 (__m128i __W, __mmask8 __U, __m128i __A)
8428 {
8429   return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A,
8430                                                          (__v4si) __W,
8431                                                          (__mmask8)
8432                                                          __U);
8433 }
8434
8435 extern __inline __m128i
8436 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8437 _mm_maskz_conflict_epi32 (__mmask8 __U, __m128i __A)
8438 {
8439   return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A,
8440                                                          (__v4si)
8441                                                          _mm_setzero_si128 (),
8442                                                          (__mmask8)
8443                                                          __U);
8444 }
8445
8446 #ifdef __DISABLE_AVX512VLCD__
8447 #pragma GCC pop_options
8448 #endif
8449
8450 extern __inline __m256d
8451 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8452 _mm256_mask_unpacklo_pd (__m256d __W, __mmask8 __U, __m256d __A,
8453                          __m256d __B)
8454 {
8455   return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
8456                                                     (__v4df) __B,
8457                                                     (__v4df) __W,
8458                                                     (__mmask8) __U);
8459 }
8460
8461 extern __inline __m256d
8462 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8463 _mm256_maskz_unpacklo_pd (__mmask8 __U, __m256d __A, __m256d __B)
8464 {
8465   return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A,
8466                                                     (__v4df) __B,
8467                                                     (__v4df)
8468                                                     _mm256_setzero_pd (),
8469                                                     (__mmask8) __U);
8470 }
8471
8472 extern __inline __m128d
8473 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8474 _mm_mask_unpacklo_pd (__m128d __W, __mmask8 __U, __m128d __A,
8475                       __m128d __B)
8476 {
8477   return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
8478                                                     (__v2df) __B,
8479                                                     (__v2df) __W,
8480                                                     (__mmask8) __U);
8481 }
8482
8483 extern __inline __m128d
8484 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8485 _mm_maskz_unpacklo_pd (__mmask8 __U, __m128d __A, __m128d __B)
8486 {
8487   return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A,
8488                                                     (__v2df) __B,
8489                                                     (__v2df)
8490                                                     _mm_setzero_pd (),
8491                                                     (__mmask8) __U);
8492 }
8493
8494 extern __inline __m256
8495 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8496 _mm256_mask_unpacklo_ps (__m256 __W, __mmask8 __U, __m256 __A,
8497                          __m256 __B)
8498 {
8499   return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
8500                                                    (__v8sf) __B,
8501                                                    (__v8sf) __W,
8502                                                    (__mmask8) __U);
8503 }
8504
8505 extern __inline __m256d
8506 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8507 _mm256_mask_unpackhi_pd (__m256d __W, __mmask8 __U, __m256d __A,
8508                          __m256d __B)
8509 {
8510   return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
8511                                                     (__v4df) __B,
8512                                                     (__v4df) __W,
8513                                                     (__mmask8) __U);
8514 }
8515
8516 extern __inline __m256d
8517 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8518 _mm256_maskz_unpackhi_pd (__mmask8 __U, __m256d __A, __m256d __B)
8519 {
8520   return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A,
8521                                                     (__v4df) __B,
8522                                                     (__v4df)
8523                                                     _mm256_setzero_pd (),
8524                                                     (__mmask8) __U);
8525 }
8526
8527 extern __inline __m128d
8528 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8529 _mm_mask_unpackhi_pd (__m128d __W, __mmask8 __U, __m128d __A,
8530                       __m128d __B)
8531 {
8532   return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
8533                                                     (__v2df) __B,
8534                                                     (__v2df) __W,
8535                                                     (__mmask8) __U);
8536 }
8537
8538 extern __inline __m128d
8539 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8540 _mm_maskz_unpackhi_pd (__mmask8 __U, __m128d __A, __m128d __B)
8541 {
8542   return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A,
8543                                                     (__v2df) __B,
8544                                                     (__v2df)
8545                                                     _mm_setzero_pd (),
8546                                                     (__mmask8) __U);
8547 }
8548
8549 extern __inline __m256
8550 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8551 _mm256_mask_unpackhi_ps (__m256 __W, __mmask8 __U, __m256 __A,
8552                          __m256 __B)
8553 {
8554   return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
8555                                                    (__v8sf) __B,
8556                                                    (__v8sf) __W,
8557                                                    (__mmask8) __U);
8558 }
8559
8560 extern __inline __m256
8561 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8562 _mm256_maskz_unpackhi_ps (__mmask8 __U, __m256 __A, __m256 __B)
8563 {
8564   return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A,
8565                                                    (__v8sf) __B,
8566                                                    (__v8sf)
8567                                                    _mm256_setzero_ps (),
8568                                                    (__mmask8) __U);
8569 }
8570
8571 extern __inline __m128
8572 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8573 _mm_mask_unpackhi_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
8574 {
8575   return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
8576                                                    (__v4sf) __B,
8577                                                    (__v4sf) __W,
8578                                                    (__mmask8) __U);
8579 }
8580
8581 extern __inline __m128
8582 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8583 _mm_maskz_unpackhi_ps (__mmask8 __U, __m128 __A, __m128 __B)
8584 {
8585   return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A,
8586                                                    (__v4sf) __B,
8587                                                    (__v4sf)
8588                                                    _mm_setzero_ps (),
8589                                                    (__mmask8) __U);
8590 }
8591
8592 extern __inline __m128
8593 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8594 _mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A)
8595 {
8596   return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A,
8597                                                  (__v4sf) __W,
8598                                                  (__mmask8) __U);
8599 }
8600
8601 extern __inline __m128
8602 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8603 _mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A)
8604 {
8605   return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A,
8606                                                  (__v4sf)
8607                                                  _mm_setzero_ps (),
8608                                                  (__mmask8) __U);
8609 }
8610
8611 extern __inline __m256
8612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8613 _mm256_maskz_unpacklo_ps (__mmask8 __U, __m256 __A, __m256 __B)
8614 {
8615   return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A,
8616                                                    (__v8sf) __B,
8617                                                    (__v8sf)
8618                                                    _mm256_setzero_ps (),
8619                                                    (__mmask8) __U);
8620 }
8621
8622 extern __inline __m256
8623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8624 _mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A)
8625 {
8626   return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A,
8627                                                     (__v8sf) __W,
8628                                                     (__mmask8) __U);
8629 }
8630
8631 extern __inline __m256
8632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8633 _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A)
8634 {
8635   return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A,
8636                                                     (__v8sf)
8637                                                     _mm256_setzero_ps (),
8638                                                     (__mmask8) __U);
8639 }
8640
8641 extern __inline __m128
8642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8643 _mm_mask_unpacklo_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
8644 {
8645   return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
8646                                                    (__v4sf) __B,
8647                                                    (__v4sf) __W,
8648                                                    (__mmask8) __U);
8649 }
8650
8651 extern __inline __m128
8652 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8653 _mm_maskz_unpacklo_ps (__mmask8 __U, __m128 __A, __m128 __B)
8654 {
8655   return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A,
8656                                                    (__v4sf) __B,
8657                                                    (__v4sf)
8658                                                    _mm_setzero_ps (),
8659                                                    (__mmask8) __U);
8660 }
8661
8662 extern __inline __m256i
8663 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8664 _mm256_mask_sra_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
8665                        __m128i __B)
8666 {
8667   return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A,
8668                                                  (__v4si) __B,
8669                                                  (__v8si) __W,
8670                                                  (__mmask8) __U);
8671 }
8672
8673 extern __inline __m256i
8674 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8675 _mm256_maskz_sra_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
8676 {
8677   return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A,
8678                                                  (__v4si) __B,
8679                                                  (__v8si)
8680                                                  _mm256_setzero_si256 (),
8681                                                  (__mmask8) __U);
8682 }
8683
8684 extern __inline __m128i
8685 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8686 _mm_mask_sra_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
8687                     __m128i __B)
8688 {
8689   return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A,
8690                                                  (__v4si) __B,
8691                                                  (__v4si) __W,
8692                                                  (__mmask8) __U);
8693 }
8694
8695 extern __inline __m128i
8696 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8697 _mm_maskz_sra_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
8698 {
8699   return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A,
8700                                                  (__v4si) __B,
8701                                                  (__v4si)
8702                                                  _mm_setzero_si128 (),
8703                                                  (__mmask8) __U);
8704 }
8705
8706 extern __inline __m256i
8707 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8708 _mm256_sra_epi64 (__m256i __A, __m128i __B)
8709 {
8710   return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
8711                                                  (__v2di) __B,
8712                                                  (__v4di)
8713                                                  _mm256_setzero_si256 (),
8714                                                  (__mmask8) -1);
8715 }
8716
8717 extern __inline __m256i
8718 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8719 _mm256_mask_sra_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
8720                        __m128i __B)
8721 {
8722   return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
8723                                                  (__v2di) __B,
8724                                                  (__v4di) __W,
8725                                                  (__mmask8) __U);
8726 }
8727
8728 extern __inline __m256i
8729 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8730 _mm256_maskz_sra_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
8731 {
8732   return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A,
8733                                                  (__v2di) __B,
8734                                                  (__v4di)
8735                                                  _mm256_setzero_si256 (),
8736                                                  (__mmask8) __U);
8737 }
8738
8739 extern __inline __m128i
8740 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8741 _mm_sra_epi64 (__m128i __A, __m128i __B)
8742 {
8743   return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
8744                                                  (__v2di) __B,
8745                                                  (__v2di)
8746                                                  _mm_setzero_si128 (),
8747                                                  (__mmask8) -1);
8748 }
8749
8750 extern __inline __m128i
8751 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8752 _mm_mask_sra_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
8753                     __m128i __B)
8754 {
8755   return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
8756                                                  (__v2di) __B,
8757                                                  (__v2di) __W,
8758                                                  (__mmask8) __U);
8759 }
8760
8761 extern __inline __m128i
8762 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8763 _mm_maskz_sra_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
8764 {
8765   return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A,
8766                                                  (__v2di) __B,
8767                                                  (__v2di)
8768                                                  _mm_setzero_si128 (),
8769                                                  (__mmask8) __U);
8770 }
8771
8772 extern __inline __m128i
8773 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8774 _mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
8775                     __m128i __B)
8776 {
8777   return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
8778                                                  (__v4si) __B,
8779                                                  (__v4si) __W,
8780                                                  (__mmask8) __U);
8781 }
8782
8783 extern __inline __m128i
8784 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8785 _mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
8786 {
8787   return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A,
8788                                                  (__v4si) __B,
8789                                                  (__v4si)
8790                                                  _mm_setzero_si128 (),
8791                                                  (__mmask8) __U);
8792 }
8793
8794 extern __inline __m128i
8795 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8796 _mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
8797                     __m128i __B)
8798 {
8799   return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
8800                                                  (__v2di) __B,
8801                                                  (__v2di) __W,
8802                                                  (__mmask8) __U);
8803 }
8804
8805 extern __inline __m128i
8806 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8807 _mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
8808 {
8809   return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A,
8810                                                  (__v2di) __B,
8811                                                  (__v2di)
8812                                                  _mm_setzero_si128 (),
8813                                                  (__mmask8) __U);
8814 }
8815
8816 extern __inline __m256i
8817 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8818 _mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
8819                        __m128i __B)
8820 {
8821   return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
8822                                                  (__v4si) __B,
8823                                                  (__v8si) __W,
8824                                                  (__mmask8) __U);
8825 }
8826
8827 extern __inline __m256i
8828 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8829 _mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B)
8830 {
8831   return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A,
8832                                                  (__v4si) __B,
8833                                                  (__v8si)
8834                                                  _mm256_setzero_si256 (),
8835                                                  (__mmask8) __U);
8836 }
8837
8838 extern __inline __m256i
8839 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8840 _mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
8841                        __m128i __B)
8842 {
8843   return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
8844                                                  (__v2di) __B,
8845                                                  (__v4di) __W,
8846                                                  (__mmask8) __U);
8847 }
8848
8849 extern __inline __m256i
8850 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8851 _mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B)
8852 {
8853   return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A,
8854                                                  (__v2di) __B,
8855                                                  (__v4di)
8856                                                  _mm256_setzero_si256 (),
8857                                                  (__mmask8) __U);
8858 }
8859
8860 extern __inline __m256
8861 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8862 _mm256_mask_permutexvar_ps (__m256 __W, __mmask8 __U, __m256i __X,
8863                             __m256 __Y)
8864 {
8865   return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y,
8866                                                     (__v8si) __X,
8867                                                     (__v8sf) __W,
8868                                                     (__mmask8) __U);
8869 }
8870
8871 extern __inline __m256
8872 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8873 _mm256_maskz_permutexvar_ps (__mmask8 __U, __m256i __X, __m256 __Y)
8874 {
8875   return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y,
8876                                                     (__v8si) __X,
8877                                                     (__v8sf)
8878                                                     _mm256_setzero_ps (),
8879                                                     (__mmask8) __U);
8880 }
8881
8882 extern __inline __m256d
8883 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8884 _mm256_permutexvar_pd (__m256i __X, __m256d __Y)
8885 {
8886   return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
8887                                                      (__v4di) __X,
8888                                                      (__v4df)
8889                                                      _mm256_setzero_pd (),
8890                                                      (__mmask8) -1);
8891 }
8892
8893 extern __inline __m256d
8894 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8895 _mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X,
8896                             __m256d __Y)
8897 {
8898   return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
8899                                                      (__v4di) __X,
8900                                                      (__v4df) __W,
8901                                                      (__mmask8) __U);
8902 }
8903
8904 extern __inline __m256d
8905 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8906 _mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y)
8907 {
8908   return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y,
8909                                                      (__v4di) __X,
8910                                                      (__v4df)
8911                                                      _mm256_setzero_pd (),
8912                                                      (__mmask8) __U);
8913 }
8914
8915 extern __inline __m256d
8916 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8917 _mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A,
8918                            __m256i __C)
8919 {
8920   return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
8921                                                         (__v4di) __C,
8922                                                         (__v4df) __W,
8923                                                         (__mmask8)
8924                                                         __U);
8925 }
8926
8927 extern __inline __m256d
8928 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8929 _mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C)
8930 {
8931   return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A,
8932                                                         (__v4di) __C,
8933                                                         (__v4df)
8934                                                         _mm256_setzero_pd (),
8935                                                         (__mmask8)
8936                                                         __U);
8937 }
8938
8939 extern __inline __m256
8940 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8941 _mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A,
8942                            __m256i __C)
8943 {
8944   return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
8945                                                        (__v8si) __C,
8946                                                        (__v8sf) __W,
8947                                                        (__mmask8) __U);
8948 }
8949
8950 extern __inline __m256
8951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8952 _mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C)
8953 {
8954   return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A,
8955                                                        (__v8si) __C,
8956                                                        (__v8sf)
8957                                                        _mm256_setzero_ps (),
8958                                                        (__mmask8) __U);
8959 }
8960
8961 extern __inline __m128d
8962 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8963 _mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A,
8964                         __m128i __C)
8965 {
8966   return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
8967                                                      (__v2di) __C,
8968                                                      (__v2df) __W,
8969                                                      (__mmask8) __U);
8970 }
8971
8972 extern __inline __m128d
8973 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8974 _mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C)
8975 {
8976   return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A,
8977                                                      (__v2di) __C,
8978                                                      (__v2df)
8979                                                      _mm_setzero_pd (),
8980                                                      (__mmask8) __U);
8981 }
8982
8983 extern __inline __m128
8984 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8985 _mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A,
8986                         __m128i __C)
8987 {
8988   return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
8989                                                     (__v4si) __C,
8990                                                     (__v4sf) __W,
8991                                                     (__mmask8) __U);
8992 }
8993
8994 extern __inline __m128
8995 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
8996 _mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C)
8997 {
8998   return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A,
8999                                                     (__v4si) __C,
9000                                                     (__v4sf)
9001                                                     _mm_setzero_ps (),
9002                                                     (__mmask8) __U);
9003 }
9004
9005 extern __inline __m256i
9006 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9007 _mm256_maskz_mullo_epi32 (__mmask8 __M, __m256i __A, __m256i __B)
9008 {
9009   return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A,
9010                                                   (__v8si) __B,
9011                                                   (__v8si)
9012                                                   _mm256_setzero_si256 (),
9013                                                   __M);
9014 }
9015
9016 extern __inline __m256i
9017 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9018 _mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y)
9019 {
9020   return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
9021                                                      (__v4di) __X,
9022                                                      (__v4di)
9023                                                      _mm256_setzero_si256 (),
9024                                                      __M);
9025 }
9026
9027 extern __inline __m256i
9028 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9029 _mm256_mask_mullo_epi32 (__m256i __W, __mmask8 __M, __m256i __A,
9030                          __m256i __B)
9031 {
9032   return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A,
9033                                                   (__v8si) __B,
9034                                                   (__v8si) __W, __M);
9035 }
9036
9037 extern __inline __m128i
9038 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9039 _mm_maskz_mullo_epi32 (__mmask8 __M, __m128i __A, __m128i __B)
9040 {
9041   return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A,
9042                                                   (__v4si) __B,
9043                                                   (__v4si)
9044                                                   _mm_setzero_si128 (),
9045                                                   __M);
9046 }
9047
9048 extern __inline __m128i
9049 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9050 _mm_mask_mullo_epi32 (__m128i __W, __mmask16 __M, __m128i __A,
9051                       __m128i __B)
9052 {
9053   return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A,
9054                                                   (__v4si) __B,
9055                                                   (__v4si) __W, __M);
9056 }
9057
9058 extern __inline __m256i
9059 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9060 _mm256_mask_mul_epi32 (__m256i __W, __mmask8 __M, __m256i __X,
9061                        __m256i __Y)
9062 {
9063   return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X,
9064                                                   (__v8si) __Y,
9065                                                   (__v4di) __W, __M);
9066 }
9067
9068 extern __inline __m256i
9069 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9070 _mm256_maskz_mul_epi32 (__mmask8 __M, __m256i __X, __m256i __Y)
9071 {
9072   return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X,
9073                                                   (__v8si) __Y,
9074                                                   (__v4di)
9075                                                   _mm256_setzero_si256 (),
9076                                                   __M);
9077 }
9078
9079 extern __inline __m128i
9080 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9081 _mm_mask_mul_epi32 (__m128i __W, __mmask8 __M, __m128i __X,
9082                     __m128i __Y)
9083 {
9084   return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X,
9085                                                   (__v4si) __Y,
9086                                                   (__v2di) __W, __M);
9087 }
9088
9089 extern __inline __m128i
9090 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9091 _mm_maskz_mul_epi32 (__mmask8 __M, __m128i __X, __m128i __Y)
9092 {
9093   return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X,
9094                                                   (__v4si) __Y,
9095                                                   (__v2di)
9096                                                   _mm_setzero_si128 (),
9097                                                   __M);
9098 }
9099
9100 extern __inline __m256i
9101 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9102 _mm256_permutexvar_epi64 (__m256i __X, __m256i __Y)
9103 {
9104   return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
9105                                                      (__v4di) __X,
9106                                                      (__v4di)
9107                                                      _mm256_setzero_si256 (),
9108                                                      (__mmask8) -1);
9109 }
9110
9111 extern __inline __m256i
9112 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9113 _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X,
9114                                __m256i __Y)
9115 {
9116   return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y,
9117                                                      (__v4di) __X,
9118                                                      (__v4di) __W,
9119                                                      __M);
9120 }
9121
9122 extern __inline __m256i
9123 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9124 _mm256_mask_mul_epu32 (__m256i __W, __mmask8 __M, __m256i __X,
9125                        __m256i __Y)
9126 {
9127   return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X,
9128                                                    (__v8si) __Y,
9129                                                    (__v4di) __W, __M);
9130 }
9131
9132 extern __inline __m256i
9133 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9134 _mm256_maskz_permutexvar_epi32 (__mmask8 __M, __m256i __X, __m256i __Y)
9135 {
9136   return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
9137                                                      (__v8si) __X,
9138                                                      (__v8si)
9139                                                      _mm256_setzero_si256 (),
9140                                                      __M);
9141 }
9142
9143 extern __inline __m256i
9144 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9145 _mm256_maskz_mul_epu32 (__mmask8 __M, __m256i __X, __m256i __Y)
9146 {
9147   return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X,
9148                                                    (__v8si) __Y,
9149                                                    (__v4di)
9150                                                    _mm256_setzero_si256 (),
9151                                                    __M);
9152 }
9153
9154 extern __inline __m128i
9155 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9156 _mm_mask_mul_epu32 (__m128i __W, __mmask8 __M, __m128i __X,
9157                     __m128i __Y)
9158 {
9159   return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X,
9160                                                    (__v4si) __Y,
9161                                                    (__v2di) __W, __M);
9162 }
9163
9164 extern __inline __m128i
9165 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9166 _mm_maskz_mul_epu32 (__mmask8 __M, __m128i __X, __m128i __Y)
9167 {
9168   return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X,
9169                                                    (__v4si) __Y,
9170                                                    (__v2di)
9171                                                    _mm_setzero_si128 (),
9172                                                    __M);
9173 }
9174
9175 extern __inline __m256i
9176 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9177 _mm256_permutexvar_epi32 (__m256i __X, __m256i __Y)
9178 {
9179   return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
9180                                                      (__v8si) __X,
9181                                                      (__v8si)
9182                                                      _mm256_setzero_si256 (),
9183                                                      (__mmask8) -1);
9184 }
9185
9186 extern __inline __m256i
9187 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9188 _mm256_mask_permutexvar_epi32 (__m256i __W, __mmask8 __M, __m256i __X,
9189                                __m256i __Y)
9190 {
9191   return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y,
9192                                                      (__v8si) __X,
9193                                                      (__v8si) __W,
9194                                                      __M);
9195 }
9196
9197 extern __inline __mmask8
9198   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9199 _mm256_mask_cmpneq_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9200 {
9201   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9202                                                   (__v8si) __Y, 4,
9203                                                   (__mmask8) __M);
9204 }
9205
9206 extern __inline __mmask8
9207   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9208 _mm256_cmpneq_epu32_mask (__m256i __X, __m256i __Y)
9209 {
9210   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9211                                                   (__v8si) __Y, 4,
9212                                                   (__mmask8) -1);
9213 }
9214
9215 extern __inline __mmask8
9216   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9217 _mm256_mask_cmplt_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9218 {
9219   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9220                                                   (__v8si) __Y, 1,
9221                                                   (__mmask8) __M);
9222 }
9223
9224 extern __inline __mmask8
9225   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9226 _mm256_cmplt_epu32_mask (__m256i __X, __m256i __Y)
9227 {
9228   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9229                                                   (__v8si) __Y, 1,
9230                                                   (__mmask8) -1);
9231 }
9232
9233 extern __inline __mmask8
9234   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9235 _mm256_mask_cmpge_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9236 {
9237   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9238                                                   (__v8si) __Y, 5,
9239                                                   (__mmask8) __M);
9240 }
9241
9242 extern __inline __mmask8
9243   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9244 _mm256_cmpge_epu32_mask (__m256i __X, __m256i __Y)
9245 {
9246   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9247                                                   (__v8si) __Y, 5,
9248                                                   (__mmask8) -1);
9249 }
9250
9251 extern __inline __mmask8
9252   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9253 _mm256_mask_cmple_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9254 {
9255   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9256                                                   (__v8si) __Y, 2,
9257                                                   (__mmask8) __M);
9258 }
9259
9260 extern __inline __mmask8
9261   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9262 _mm256_cmple_epu32_mask (__m256i __X, __m256i __Y)
9263 {
9264   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
9265                                                   (__v8si) __Y, 2,
9266                                                   (__mmask8) -1);
9267 }
9268
9269 extern __inline __mmask8
9270   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9271 _mm256_mask_cmpneq_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9272 {
9273   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9274                                                   (__v4di) __Y, 4,
9275                                                   (__mmask8) __M);
9276 }
9277
9278 extern __inline __mmask8
9279   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9280 _mm256_cmpneq_epu64_mask (__m256i __X, __m256i __Y)
9281 {
9282   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9283                                                   (__v4di) __Y, 4,
9284                                                   (__mmask8) -1);
9285 }
9286
9287 extern __inline __mmask8
9288   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9289 _mm256_mask_cmplt_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9290 {
9291   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9292                                                   (__v4di) __Y, 1,
9293                                                   (__mmask8) __M);
9294 }
9295
9296 extern __inline __mmask8
9297   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9298 _mm256_cmplt_epu64_mask (__m256i __X, __m256i __Y)
9299 {
9300   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9301                                                   (__v4di) __Y, 1,
9302                                                   (__mmask8) -1);
9303 }
9304
9305 extern __inline __mmask8
9306   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9307 _mm256_mask_cmpge_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9308 {
9309   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9310                                                   (__v4di) __Y, 5,
9311                                                   (__mmask8) __M);
9312 }
9313
9314 extern __inline __mmask8
9315   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9316 _mm256_cmpge_epu64_mask (__m256i __X, __m256i __Y)
9317 {
9318   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9319                                                   (__v4di) __Y, 5,
9320                                                   (__mmask8) -1);
9321 }
9322
9323 extern __inline __mmask8
9324   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9325 _mm256_mask_cmple_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9326 {
9327   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9328                                                   (__v4di) __Y, 2,
9329                                                   (__mmask8) __M);
9330 }
9331
9332 extern __inline __mmask8
9333   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9334 _mm256_cmple_epu64_mask (__m256i __X, __m256i __Y)
9335 {
9336   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
9337                                                   (__v4di) __Y, 2,
9338                                                   (__mmask8) -1);
9339 }
9340
9341 extern __inline __mmask8
9342   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9343 _mm256_mask_cmpneq_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9344 {
9345   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9346                                                  (__v8si) __Y, 4,
9347                                                  (__mmask8) __M);
9348 }
9349
9350 extern __inline __mmask8
9351   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9352 _mm256_cmpneq_epi32_mask (__m256i __X, __m256i __Y)
9353 {
9354   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9355                                                  (__v8si) __Y, 4,
9356                                                  (__mmask8) -1);
9357 }
9358
9359 extern __inline __mmask8
9360   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9361 _mm256_mask_cmplt_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9362 {
9363   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9364                                                  (__v8si) __Y, 1,
9365                                                  (__mmask8) __M);
9366 }
9367
9368 extern __inline __mmask8
9369   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9370 _mm256_cmplt_epi32_mask (__m256i __X, __m256i __Y)
9371 {
9372   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9373                                                  (__v8si) __Y, 1,
9374                                                  (__mmask8) -1);
9375 }
9376
9377 extern __inline __mmask8
9378   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9379 _mm256_mask_cmpge_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9380 {
9381   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9382                                                  (__v8si) __Y, 5,
9383                                                  (__mmask8) __M);
9384 }
9385
9386 extern __inline __mmask8
9387   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9388 _mm256_cmpge_epi32_mask (__m256i __X, __m256i __Y)
9389 {
9390   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9391                                                  (__v8si) __Y, 5,
9392                                                  (__mmask8) -1);
9393 }
9394
9395 extern __inline __mmask8
9396   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9397 _mm256_mask_cmple_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9398 {
9399   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9400                                                  (__v8si) __Y, 2,
9401                                                  (__mmask8) __M);
9402 }
9403
9404 extern __inline __mmask8
9405   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9406 _mm256_cmple_epi32_mask (__m256i __X, __m256i __Y)
9407 {
9408   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
9409                                                  (__v8si) __Y, 2,
9410                                                  (__mmask8) -1);
9411 }
9412
9413 extern __inline __mmask8
9414   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9415 _mm256_mask_cmpneq_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9416 {
9417   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9418                                                  (__v4di) __Y, 4,
9419                                                  (__mmask8) __M);
9420 }
9421
9422 extern __inline __mmask8
9423   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9424 _mm256_cmpneq_epi64_mask (__m256i __X, __m256i __Y)
9425 {
9426   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9427                                                  (__v4di) __Y, 4,
9428                                                  (__mmask8) -1);
9429 }
9430
9431 extern __inline __mmask8
9432   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9433 _mm256_mask_cmplt_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9434 {
9435   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9436                                                  (__v4di) __Y, 1,
9437                                                  (__mmask8) __M);
9438 }
9439
9440 extern __inline __mmask8
9441   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9442 _mm256_cmplt_epi64_mask (__m256i __X, __m256i __Y)
9443 {
9444   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9445                                                  (__v4di) __Y, 1,
9446                                                  (__mmask8) -1);
9447 }
9448
9449 extern __inline __mmask8
9450   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9451 _mm256_mask_cmpge_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9452 {
9453   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9454                                                  (__v4di) __Y, 5,
9455                                                  (__mmask8) __M);
9456 }
9457
9458 extern __inline __mmask8
9459   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9460 _mm256_cmpge_epi64_mask (__m256i __X, __m256i __Y)
9461 {
9462   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9463                                                  (__v4di) __Y, 5,
9464                                                  (__mmask8) -1);
9465 }
9466
9467 extern __inline __mmask8
9468   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9469 _mm256_mask_cmple_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y)
9470 {
9471   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9472                                                  (__v4di) __Y, 2,
9473                                                  (__mmask8) __M);
9474 }
9475
9476 extern __inline __mmask8
9477   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9478 _mm256_cmple_epi64_mask (__m256i __X, __m256i __Y)
9479 {
9480   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
9481                                                  (__v4di) __Y, 2,
9482                                                  (__mmask8) -1);
9483 }
9484
9485 extern __inline __mmask8
9486   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9487 _mm_mask_cmpneq_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9488 {
9489   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9490                                                   (__v4si) __Y, 4,
9491                                                   (__mmask8) __M);
9492 }
9493
9494 extern __inline __mmask8
9495   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9496 _mm_cmpneq_epu32_mask (__m128i __X, __m128i __Y)
9497 {
9498   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9499                                                   (__v4si) __Y, 4,
9500                                                   (__mmask8) -1);
9501 }
9502
9503 extern __inline __mmask8
9504   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9505 _mm_mask_cmplt_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9506 {
9507   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9508                                                   (__v4si) __Y, 1,
9509                                                   (__mmask8) __M);
9510 }
9511
9512 extern __inline __mmask8
9513   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9514 _mm_cmplt_epu32_mask (__m128i __X, __m128i __Y)
9515 {
9516   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9517                                                   (__v4si) __Y, 1,
9518                                                   (__mmask8) -1);
9519 }
9520
9521 extern __inline __mmask8
9522   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9523 _mm_mask_cmpge_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9524 {
9525   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9526                                                   (__v4si) __Y, 5,
9527                                                   (__mmask8) __M);
9528 }
9529
9530 extern __inline __mmask8
9531   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9532 _mm_cmpge_epu32_mask (__m128i __X, __m128i __Y)
9533 {
9534   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9535                                                   (__v4si) __Y, 5,
9536                                                   (__mmask8) -1);
9537 }
9538
9539 extern __inline __mmask8
9540   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9541 _mm_mask_cmple_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9542 {
9543   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9544                                                   (__v4si) __Y, 2,
9545                                                   (__mmask8) __M);
9546 }
9547
9548 extern __inline __mmask8
9549   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9550 _mm_cmple_epu32_mask (__m128i __X, __m128i __Y)
9551 {
9552   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
9553                                                   (__v4si) __Y, 2,
9554                                                   (__mmask8) -1);
9555 }
9556
9557 extern __inline __mmask8
9558   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9559 _mm_mask_cmpneq_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9560 {
9561   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9562                                                   (__v2di) __Y, 4,
9563                                                   (__mmask8) __M);
9564 }
9565
9566 extern __inline __mmask8
9567   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9568 _mm_cmpneq_epu64_mask (__m128i __X, __m128i __Y)
9569 {
9570   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9571                                                   (__v2di) __Y, 4,
9572                                                   (__mmask8) -1);
9573 }
9574
9575 extern __inline __mmask8
9576   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9577 _mm_mask_cmplt_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9578 {
9579   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9580                                                   (__v2di) __Y, 1,
9581                                                   (__mmask8) __M);
9582 }
9583
9584 extern __inline __mmask8
9585   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9586 _mm_cmplt_epu64_mask (__m128i __X, __m128i __Y)
9587 {
9588   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9589                                                   (__v2di) __Y, 1,
9590                                                   (__mmask8) -1);
9591 }
9592
9593 extern __inline __mmask8
9594   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9595 _mm_mask_cmpge_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9596 {
9597   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9598                                                   (__v2di) __Y, 5,
9599                                                   (__mmask8) __M);
9600 }
9601
9602 extern __inline __mmask8
9603   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9604 _mm_cmpge_epu64_mask (__m128i __X, __m128i __Y)
9605 {
9606   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9607                                                   (__v2di) __Y, 5,
9608                                                   (__mmask8) -1);
9609 }
9610
9611 extern __inline __mmask8
9612   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9613 _mm_mask_cmple_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9614 {
9615   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9616                                                   (__v2di) __Y, 2,
9617                                                   (__mmask8) __M);
9618 }
9619
9620 extern __inline __mmask8
9621   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9622 _mm_cmple_epu64_mask (__m128i __X, __m128i __Y)
9623 {
9624   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
9625                                                   (__v2di) __Y, 2,
9626                                                   (__mmask8) -1);
9627 }
9628
9629 extern __inline __mmask8
9630   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9631 _mm_mask_cmpneq_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9632 {
9633   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9634                                                  (__v4si) __Y, 4,
9635                                                  (__mmask8) __M);
9636 }
9637
9638 extern __inline __mmask8
9639   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9640 _mm_cmpneq_epi32_mask (__m128i __X, __m128i __Y)
9641 {
9642   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9643                                                  (__v4si) __Y, 4,
9644                                                  (__mmask8) -1);
9645 }
9646
9647 extern __inline __mmask8
9648   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9649 _mm_mask_cmplt_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9650 {
9651   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9652                                                  (__v4si) __Y, 1,
9653                                                  (__mmask8) __M);
9654 }
9655
9656 extern __inline __mmask8
9657   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9658 _mm_cmplt_epi32_mask (__m128i __X, __m128i __Y)
9659 {
9660   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9661                                                  (__v4si) __Y, 1,
9662                                                  (__mmask8) -1);
9663 }
9664
9665 extern __inline __mmask8
9666   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9667 _mm_mask_cmpge_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9668 {
9669   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9670                                                  (__v4si) __Y, 5,
9671                                                  (__mmask8) __M);
9672 }
9673
9674 extern __inline __mmask8
9675   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9676 _mm_cmpge_epi32_mask (__m128i __X, __m128i __Y)
9677 {
9678   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9679                                                  (__v4si) __Y, 5,
9680                                                  (__mmask8) -1);
9681 }
9682
9683 extern __inline __mmask8
9684   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9685 _mm_mask_cmple_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9686 {
9687   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9688                                                  (__v4si) __Y, 2,
9689                                                  (__mmask8) __M);
9690 }
9691
9692 extern __inline __mmask8
9693   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9694 _mm_cmple_epi32_mask (__m128i __X, __m128i __Y)
9695 {
9696   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
9697                                                  (__v4si) __Y, 2,
9698                                                  (__mmask8) -1);
9699 }
9700
9701 extern __inline __mmask8
9702   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9703 _mm_mask_cmpneq_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9704 {
9705   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9706                                                  (__v2di) __Y, 4,
9707                                                  (__mmask8) __M);
9708 }
9709
9710 extern __inline __mmask8
9711   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9712 _mm_cmpneq_epi64_mask (__m128i __X, __m128i __Y)
9713 {
9714   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9715                                                  (__v2di) __Y, 4,
9716                                                  (__mmask8) -1);
9717 }
9718
9719 extern __inline __mmask8
9720   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9721 _mm_mask_cmplt_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9722 {
9723   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9724                                                  (__v2di) __Y, 1,
9725                                                  (__mmask8) __M);
9726 }
9727
9728 extern __inline __mmask8
9729   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9730 _mm_cmplt_epi64_mask (__m128i __X, __m128i __Y)
9731 {
9732   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9733                                                  (__v2di) __Y, 1,
9734                                                  (__mmask8) -1);
9735 }
9736
9737 extern __inline __mmask8
9738   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9739 _mm_mask_cmpge_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9740 {
9741   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9742                                                  (__v2di) __Y, 5,
9743                                                  (__mmask8) __M);
9744 }
9745
9746 extern __inline __mmask8
9747   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9748 _mm_cmpge_epi64_mask (__m128i __X, __m128i __Y)
9749 {
9750   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9751                                                  (__v2di) __Y, 5,
9752                                                  (__mmask8) -1);
9753 }
9754
9755 extern __inline __mmask8
9756   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9757 _mm_mask_cmple_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y)
9758 {
9759   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9760                                                  (__v2di) __Y, 2,
9761                                                  (__mmask8) __M);
9762 }
9763
9764 extern __inline __mmask8
9765   __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9766 _mm_cmple_epi64_mask (__m128i __X, __m128i __Y)
9767 {
9768   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
9769                                                  (__v2di) __Y, 2,
9770                                                  (__mmask8) -1);
9771 }
9772
9773 #ifdef __OPTIMIZE__
9774 extern __inline __m256i
9775 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9776 _mm256_permutex_epi64 (__m256i __X, const int __I)
9777 {
9778   return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X,
9779                                               __I,
9780                                               (__v4di)
9781                                               _mm256_setzero_si256(),
9782                                               (__mmask8) -1);
9783 }
9784
9785 extern __inline __m256i
9786 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9787 _mm256_mask_permutex_epi64 (__m256i __W, __mmask8 __M,
9788                             __m256i __X, const int __I)
9789 {
9790   return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X,
9791                                                   __I,
9792                                                   (__v4di) __W,
9793                                                   (__mmask8) __M);
9794 }
9795
9796 extern __inline __m256i
9797 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9798 _mm256_maskz_permutex_epi64 (__mmask8 __M, __m256i __X, const int __I)
9799 {
9800   return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X,
9801                                                   __I,
9802                                                   (__v4di)
9803                                                   _mm256_setzero_si256 (),
9804                                                   (__mmask8) __M);
9805 }
9806
9807 extern __inline __m256d
9808 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9809 _mm256_mask_shuffle_pd (__m256d __W, __mmask8 __U, __m256d __A,
9810                         __m256d __B, const int __imm)
9811 {
9812   return (__m256d) __builtin_ia32_shufpd256_mask ((__v4df) __A,
9813                                                   (__v4df) __B, __imm,
9814                                                   (__v4df) __W,
9815                                                   (__mmask8) __U);
9816 }
9817
9818 extern __inline __m256d
9819 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9820 _mm256_maskz_shuffle_pd (__mmask8 __U, __m256d __A, __m256d __B,
9821                          const int __imm)
9822 {
9823   return (__m256d) __builtin_ia32_shufpd256_mask ((__v4df) __A,
9824                                                   (__v4df) __B, __imm,
9825                                                   (__v4df)
9826                                                   _mm256_setzero_pd (),
9827                                                   (__mmask8) __U);
9828 }
9829
9830 extern __inline __m128d
9831 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9832 _mm_mask_shuffle_pd (__m128d __W, __mmask8 __U, __m128d __A,
9833                      __m128d __B, const int __imm)
9834 {
9835   return (__m128d) __builtin_ia32_shufpd128_mask ((__v2df) __A,
9836                                                   (__v2df) __B, __imm,
9837                                                   (__v2df) __W,
9838                                                   (__mmask8) __U);
9839 }
9840
9841 extern __inline __m128d
9842 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9843 _mm_maskz_shuffle_pd (__mmask8 __U, __m128d __A, __m128d __B,
9844                       const int __imm)
9845 {
9846   return (__m128d) __builtin_ia32_shufpd128_mask ((__v2df) __A,
9847                                                   (__v2df) __B, __imm,
9848                                                   (__v2df)
9849                                                   _mm_setzero_pd (),
9850                                                   (__mmask8) __U);
9851 }
9852
9853 extern __inline __m256
9854 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9855 _mm256_mask_shuffle_ps (__m256 __W, __mmask8 __U, __m256 __A,
9856                         __m256 __B, const int __imm)
9857 {
9858   return (__m256) __builtin_ia32_shufps256_mask ((__v8sf) __A,
9859                                                  (__v8sf) __B, __imm,
9860                                                  (__v8sf) __W,
9861                                                  (__mmask8) __U);
9862 }
9863
9864 extern __inline __m256
9865 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9866 _mm256_maskz_shuffle_ps (__mmask8 __U, __m256 __A, __m256 __B,
9867                          const int __imm)
9868 {
9869   return (__m256) __builtin_ia32_shufps256_mask ((__v8sf) __A,
9870                                                  (__v8sf) __B, __imm,
9871                                                  (__v8sf)
9872                                                  _mm256_setzero_ps (),
9873                                                  (__mmask8) __U);
9874 }
9875
9876 extern __inline __m128
9877 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9878 _mm_mask_shuffle_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B,
9879                      const int __imm)
9880 {
9881   return (__m128) __builtin_ia32_shufps128_mask ((__v4sf) __A,
9882                                                  (__v4sf) __B, __imm,
9883                                                  (__v4sf) __W,
9884                                                  (__mmask8) __U);
9885 }
9886
9887 extern __inline __m128
9888 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9889 _mm_maskz_shuffle_ps (__mmask8 __U, __m128 __A, __m128 __B,
9890                       const int __imm)
9891 {
9892   return (__m128) __builtin_ia32_shufps128_mask ((__v4sf) __A,
9893                                                  (__v4sf) __B, __imm,
9894                                                  (__v4sf)
9895                                                  _mm_setzero_ps (),
9896                                                  (__mmask8) __U);
9897 }
9898
9899 extern __inline __m256i
9900 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9901 _mm256_inserti32x4 (__m256i __A, __m128i __B, const int __imm)
9902 {
9903   return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A,
9904                                                         (__v4si) __B,
9905                                                         __imm,
9906                                                         (__v8si)
9907                                                         _mm256_setzero_si256 (),
9908                                                         (__mmask8) -1);
9909 }
9910
9911 extern __inline __m256i
9912 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9913 _mm256_mask_inserti32x4 (__m256i __W, __mmask8 __U, __m256i __A,
9914                          __m128i __B, const int __imm)
9915 {
9916   return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A,
9917                                                         (__v4si) __B,
9918                                                         __imm,
9919                                                         (__v8si) __W,
9920                                                         (__mmask8)
9921                                                         __U);
9922 }
9923
9924 extern __inline __m256i
9925 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9926 _mm256_maskz_inserti32x4 (__mmask8 __U, __m256i __A, __m128i __B,
9927                           const int __imm)
9928 {
9929   return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A,
9930                                                         (__v4si) __B,
9931                                                         __imm,
9932                                                         (__v8si)
9933                                                         _mm256_setzero_si256 (),
9934                                                         (__mmask8)
9935                                                         __U);
9936 }
9937
9938 extern __inline __m256
9939 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9940 _mm256_insertf32x4 (__m256 __A, __m128 __B, const int __imm)
9941 {
9942   return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A,
9943                                                        (__v4sf) __B,
9944                                                        __imm,
9945                                                        (__v8sf)
9946                                                        _mm256_setzero_ps (),
9947                                                        (__mmask8) -1);
9948 }
9949
9950 extern __inline __m256
9951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9952 _mm256_mask_insertf32x4 (__m256 __W, __mmask8 __U, __m256 __A,
9953                          __m128 __B, const int __imm)
9954 {
9955   return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A,
9956                                                        (__v4sf) __B,
9957                                                        __imm,
9958                                                        (__v8sf) __W,
9959                                                        (__mmask8) __U);
9960 }
9961
9962 extern __inline __m256
9963 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9964 _mm256_maskz_insertf32x4 (__mmask8 __U, __m256 __A, __m128 __B,
9965                           const int __imm)
9966 {
9967   return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A,
9968                                                        (__v4sf) __B,
9969                                                        __imm,
9970                                                        (__v8sf)
9971                                                        _mm256_setzero_ps (),
9972                                                        (__mmask8) __U);
9973 }
9974
9975 extern __inline __m128i
9976 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9977 _mm256_extracti32x4_epi32 (__m256i __A, const int __imm)
9978 {
9979   return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A,
9980                                                          __imm,
9981                                                          (__v4si)
9982                                                          _mm_setzero_si128 (),
9983                                                          (__mmask8) -1);
9984 }
9985
9986 extern __inline __m128i
9987 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
9988 _mm256_mask_extracti32x4_epi32 (__m128i __W, __mmask8 __U, __m256i __A,
9989                                 const int __imm)
9990 {
9991   return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A,
9992                                                          __imm,
9993                                                          (__v4si) __W,
9994                                                          (__mmask8)
9995                                                          __U);
9996 }
9997
9998 extern __inline __m128i
9999 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10000 _mm256_maskz_extracti32x4_epi32 (__mmask8 __U, __m256i __A,
10001                                  const int __imm)
10002 {
10003   return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A,
10004                                                          __imm,
10005                                                          (__v4si)
10006                                                          _mm_setzero_si128 (),
10007                                                          (__mmask8)
10008                                                          __U);
10009 }
10010
10011 extern __inline __m128
10012 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10013 _mm256_extractf32x4_ps (__m256 __A, const int __imm)
10014 {
10015   return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A,
10016                                                         __imm,
10017                                                         (__v4sf)
10018                                                         _mm_setzero_ps (),
10019                                                         (__mmask8) -1);
10020 }
10021
10022 extern __inline __m128
10023 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10024 _mm256_mask_extractf32x4_ps (__m128 __W, __mmask8 __U, __m256 __A,
10025                              const int __imm)
10026 {
10027   return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A,
10028                                                         __imm,
10029                                                         (__v4sf) __W,
10030                                                         (__mmask8)
10031                                                         __U);
10032 }
10033
10034 extern __inline __m128
10035 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10036 _mm256_maskz_extractf32x4_ps (__mmask8 __U, __m256 __A,
10037                               const int __imm)
10038 {
10039   return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A,
10040                                                         __imm,
10041                                                         (__v4sf)
10042                                                         _mm_setzero_ps (),
10043                                                         (__mmask8)
10044                                                         __U);
10045 }
10046
10047 extern __inline __m256i
10048 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10049 _mm256_shuffle_i64x2 (__m256i __A, __m256i __B, const int __imm)
10050 {
10051   return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A,
10052                                                        (__v4di) __B,
10053                                                        __imm,
10054                                                        (__v4di)
10055                                                        _mm256_setzero_si256 (),
10056                                                        (__mmask8) -1);
10057 }
10058
10059 extern __inline __m256i
10060 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10061 _mm256_mask_shuffle_i64x2 (__m256i __W, __mmask8 __U, __m256i __A,
10062                            __m256i __B, const int __imm)
10063 {
10064   return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A,
10065                                                        (__v4di) __B,
10066                                                        __imm,
10067                                                        (__v4di) __W,
10068                                                        (__mmask8) __U);
10069 }
10070
10071 extern __inline __m256i
10072 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10073 _mm256_maskz_shuffle_i64x2 (__mmask8 __U, __m256i __A, __m256i __B,
10074                             const int __imm)
10075 {
10076   return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A,
10077                                                        (__v4di) __B,
10078                                                        __imm,
10079                                                        (__v4di)
10080                                                        _mm256_setzero_si256 (),
10081                                                        (__mmask8) __U);
10082 }
10083
10084 extern __inline __m256i
10085 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10086 _mm256_shuffle_i32x4 (__m256i __A, __m256i __B, const int __imm)
10087 {
10088   return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A,
10089                                                        (__v8si) __B,
10090                                                        __imm,
10091                                                        (__v8si)
10092                                                        _mm256_setzero_si256 (),
10093                                                        (__mmask8) -1);
10094 }
10095
10096 extern __inline __m256i
10097 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10098 _mm256_mask_shuffle_i32x4 (__m256i __W, __mmask8 __U, __m256i __A,
10099                            __m256i __B, const int __imm)
10100 {
10101   return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A,
10102                                                        (__v8si) __B,
10103                                                        __imm,
10104                                                        (__v8si) __W,
10105                                                        (__mmask8) __U);
10106 }
10107
10108 extern __inline __m256i
10109 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10110 _mm256_maskz_shuffle_i32x4 (__mmask8 __U, __m256i __A, __m256i __B,
10111                             const int __imm)
10112 {
10113   return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A,
10114                                                        (__v8si) __B,
10115                                                        __imm,
10116                                                        (__v8si)
10117                                                        _mm256_setzero_si256 (),
10118                                                        (__mmask8) __U);
10119 }
10120
10121 extern __inline __m256d
10122 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10123 _mm256_shuffle_f64x2 (__m256d __A, __m256d __B, const int __imm)
10124 {
10125   return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A,
10126                                                        (__v4df) __B,
10127                                                        __imm,
10128                                                        (__v4df)
10129                                                        _mm256_setzero_pd (),
10130                                                        (__mmask8) -1);
10131 }
10132
10133 extern __inline __m256d
10134 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10135 _mm256_mask_shuffle_f64x2 (__m256d __W, __mmask8 __U, __m256d __A,
10136                            __m256d __B, const int __imm)
10137 {
10138   return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A,
10139                                                        (__v4df) __B,
10140                                                        __imm,
10141                                                        (__v4df) __W,
10142                                                        (__mmask8) __U);
10143 }
10144
10145 extern __inline __m256d
10146 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10147 _mm256_maskz_shuffle_f64x2 (__mmask8 __U, __m256d __A, __m256d __B,
10148                             const int __imm)
10149 {
10150   return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A,
10151                                                        (__v4df) __B,
10152                                                        __imm,
10153                                                        (__v4df)
10154                                                        _mm256_setzero_pd (),
10155                                                        (__mmask8) __U);
10156 }
10157
10158 extern __inline __m256
10159 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10160 _mm256_shuffle_f32x4 (__m256 __A, __m256 __B, const int __imm)
10161 {
10162   return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A,
10163                                                       (__v8sf) __B,
10164                                                       __imm,
10165                                                       (__v8sf)
10166                                                       _mm256_setzero_ps (),
10167                                                       (__mmask8) -1);
10168 }
10169
10170 extern __inline __m256
10171 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10172 _mm256_mask_shuffle_f32x4 (__m256 __W, __mmask8 __U, __m256 __A,
10173                            __m256 __B, const int __imm)
10174 {
10175   return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A,
10176                                                       (__v8sf) __B,
10177                                                       __imm,
10178                                                       (__v8sf) __W,
10179                                                       (__mmask8) __U);
10180 }
10181
10182 extern __inline __m256
10183 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10184 _mm256_maskz_shuffle_f32x4 (__mmask8 __U, __m256 __A, __m256 __B,
10185                             const int __imm)
10186 {
10187   return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A,
10188                                                       (__v8sf) __B,
10189                                                       __imm,
10190                                                       (__v8sf)
10191                                                       _mm256_setzero_ps (),
10192                                                       (__mmask8) __U);
10193 }
10194
10195 extern __inline __m256d
10196 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10197 _mm256_fixupimm_pd (__m256d __A, __m256d __B, __m256i __C,
10198                     const int __imm)
10199 {
10200   return (__m256d) __builtin_ia32_fixupimmpd256_mask ((__v4df) __A,
10201                                                       (__v4df) __B,
10202                                                       (__v4di) __C,
10203                                                       __imm,
10204                                                       (__mmask8) -1);
10205 }
10206
10207 extern __inline __m256d
10208 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10209 _mm256_mask_fixupimm_pd (__m256d __A, __mmask8 __U, __m256d __B,
10210                          __m256i __C, const int __imm)
10211 {
10212   return (__m256d) __builtin_ia32_fixupimmpd256_mask ((__v4df) __A,
10213                                                       (__v4df) __B,
10214                                                       (__v4di) __C,
10215                                                       __imm,
10216                                                       (__mmask8) __U);
10217 }
10218
10219 extern __inline __m256d
10220 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10221 _mm256_maskz_fixupimm_pd (__mmask8 __U, __m256d __A, __m256d __B,
10222                           __m256i __C, const int __imm)
10223 {
10224   return (__m256d) __builtin_ia32_fixupimmpd256_maskz ((__v4df) __A,
10225                                                        (__v4df) __B,
10226                                                        (__v4di) __C,
10227                                                        __imm,
10228                                                        (__mmask8) __U);
10229 }
10230
10231 extern __inline __m256
10232 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10233 _mm256_fixupimm_ps (__m256 __A, __m256 __B, __m256i __C,
10234                     const int __imm)
10235 {
10236   return (__m256) __builtin_ia32_fixupimmps256_mask ((__v8sf) __A,
10237                                                      (__v8sf) __B,
10238                                                      (__v8si) __C,
10239                                                      __imm,
10240                                                      (__mmask8) -1);
10241 }
10242
10243 extern __inline __m256
10244 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10245 _mm256_mask_fixupimm_ps (__m256 __A, __mmask8 __U, __m256 __B,
10246                          __m256i __C, const int __imm)
10247 {
10248   return (__m256) __builtin_ia32_fixupimmps256_mask ((__v8sf) __A,
10249                                                      (__v8sf) __B,
10250                                                      (__v8si) __C,
10251                                                      __imm,
10252                                                      (__mmask8) __U);
10253 }
10254
10255 extern __inline __m256
10256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10257 _mm256_maskz_fixupimm_ps (__mmask8 __U, __m256 __A, __m256 __B,
10258                           __m256i __C, const int __imm)
10259 {
10260   return (__m256) __builtin_ia32_fixupimmps256_maskz ((__v8sf) __A,
10261                                                       (__v8sf) __B,
10262                                                       (__v8si) __C,
10263                                                       __imm,
10264                                                       (__mmask8) __U);
10265 }
10266
10267 extern __inline __m128d
10268 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10269 _mm_fixupimm_pd (__m128d __A, __m128d __B, __m128i __C,
10270                  const int __imm)
10271 {
10272   return (__m128d) __builtin_ia32_fixupimmpd128_mask ((__v2df) __A,
10273                                                       (__v2df) __B,
10274                                                       (__v2di) __C,
10275                                                       __imm,
10276                                                       (__mmask8) -1);
10277 }
10278
10279 extern __inline __m128d
10280 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10281 _mm_mask_fixupimm_pd (__m128d __A, __mmask8 __U, __m128d __B,
10282                       __m128i __C, const int __imm)
10283 {
10284   return (__m128d) __builtin_ia32_fixupimmpd128_mask ((__v2df) __A,
10285                                                       (__v2df) __B,
10286                                                       (__v2di) __C,
10287                                                       __imm,
10288                                                       (__mmask8) __U);
10289 }
10290
10291 extern __inline __m128d
10292 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10293 _mm_maskz_fixupimm_pd (__mmask8 __U, __m128d __A, __m128d __B,
10294                        __m128i __C, const int __imm)
10295 {
10296   return (__m128d) __builtin_ia32_fixupimmpd128_maskz ((__v2df) __A,
10297                                                        (__v2df) __B,
10298                                                        (__v2di) __C,
10299                                                        __imm,
10300                                                        (__mmask8) __U);
10301 }
10302
10303 extern __inline __m128
10304 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10305 _mm_fixupimm_ps (__m128 __A, __m128 __B, __m128i __C, const int __imm)
10306 {
10307   return (__m128) __builtin_ia32_fixupimmps128_mask ((__v4sf) __A,
10308                                                      (__v4sf) __B,
10309                                                      (__v4si) __C,
10310                                                      __imm,
10311                                                      (__mmask8) -1);
10312 }
10313
10314 extern __inline __m128
10315 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10316 _mm_mask_fixupimm_ps (__m128 __A, __mmask8 __U, __m128 __B,
10317                       __m128i __C, const int __imm)
10318 {
10319   return (__m128) __builtin_ia32_fixupimmps128_mask ((__v4sf) __A,
10320                                                      (__v4sf) __B,
10321                                                      (__v4si) __C,
10322                                                      __imm,
10323                                                      (__mmask8) __U);
10324 }
10325
10326 extern __inline __m128
10327 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10328 _mm_maskz_fixupimm_ps (__mmask8 __U, __m128 __A, __m128 __B,
10329                        __m128i __C, const int __imm)
10330 {
10331   return (__m128) __builtin_ia32_fixupimmps128_maskz ((__v4sf) __A,
10332                                                       (__v4sf) __B,
10333                                                       (__v4si) __C,
10334                                                       __imm,
10335                                                       (__mmask8) __U);
10336 }
10337
10338 extern __inline __m256i
10339 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10340 _mm256_mask_srli_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
10341                         const int __imm)
10342 {
10343   return (__m256i) __builtin_ia32_psrldi256_mask ((__v8si) __A, __imm,
10344                                                   (__v8si) __W,
10345                                                   (__mmask8) __U);
10346 }
10347
10348 extern __inline __m256i
10349 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10350 _mm256_maskz_srli_epi32 (__mmask8 __U, __m256i __A, const int __imm)
10351 {
10352   return (__m256i) __builtin_ia32_psrldi256_mask ((__v8si) __A, __imm,
10353                                                   (__v8si)
10354                                                   _mm256_setzero_si256 (),
10355                                                   (__mmask8) __U);
10356 }
10357
10358 extern __inline __m128i
10359 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10360 _mm_mask_srli_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
10361                      const int __imm)
10362 {
10363   return (__m128i) __builtin_ia32_psrldi128_mask ((__v4si) __A, __imm,
10364                                                   (__v4si) __W,
10365                                                   (__mmask8) __U);
10366 }
10367
10368 extern __inline __m128i
10369 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10370 _mm_maskz_srli_epi32 (__mmask8 __U, __m128i __A, const int __imm)
10371 {
10372   return (__m128i) __builtin_ia32_psrldi128_mask ((__v4si) __A, __imm,
10373                                                   (__v4si)
10374                                                   _mm_setzero_si128 (),
10375                                                   (__mmask8) __U);
10376 }
10377
10378 extern __inline __m256i
10379 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10380 _mm256_mask_srli_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
10381                         const int __imm)
10382 {
10383   return (__m256i) __builtin_ia32_psrlqi256_mask ((__v4di) __A, __imm,
10384                                                   (__v4di) __W,
10385                                                   (__mmask8) __U);
10386 }
10387
10388 extern __inline __m256i
10389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10390 _mm256_maskz_srli_epi64 (__mmask8 __U, __m256i __A, const int __imm)
10391 {
10392   return (__m256i) __builtin_ia32_psrlqi256_mask ((__v4di) __A, __imm,
10393                                                   (__v4di)
10394                                                   _mm256_setzero_si256 (),
10395                                                   (__mmask8) __U);
10396 }
10397
10398 extern __inline __m128i
10399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10400 _mm_mask_srli_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
10401                      const int __imm)
10402 {
10403   return (__m128i) __builtin_ia32_psrlqi128_mask ((__v2di) __A, __imm,
10404                                                   (__v2di) __W,
10405                                                   (__mmask8) __U);
10406 }
10407
10408 extern __inline __m128i
10409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10410 _mm_maskz_srli_epi64 (__mmask8 __U, __m128i __A, const int __imm)
10411 {
10412   return (__m128i) __builtin_ia32_psrlqi128_mask ((__v2di) __A, __imm,
10413                                                   (__v2di)
10414                                                   _mm_setzero_si128 (),
10415                                                   (__mmask8) __U);
10416 }
10417
10418 extern __inline __m256i
10419 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10420 _mm256_ternarylogic_epi64 (__m256i __A, __m256i __B, __m256i __C,
10421                            const int __imm)
10422 {
10423   return (__m256i) __builtin_ia32_pternlogq256_mask ((__v4di) __A,
10424                                                      (__v4di) __B,
10425                                                      (__v4di) __C, __imm,
10426                                                      (__mmask8) -1);
10427 }
10428
10429 extern __inline __m256i
10430 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10431 _mm256_mask_ternarylogic_epi64 (__m256i __A, __mmask8 __U,
10432                                 __m256i __B, __m256i __C,
10433                                 const int __imm)
10434 {
10435   return (__m256i) __builtin_ia32_pternlogq256_mask ((__v4di) __A,
10436                                                      (__v4di) __B,
10437                                                      (__v4di) __C, __imm,
10438                                                      (__mmask8) __U);
10439 }
10440
10441 extern __inline __m256i
10442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10443 _mm256_maskz_ternarylogic_epi64 (__mmask8 __U, __m256i __A,
10444                                  __m256i __B, __m256i __C,
10445                                  const int __imm)
10446 {
10447   return (__m256i) __builtin_ia32_pternlogq256_maskz ((__v4di) __A,
10448                                                       (__v4di) __B,
10449                                                       (__v4di) __C,
10450                                                       __imm,
10451                                                       (__mmask8) __U);
10452 }
10453
10454 extern __inline __m256i
10455 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10456 _mm256_ternarylogic_epi32 (__m256i __A, __m256i __B, __m256i __C,
10457                            const int __imm)
10458 {
10459   return (__m256i) __builtin_ia32_pternlogd256_mask ((__v8si) __A,
10460                                                      (__v8si) __B,
10461                                                      (__v8si) __C, __imm,
10462                                                      (__mmask8) -1);
10463 }
10464
10465 extern __inline __m256i
10466 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10467 _mm256_mask_ternarylogic_epi32 (__m256i __A, __mmask8 __U,
10468                                 __m256i __B, __m256i __C,
10469                                 const int __imm)
10470 {
10471   return (__m256i) __builtin_ia32_pternlogd256_mask ((__v8si) __A,
10472                                                      (__v8si) __B,
10473                                                      (__v8si) __C, __imm,
10474                                                      (__mmask8) __U);
10475 }
10476
10477 extern __inline __m256i
10478 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10479 _mm256_maskz_ternarylogic_epi32 (__mmask8 __U, __m256i __A,
10480                                  __m256i __B, __m256i __C,
10481                                  const int __imm)
10482 {
10483   return (__m256i) __builtin_ia32_pternlogd256_maskz ((__v8si) __A,
10484                                                       (__v8si) __B,
10485                                                       (__v8si) __C,
10486                                                       __imm,
10487                                                       (__mmask8) __U);
10488 }
10489
10490 extern __inline __m128i
10491 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10492 _mm_ternarylogic_epi64 (__m128i __A, __m128i __B, __m128i __C,
10493                         const int __imm)
10494 {
10495   return (__m128i) __builtin_ia32_pternlogq128_mask ((__v2di) __A,
10496                                                      (__v2di) __B,
10497                                                      (__v2di) __C, __imm,
10498                                                      (__mmask8) -1);
10499 }
10500
10501 extern __inline __m128i
10502 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10503 _mm_mask_ternarylogic_epi64 (__m128i __A, __mmask8 __U,
10504                              __m128i __B, __m128i __C, const int __imm)
10505 {
10506   return (__m128i) __builtin_ia32_pternlogq128_mask ((__v2di) __A,
10507                                                      (__v2di) __B,
10508                                                      (__v2di) __C, __imm,
10509                                                      (__mmask8) __U);
10510 }
10511
10512 extern __inline __m128i
10513 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10514 _mm_maskz_ternarylogic_epi64 (__mmask8 __U, __m128i __A,
10515                               __m128i __B, __m128i __C, const int __imm)
10516 {
10517   return (__m128i) __builtin_ia32_pternlogq128_maskz ((__v2di) __A,
10518                                                       (__v2di) __B,
10519                                                       (__v2di) __C,
10520                                                       __imm,
10521                                                       (__mmask8) __U);
10522 }
10523
10524 extern __inline __m128i
10525 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10526 _mm_ternarylogic_epi32 (__m128i __A, __m128i __B, __m128i __C,
10527                         const int __imm)
10528 {
10529   return (__m128i) __builtin_ia32_pternlogd128_mask ((__v4si) __A,
10530                                                      (__v4si) __B,
10531                                                      (__v4si) __C, __imm,
10532                                                      (__mmask8) -1);
10533 }
10534
10535 extern __inline __m128i
10536 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10537 _mm_mask_ternarylogic_epi32 (__m128i __A, __mmask8 __U,
10538                              __m128i __B, __m128i __C, const int __imm)
10539 {
10540   return (__m128i) __builtin_ia32_pternlogd128_mask ((__v4si) __A,
10541                                                      (__v4si) __B,
10542                                                      (__v4si) __C, __imm,
10543                                                      (__mmask8) __U);
10544 }
10545
10546 extern __inline __m128i
10547 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10548 _mm_maskz_ternarylogic_epi32 (__mmask8 __U, __m128i __A,
10549                               __m128i __B, __m128i __C, const int __imm)
10550 {
10551   return (__m128i) __builtin_ia32_pternlogd128_maskz ((__v4si) __A,
10552                                                       (__v4si) __B,
10553                                                       (__v4si) __C,
10554                                                       __imm,
10555                                                       (__mmask8) __U);
10556 }
10557
10558 extern __inline __m256
10559 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10560 _mm256_roundscale_ps (__m256 __A, const int __imm)
10561 {
10562   return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,
10563                                                       __imm,
10564                                                       (__v8sf)
10565                                                       _mm256_setzero_ps (),
10566                                                       (__mmask8) -1);
10567 }
10568
10569 extern __inline __m256
10570 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10571 _mm256_mask_roundscale_ps (__m256 __W, __mmask8 __U, __m256 __A,
10572                            const int __imm)
10573 {
10574   return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,
10575                                                       __imm,
10576                                                       (__v8sf) __W,
10577                                                       (__mmask8) __U);
10578 }
10579
10580 extern __inline __m256
10581 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10582 _mm256_maskz_roundscale_ps (__mmask8 __U, __m256 __A, const int __imm)
10583 {
10584   return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A,
10585                                                       __imm,
10586                                                       (__v8sf)
10587                                                       _mm256_setzero_ps (),
10588                                                       (__mmask8) __U);
10589 }
10590
10591 extern __inline __m256d
10592 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10593 _mm256_roundscale_pd (__m256d __A, const int __imm)
10594 {
10595   return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A,
10596                                                        __imm,
10597                                                        (__v4df)
10598                                                        _mm256_setzero_pd (),
10599                                                        (__mmask8) -1);
10600 }
10601
10602 extern __inline __m256d
10603 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10604 _mm256_mask_roundscale_pd (__m256d __W, __mmask8 __U, __m256d __A,
10605                            const int __imm)
10606 {
10607   return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A,
10608                                                        __imm,
10609                                                        (__v4df) __W,
10610                                                        (__mmask8) __U);
10611 }
10612
10613 extern __inline __m256d
10614 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10615 _mm256_maskz_roundscale_pd (__mmask8 __U, __m256d __A, const int __imm)
10616 {
10617   return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A,
10618                                                        __imm,
10619                                                        (__v4df)
10620                                                        _mm256_setzero_pd (),
10621                                                        (__mmask8) __U);
10622 }
10623
10624 extern __inline __m128
10625 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10626 _mm_roundscale_ps (__m128 __A, const int __imm)
10627 {
10628   return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A,
10629                                                       __imm,
10630                                                       (__v4sf)
10631                                                       _mm_setzero_ps (),
10632                                                       (__mmask8) -1);
10633 }
10634
10635 extern __inline __m128
10636 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10637 _mm_mask_roundscale_ps (__m128 __W, __mmask8 __U, __m128 __A,
10638                         const int __imm)
10639 {
10640   return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A,
10641                                                       __imm,
10642                                                       (__v4sf) __W,
10643                                                       (__mmask8) __U);
10644 }
10645
10646 extern __inline __m128
10647 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10648 _mm_maskz_roundscale_ps (__mmask8 __U, __m128 __A, const int __imm)
10649 {
10650   return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A,
10651                                                       __imm,
10652                                                       (__v4sf)
10653                                                       _mm_setzero_ps (),
10654                                                       (__mmask8) __U);
10655 }
10656
10657 extern __inline __m128d
10658 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10659 _mm_roundscale_pd (__m128d __A, const int __imm)
10660 {
10661   return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A,
10662                                                        __imm,
10663                                                        (__v2df)
10664                                                        _mm_setzero_pd (),
10665                                                        (__mmask8) -1);
10666 }
10667
10668 extern __inline __m128d
10669 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10670 _mm_mask_roundscale_pd (__m128d __W, __mmask8 __U, __m128d __A,
10671                         const int __imm)
10672 {
10673   return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A,
10674                                                        __imm,
10675                                                        (__v2df) __W,
10676                                                        (__mmask8) __U);
10677 }
10678
10679 extern __inline __m128d
10680 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10681 _mm_maskz_roundscale_pd (__mmask8 __U, __m128d __A, const int __imm)
10682 {
10683   return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A,
10684                                                        __imm,
10685                                                        (__v2df)
10686                                                        _mm_setzero_pd (),
10687                                                        (__mmask8) __U);
10688 }
10689
10690 extern __inline __m256
10691 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10692 _mm256_getmant_ps (__m256 __A, _MM_MANTISSA_NORM_ENUM __B,
10693                    _MM_MANTISSA_SIGN_ENUM __C)
10694 {
10695   return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A,
10696                                                     (__C << 2) | __B,
10697                                                     (__v8sf)
10698                                                     _mm256_setzero_ps (),
10699                                                     (__mmask8) -1);
10700 }
10701
10702 extern __inline __m256
10703 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10704 _mm256_mask_getmant_ps (__m256 __W, __mmask8 __U, __m256 __A,
10705                         _MM_MANTISSA_NORM_ENUM __B,
10706                         _MM_MANTISSA_SIGN_ENUM __C)
10707 {
10708   return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A,
10709                                                     (__C << 2) | __B,
10710                                                     (__v8sf) __W,
10711                                                     (__mmask8) __U);
10712 }
10713
10714 extern __inline __m256
10715 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10716 _mm256_maskz_getmant_ps (__mmask8 __U, __m256 __A,
10717                          _MM_MANTISSA_NORM_ENUM __B,
10718                          _MM_MANTISSA_SIGN_ENUM __C)
10719 {
10720   return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A,
10721                                                     (__C << 2) | __B,
10722                                                     (__v8sf)
10723                                                     _mm256_setzero_ps (),
10724                                                     (__mmask8) __U);
10725 }
10726
10727 extern __inline __m128
10728 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10729 _mm_getmant_ps (__m128 __A, _MM_MANTISSA_NORM_ENUM __B,
10730                 _MM_MANTISSA_SIGN_ENUM __C)
10731 {
10732   return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A,
10733                                                     (__C << 2) | __B,
10734                                                     (__v4sf)
10735                                                     _mm_setzero_ps (),
10736                                                     (__mmask8) -1);
10737 }
10738
10739 extern __inline __m128
10740 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10741 _mm_mask_getmant_ps (__m128 __W, __mmask8 __U, __m128 __A,
10742                      _MM_MANTISSA_NORM_ENUM __B,
10743                      _MM_MANTISSA_SIGN_ENUM __C)
10744 {
10745   return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A,
10746                                                     (__C << 2) | __B,
10747                                                     (__v4sf) __W,
10748                                                     (__mmask8) __U);
10749 }
10750
10751 extern __inline __m128
10752 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10753 _mm_maskz_getmant_ps (__mmask8 __U, __m128 __A,
10754                       _MM_MANTISSA_NORM_ENUM __B,
10755                       _MM_MANTISSA_SIGN_ENUM __C)
10756 {
10757   return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A,
10758                                                     (__C << 2) | __B,
10759                                                     (__v4sf)
10760                                                     _mm_setzero_ps (),
10761                                                     (__mmask8) __U);
10762 }
10763
10764 extern __inline __m256d
10765 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10766 _mm256_getmant_pd (__m256d __A, _MM_MANTISSA_NORM_ENUM __B,
10767                    _MM_MANTISSA_SIGN_ENUM __C)
10768 {
10769   return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A,
10770                                                      (__C << 2) | __B,
10771                                                      (__v4df)
10772                                                      _mm256_setzero_pd (),
10773                                                      (__mmask8) -1);
10774 }
10775
10776 extern __inline __m256d
10777 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10778 _mm256_mask_getmant_pd (__m256d __W, __mmask8 __U, __m256d __A,
10779                         _MM_MANTISSA_NORM_ENUM __B,
10780                         _MM_MANTISSA_SIGN_ENUM __C)
10781 {
10782   return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A,
10783                                                      (__C << 2) | __B,
10784                                                      (__v4df) __W,
10785                                                      (__mmask8) __U);
10786 }
10787
10788 extern __inline __m256d
10789 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10790 _mm256_maskz_getmant_pd (__mmask8 __U, __m256d __A,
10791                          _MM_MANTISSA_NORM_ENUM __B,
10792                          _MM_MANTISSA_SIGN_ENUM __C)
10793 {
10794   return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A,
10795                                                      (__C << 2) | __B,
10796                                                      (__v4df)
10797                                                      _mm256_setzero_pd (),
10798                                                      (__mmask8) __U);
10799 }
10800
10801 extern __inline __m128d
10802 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10803 _mm_getmant_pd (__m128d __A, _MM_MANTISSA_NORM_ENUM __B,
10804                 _MM_MANTISSA_SIGN_ENUM __C)
10805 {
10806   return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A,
10807                                                      (__C << 2) | __B,
10808                                                      (__v2df)
10809                                                      _mm_setzero_pd (),
10810                                                      (__mmask8) -1);
10811 }
10812
10813 extern __inline __m128d
10814 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10815 _mm_mask_getmant_pd (__m128d __W, __mmask8 __U, __m128d __A,
10816                      _MM_MANTISSA_NORM_ENUM __B,
10817                      _MM_MANTISSA_SIGN_ENUM __C)
10818 {
10819   return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A,
10820                                                      (__C << 2) | __B,
10821                                                      (__v2df) __W,
10822                                                      (__mmask8) __U);
10823 }
10824
10825 extern __inline __m128d
10826 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10827 _mm_maskz_getmant_pd (__mmask8 __U, __m128d __A,
10828                       _MM_MANTISSA_NORM_ENUM __B,
10829                       _MM_MANTISSA_SIGN_ENUM __C)
10830 {
10831   return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A,
10832                                                      (__C << 2) | __B,
10833                                                      (__v2df)
10834                                                      _mm_setzero_pd (),
10835                                                      (__mmask8) __U);
10836 }
10837
10838 extern __inline __m256
10839 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10840 _mm256_mmask_i32gather_ps (__m256 __v1_old, __mmask8 __mask,
10841                            __m256i __index, void const *__addr,
10842                            int __scale)
10843 {
10844   return (__m256) __builtin_ia32_gather3siv8sf ((__v8sf) __v1_old,
10845                                                 __addr,
10846                                                 (__v8si) __index,
10847                                                 __mask, __scale);
10848 }
10849
10850 extern __inline __m128
10851 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10852 _mm_mmask_i32gather_ps (__m128 __v1_old, __mmask8 __mask,
10853                         __m128i __index, void const *__addr,
10854                         int __scale)
10855 {
10856   return (__m128) __builtin_ia32_gather3siv4sf ((__v4sf) __v1_old,
10857                                                 __addr,
10858                                                 (__v4si) __index,
10859                                                 __mask, __scale);
10860 }
10861
10862 extern __inline __m256d
10863 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10864 _mm256_mmask_i32gather_pd (__m256d __v1_old, __mmask8 __mask,
10865                            __m128i __index, void const *__addr,
10866                            int __scale)
10867 {
10868   return (__m256d) __builtin_ia32_gather3siv4df ((__v4df) __v1_old,
10869                                                  __addr,
10870                                                  (__v4si) __index,
10871                                                  __mask, __scale);
10872 }
10873
10874 extern __inline __m128d
10875 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10876 _mm_mmask_i32gather_pd (__m128d __v1_old, __mmask8 __mask,
10877                         __m128i __index, void const *__addr,
10878                         int __scale)
10879 {
10880   return (__m128d) __builtin_ia32_gather3siv2df ((__v2df) __v1_old,
10881                                                  __addr,
10882                                                  (__v4si) __index,
10883                                                  __mask, __scale);
10884 }
10885
10886 extern __inline __m128
10887 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10888 _mm256_mmask_i64gather_ps (__m128 __v1_old, __mmask8 __mask,
10889                            __m256i __index, void const *__addr,
10890                            int __scale)
10891 {
10892   return (__m128) __builtin_ia32_gather3div8sf ((__v4sf) __v1_old,
10893                                                 __addr,
10894                                                 (__v4di) __index,
10895                                                 __mask, __scale);
10896 }
10897
10898 extern __inline __m128
10899 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10900 _mm_mmask_i64gather_ps (__m128 __v1_old, __mmask8 __mask,
10901                         __m128i __index, void const *__addr,
10902                         int __scale)
10903 {
10904   return (__m128) __builtin_ia32_gather3div4sf ((__v4sf) __v1_old,
10905                                                 __addr,
10906                                                 (__v2di) __index,
10907                                                 __mask, __scale);
10908 }
10909
10910 extern __inline __m256d
10911 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10912 _mm256_mmask_i64gather_pd (__m256d __v1_old, __mmask8 __mask,
10913                            __m256i __index, void const *__addr,
10914                            int __scale)
10915 {
10916   return (__m256d) __builtin_ia32_gather3div4df ((__v4df) __v1_old,
10917                                                  __addr,
10918                                                  (__v4di) __index,
10919                                                  __mask, __scale);
10920 }
10921
10922 extern __inline __m128d
10923 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10924 _mm_mmask_i64gather_pd (__m128d __v1_old, __mmask8 __mask,
10925                         __m128i __index, void const *__addr,
10926                         int __scale)
10927 {
10928   return (__m128d) __builtin_ia32_gather3div2df ((__v2df) __v1_old,
10929                                                  __addr,
10930                                                  (__v2di) __index,
10931                                                  __mask, __scale);
10932 }
10933
10934 extern __inline __m256i
10935 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10936 _mm256_mmask_i32gather_epi32 (__m256i __v1_old, __mmask8 __mask,
10937                               __m256i __index, void const *__addr,
10938                               int __scale)
10939 {
10940   return (__m256i) __builtin_ia32_gather3siv8si ((__v8si) __v1_old,
10941                                                  __addr,
10942                                                  (__v8si) __index,
10943                                                  __mask, __scale);
10944 }
10945
10946 extern __inline __m128i
10947 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10948 _mm_mmask_i32gather_epi32 (__m128i __v1_old, __mmask8 __mask,
10949                            __m128i __index, void const *__addr,
10950                            int __scale)
10951 {
10952   return (__m128i) __builtin_ia32_gather3siv4si ((__v4si) __v1_old,
10953                                                  __addr,
10954                                                  (__v4si) __index,
10955                                                  __mask, __scale);
10956 }
10957
10958 extern __inline __m256i
10959 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10960 _mm256_mmask_i32gather_epi64 (__m256i __v1_old, __mmask8 __mask,
10961                               __m128i __index, void const *__addr,
10962                               int __scale)
10963 {
10964   return (__m256i) __builtin_ia32_gather3siv4di ((__v4di) __v1_old,
10965                                                  __addr,
10966                                                  (__v4si) __index,
10967                                                  __mask, __scale);
10968 }
10969
10970 extern __inline __m128i
10971 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10972 _mm_mmask_i32gather_epi64 (__m128i __v1_old, __mmask8 __mask,
10973                            __m128i __index, void const *__addr,
10974                            int __scale)
10975 {
10976   return (__m128i) __builtin_ia32_gather3siv2di ((__v2di) __v1_old,
10977                                                  __addr,
10978                                                  (__v4si) __index,
10979                                                  __mask, __scale);
10980 }
10981
10982 extern __inline __m128i
10983 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10984 _mm256_mmask_i64gather_epi32 (__m128i __v1_old, __mmask8 __mask,
10985                               __m256i __index, void const *__addr,
10986                               int __scale)
10987 {
10988   return (__m128i) __builtin_ia32_gather3div8si ((__v4si) __v1_old,
10989                                                  __addr,
10990                                                  (__v4di) __index,
10991                                                  __mask, __scale);
10992 }
10993
10994 extern __inline __m128i
10995 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
10996 _mm_mmask_i64gather_epi32 (__m128i __v1_old, __mmask8 __mask,
10997                            __m128i __index, void const *__addr,
10998                            int __scale)
10999 {
11000   return (__m128i) __builtin_ia32_gather3div4si ((__v4si) __v1_old,
11001                                                  __addr,
11002                                                  (__v2di) __index,
11003                                                  __mask, __scale);
11004 }
11005
11006 extern __inline __m256i
11007 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11008 _mm256_mmask_i64gather_epi64 (__m256i __v1_old, __mmask8 __mask,
11009                               __m256i __index, void const *__addr,
11010                               int __scale)
11011 {
11012   return (__m256i) __builtin_ia32_gather3div4di ((__v4di) __v1_old,
11013                                                  __addr,
11014                                                  (__v4di) __index,
11015                                                  __mask, __scale);
11016 }
11017
11018 extern __inline __m128i
11019 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11020 _mm_mmask_i64gather_epi64 (__m128i __v1_old, __mmask8 __mask,
11021                            __m128i __index, void const *__addr,
11022                            int __scale)
11023 {
11024   return (__m128i) __builtin_ia32_gather3div2di ((__v2di) __v1_old,
11025                                                  __addr,
11026                                                  (__v2di) __index,
11027                                                  __mask, __scale);
11028 }
11029
11030 extern __inline void
11031 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11032 _mm256_i32scatter_ps (void *__addr, __m256i __index,
11033                       __m256 __v1, const int __scale)
11034 {
11035   __builtin_ia32_scattersiv8sf (__addr, (__mmask8) 0xFF,
11036                                 (__v8si) __index, (__v8sf) __v1,
11037                                 __scale);
11038 }
11039
11040 extern __inline void
11041 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11042 _mm256_mask_i32scatter_ps (void *__addr, __mmask8 __mask,
11043                            __m256i __index, __m256 __v1,
11044                            const int __scale)
11045 {
11046   __builtin_ia32_scattersiv8sf (__addr, __mask, (__v8si) __index,
11047                                 (__v8sf) __v1, __scale);
11048 }
11049
11050 extern __inline void
11051 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11052 _mm_i32scatter_ps (void *__addr, __m128i __index, __m128 __v1,
11053                    const int __scale)
11054 {
11055   __builtin_ia32_scattersiv4sf (__addr, (__mmask8) 0xFF,
11056                                 (__v4si) __index, (__v4sf) __v1,
11057                                 __scale);
11058 }
11059
11060 extern __inline void
11061 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11062 _mm_mask_i32scatter_ps (void *__addr, __mmask8 __mask,
11063                         __m128i __index, __m128 __v1,
11064                         const int __scale)
11065 {
11066   __builtin_ia32_scattersiv4sf (__addr, __mask, (__v4si) __index,
11067                                 (__v4sf) __v1, __scale);
11068 }
11069
11070 extern __inline void
11071 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11072 _mm256_i32scatter_pd (void *__addr, __m128i __index,
11073                       __m256d __v1, const int __scale)
11074 {
11075   __builtin_ia32_scattersiv4df (__addr, (__mmask8) 0xFF,
11076                                 (__v4si) __index, (__v4df) __v1,
11077                                 __scale);
11078 }
11079
11080 extern __inline void
11081 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11082 _mm256_mask_i32scatter_pd (void *__addr, __mmask8 __mask,
11083                            __m128i __index, __m256d __v1,
11084                            const int __scale)
11085 {
11086   __builtin_ia32_scattersiv4df (__addr, __mask, (__v4si) __index,
11087                                 (__v4df) __v1, __scale);
11088 }
11089
11090 extern __inline void
11091 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11092 _mm_i32scatter_pd (void *__addr, __m128i __index,
11093                    __m128d __v1, const int __scale)
11094 {
11095   __builtin_ia32_scattersiv2df (__addr, (__mmask8) 0xFF,
11096                                 (__v4si) __index, (__v2df) __v1,
11097                                 __scale);
11098 }
11099
11100 extern __inline void
11101 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11102 _mm_mask_i32scatter_pd (void *__addr, __mmask8 __mask,
11103                         __m128i __index, __m128d __v1,
11104                         const int __scale)
11105 {
11106   __builtin_ia32_scattersiv2df (__addr, __mask, (__v4si) __index,
11107                                 (__v2df) __v1, __scale);
11108 }
11109
11110 extern __inline void
11111 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11112 _mm256_i64scatter_ps (void *__addr, __m256i __index,
11113                       __m128 __v1, const int __scale)
11114 {
11115   __builtin_ia32_scatterdiv8sf (__addr, (__mmask8) 0xFF,
11116                                 (__v4di) __index, (__v4sf) __v1,
11117                                 __scale);
11118 }
11119
11120 extern __inline void
11121 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11122 _mm256_mask_i64scatter_ps (void *__addr, __mmask8 __mask,
11123                            __m256i __index, __m128 __v1,
11124                            const int __scale)
11125 {
11126   __builtin_ia32_scatterdiv8sf (__addr, __mask, (__v4di) __index,
11127                                 (__v4sf) __v1, __scale);
11128 }
11129
11130 extern __inline void
11131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11132 _mm_i64scatter_ps (void *__addr, __m128i __index, __m128 __v1,
11133                    const int __scale)
11134 {
11135   __builtin_ia32_scatterdiv4sf (__addr, (__mmask8) 0xFF,
11136                                 (__v2di) __index, (__v4sf) __v1,
11137                                 __scale);
11138 }
11139
11140 extern __inline void
11141 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11142 _mm_mask_i64scatter_ps (void *__addr, __mmask8 __mask,
11143                         __m128i __index, __m128 __v1,
11144                         const int __scale)
11145 {
11146   __builtin_ia32_scatterdiv4sf (__addr, __mask, (__v2di) __index,
11147                                 (__v4sf) __v1, __scale);
11148 }
11149
11150 extern __inline void
11151 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11152 _mm256_i64scatter_pd (void *__addr, __m256i __index,
11153                       __m256d __v1, const int __scale)
11154 {
11155   __builtin_ia32_scatterdiv4df (__addr, (__mmask8) 0xFF,
11156                                 (__v4di) __index, (__v4df) __v1,
11157                                 __scale);
11158 }
11159
11160 extern __inline void
11161 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11162 _mm256_mask_i64scatter_pd (void *__addr, __mmask8 __mask,
11163                            __m256i __index, __m256d __v1,
11164                            const int __scale)
11165 {
11166   __builtin_ia32_scatterdiv4df (__addr, __mask, (__v4di) __index,
11167                                 (__v4df) __v1, __scale);
11168 }
11169
11170 extern __inline void
11171 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11172 _mm_i64scatter_pd (void *__addr, __m128i __index,
11173                    __m128d __v1, const int __scale)
11174 {
11175   __builtin_ia32_scatterdiv2df (__addr, (__mmask8) 0xFF,
11176                                 (__v2di) __index, (__v2df) __v1,
11177                                 __scale);
11178 }
11179
11180 extern __inline void
11181 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11182 _mm_mask_i64scatter_pd (void *__addr, __mmask8 __mask,
11183                         __m128i __index, __m128d __v1,
11184                         const int __scale)
11185 {
11186   __builtin_ia32_scatterdiv2df (__addr, __mask, (__v2di) __index,
11187                                 (__v2df) __v1, __scale);
11188 }
11189
11190 extern __inline void
11191 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11192 _mm256_i32scatter_epi32 (void *__addr, __m256i __index,
11193                          __m256i __v1, const int __scale)
11194 {
11195   __builtin_ia32_scattersiv8si (__addr, (__mmask8) 0xFF,
11196                                 (__v8si) __index, (__v8si) __v1,
11197                                 __scale);
11198 }
11199
11200 extern __inline void
11201 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11202 _mm256_mask_i32scatter_epi32 (void *__addr, __mmask8 __mask,
11203                               __m256i __index, __m256i __v1,
11204                               const int __scale)
11205 {
11206   __builtin_ia32_scattersiv8si (__addr, __mask, (__v8si) __index,
11207                                 (__v8si) __v1, __scale);
11208 }
11209
11210 extern __inline void
11211 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11212 _mm_i32scatter_epi32 (void *__addr, __m128i __index,
11213                       __m128i __v1, const int __scale)
11214 {
11215   __builtin_ia32_scattersiv4si (__addr, (__mmask8) 0xFF,
11216                                 (__v4si) __index, (__v4si) __v1,
11217                                 __scale);
11218 }
11219
11220 extern __inline void
11221 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11222 _mm_mask_i32scatter_epi32 (void *__addr, __mmask8 __mask,
11223                            __m128i __index, __m128i __v1,
11224                            const int __scale)
11225 {
11226   __builtin_ia32_scattersiv4si (__addr, __mask, (__v4si) __index,
11227                                 (__v4si) __v1, __scale);
11228 }
11229
11230 extern __inline void
11231 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11232 _mm256_i32scatter_epi64 (void *__addr, __m128i __index,
11233                          __m256i __v1, const int __scale)
11234 {
11235   __builtin_ia32_scattersiv4di (__addr, (__mmask8) 0xFF,
11236                                 (__v4si) __index, (__v4di) __v1,
11237                                 __scale);
11238 }
11239
11240 extern __inline void
11241 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11242 _mm256_mask_i32scatter_epi64 (void *__addr, __mmask8 __mask,
11243                               __m128i __index, __m256i __v1,
11244                               const int __scale)
11245 {
11246   __builtin_ia32_scattersiv4di (__addr, __mask, (__v4si) __index,
11247                                 (__v4di) __v1, __scale);
11248 }
11249
11250 extern __inline void
11251 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11252 _mm_i32scatter_epi64 (void *__addr, __m128i __index,
11253                       __m128i __v1, const int __scale)
11254 {
11255   __builtin_ia32_scattersiv2di (__addr, (__mmask8) 0xFF,
11256                                 (__v4si) __index, (__v2di) __v1,
11257                                 __scale);
11258 }
11259
11260 extern __inline void
11261 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11262 _mm_mask_i32scatter_epi64 (void *__addr, __mmask8 __mask,
11263                            __m128i __index, __m128i __v1,
11264                            const int __scale)
11265 {
11266   __builtin_ia32_scattersiv2di (__addr, __mask, (__v4si) __index,
11267                                 (__v2di) __v1, __scale);
11268 }
11269
11270 extern __inline void
11271 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11272 _mm256_i64scatter_epi32 (void *__addr, __m256i __index,
11273                          __m128i __v1, const int __scale)
11274 {
11275   __builtin_ia32_scatterdiv8si (__addr, (__mmask8) 0xFF,
11276                                 (__v4di) __index, (__v4si) __v1,
11277                                 __scale);
11278 }
11279
11280 extern __inline void
11281 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11282 _mm256_mask_i64scatter_epi32 (void *__addr, __mmask8 __mask,
11283                               __m256i __index, __m128i __v1,
11284                               const int __scale)
11285 {
11286   __builtin_ia32_scatterdiv8si (__addr, __mask, (__v4di) __index,
11287                                 (__v4si) __v1, __scale);
11288 }
11289
11290 extern __inline void
11291 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11292 _mm_i64scatter_epi32 (void *__addr, __m128i __index,
11293                       __m128i __v1, const int __scale)
11294 {
11295   __builtin_ia32_scatterdiv4si (__addr, (__mmask8) 0xFF,
11296                                 (__v2di) __index, (__v4si) __v1,
11297                                 __scale);
11298 }
11299
11300 extern __inline void
11301 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11302 _mm_mask_i64scatter_epi32 (void *__addr, __mmask8 __mask,
11303                            __m128i __index, __m128i __v1,
11304                            const int __scale)
11305 {
11306   __builtin_ia32_scatterdiv4si (__addr, __mask, (__v2di) __index,
11307                                 (__v4si) __v1, __scale);
11308 }
11309
11310 extern __inline void
11311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11312 _mm256_i64scatter_epi64 (void *__addr, __m256i __index,
11313                          __m256i __v1, const int __scale)
11314 {
11315   __builtin_ia32_scatterdiv4di (__addr, (__mmask8) 0xFF,
11316                                 (__v4di) __index, (__v4di) __v1,
11317                                 __scale);
11318 }
11319
11320 extern __inline void
11321 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11322 _mm256_mask_i64scatter_epi64 (void *__addr, __mmask8 __mask,
11323                               __m256i __index, __m256i __v1,
11324                               const int __scale)
11325 {
11326   __builtin_ia32_scatterdiv4di (__addr, __mask, (__v4di) __index,
11327                                 (__v4di) __v1, __scale);
11328 }
11329
11330 extern __inline void
11331 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11332 _mm_i64scatter_epi64 (void *__addr, __m128i __index,
11333                       __m128i __v1, const int __scale)
11334 {
11335   __builtin_ia32_scatterdiv2di (__addr, (__mmask8) 0xFF,
11336                                 (__v2di) __index, (__v2di) __v1,
11337                                 __scale);
11338 }
11339
11340 extern __inline void
11341 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11342 _mm_mask_i64scatter_epi64 (void *__addr, __mmask8 __mask,
11343                            __m128i __index, __m128i __v1,
11344                            const int __scale)
11345 {
11346   __builtin_ia32_scatterdiv2di (__addr, __mask, (__v2di) __index,
11347                                 (__v2di) __v1, __scale);
11348 }
11349
11350 extern __inline __m256i
11351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11352 _mm256_mask_shuffle_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11353                            _MM_PERM_ENUM __mask)
11354 {
11355   return (__m256i) __builtin_ia32_pshufd256_mask ((__v8si) __A, __mask,
11356                                                   (__v8si) __W,
11357                                                   (__mmask8) __U);
11358 }
11359
11360 extern __inline __m256i
11361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11362 _mm256_maskz_shuffle_epi32 (__mmask8 __U, __m256i __A,
11363                             _MM_PERM_ENUM __mask)
11364 {
11365   return (__m256i) __builtin_ia32_pshufd256_mask ((__v8si) __A, __mask,
11366                                                   (__v8si)
11367                                                   _mm256_setzero_si256 (),
11368                                                   (__mmask8) __U);
11369 }
11370
11371 extern __inline __m128i
11372 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11373 _mm_mask_shuffle_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
11374                         _MM_PERM_ENUM __mask)
11375 {
11376   return (__m128i) __builtin_ia32_pshufd128_mask ((__v4si) __A, __mask,
11377                                                   (__v4si) __W,
11378                                                   (__mmask8) __U);
11379 }
11380
11381 extern __inline __m128i
11382 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11383 _mm_maskz_shuffle_epi32 (__mmask8 __U, __m128i __A,
11384                          _MM_PERM_ENUM __mask)
11385 {
11386   return (__m128i) __builtin_ia32_pshufd128_mask ((__v4si) __A, __mask,
11387                                                   (__v4si)
11388                                                   _mm_setzero_si128 (),
11389                                                   (__mmask8) __U);
11390 }
11391
11392 extern __inline __m256i
11393 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11394 _mm256_rol_epi32 (__m256i __A, const int __B)
11395 {
11396   return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B,
11397                                                  (__v8si)
11398                                                  _mm256_setzero_si256 (),
11399                                                  (__mmask8) -1);
11400 }
11401
11402 extern __inline __m256i
11403 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11404 _mm256_mask_rol_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11405                        const int __B)
11406 {
11407   return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B,
11408                                                  (__v8si) __W,
11409                                                  (__mmask8) __U);
11410 }
11411
11412 extern __inline __m256i
11413 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11414 _mm256_maskz_rol_epi32 (__mmask8 __U, __m256i __A, const int __B)
11415 {
11416   return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B,
11417                                                  (__v8si)
11418                                                  _mm256_setzero_si256 (),
11419                                                  (__mmask8) __U);
11420 }
11421
11422 extern __inline __m128i
11423 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11424 _mm_rol_epi32 (__m128i __A, const int __B)
11425 {
11426   return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B,
11427                                                  (__v4si)
11428                                                  _mm_setzero_si128 (),
11429                                                  (__mmask8) -1);
11430 }
11431
11432 extern __inline __m128i
11433 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11434 _mm_mask_rol_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
11435                     const int __B)
11436 {
11437   return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B,
11438                                                  (__v4si) __W,
11439                                                  (__mmask8) __U);
11440 }
11441
11442 extern __inline __m128i
11443 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11444 _mm_maskz_rol_epi32 (__mmask8 __U, __m128i __A, const int __B)
11445 {
11446   return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B,
11447                                                  (__v4si)
11448                                                  _mm_setzero_si128 (),
11449                                                  (__mmask8) __U);
11450 }
11451
11452 extern __inline __m256i
11453 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11454 _mm256_ror_epi32 (__m256i __A, const int __B)
11455 {
11456   return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B,
11457                                                  (__v8si)
11458                                                  _mm256_setzero_si256 (),
11459                                                  (__mmask8) -1);
11460 }
11461
11462 extern __inline __m256i
11463 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11464 _mm256_mask_ror_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11465                        const int __B)
11466 {
11467   return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B,
11468                                                  (__v8si) __W,
11469                                                  (__mmask8) __U);
11470 }
11471
11472 extern __inline __m256i
11473 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11474 _mm256_maskz_ror_epi32 (__mmask8 __U, __m256i __A, const int __B)
11475 {
11476   return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B,
11477                                                  (__v8si)
11478                                                  _mm256_setzero_si256 (),
11479                                                  (__mmask8) __U);
11480 }
11481
11482 extern __inline __m128i
11483 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11484 _mm_ror_epi32 (__m128i __A, const int __B)
11485 {
11486   return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B,
11487                                                  (__v4si)
11488                                                  _mm_setzero_si128 (),
11489                                                  (__mmask8) -1);
11490 }
11491
11492 extern __inline __m128i
11493 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11494 _mm_mask_ror_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
11495                     const int __B)
11496 {
11497   return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B,
11498                                                  (__v4si) __W,
11499                                                  (__mmask8) __U);
11500 }
11501
11502 extern __inline __m128i
11503 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11504 _mm_maskz_ror_epi32 (__mmask8 __U, __m128i __A, const int __B)
11505 {
11506   return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B,
11507                                                  (__v4si)
11508                                                  _mm_setzero_si128 (),
11509                                                  (__mmask8) __U);
11510 }
11511
11512 extern __inline __m256i
11513 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11514 _mm256_rol_epi64 (__m256i __A, const int __B)
11515 {
11516   return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B,
11517                                                  (__v4di)
11518                                                  _mm256_setzero_si256 (),
11519                                                  (__mmask8) -1);
11520 }
11521
11522 extern __inline __m256i
11523 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11524 _mm256_mask_rol_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
11525                        const int __B)
11526 {
11527   return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B,
11528                                                  (__v4di) __W,
11529                                                  (__mmask8) __U);
11530 }
11531
11532 extern __inline __m256i
11533 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11534 _mm256_maskz_rol_epi64 (__mmask8 __U, __m256i __A, const int __B)
11535 {
11536   return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B,
11537                                                  (__v4di)
11538                                                  _mm256_setzero_si256 (),
11539                                                  (__mmask8) __U);
11540 }
11541
11542 extern __inline __m128i
11543 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11544 _mm_rol_epi64 (__m128i __A, const int __B)
11545 {
11546   return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B,
11547                                                  (__v2di)
11548                                                  _mm_setzero_si128 (),
11549                                                  (__mmask8) -1);
11550 }
11551
11552 extern __inline __m128i
11553 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11554 _mm_mask_rol_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
11555                     const int __B)
11556 {
11557   return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B,
11558                                                  (__v2di) __W,
11559                                                  (__mmask8) __U);
11560 }
11561
11562 extern __inline __m128i
11563 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11564 _mm_maskz_rol_epi64 (__mmask8 __U, __m128i __A, const int __B)
11565 {
11566   return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B,
11567                                                  (__v2di)
11568                                                  _mm_setzero_si128 (),
11569                                                  (__mmask8) __U);
11570 }
11571
11572 extern __inline __m256i
11573 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11574 _mm256_ror_epi64 (__m256i __A, const int __B)
11575 {
11576   return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B,
11577                                                  (__v4di)
11578                                                  _mm256_setzero_si256 (),
11579                                                  (__mmask8) -1);
11580 }
11581
11582 extern __inline __m256i
11583 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11584 _mm256_mask_ror_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
11585                        const int __B)
11586 {
11587   return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B,
11588                                                  (__v4di) __W,
11589                                                  (__mmask8) __U);
11590 }
11591
11592 extern __inline __m256i
11593 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11594 _mm256_maskz_ror_epi64 (__mmask8 __U, __m256i __A, const int __B)
11595 {
11596   return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B,
11597                                                  (__v4di)
11598                                                  _mm256_setzero_si256 (),
11599                                                  (__mmask8) __U);
11600 }
11601
11602 extern __inline __m128i
11603 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11604 _mm_ror_epi64 (__m128i __A, const int __B)
11605 {
11606   return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B,
11607                                                  (__v2di)
11608                                                  _mm_setzero_si128 (),
11609                                                  (__mmask8) -1);
11610 }
11611
11612 extern __inline __m128i
11613 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11614 _mm_mask_ror_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
11615                     const int __B)
11616 {
11617   return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B,
11618                                                  (__v2di) __W,
11619                                                  (__mmask8) __U);
11620 }
11621
11622 extern __inline __m128i
11623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11624 _mm_maskz_ror_epi64 (__mmask8 __U, __m128i __A, const int __B)
11625 {
11626   return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B,
11627                                                  (__v2di)
11628                                                  _mm_setzero_si128 (),
11629                                                  (__mmask8) __U);
11630 }
11631
11632 extern __inline __m128i
11633 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11634 _mm_alignr_epi32 (__m128i __A, __m128i __B, const int __imm)
11635 {
11636   return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A,
11637                                                   (__v4si) __B, __imm,
11638                                                   (__v4si)
11639                                                   _mm_setzero_si128 (),
11640                                                   (__mmask8) -1);
11641 }
11642
11643 extern __inline __m128i
11644 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11645 _mm_mask_alignr_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
11646                        __m128i __B, const int __imm)
11647 {
11648   return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A,
11649                                                   (__v4si) __B, __imm,
11650                                                   (__v4si) __W,
11651                                                   (__mmask8) __U);
11652 }
11653
11654 extern __inline __m128i
11655 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11656 _mm_maskz_alignr_epi32 (__mmask8 __U, __m128i __A, __m128i __B,
11657                         const int __imm)
11658 {
11659   return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A,
11660                                                   (__v4si) __B, __imm,
11661                                                   (__v4si)
11662                                                   _mm_setzero_si128 (),
11663                                                   (__mmask8) __U);
11664 }
11665
11666 extern __inline __m128i
11667 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11668 _mm_alignr_epi64 (__m128i __A, __m128i __B, const int __imm)
11669 {
11670   return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A,
11671                                                   (__v2di) __B, __imm,
11672                                                   (__v2di)
11673                                                   _mm_setzero_si128 (),
11674                                                   (__mmask8) -1);
11675 }
11676
11677 extern __inline __m128i
11678 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11679 _mm_mask_alignr_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
11680                        __m128i __B, const int __imm)
11681 {
11682   return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A,
11683                                                   (__v2di) __B, __imm,
11684                                                   (__v2di) __W,
11685                                                   (__mmask8) __U);
11686 }
11687
11688 extern __inline __m128i
11689 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11690 _mm_maskz_alignr_epi64 (__mmask8 __U, __m128i __A, __m128i __B,
11691                         const int __imm)
11692 {
11693   return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A,
11694                                                   (__v2di) __B, __imm,
11695                                                   (__v2di)
11696                                                   _mm_setzero_si128 (),
11697                                                   (__mmask8) __U);
11698 }
11699
11700 extern __inline __m256i
11701 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11702 _mm256_alignr_epi32 (__m256i __A, __m256i __B, const int __imm)
11703 {
11704   return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A,
11705                                                   (__v8si) __B, __imm,
11706                                                   (__v8si)
11707                                                   _mm256_setzero_si256 (),
11708                                                   (__mmask8) -1);
11709 }
11710
11711 extern __inline __m256i
11712 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11713 _mm256_mask_alignr_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11714                           __m256i __B, const int __imm)
11715 {
11716   return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A,
11717                                                   (__v8si) __B, __imm,
11718                                                   (__v8si) __W,
11719                                                   (__mmask8) __U);
11720 }
11721
11722 extern __inline __m256i
11723 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11724 _mm256_maskz_alignr_epi32 (__mmask8 __U, __m256i __A, __m256i __B,
11725                            const int __imm)
11726 {
11727   return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A,
11728                                                   (__v8si) __B, __imm,
11729                                                   (__v8si)
11730                                                   _mm256_setzero_si256 (),
11731                                                   (__mmask8) __U);
11732 }
11733
11734 extern __inline __m256i
11735 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11736 _mm256_alignr_epi64 (__m256i __A, __m256i __B, const int __imm)
11737 {
11738   return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A,
11739                                                   (__v4di) __B, __imm,
11740                                                   (__v4di)
11741                                                   _mm256_setzero_si256 (),
11742                                                   (__mmask8) -1);
11743 }
11744
11745 extern __inline __m256i
11746 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11747 _mm256_mask_alignr_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
11748                           __m256i __B, const int __imm)
11749 {
11750   return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A,
11751                                                   (__v4di) __B, __imm,
11752                                                   (__v4di) __W,
11753                                                   (__mmask8) __U);
11754 }
11755
11756 extern __inline __m256i
11757 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11758 _mm256_maskz_alignr_epi64 (__mmask8 __U, __m256i __A, __m256i __B,
11759                            const int __imm)
11760 {
11761   return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A,
11762                                                   (__v4di) __B, __imm,
11763                                                   (__v4di)
11764                                                   _mm256_setzero_si256 (),
11765                                                   (__mmask8) __U);
11766 }
11767
11768 extern __inline __m128i
11769 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11770 _mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A,
11771                    const int __I)
11772 {
11773   return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, __I,
11774                                                   (__v8hi) __W,
11775                                                   (__mmask8) __U);
11776 }
11777
11778 extern __inline __m128i
11779 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11780 _mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A, const int __I)
11781 {
11782   return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, __I,
11783                                                   (__v8hi)
11784                                                   _mm_setzero_si128 (),
11785                                                   (__mmask8) __U);
11786 }
11787
11788 extern __inline __m128i
11789 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11790 _mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A,
11791                       const int __I)
11792 {
11793   return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, __I,
11794                                                      (__v8hi) __W,
11795                                                      (__mmask8) __U);
11796 }
11797
11798 extern __inline __m128i
11799 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11800 _mm256_maskz_cvtps_ph (__mmask8 __U, __m256 __A, const int __I)
11801 {
11802   return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, __I,
11803                                                      (__v8hi)
11804                                                      _mm_setzero_si128 (),
11805                                                      (__mmask8) __U);
11806 }
11807
11808 extern __inline __m256i
11809 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11810 _mm256_mask_srai_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11811                         const int __imm)
11812 {
11813   return (__m256i) __builtin_ia32_psradi256_mask ((__v8si) __A, __imm,
11814                                                   (__v8si) __W,
11815                                                   (__mmask8) __U);
11816 }
11817
11818 extern __inline __m256i
11819 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11820 _mm256_maskz_srai_epi32 (__mmask8 __U, __m256i __A, const int __imm)
11821 {
11822   return (__m256i) __builtin_ia32_psradi256_mask ((__v8si) __A, __imm,
11823                                                   (__v8si)
11824                                                   _mm256_setzero_si256 (),
11825                                                   (__mmask8) __U);
11826 }
11827
11828 extern __inline __m128i
11829 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11830 _mm_mask_srai_epi32 (__m128i __W, __mmask8 __U, __m128i __A,
11831                      const int __imm)
11832 {
11833   return (__m128i) __builtin_ia32_psradi128_mask ((__v4si) __A, __imm,
11834                                                   (__v4si) __W,
11835                                                   (__mmask8) __U);
11836 }
11837
11838 extern __inline __m128i
11839 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11840 _mm_maskz_srai_epi32 (__mmask8 __U, __m128i __A, const int __imm)
11841 {
11842   return (__m128i) __builtin_ia32_psradi128_mask ((__v4si) __A, __imm,
11843                                                   (__v4si)
11844                                                   _mm_setzero_si128 (),
11845                                                   (__mmask8) __U);
11846 }
11847
11848 extern __inline __m256i
11849 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11850 _mm256_srai_epi64 (__m256i __A, const int __imm)
11851 {
11852   return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm,
11853                                                   (__v4di)
11854                                                   _mm256_setzero_si256 (),
11855                                                   (__mmask8) -1);
11856 }
11857
11858 extern __inline __m256i
11859 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11860 _mm256_mask_srai_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
11861                         const int __imm)
11862 {
11863   return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm,
11864                                                   (__v4di) __W,
11865                                                   (__mmask8) __U);
11866 }
11867
11868 extern __inline __m256i
11869 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11870 _mm256_maskz_srai_epi64 (__mmask8 __U, __m256i __A, const int __imm)
11871 {
11872   return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm,
11873                                                   (__v4di)
11874                                                   _mm256_setzero_si256 (),
11875                                                   (__mmask8) __U);
11876 }
11877
11878 extern __inline __m128i
11879 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11880 _mm_srai_epi64 (__m128i __A, const int __imm)
11881 {
11882   return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm,
11883                                                   (__v2di)
11884                                                   _mm_setzero_si128 (),
11885                                                   (__mmask8) -1);
11886 }
11887
11888 extern __inline __m128i
11889 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11890 _mm_mask_srai_epi64 (__m128i __W, __mmask8 __U, __m128i __A,
11891                      const int __imm)
11892 {
11893   return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm,
11894                                                   (__v2di) __W,
11895                                                   (__mmask8) __U);
11896 }
11897
11898 extern __inline __m128i
11899 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11900 _mm_maskz_srai_epi64 (__mmask8 __U, __m128i __A, const int __imm)
11901 {
11902   return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm,
11903                                                   (__v2di)
11904                                                   _mm_setzero_si128 (),
11905                                                   (__mmask8) __U);
11906 }
11907
11908 extern __inline __m128i
11909 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11910 _mm_mask_slli_epi32 (__m128i __W, __mmask8 __U, __m128i __A, int __B)
11911 {
11912   return (__m128i) __builtin_ia32_pslldi128_mask ((__v4si) __A, __B,
11913                                                   (__v4si) __W,
11914                                                   (__mmask8) __U);
11915 }
11916
11917 extern __inline __m128i
11918 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11919 _mm_maskz_slli_epi32 (__mmask8 __U, __m128i __A, int __B)
11920 {
11921   return (__m128i) __builtin_ia32_pslldi128_mask ((__v4si) __A, __B,
11922                                                   (__v4si)
11923                                                   _mm_setzero_si128 (),
11924                                                   (__mmask8) __U);
11925 }
11926
11927 extern __inline __m128i
11928 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11929 _mm_mask_slli_epi64 (__m128i __W, __mmask8 __U, __m128i __A, int __B)
11930 {
11931   return (__m128i) __builtin_ia32_psllqi128_mask ((__v2di) __A, __B,
11932                                                   (__v2di) __W,
11933                                                   (__mmask8) __U);
11934 }
11935
11936 extern __inline __m128i
11937 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11938 _mm_maskz_slli_epi64 (__mmask8 __U, __m128i __A, int __B)
11939 {
11940   return (__m128i) __builtin_ia32_psllqi128_mask ((__v2di) __A, __B,
11941                                                   (__v2di)
11942                                                   _mm_setzero_si128 (),
11943                                                   (__mmask8) __U);
11944 }
11945
11946 extern __inline __m256i
11947 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11948 _mm256_mask_slli_epi32 (__m256i __W, __mmask8 __U, __m256i __A,
11949                         int __B)
11950 {
11951   return (__m256i) __builtin_ia32_pslldi256_mask ((__v8si) __A, __B,
11952                                                   (__v8si) __W,
11953                                                   (__mmask8) __U);
11954 }
11955
11956 extern __inline __m256i
11957 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11958 _mm256_maskz_slli_epi32 (__mmask8 __U, __m256i __A, int __B)
11959 {
11960   return (__m256i) __builtin_ia32_pslldi256_mask ((__v8si) __A, __B,
11961                                                   (__v8si)
11962                                                   _mm256_setzero_si256 (),
11963                                                   (__mmask8) __U);
11964 }
11965
11966 extern __inline __m256i
11967 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11968 _mm256_mask_slli_epi64 (__m256i __W, __mmask8 __U, __m256i __A,
11969                         int __B)
11970 {
11971   return (__m256i) __builtin_ia32_psllqi256_mask ((__v4di) __A, __B,
11972                                                   (__v4di) __W,
11973                                                   (__mmask8) __U);
11974 }
11975
11976 extern __inline __m256i
11977 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11978 _mm256_maskz_slli_epi64 (__mmask8 __U, __m256i __A, int __B)
11979 {
11980   return (__m256i) __builtin_ia32_psllqi256_mask ((__v4di) __A, __B,
11981                                                   (__v4di)
11982                                                   _mm256_setzero_si256 (),
11983                                                   (__mmask8) __U);
11984 }
11985
11986 extern __inline __m256d
11987 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11988 _mm256_mask_permutex_pd (__m256d __W, __mmask8 __U, __m256d __X,
11989                          const int __imm)
11990 {
11991   return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __imm,
11992                                                   (__v4df) __W,
11993                                                   (__mmask8) __U);
11994 }
11995
11996 extern __inline __m256d
11997 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
11998 _mm256_maskz_permutex_pd (__mmask8 __U, __m256d __X, const int __imm)
11999 {
12000   return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __imm,
12001                                                   (__v4df)
12002                                                   _mm256_setzero_pd (),
12003                                                   (__mmask8) __U);
12004 }
12005
12006 extern __inline __m256d
12007 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12008 _mm256_mask_permute_pd (__m256d __W, __mmask8 __U, __m256d __X,
12009                         const int __C)
12010 {
12011   return (__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df) __X, __C,
12012                                                      (__v4df) __W,
12013                                                      (__mmask8) __U);
12014 }
12015
12016 extern __inline __m256d
12017 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12018 _mm256_maskz_permute_pd (__mmask8 __U, __m256d __X, const int __C)
12019 {
12020   return (__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df) __X, __C,
12021                                                      (__v4df)
12022                                                      _mm256_setzero_pd (),
12023                                                      (__mmask8) __U);
12024 }
12025
12026 extern __inline __m128d
12027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12028 _mm_mask_permute_pd (__m128d __W, __mmask8 __U, __m128d __X,
12029                      const int __C)
12030 {
12031   return (__m128d) __builtin_ia32_vpermilpd_mask ((__v2df) __X, __C,
12032                                                   (__v2df) __W,
12033                                                   (__mmask8) __U);
12034 }
12035
12036 extern __inline __m128d
12037 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12038 _mm_maskz_permute_pd (__mmask8 __U, __m128d __X, const int __C)
12039 {
12040   return (__m128d) __builtin_ia32_vpermilpd_mask ((__v2df) __X, __C,
12041                                                   (__v2df)
12042                                                   _mm_setzero_pd (),
12043                                                   (__mmask8) __U);
12044 }
12045
12046 extern __inline __m256
12047 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12048 _mm256_mask_permute_ps (__m256 __W, __mmask8 __U, __m256 __X,
12049                         const int __C)
12050 {
12051   return (__m256) __builtin_ia32_vpermilps256_mask ((__v8sf) __X, __C,
12052                                                     (__v8sf) __W,
12053                                                     (__mmask8) __U);
12054 }
12055
12056 extern __inline __m256
12057 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12058 _mm256_maskz_permute_ps (__mmask8 __U, __m256 __X, const int __C)
12059 {
12060   return (__m256) __builtin_ia32_vpermilps256_mask ((__v8sf) __X, __C,
12061                                                     (__v8sf)
12062                                                     _mm256_setzero_ps (),
12063                                                     (__mmask8) __U);
12064 }
12065
12066 extern __inline __m128
12067 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12068 _mm_mask_permute_ps (__m128 __W, __mmask8 __U, __m128 __X,
12069                      const int __C)
12070 {
12071   return (__m128) __builtin_ia32_vpermilps_mask ((__v4sf) __X, __C,
12072                                                  (__v4sf) __W,
12073                                                  (__mmask8) __U);
12074 }
12075
12076 extern __inline __m128
12077 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12078 _mm_maskz_permute_ps (__mmask8 __U, __m128 __X, const int __C)
12079 {
12080   return (__m128) __builtin_ia32_vpermilps_mask ((__v4sf) __X, __C,
12081                                                  (__v4sf)
12082                                                  _mm_setzero_ps (),
12083                                                  (__mmask8) __U);
12084 }
12085
12086 extern __inline __m256d
12087 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12088 _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W)
12089 {
12090   return (__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) __A,
12091                                                      (__v4df) __W,
12092                                                      (__mmask8) __U);
12093 }
12094
12095 extern __inline __m256
12096 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12097 _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W)
12098 {
12099   return (__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) __A,
12100                                                     (__v8sf) __W,
12101                                                     (__mmask8) __U);
12102 }
12103
12104 extern __inline __m256i
12105 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12106 _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W)
12107 {
12108   return (__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) __A,
12109                                                     (__v4di) __W,
12110                                                     (__mmask8) __U);
12111 }
12112
12113 extern __inline __m256i
12114 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12115 _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W)
12116 {
12117   return (__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) __A,
12118                                                     (__v8si) __W,
12119                                                     (__mmask8) __U);
12120 }
12121
12122 extern __inline __m128d
12123 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12124 _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W)
12125 {
12126   return (__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) __A,
12127                                                      (__v2df) __W,
12128                                                      (__mmask8) __U);
12129 }
12130
12131 extern __inline __m128
12132 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12133 _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W)
12134 {
12135   return (__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) __A,
12136                                                     (__v4sf) __W,
12137                                                     (__mmask8) __U);
12138 }
12139
12140 extern __inline __m128i
12141 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12142 _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W)
12143 {
12144   return (__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) __A,
12145                                                     (__v2di) __W,
12146                                                     (__mmask8) __U);
12147 }
12148
12149 extern __inline __m128i
12150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12151 _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W)
12152 {
12153   return (__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) __A,
12154                                                     (__v4si) __W,
12155                                                     (__mmask8) __U);
12156 }
12157
12158 extern __inline __mmask8
12159 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12160 _mm256_cmp_epi64_mask (__m256i __X, __m256i __Y, const int __P)
12161 {
12162   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
12163                                                  (__v4di) __Y, __P,
12164                                                  (__mmask8) -1);
12165 }
12166
12167 extern __inline __mmask8
12168 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12169 _mm256_cmp_epi32_mask (__m256i __X, __m256i __Y, const int __P)
12170 {
12171   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
12172                                                  (__v8si) __Y, __P,
12173                                                  (__mmask8) -1);
12174 }
12175
12176 extern __inline __mmask8
12177 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12178 _mm256_cmp_epu64_mask (__m256i __X, __m256i __Y, const int __P)
12179 {
12180   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
12181                                                   (__v4di) __Y, __P,
12182                                                   (__mmask8) -1);
12183 }
12184
12185 extern __inline __mmask8
12186 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12187 _mm256_cmp_epu32_mask (__m256i __X, __m256i __Y, const int __P)
12188 {
12189   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
12190                                                   (__v8si) __Y, __P,
12191                                                   (__mmask8) -1);
12192 }
12193
12194 extern __inline __mmask8
12195 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12196 _mm256_cmp_pd_mask (__m256d __X, __m256d __Y, const int __P)
12197 {
12198   return (__mmask8) __builtin_ia32_cmppd256_mask ((__v4df) __X,
12199                                                   (__v4df) __Y, __P,
12200                                                   (__mmask8) -1);
12201 }
12202
12203 extern __inline __mmask8
12204 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12205 _mm256_cmp_ps_mask (__m256 __X, __m256 __Y, const int __P)
12206 {
12207   return (__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf) __X,
12208                                                   (__v8sf) __Y, __P,
12209                                                   (__mmask8) -1);
12210 }
12211
12212 extern __inline __mmask8
12213 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12214 _mm256_mask_cmp_epi64_mask (__mmask8 __U, __m256i __X, __m256i __Y,
12215                             const int __P)
12216 {
12217   return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X,
12218                                                  (__v4di) __Y, __P,
12219                                                  (__mmask8) __U);
12220 }
12221
12222 extern __inline __mmask8
12223 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12224 _mm256_mask_cmp_epi32_mask (__mmask8 __U, __m256i __X, __m256i __Y,
12225                             const int __P)
12226 {
12227   return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X,
12228                                                  (__v8si) __Y, __P,
12229                                                  (__mmask8) __U);
12230 }
12231
12232 extern __inline __mmask8
12233 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12234 _mm256_mask_cmp_epu64_mask (__mmask8 __U, __m256i __X, __m256i __Y,
12235                             const int __P)
12236 {
12237   return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X,
12238                                                   (__v4di) __Y, __P,
12239                                                   (__mmask8) __U);
12240 }
12241
12242 extern __inline __mmask8
12243 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12244 _mm256_mask_cmp_epu32_mask (__mmask8 __U, __m256i __X, __m256i __Y,
12245                             const int __P)
12246 {
12247   return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X,
12248                                                   (__v8si) __Y, __P,
12249                                                   (__mmask8) __U);
12250 }
12251
12252 extern __inline __mmask8
12253 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12254 _mm256_mask_cmp_pd_mask (__mmask8 __U, __m256d __X, __m256d __Y,
12255                          const int __P)
12256 {
12257   return (__mmask8) __builtin_ia32_cmppd256_mask ((__v4df) __X,
12258                                                   (__v4df) __Y, __P,
12259                                                   (__mmask8) __U);
12260 }
12261
12262 extern __inline __mmask8
12263 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12264 _mm256_mask_cmp_ps_mask (__mmask8 __U, __m256 __X, __m256 __Y,
12265                          const int __P)
12266 {
12267   return (__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf) __X,
12268                                                   (__v8sf) __Y, __P,
12269                                                   (__mmask8) __U);
12270 }
12271
12272 extern __inline __mmask8
12273 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12274 _mm_cmp_epi64_mask (__m128i __X, __m128i __Y, const int __P)
12275 {
12276   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
12277                                                  (__v2di) __Y, __P,
12278                                                  (__mmask8) -1);
12279 }
12280
12281 extern __inline __mmask8
12282 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12283 _mm_cmp_epi32_mask (__m128i __X, __m128i __Y, const int __P)
12284 {
12285   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
12286                                                  (__v4si) __Y, __P,
12287                                                  (__mmask8) -1);
12288 }
12289
12290 extern __inline __mmask8
12291 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12292 _mm_cmp_epu64_mask (__m128i __X, __m128i __Y, const int __P)
12293 {
12294   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
12295                                                   (__v2di) __Y, __P,
12296                                                   (__mmask8) -1);
12297 }
12298
12299 extern __inline __mmask8
12300 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12301 _mm_cmp_epu32_mask (__m128i __X, __m128i __Y, const int __P)
12302 {
12303   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
12304                                                   (__v4si) __Y, __P,
12305                                                   (__mmask8) -1);
12306 }
12307
12308 extern __inline __mmask8
12309 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12310 _mm_cmp_pd_mask (__m128d __X, __m128d __Y, const int __P)
12311 {
12312   return (__mmask8) __builtin_ia32_cmppd128_mask ((__v2df) __X,
12313                                                   (__v2df) __Y, __P,
12314                                                   (__mmask8) -1);
12315 }
12316
12317 extern __inline __mmask8
12318 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12319 _mm_cmp_ps_mask (__m128 __X, __m128 __Y, const int __P)
12320 {
12321   return (__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf) __X,
12322                                                   (__v4sf) __Y, __P,
12323                                                   (__mmask8) -1);
12324 }
12325
12326 extern __inline __mmask8
12327 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12328 _mm_mask_cmp_epi64_mask (__mmask8 __U, __m128i __X, __m128i __Y,
12329                          const int __P)
12330 {
12331   return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X,
12332                                                  (__v2di) __Y, __P,
12333                                                  (__mmask8) __U);
12334 }
12335
12336 extern __inline __mmask8
12337 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12338 _mm_mask_cmp_epi32_mask (__mmask8 __U, __m128i __X, __m128i __Y,
12339                          const int __P)
12340 {
12341   return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X,
12342                                                  (__v4si) __Y, __P,
12343                                                  (__mmask8) __U);
12344 }
12345
12346 extern __inline __mmask8
12347 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12348 _mm_mask_cmp_epu64_mask (__mmask8 __U, __m128i __X, __m128i __Y,
12349                          const int __P)
12350 {
12351   return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X,
12352                                                   (__v2di) __Y, __P,
12353                                                   (__mmask8) __U);
12354 }
12355
12356 extern __inline __mmask8
12357 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12358 _mm_mask_cmp_epu32_mask (__mmask8 __U, __m128i __X, __m128i __Y,
12359                          const int __P)
12360 {
12361   return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X,
12362                                                   (__v4si) __Y, __P,
12363                                                   (__mmask8) __U);
12364 }
12365
12366 extern __inline __mmask8
12367 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12368 _mm_mask_cmp_pd_mask (__mmask8 __U, __m128d __X, __m128d __Y,
12369                       const int __P)
12370 {
12371   return (__mmask8) __builtin_ia32_cmppd128_mask ((__v2df) __X,
12372                                                   (__v2df) __Y, __P,
12373                                                   (__mmask8) __U);
12374 }
12375
12376 extern __inline __mmask8
12377 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12378 _mm_mask_cmp_ps_mask (__mmask8 __U, __m128 __X, __m128 __Y,
12379                       const int __P)
12380 {
12381   return (__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf) __X,
12382                                                   (__v4sf) __Y, __P,
12383                                                   (__mmask8) __U);
12384 }
12385
12386 extern __inline __m256d
12387 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
12388 _mm256_permutex_pd (__m256d __X, const int __M)
12389 {
12390   return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __M,
12391                                                   (__v4df)
12392                                                   _mm256_undefined_pd (),
12393                                                   (__mmask8) -1);
12394 }
12395
12396 #else
12397 #define _mm256_permutex_pd(X, M)                                                \
12398   ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(X), (int)(M),     \
12399                                             (__v4df)(__m256d)                   \
12400                                             _mm256_undefined_pd (),             \
12401                                             (__mmask8)-1))
12402
12403 #define _mm256_permutex_epi64(X, I)               \
12404   ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X), \
12405                                             (int)(I),           \
12406                                             (__v4di)(__m256i)   \
12407                                             (_mm256_setzero_si256 ()),\
12408                                             (__mmask8) -1))
12409
12410 #define _mm256_maskz_permutex_epi64(M, X, I)                    \
12411   ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X),    \
12412                                             (int)(I),                \
12413                                             (__v4di)(__m256i)        \
12414                                             (_mm256_setzero_si256 ()),\
12415                                             (__mmask8)(M)))
12416
12417 #define _mm256_mask_permutex_epi64(W, M, X, I)               \
12418   ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X), \
12419                                             (int)(I),             \
12420                                             (__v4di)(__m256i)(W), \
12421                                             (__mmask8)(M)))
12422
12423 #define _mm256_insertf32x4(X, Y, C)                                     \
12424   ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X),  \
12425     (__v4sf)(__m128) (Y), (int) (C),                                    \
12426     (__v8sf)(__m256)_mm256_setzero_ps (),                               \
12427     (__mmask8)-1))
12428
12429 #define _mm256_mask_insertf32x4(W, U, X, Y, C)                          \
12430   ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X),  \
12431     (__v4sf)(__m128) (Y), (int) (C),                                    \
12432     (__v8sf)(__m256)(W),                                                \
12433     (__mmask8)(U)))
12434
12435 #define _mm256_maskz_insertf32x4(U, X, Y, C)                            \
12436   ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X),  \
12437     (__v4sf)(__m128) (Y), (int) (C),                                    \
12438     (__v8sf)(__m256)_mm256_setzero_ps (),                               \
12439     (__mmask8)(U)))
12440
12441 #define _mm256_inserti32x4(X, Y, C)                                     \
12442   ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\
12443     (__v4si)(__m128i) (Y), (int) (C),                                   \
12444     (__v8si)(__m256i)_mm256_setzero_si256 (),                           \
12445     (__mmask8)-1))
12446
12447 #define _mm256_mask_inserti32x4(W, U, X, Y, C)                          \
12448   ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\
12449     (__v4si)(__m128i) (Y), (int) (C),                                   \
12450     (__v8si)(__m256i)(W),                                               \
12451     (__mmask8)(U)))
12452
12453 #define _mm256_maskz_inserti32x4(U, X, Y, C)                            \
12454   ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\
12455     (__v4si)(__m128i) (Y), (int) (C),                                   \
12456     (__v8si)(__m256i)_mm256_setzero_si256 (),                           \
12457     (__mmask8)(U)))
12458
12459 #define _mm256_extractf32x4_ps(X, C)                                    \
12460   ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \
12461     (int) (C),                                                          \
12462     (__v4sf)(__m128)_mm_setzero_ps (),                                  \
12463     (__mmask8)-1))
12464
12465 #define _mm256_mask_extractf32x4_ps(W, U, X, C)                         \
12466   ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \
12467     (int) (C),                                                          \
12468     (__v4sf)(__m128)(W),                                                \
12469     (__mmask8)(U)))
12470
12471 #define _mm256_maskz_extractf32x4_ps(U, X, C)                           \
12472   ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \
12473     (int) (C),                                                          \
12474     (__v4sf)(__m128)_mm_setzero_ps (),                                  \
12475     (__mmask8)(U)))
12476
12477 #define _mm256_extracti32x4_epi32(X, C)                                 \
12478   ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\
12479     (int) (C), (__v4si)(__m128i)_mm_setzero_si128 (), (__mmask8)-1))
12480
12481 #define _mm256_mask_extracti32x4_epi32(W, U, X, C)                      \
12482   ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\
12483     (int) (C), (__v4si)(__m128i)(W), (__mmask8)(U)))
12484
12485 #define _mm256_maskz_extracti32x4_epi32(U, X, C)                        \
12486   ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\
12487     (int) (C), (__v4si)(__m128i)_mm_setzero_si128 (), (__mmask8)(U)))
12488
12489 #define _mm256_shuffle_i64x2(X, Y, C)                                                   \
12490   ((__m256i)  __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X),                 \
12491                                                   (__v4di)(__m256i)(Y), (int)(C),       \
12492                                                   (__v4di)(__m256i)_mm256_setzero_si256 (), \
12493                                                   (__mmask8)-1))
12494
12495 #define _mm256_mask_shuffle_i64x2(W, U, X, Y, C)                                        \
12496   ((__m256i)  __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X),                 \
12497                                                   (__v4di)(__m256i)(Y), (int)(C),       \
12498                                                   (__v4di)(__m256i)(W),\
12499                                                   (__mmask8)(U)))
12500
12501 #define _mm256_maskz_shuffle_i64x2(U, X, Y, C)                                          \
12502   ((__m256i)  __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X),                 \
12503                                                   (__v4di)(__m256i)(Y), (int)(C),       \
12504                                                   (__v4di)(__m256i)_mm256_setzero_si256 (), \
12505                                                   (__mmask8)(U)))
12506
12507 #define _mm256_shuffle_i32x4(X, Y, C)                                                   \
12508   ((__m256i)  __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X),                 \
12509                                                   (__v8si)(__m256i)(Y), (int)(C),       \
12510                                                   (__v8si)(__m256i)                     \
12511                                                   _mm256_setzero_si256 (),              \
12512                                                   (__mmask8)-1))
12513
12514 #define _mm256_mask_shuffle_i32x4(W, U, X, Y, C)                                        \
12515   ((__m256i)  __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X),                 \
12516                                                   (__v8si)(__m256i)(Y), (int)(C),       \
12517                                                   (__v8si)(__m256i)(W),                 \
12518                                                   (__mmask8)(U)))
12519
12520 #define _mm256_maskz_shuffle_i32x4(U, X, Y, C)                                          \
12521   ((__m256i)  __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X),                 \
12522                                                   (__v8si)(__m256i)(Y), (int)(C),       \
12523                                                   (__v8si)(__m256i)                     \
12524                                                   _mm256_setzero_si256 (),              \
12525                                                   (__mmask8)(U)))
12526
12527 #define _mm256_shuffle_f64x2(X, Y, C)                                                   \
12528   ((__m256d)  __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X),                 \
12529                                                   (__v4df)(__m256d)(Y), (int)(C),       \
12530                                                   (__v4df)(__m256d)_mm256_setzero_pd (),\
12531                                                   (__mmask8)-1))
12532
12533 #define _mm256_mask_shuffle_f64x2(W, U, X, Y, C)                                        \
12534   ((__m256d)  __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X),                 \
12535                                                   (__v4df)(__m256d)(Y), (int)(C),       \
12536                                                   (__v4df)(__m256d)(W),                 \
12537                                                   (__mmask8)(U)))
12538
12539 #define _mm256_maskz_shuffle_f64x2(U, X, Y, C)                                          \
12540   ((__m256d)  __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X),                 \
12541                                                   (__v4df)(__m256d)(Y), (int)(C),       \
12542                                                   (__v4df)(__m256d)_mm256_setzero_pd( ),\
12543                                                   (__mmask8)(U)))
12544
12545 #define _mm256_shuffle_f32x4(X, Y, C)                                                   \
12546   ((__m256)  __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X),                   \
12547                                                  (__v8sf)(__m256)(Y), (int)(C),         \
12548                                                  (__v8sf)(__m256)_mm256_setzero_ps (),  \
12549                                                  (__mmask8)-1))
12550
12551 #define _mm256_mask_shuffle_f32x4(W, U, X, Y, C)                                        \
12552   ((__m256)  __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X),                   \
12553                                                  (__v8sf)(__m256)(Y), (int)(C),         \
12554                                                  (__v8sf)(__m256)(W),                   \
12555                                                  (__mmask8)(U)))
12556
12557 #define _mm256_maskz_shuffle_f32x4(U, X, Y, C)                                          \
12558   ((__m256)  __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X),                   \
12559                                                  (__v8sf)(__m256)(Y), (int)(C),         \
12560                                                  (__v8sf)(__m256)_mm256_setzero_ps (),  \
12561                                                  (__mmask8)(U)))
12562
12563 #define _mm256_mask_shuffle_pd(W, U, A, B, C)                                   \
12564   ((__m256d)__builtin_ia32_shufpd256_mask ((__v4df)(__m256d)(A),                \
12565                                            (__v4df)(__m256d)(B), (int)(C),      \
12566                                            (__v4df)(__m256d)(W),                \
12567                                            (__mmask8)(U)))
12568
12569 #define _mm256_maskz_shuffle_pd(U, A, B, C)                                     \
12570   ((__m256d)__builtin_ia32_shufpd256_mask ((__v4df)(__m256d)(A),                \
12571                                            (__v4df)(__m256d)(B), (int)(C),      \
12572                                            (__v4df)(__m256d)                    \
12573                                            _mm256_setzero_pd (),                \
12574                                            (__mmask8)(U)))
12575
12576 #define _mm_mask_shuffle_pd(W, U, A, B, C)                                      \
12577   ((__m128d)__builtin_ia32_shufpd128_mask ((__v2df)(__m128d)(A),                \
12578                                            (__v2df)(__m128d)(B), (int)(C),      \
12579                                            (__v2df)(__m128d)(W),                \
12580                                            (__mmask8)(U)))
12581
12582 #define _mm_maskz_shuffle_pd(U, A, B, C)                                        \
12583   ((__m128d)__builtin_ia32_shufpd128_mask ((__v2df)(__m128d)(A),                \
12584                                            (__v2df)(__m128d)(B), (int)(C),      \
12585                                            (__v2df)(__m128d)_mm_setzero_pd (),  \
12586                                            (__mmask8)(U)))
12587
12588 #define _mm256_mask_shuffle_ps(W, U, A, B, C)                                   \
12589   ((__m256) __builtin_ia32_shufps256_mask ((__v8sf)(__m256)(A),                 \
12590                                            (__v8sf)(__m256)(B), (int)(C),       \
12591                                            (__v8sf)(__m256)(W),                 \
12592                                            (__mmask8)(U)))
12593
12594 #define _mm256_maskz_shuffle_ps(U, A, B, C)                                     \
12595   ((__m256) __builtin_ia32_shufps256_mask ((__v8sf)(__m256)(A),                 \
12596                                            (__v8sf)(__m256)(B), (int)(C),       \
12597                                            (__v8sf)(__m256)_mm256_setzero_ps (),\
12598                                            (__mmask8)(U)))
12599
12600 #define _mm_mask_shuffle_ps(W, U, A, B, C)                                      \
12601   ((__m128) __builtin_ia32_shufps128_mask ((__v4sf)(__m128)(A),                 \
12602                                            (__v4sf)(__m128)(B), (int)(C),       \
12603                                            (__v4sf)(__m128)(W),                 \
12604                                            (__mmask8)(U)))
12605
12606 #define _mm_maskz_shuffle_ps(U, A, B, C)                                        \
12607   ((__m128) __builtin_ia32_shufps128_mask ((__v4sf)(__m128)(A),                 \
12608                                            (__v4sf)(__m128)(B), (int)(C),       \
12609                                            (__v4sf)(__m128)_mm_setzero_ps (),   \
12610                                            (__mmask8)(U)))
12611
12612 #define _mm256_fixupimm_pd(X, Y, Z, C)                                          \
12613   ((__m256d)__builtin_ia32_fixupimmpd256_mask ((__v4df)(__m256d)(X),            \
12614                                                (__v4df)(__m256d)(Y),            \
12615                                                (__v4di)(__m256i)(Z), (int)(C),  \
12616                                                (__mmask8)(-1)))
12617
12618 #define _mm256_mask_fixupimm_pd(X, U, Y, Z, C)                                  \
12619    ((__m256d)__builtin_ia32_fixupimmpd256_mask ((__v4df)(__m256d)(X),           \
12620                                                 (__v4df)(__m256d)(Y),           \
12621                                                 (__v4di)(__m256i)(Z), (int)(C), \
12622                                                 (__mmask8)(U)))
12623
12624 #define _mm256_maskz_fixupimm_pd(U, X, Y, Z, C)                                 \
12625    ((__m256d)__builtin_ia32_fixupimmpd256_maskz ((__v4df)(__m256d)(X),          \
12626                                                  (__v4df)(__m256d)(Y),          \
12627                                                  (__v4di)(__m256i)(Z), (int)(C),\
12628                                                  (__mmask8)(U)))
12629
12630 #define _mm256_fixupimm_ps(X, Y, Z, C)                                          \
12631   ((__m256)__builtin_ia32_fixupimmps256_mask ((__v8sf)(__m256)(X),              \
12632                                               (__v8sf)(__m256)(Y),              \
12633                                               (__v8si)(__m256i)(Z), (int)(C),   \
12634                                               (__mmask8)(-1)))
12635
12636
12637 #define _mm256_mask_fixupimm_ps(X, U, Y, Z, C)                                  \
12638     ((__m256)__builtin_ia32_fixupimmps256_mask ((__v8sf)(__m256)(X),            \
12639                                                 (__v8sf)(__m256)(Y),            \
12640                                                 (__v8si)(__m256i)(Z), (int)(C), \
12641                                                 (__mmask8)(U)))
12642
12643 #define _mm256_maskz_fixupimm_ps(U, X, Y, Z, C)                                 \
12644     ((__m256)__builtin_ia32_fixupimmps256_maskz ((__v8sf)(__m256)(X),           \
12645                                                  (__v8sf)(__m256)(Y),           \
12646                                                  (__v8si)(__m256i)(Z), (int)(C),\
12647                                                  (__mmask8)(U)))
12648
12649 #define _mm_fixupimm_pd(X, Y, Z, C)                                             \
12650   ((__m128d)__builtin_ia32_fixupimmpd128_mask ((__v2df)(__m128d)(X),            \
12651                                                (__v2df)(__m128d)(Y),            \
12652                                                (__v2di)(__m128i)(Z), (int)(C),  \
12653                                                (__mmask8)(-1)))
12654
12655
12656 #define _mm_mask_fixupimm_pd(X, U, Y, Z, C)                                       \
12657      ((__m128d)__builtin_ia32_fixupimmpd128_mask ((__v2df)(__m128d)(X),           \
12658                                                   (__v2df)(__m128d)(Y),           \
12659                                                   (__v2di)(__m128i)(Z), (int)(C), \
12660                                                   (__mmask8)(U)))
12661
12662 #define _mm_maskz_fixupimm_pd(U, X, Y, Z, C)                                      \
12663      ((__m128d)__builtin_ia32_fixupimmpd128_maskz ((__v2df)(__m128d)(X),          \
12664                                                    (__v2df)(__m128d)(Y),          \
12665                                                    (__v2di)(__m128i)(Z), (int)(C),\
12666                                                    (__mmask8)(U)))
12667
12668 #define _mm_fixupimm_ps(X, Y, Z, C)                                             \
12669    ((__m128)__builtin_ia32_fixupimmps128_mask ((__v4sf)(__m128)(X),             \
12670                                                (__v4sf)(__m128)(Y),             \
12671                                                (__v4si)(__m128i)(Z), (int)(C),  \
12672                                                (__mmask8)(-1)))
12673
12674 #define _mm_mask_fixupimm_ps(X, U, Y, Z, C)                                      \
12675       ((__m128)__builtin_ia32_fixupimmps128_mask ((__v4sf)(__m128)(X),           \
12676                                                   (__v4sf)(__m128)(Y),           \
12677                                                   (__v4si)(__m128i)(Z), (int)(C),\
12678                                                   (__mmask8)(U)))
12679
12680 #define _mm_maskz_fixupimm_ps(U, X, Y, Z, C)                                      \
12681       ((__m128)__builtin_ia32_fixupimmps128_maskz ((__v4sf)(__m128)(X),           \
12682                                                    (__v4sf)(__m128)(Y),           \
12683                                                    (__v4si)(__m128i)(Z), (int)(C),\
12684                                                    (__mmask8)(U)))
12685
12686 #define _mm256_mask_srli_epi32(W, U, A, B)                              \
12687   ((__m256i) __builtin_ia32_psrldi256_mask ((__v8si)(__m256i)(A),       \
12688     (int)(B), (__v8si)(__m256i)(W), (__mmask8)(U)))
12689
12690 #define _mm256_maskz_srli_epi32(U, A, B)                                \
12691   ((__m256i) __builtin_ia32_psrldi256_mask ((__v8si)(__m256i)(A),       \
12692     (int)(B), (__v8si)_mm256_setzero_si256 (), (__mmask8)(U)))
12693
12694 #define _mm_mask_srli_epi32(W, U, A, B)                                 \
12695   ((__m128i) __builtin_ia32_psrldi128_mask ((__v4si)(__m128i)(A),       \
12696     (int)(B), (__v4si)(__m128i)(W), (__mmask8)(U)))
12697
12698 #define _mm_maskz_srli_epi32(U, A, B)                                   \
12699   ((__m128i) __builtin_ia32_psrldi128_mask ((__v4si)(__m128i)(A),       \
12700     (int)(B), (__v4si)_mm_setzero_si128 (), (__mmask8)(U)))
12701
12702 #define _mm256_mask_srli_epi64(W, U, A, B)                              \
12703   ((__m256i) __builtin_ia32_psrlqi256_mask ((__v4di)(__m256i)(A),       \
12704     (int)(B), (__v4di)(__m256i)(W), (__mmask8)(U)))
12705
12706 #define _mm256_maskz_srli_epi64(U, A, B)                                \
12707   ((__m256i) __builtin_ia32_psrlqi256_mask ((__v4di)(__m256i)(A),       \
12708     (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)(U)))
12709
12710 #define _mm_mask_srli_epi64(W, U, A, B)                                 \
12711   ((__m128i) __builtin_ia32_psrlqi128_mask ((__v2di)(__m128i)(A),       \
12712     (int)(B), (__v2di)(__m128i)(W), (__mmask8)(U)))
12713
12714 #define _mm_maskz_srli_epi64(U, A, B)                                   \
12715   ((__m128i) __builtin_ia32_psrlqi128_mask ((__v2di)(__m128i)(A),       \
12716     (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)(U)))
12717
12718 #define _mm256_mask_slli_epi32(W, U, X, C)                                \
12719   ((__m256i)__builtin_ia32_pslldi256_mask ((__v8si)(__m256i)(X), (int)(C),\
12720     (__v8si)(__m256i)(W),                                                 \
12721     (__mmask8)(U)))
12722
12723 #define _mm256_maskz_slli_epi32(U, X, C)                                  \
12724   ((__m256i)__builtin_ia32_pslldi256_mask ((__v8si)(__m256i)(X), (int)(C),\
12725     (__v8si)(__m256i)_mm256_setzero_si256 (),                             \
12726     (__mmask8)(U)))
12727
12728 #define _mm256_mask_slli_epi64(W, U, X, C)                                \
12729   ((__m256i)__builtin_ia32_psllqi256_mask ((__v4di)(__m256i)(X), (int)(C),\
12730     (__v4di)(__m256i)(W),                                                 \
12731     (__mmask8)(U)))
12732
12733 #define _mm256_maskz_slli_epi64(U, X, C)                                  \
12734   ((__m256i)__builtin_ia32_psllqi256_mask ((__v4di)(__m256i)(X), (int)(C),\
12735     (__v4di)(__m256i)_mm256_setzero_si256 (),                             \
12736     (__mmask8)(U)))
12737
12738 #define _mm_mask_slli_epi32(W, U, X, C)                                   \
12739   ((__m128i)__builtin_ia32_pslldi128_mask ((__v4si)(__m128i)(X), (int)(C),\
12740     (__v4si)(__m128i)(W),\
12741     (__mmask8)(U)))
12742
12743 #define _mm_maskz_slli_epi32(U, X, C)                                     \
12744   ((__m128i)__builtin_ia32_pslldi128_mask ((__v4si)(__m128i)(X), (int)(C),\
12745     (__v4si)(__m128i)_mm_setzero_si128 (),\
12746     (__mmask8)(U)))
12747
12748 #define _mm_mask_slli_epi64(W, U, X, C)                                   \
12749   ((__m128i)__builtin_ia32_psllqi128_mask ((__v2di)(__m128i)(X), (int)(C),\
12750     (__v2di)(__m128i)(W),\
12751     (__mmask8)(U)))
12752
12753 #define _mm_maskz_slli_epi64(U, X, C)                                     \
12754   ((__m128i)__builtin_ia32_psllqi128_mask ((__v2di)(__m128i)(X), (int)(C),\
12755     (__v2di)(__m128i)_mm_setzero_si128 (),\
12756     (__mmask8)(U)))
12757
12758 #define _mm256_ternarylogic_epi64(A, B, C, I)                           \
12759   ((__m256i) __builtin_ia32_pternlogq256_mask ((__v4di)(__m256i)(A),    \
12760     (__v4di)(__m256i)(B), (__v4di)(__m256i)(C), (int)(I), (__mmask8)-1))
12761
12762 #define _mm256_mask_ternarylogic_epi64(A, U, B, C, I)                   \
12763   ((__m256i) __builtin_ia32_pternlogq256_mask ((__v4di)(__m256i)(A),    \
12764     (__v4di)(__m256i)(B), (__v4di)(__m256i)(C), (int)(I), (__mmask8)(U)))
12765
12766 #define _mm256_maskz_ternarylogic_epi64(U, A, B, C, I)                  \
12767   ((__m256i) __builtin_ia32_pternlogq256_maskz ((__v4di)(__m256i)(A),   \
12768     (__v4di)(__m256i)(B), (__v4di)(__m256i)(C), (int)(I), (__mmask8)(U)))
12769
12770 #define _mm256_ternarylogic_epi32(A, B, C, I)                           \
12771   ((__m256i) __builtin_ia32_pternlogd256_mask ((__v8si)(__m256i)(A),    \
12772     (__v8si)(__m256i)(B), (__v8si)(__m256i)(C), (int)(I), (__mmask8)-1))
12773
12774 #define _mm256_mask_ternarylogic_epi32(A, U, B, C, I)                   \
12775   ((__m256i) __builtin_ia32_pternlogd256_mask ((__v8si)(__m256i)(A),    \
12776     (__v8si)(__m256i)(B), (__v8si)(__m256i)(C), (int)(I), (__mmask8)(U)))
12777
12778 #define _mm256_maskz_ternarylogic_epi32(U, A, B, C, I)                  \
12779   ((__m256i) __builtin_ia32_pternlogd256_maskz ((__v8si)(__m256i)(A),   \
12780     (__v8si)(__m256i)(B), (__v8si)(__m256i)(C), (int)(I), (__mmask8)(U)))
12781
12782 #define _mm_ternarylogic_epi64(A, B, C, I)                              \
12783   ((__m128i) __builtin_ia32_pternlogq128_mask ((__v2di)(__m128i)(A),    \
12784     (__v2di)(__m128i)(B), (__v2di)(__m128i)(C), (int)(I), (__mmask8)-1))
12785
12786 #define _mm_mask_ternarylogic_epi64(A, U, B, C, I)                      \
12787   ((__m128i) __builtin_ia32_pternlogq128_mask ((__v2di)(__m128i)(A),    \
12788     (__v2di)(__m128i)(B), (__v2di)(__m128i)(C), (int)(I), (__mmask8)(U)))
12789
12790 #define _mm_maskz_ternarylogic_epi64(U, A, B, C, I)                     \
12791   ((__m128i) __builtin_ia32_pternlogq128_maskz ((__v2di)(__m128i)(A),   \
12792     (__v2di)(__m128i)(B), (__v2di)(__m128i)(C), (int)(I), (__mmask8)(U)))
12793
12794 #define _mm_ternarylogic_epi32(A, B, C, I)                              \
12795   ((__m128i) __builtin_ia32_pternlogd128_mask ((__v4si)(__m128i)(A),    \
12796     (__v4si)(__m128i)(B), (__v4si)(__m128i)(C), (int)(I), (__mmask8)-1))
12797
12798 #define _mm_mask_ternarylogic_epi32(A, U, B, C, I)                      \
12799   ((__m128i) __builtin_ia32_pternlogd128_mask ((__v4si)(__m128i)(A),    \
12800     (__v4si)(__m128i)(B), (__v4si)(__m128i)(C), (int)(I), (__mmask8)(U)))
12801
12802 #define _mm_maskz_ternarylogic_epi32(U, A, B, C, I)                     \
12803   ((__m128i) __builtin_ia32_pternlogd128_maskz ((__v4si)(__m128i)(A),   \
12804     (__v4si)(__m128i)(B), (__v4si)(__m128i)(C), (int)(I), (__mmask8)(U)))
12805
12806 #define _mm256_roundscale_ps(A, B)                                      \
12807   ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A),    \
12808     (int)(B), (__v8sf)(__m256)_mm256_setzero_ps (), (__mmask8)-1))
12809
12810 #define _mm256_mask_roundscale_ps(W, U, A, B)                           \
12811   ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A),    \
12812     (int)(B), (__v8sf)(__m256)(W), (__mmask8)(U)))
12813
12814 #define _mm256_maskz_roundscale_ps(U, A, B)                             \
12815   ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A),    \
12816     (int)(B), (__v8sf)(__m256)_mm256_setzero_ps (), (__mmask8)(U)))
12817
12818 #define _mm256_roundscale_pd(A, B)                                      \
12819   ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A),  \
12820     (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)-1))
12821
12822 #define _mm256_mask_roundscale_pd(W, U, A, B)                           \
12823   ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A),  \
12824     (int)(B), (__v4df)(__m256d)(W), (__mmask8)(U)))
12825
12826 #define _mm256_maskz_roundscale_pd(U, A, B)                             \
12827   ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A),  \
12828     (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)(U)))
12829
12830 #define _mm_roundscale_ps(A, B)                                         \
12831   ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A),    \
12832     (int)(B), (__v4sf)(__m128)_mm_setzero_ps (), (__mmask8)-1))
12833
12834 #define _mm_mask_roundscale_ps(W, U, A, B)                              \
12835   ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A),    \
12836     (int)(B), (__v4sf)(__m128)(W), (__mmask8)(U)))
12837
12838 #define _mm_maskz_roundscale_ps(U, A, B)                                \
12839   ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A),    \
12840     (int)(B), (__v4sf)(__m128)_mm_setzero_ps (), (__mmask8)(U)))
12841
12842 #define _mm_roundscale_pd(A, B)                                         \
12843   ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A),  \
12844     (int)(B), (__v2df)(__m128d)_mm_setzero_pd (), (__mmask8)-1))
12845
12846 #define _mm_mask_roundscale_pd(W, U, A, B)                              \
12847   ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A),  \
12848     (int)(B), (__v2df)(__m128d)(W), (__mmask8)(U)))
12849
12850 #define _mm_maskz_roundscale_pd(U, A, B)                                \
12851   ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A),  \
12852     (int)(B), (__v2df)(__m128d)_mm_setzero_pd (), (__mmask8)(U)))
12853
12854 #define _mm256_getmant_ps(X, B, C)                                              \
12855   ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X),             \
12856                                          (int)(((C)<<2) | (B)),                 \
12857                                          (__v8sf)(__m256)_mm256_setzero_ps (),  \
12858                                          (__mmask8)-1))
12859
12860 #define _mm256_mask_getmant_ps(W, U, X, B, C)                                   \
12861   ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X),             \
12862                                          (int)(((C)<<2) | (B)),                 \
12863                                          (__v8sf)(__m256)(W),                   \
12864                                          (__mmask8)(U)))
12865
12866 #define _mm256_maskz_getmant_ps(U, X, B, C)                                     \
12867   ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X),             \
12868                                          (int)(((C)<<2) | (B)),                 \
12869                                          (__v8sf)(__m256)_mm256_setzero_ps (),  \
12870                                          (__mmask8)(U)))
12871
12872 #define _mm_getmant_ps(X, B, C)                                                 \
12873   ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X),             \
12874                                          (int)(((C)<<2) | (B)),                 \
12875                                          (__v4sf)(__m128)_mm_setzero_ps (),     \
12876                                          (__mmask8)-1))
12877
12878 #define _mm_mask_getmant_ps(W, U, X, B, C)                                      \
12879   ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X),             \
12880                                          (int)(((C)<<2) | (B)),                 \
12881                                          (__v4sf)(__m128)(W),                   \
12882                                          (__mmask8)(U)))
12883
12884 #define _mm_maskz_getmant_ps(U, X, B, C)                                        \
12885   ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X),             \
12886                                          (int)(((C)<<2) | (B)),                 \
12887                                          (__v4sf)(__m128)_mm_setzero_ps (),     \
12888                                          (__mmask8)(U)))
12889
12890 #define _mm256_getmant_pd(X, B, C)                                              \
12891   ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X),           \
12892                                          (int)(((C)<<2) | (B)),                 \
12893                                           (__v4df)(__m256d)_mm256_setzero_pd (),\
12894                                           (__mmask8)-1))
12895
12896 #define _mm256_mask_getmant_pd(W, U, X, B, C)                                   \
12897   ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X),           \
12898                                          (int)(((C)<<2) | (B)),                 \
12899                                           (__v4df)(__m256d)(W),                 \
12900                                           (__mmask8)(U)))
12901
12902 #define _mm256_maskz_getmant_pd(U, X, B, C)                                     \
12903   ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X),           \
12904                                          (int)(((C)<<2) | (B)),                 \
12905                                           (__v4df)(__m256d)_mm256_setzero_pd (),\
12906                                           (__mmask8)(U)))
12907
12908 #define _mm_getmant_pd(X, B, C)                                                 \
12909   ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X),           \
12910                                          (int)(((C)<<2) | (B)),                 \
12911                                           (__v2df)(__m128d)_mm_setzero_pd (),   \
12912                                           (__mmask8)-1))
12913
12914 #define _mm_mask_getmant_pd(W, U, X, B, C)                                      \
12915   ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X),           \
12916                                          (int)(((C)<<2) | (B)),                 \
12917                                           (__v2df)(__m128d)(W),                 \
12918                                           (__mmask8)(U)))
12919
12920 #define _mm_maskz_getmant_pd(U, X, B, C)                                        \
12921   ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X),           \
12922                                          (int)(((C)<<2) | (B)),                 \
12923                                           (__v2df)(__m128d)_mm_setzero_pd (),   \
12924                                           (__mmask8)(U)))
12925
12926 #define _mm256_mmask_i32gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12927   (__m256) __builtin_ia32_gather3siv8sf ((__v8sf)(__m256)V1OLD,         \
12928                                          (void const *)ADDR,            \
12929                                          (__v8si)(__m256i)INDEX,        \
12930                                          (__mmask8)MASK, (int)SCALE)
12931
12932 #define _mm_mmask_i32gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE)         \
12933   (__m128) __builtin_ia32_gather3siv4sf ((__v4sf)(__m128)V1OLD,         \
12934                                          (void const *)ADDR,            \
12935                                          (__v4si)(__m128i)INDEX,        \
12936                                          (__mmask8)MASK, (int)SCALE)
12937
12938 #define _mm256_mmask_i32gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12939   (__m256d) __builtin_ia32_gather3siv4df ((__v4df)(__m256d)V1OLD,       \
12940                                           (void const *)ADDR,           \
12941                                           (__v4si)(__m128i)INDEX,       \
12942                                           (__mmask8)MASK, (int)SCALE)
12943
12944 #define _mm_mmask_i32gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE)         \
12945   (__m128d) __builtin_ia32_gather3siv2df ((__v2df)(__m128d)V1OLD,       \
12946                                           (void const *)ADDR,           \
12947                                           (__v4si)(__m128i)INDEX,       \
12948                                           (__mmask8)MASK, (int)SCALE)
12949
12950 #define _mm256_mmask_i64gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12951   (__m128) __builtin_ia32_gather3div8sf ((__v4sf)(__m128)V1OLD,         \
12952                                          (void const *)ADDR,            \
12953                                          (__v4di)(__m256i)INDEX,        \
12954                                          (__mmask8)MASK, (int)SCALE)
12955
12956 #define _mm_mmask_i64gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE)         \
12957   (__m128) __builtin_ia32_gather3div4sf ((__v4sf)(__m128)V1OLD,         \
12958                                          (void const *)ADDR,            \
12959                                          (__v2di)(__m128i)INDEX,        \
12960                                          (__mmask8)MASK, (int)SCALE)
12961
12962 #define _mm256_mmask_i64gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12963   (__m256d) __builtin_ia32_gather3div4df ((__v4df)(__m256d)V1OLD,       \
12964                                           (void const *)ADDR,           \
12965                                           (__v4di)(__m256i)INDEX,       \
12966                                           (__mmask8)MASK, (int)SCALE)
12967
12968 #define _mm_mmask_i64gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE)         \
12969   (__m128d) __builtin_ia32_gather3div2df ((__v2df)(__m128d)V1OLD,       \
12970                                           (void const *)ADDR,           \
12971                                           (__v2di)(__m128i)INDEX,       \
12972                                           (__mmask8)MASK, (int)SCALE)
12973
12974 #define _mm256_mmask_i32gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE)   \
12975   (__m256i) __builtin_ia32_gather3siv8si ((__v8si)(__m256i)V1OLD,       \
12976                                           (void const *)ADDR,           \
12977                                           (__v8si)(__m256i)INDEX,       \
12978                                           (__mmask8)MASK, (int)SCALE)
12979
12980 #define _mm_mmask_i32gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12981   (__m128i) __builtin_ia32_gather3siv4si ((__v4si)(__m128i)V1OLD,       \
12982                                           (void const *)ADDR,           \
12983                                           (__v4si)(__m128i)INDEX,       \
12984                                           (__mmask8)MASK, (int)SCALE)
12985
12986 #define _mm256_mmask_i32gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE)   \
12987   (__m256i) __builtin_ia32_gather3siv4di ((__v4di)(__m256i)V1OLD,       \
12988                                           (void const *)ADDR,           \
12989                                           (__v4si)(__m128i)INDEX,       \
12990                                           (__mmask8)MASK, (int)SCALE)
12991
12992 #define _mm_mmask_i32gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE)      \
12993   (__m128i) __builtin_ia32_gather3siv2di ((__v2di)(__m128i)V1OLD,       \
12994                                           (void const *)ADDR,           \
12995                                           (__v4si)(__m128i)INDEX,       \
12996                                           (__mmask8)MASK, (int)SCALE)
12997
12998 #define _mm256_mmask_i64gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE)   \
12999   (__m128i) __builtin_ia32_gather3div8si ((__v4si)(__m128i)V1OLD,       \
13000                                           (void const *)ADDR,           \
13001                                           (__v4di)(__m256i)INDEX,       \
13002                                           (__mmask8)MASK, (int)SCALE)
13003
13004 #define _mm_mmask_i64gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE)      \
13005   (__m128i) __builtin_ia32_gather3div4si ((__v4si)(__m128i)V1OLD,       \
13006                                           (void const *)ADDR,           \
13007                                           (__v2di)(__m128i)INDEX,       \
13008                                           (__mmask8)MASK, (int)SCALE)
13009
13010 #define _mm256_mmask_i64gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE)   \
13011   (__m256i) __builtin_ia32_gather3div4di ((__v4di)(__m256i)V1OLD,       \
13012                                           (void const *)ADDR,           \
13013                                           (__v4di)(__m256i)INDEX,       \
13014                                           (__mmask8)MASK, (int)SCALE)
13015
13016 #define _mm_mmask_i64gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE)      \
13017   (__m128i) __builtin_ia32_gather3div2di ((__v2di)(__m128i)V1OLD,       \
13018                                           (void const *)ADDR,           \
13019                                           (__v2di)(__m128i)INDEX,       \
13020                                           (__mmask8)MASK, (int)SCALE)
13021
13022 #define _mm256_i32scatter_ps(ADDR, INDEX, V1, SCALE)                    \
13023   __builtin_ia32_scattersiv8sf ((void *)ADDR, (__mmask8)0xFF,           \
13024                                 (__v8si)(__m256i)INDEX,                 \
13025                                 (__v8sf)(__m256)V1, (int)SCALE)
13026
13027 #define _mm256_mask_i32scatter_ps(ADDR, MASK, INDEX, V1, SCALE)         \
13028   __builtin_ia32_scattersiv8sf ((void *)ADDR, (__mmask8)MASK,           \
13029                                 (__v8si)(__m256i)INDEX,                 \
13030                                 (__v8sf)(__m256)V1, (int)SCALE)
13031
13032 #define _mm_i32scatter_ps(ADDR, INDEX, V1, SCALE)                       \
13033   __builtin_ia32_scattersiv4sf ((void *)ADDR, (__mmask8)0xFF,           \
13034                                 (__v4si)(__m128i)INDEX,                 \
13035                                 (__v4sf)(__m128)V1, (int)SCALE)
13036
13037 #define _mm_mask_i32scatter_ps(ADDR, MASK, INDEX, V1, SCALE)            \
13038   __builtin_ia32_scattersiv4sf ((void *)ADDR, (__mmask8)MASK,           \
13039                                 (__v4si)(__m128i)INDEX,                 \
13040                                 (__v4sf)(__m128)V1, (int)SCALE)
13041
13042 #define _mm256_i32scatter_pd(ADDR, INDEX, V1, SCALE)                    \
13043   __builtin_ia32_scattersiv4df ((void *)ADDR, (__mmask8)0xFF,           \
13044                                 (__v4si)(__m128i)INDEX,                 \
13045                                 (__v4df)(__m256d)V1, (int)SCALE)
13046
13047 #define _mm256_mask_i32scatter_pd(ADDR, MASK, INDEX, V1, SCALE)         \
13048   __builtin_ia32_scattersiv4df ((void *)ADDR, (__mmask8)MASK,           \
13049                                 (__v4si)(__m128i)INDEX,                 \
13050                                 (__v4df)(__m256d)V1, (int)SCALE)
13051
13052 #define _mm_i32scatter_pd(ADDR, INDEX, V1, SCALE)                       \
13053   __builtin_ia32_scattersiv2df ((void *)ADDR, (__mmask8)0xFF,           \
13054                                 (__v4si)(__m128i)INDEX,                 \
13055                                 (__v2df)(__m128d)V1, (int)SCALE)
13056
13057 #define _mm_mask_i32scatter_pd(ADDR, MASK, INDEX, V1, SCALE)            \
13058   __builtin_ia32_scattersiv2df ((void *)ADDR, (__mmask8)MASK,           \
13059                                 (__v4si)(__m128i)INDEX,                 \
13060                                 (__v2df)(__m128d)V1, (int)SCALE)
13061
13062 #define _mm256_i64scatter_ps(ADDR, INDEX, V1, SCALE)                    \
13063   __builtin_ia32_scatterdiv8sf ((void *)ADDR, (__mmask8)0xFF,           \
13064                                 (__v4di)(__m256i)INDEX,                 \
13065                                 (__v4sf)(__m128)V1, (int)SCALE)
13066
13067 #define _mm256_mask_i64scatter_ps(ADDR, MASK, INDEX, V1, SCALE)         \
13068   __builtin_ia32_scatterdiv8sf ((void *)ADDR, (__mmask8)MASK,           \
13069                                 (__v4di)(__m256i)INDEX,                 \
13070                                 (__v4sf)(__m128)V1, (int)SCALE)
13071
13072 #define _mm_i64scatter_ps(ADDR, INDEX, V1, SCALE)                       \
13073   __builtin_ia32_scatterdiv4sf ((void *)ADDR, (__mmask8)0xFF,           \
13074                                 (__v2di)(__m128i)INDEX,                 \
13075                                 (__v4sf)(__m128)V1, (int)SCALE)
13076
13077 #define _mm_mask_i64scatter_ps(ADDR, MASK, INDEX, V1, SCALE)            \
13078   __builtin_ia32_scatterdiv4sf ((void *)ADDR, (__mmask8)MASK,           \
13079                                 (__v2di)(__m128i)INDEX,                 \
13080                                 (__v4sf)(__m128)V1, (int)SCALE)
13081
13082 #define _mm256_i64scatter_pd(ADDR, INDEX, V1, SCALE)                    \
13083   __builtin_ia32_scatterdiv4df ((void *)ADDR, (__mmask8)0xFF,           \
13084                                 (__v4di)(__m256i)INDEX,                 \
13085                                 (__v4df)(__m256d)V1, (int)SCALE)
13086
13087 #define _mm256_mask_i64scatter_pd(ADDR, MASK, INDEX, V1, SCALE)         \
13088   __builtin_ia32_scatterdiv4df ((void *)ADDR, (__mmask8)MASK,           \
13089                                 (__v4di)(__m256i)INDEX,                 \
13090                                 (__v4df)(__m256d)V1, (int)SCALE)
13091
13092 #define _mm_i64scatter_pd(ADDR, INDEX, V1, SCALE)                       \
13093   __builtin_ia32_scatterdiv2df ((void *)ADDR, (__mmask8)0xFF,           \
13094                                 (__v2di)(__m128i)INDEX,                 \
13095                                 (__v2df)(__m128d)V1, (int)SCALE)
13096
13097 #define _mm_mask_i64scatter_pd(ADDR, MASK, INDEX, V1, SCALE)            \
13098   __builtin_ia32_scatterdiv2df ((void *)ADDR, (__mmask8)MASK,           \
13099                                 (__v2di)(__m128i)INDEX,                 \
13100                                 (__v2df)(__m128d)V1, (int)SCALE)
13101
13102 #define _mm256_i32scatter_epi32(ADDR, INDEX, V1, SCALE)                 \
13103   __builtin_ia32_scattersiv8si ((void *)ADDR, (__mmask8)0xFF,           \
13104                                 (__v8si)(__m256i)INDEX,                 \
13105                                 (__v8si)(__m256i)V1, (int)SCALE)
13106
13107 #define _mm256_mask_i32scatter_epi32(ADDR, MASK, INDEX, V1, SCALE)      \
13108   __builtin_ia32_scattersiv8si ((void *)ADDR, (__mmask8)MASK,           \
13109                                 (__v8si)(__m256i)INDEX,                 \
13110                                 (__v8si)(__m256i)V1, (int)SCALE)
13111
13112 #define _mm_i32scatter_epi32(ADDR, INDEX, V1, SCALE)                    \
13113   __builtin_ia32_scattersiv4si ((void *)ADDR, (__mmask8)0xFF,           \
13114                                 (__v4si)(__m128i)INDEX,                 \
13115                                 (__v4si)(__m128i)V1, (int)SCALE)
13116
13117 #define _mm_mask_i32scatter_epi32(ADDR, MASK, INDEX, V1, SCALE)         \
13118   __builtin_ia32_scattersiv4si ((void *)ADDR, (__mmask8)MASK,           \
13119                                 (__v4si)(__m128i)INDEX,                 \
13120                                 (__v4si)(__m128i)V1, (int)SCALE)
13121
13122 #define _mm256_i32scatter_epi64(ADDR, INDEX, V1, SCALE)                 \
13123   __builtin_ia32_scattersiv4di ((void *)ADDR, (__mmask8)0xFF,           \
13124                                 (__v4si)(__m128i)INDEX,                 \
13125                                 (__v4di)(__m256i)V1, (int)SCALE)
13126
13127 #define _mm256_mask_i32scatter_epi64(ADDR, MASK, INDEX, V1, SCALE)      \
13128   __builtin_ia32_scattersiv4di ((void *)ADDR, (__mmask8)MASK,           \
13129                                 (__v4si)(__m128i)INDEX,                 \
13130                                 (__v4di)(__m256i)V1, (int)SCALE)
13131
13132 #define _mm_i32scatter_epi64(ADDR, INDEX, V1, SCALE)                    \
13133   __builtin_ia32_scattersiv2di ((void *)ADDR, (__mmask8)0xFF,           \
13134                                 (__v4si)(__m128i)INDEX,                 \
13135                                 (__v2di)(__m128i)V1, (int)SCALE)
13136
13137 #define _mm_mask_i32scatter_epi64(ADDR, MASK, INDEX, V1, SCALE)         \
13138   __builtin_ia32_scattersiv2di ((void *)ADDR, (__mmask8)MASK,           \
13139                                 (__v4si)(__m128i)INDEX,                 \
13140                                 (__v2di)(__m128i)V1, (int)SCALE)
13141
13142 #define _mm256_i64scatter_epi32(ADDR, INDEX, V1, SCALE)                 \
13143   __builtin_ia32_scatterdiv8si ((void *)ADDR, (__mmask8)0xFF,           \
13144                                 (__v4di)(__m256i)INDEX,                 \
13145                                 (__v4si)(__m128i)V1, (int)SCALE)
13146
13147 #define _mm256_mask_i64scatter_epi32(ADDR, MASK, INDEX, V1, SCALE)      \
13148   __builtin_ia32_scatterdiv8si ((void *)ADDR, (__mmask8)MASK,           \
13149                                 (__v4di)(__m256i)INDEX,                 \
13150                                 (__v4si)(__m128i)V1, (int)SCALE)
13151
13152 #define _mm_i64scatter_epi32(ADDR, INDEX, V1, SCALE)                    \
13153   __builtin_ia32_scatterdiv4si ((void *)ADDR, (__mmask8)0xFF,           \
13154                                 (__v2di)(__m128i)INDEX,                 \
13155                                 (__v4si)(__m128i)V1, (int)SCALE)
13156
13157 #define _mm_mask_i64scatter_epi32(ADDR, MASK, INDEX, V1, SCALE)         \
13158   __builtin_ia32_scatterdiv4si ((void *)ADDR, (__mmask8)MASK,           \
13159                                 (__v2di)(__m128i)INDEX,                 \
13160                                 (__v4si)(__m128i)V1, (int)SCALE)
13161
13162 #define _mm256_i64scatter_epi64(ADDR, INDEX, V1, SCALE)                 \
13163   __builtin_ia32_scatterdiv4di ((void *)ADDR, (__mmask8)0xFF,           \
13164                                 (__v4di)(__m256i)INDEX,                 \
13165                                 (__v4di)(__m256i)V1, (int)SCALE)
13166
13167 #define _mm256_mask_i64scatter_epi64(ADDR, MASK, INDEX, V1, SCALE)      \
13168   __builtin_ia32_scatterdiv4di ((void *)ADDR, (__mmask8)MASK,           \
13169                                 (__v4di)(__m256i)INDEX,                 \
13170                                 (__v4di)(__m256i)V1, (int)SCALE)
13171
13172 #define _mm_i64scatter_epi64(ADDR, INDEX, V1, SCALE)                    \
13173   __builtin_ia32_scatterdiv2di ((void *)ADDR, (__mmask8)0xFF,           \
13174                                 (__v2di)(__m128i)INDEX,                 \
13175                                 (__v2di)(__m128i)V1, (int)SCALE)
13176
13177 #define _mm_mask_i64scatter_epi64(ADDR, MASK, INDEX, V1, SCALE)         \
13178   __builtin_ia32_scatterdiv2di ((void *)ADDR, (__mmask8)MASK,           \
13179                                 (__v2di)(__m128i)INDEX,                 \
13180                                 (__v2di)(__m128i)V1, (int)SCALE)
13181
13182 #define _mm256_mask_shuffle_epi32(W, U, X, C)                                       \
13183   ((__m256i)  __builtin_ia32_pshufd256_mask ((__v8si)(__m256i)(X), (int)(C),        \
13184                                              (__v8si)(__m256i)(W),                  \
13185                                              (__mmask8)(U)))
13186
13187 #define _mm256_maskz_shuffle_epi32(U, X, C)                                         \
13188   ((__m256i)  __builtin_ia32_pshufd256_mask ((__v8si)(__m256i)(X), (int)(C),        \
13189                                              (__v8si)(__m256i)                      \
13190                                              _mm256_setzero_si256 (),               \
13191                                              (__mmask8)(U)))
13192
13193 #define _mm_mask_shuffle_epi32(W, U, X, C)                                          \
13194   ((__m128i)  __builtin_ia32_pshufd128_mask ((__v4si)(__m128i)(X), (int)(C),        \
13195                                              (__v4si)(__m128i)(W),                  \
13196                                              (__mmask8)(U)))
13197
13198 #define _mm_maskz_shuffle_epi32(U, X, C)                                            \
13199   ((__m128i)  __builtin_ia32_pshufd128_mask ((__v4si)(__m128i)(X), (int)(C),        \
13200                                              (__v4si)(__m128i)_mm_setzero_si128 (), \
13201                                              (__mmask8)(U)))
13202
13203 #define _mm256_rol_epi64(A, B)                                                 \
13204   ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13205                                           (__v4di)(__m256i)_mm256_setzero_si256 (),\
13206                                           (__mmask8)-1))
13207
13208 #define _mm256_mask_rol_epi64(W, U, A, B)                                      \
13209   ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13210                                           (__v4di)(__m256i)(W),                \
13211                                           (__mmask8)(U)))
13212
13213 #define _mm256_maskz_rol_epi64(U, A, B)                                        \
13214   ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13215                                           (__v4di)(__m256i)_mm256_setzero_si256 (),\
13216                                           (__mmask8)(U)))
13217
13218 #define _mm_rol_epi64(A, B)                                                    \
13219   ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13220                                           (__v2di)(__m128i)_mm_setzero_si128 (),\
13221                                           (__mmask8)-1))
13222
13223 #define _mm_mask_rol_epi64(W, U, A, B)                                         \
13224   ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13225                                           (__v2di)(__m128i)(W),                \
13226                                           (__mmask8)(U)))
13227
13228 #define _mm_maskz_rol_epi64(U, A, B)                                           \
13229   ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13230                                           (__v2di)(__m128i)_mm_setzero_si128 (),\
13231                                           (__mmask8)(U)))
13232
13233 #define _mm256_ror_epi64(A, B)                                                 \
13234   ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13235                                           (__v4di)(__m256i)_mm256_setzero_si256 (),\
13236                                           (__mmask8)-1))
13237
13238 #define _mm256_mask_ror_epi64(W, U, A, B)                                      \
13239   ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13240                                           (__v4di)(__m256i)(W),                \
13241                                           (__mmask8)(U)))
13242
13243 #define _mm256_maskz_ror_epi64(U, A, B)                                        \
13244   ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B),      \
13245                                           (__v4di)(__m256i)_mm256_setzero_si256 (),\
13246                                           (__mmask8)(U)))
13247
13248 #define _mm_ror_epi64(A, B)                                                    \
13249   ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13250                                           (__v2di)(__m128i)_mm_setzero_si128 (),\
13251                                           (__mmask8)-1))
13252
13253 #define _mm_mask_ror_epi64(W, U, A, B)                                         \
13254   ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13255                                           (__v2di)(__m128i)(W),                \
13256                                           (__mmask8)(U)))
13257
13258 #define _mm_maskz_ror_epi64(U, A, B)                                           \
13259   ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B),      \
13260                                           (__v2di)(__m128i)_mm_setzero_si128 (),\
13261                                           (__mmask8)(U)))
13262
13263 #define _mm256_rol_epi32(A, B)                                                 \
13264   ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13265                                           (__v8si)(__m256i)_mm256_setzero_si256 (),\
13266                                           (__mmask8)-1))
13267
13268 #define _mm256_mask_rol_epi32(W, U, A, B)                                      \
13269   ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13270                                           (__v8si)(__m256i)(W),                \
13271                                           (__mmask8)(U)))
13272
13273 #define _mm256_maskz_rol_epi32(U, A, B)                                        \
13274   ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13275                                           (__v8si)(__m256i)_mm256_setzero_si256 (),\
13276                                           (__mmask8)(U)))
13277
13278 #define _mm_rol_epi32(A, B)                                                    \
13279   ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13280                                           (__v4si)(__m128i)_mm_setzero_si128 (),\
13281                                           (__mmask8)-1))
13282
13283 #define _mm_mask_rol_epi32(W, U, A, B)                                         \
13284   ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13285                                           (__v4si)(__m128i)(W),                \
13286                                           (__mmask8)(U)))
13287
13288 #define _mm_maskz_rol_epi32(U, A, B)                                           \
13289   ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13290                                           (__v4si)(__m128i)_mm_setzero_si128 (),\
13291                                           (__mmask8)(U)))
13292
13293 #define _mm256_ror_epi32(A, B)                                                 \
13294   ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13295                                           (__v8si)(__m256i)_mm256_setzero_si256 (),\
13296                                           (__mmask8)-1))
13297
13298 #define _mm256_mask_ror_epi32(W, U, A, B)                                      \
13299   ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13300                                           (__v8si)(__m256i)(W),                \
13301                                           (__mmask8)(U)))
13302
13303 #define _mm256_maskz_ror_epi32(U, A, B)                                        \
13304   ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B),      \
13305                                           (__v8si)(__m256i)                    \
13306                                           _mm256_setzero_si256 (),             \
13307                                           (__mmask8)(U)))
13308
13309 #define _mm_ror_epi32(A, B)                                                    \
13310   ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13311                                           (__v4si)(__m128i)_mm_setzero_si128 (),\
13312                                           (__mmask8)-1))
13313
13314 #define _mm_mask_ror_epi32(W, U, A, B)                                         \
13315   ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13316                                           (__v4si)(__m128i)(W),                \
13317                                           (__mmask8)(U)))
13318
13319 #define _mm_maskz_ror_epi32(U, A, B)                                           \
13320   ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B),      \
13321                                           (__v4si)(__m128i)_mm_setzero_si128 (),\
13322                                           (__mmask8)(U)))
13323
13324 #define _mm256_alignr_epi32(X, Y, C)                                        \
13325     ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X),          \
13326         (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)(X), (__mmask8)-1))
13327
13328 #define _mm256_mask_alignr_epi32(W, U, X, Y, C)                             \
13329     ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X),          \
13330         (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)(W), (__mmask8)(U)))
13331
13332 #define _mm256_maskz_alignr_epi32(U, X, Y, C)                               \
13333     ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X),          \
13334         (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)_mm256_setzero_si256 (),\
13335         (__mmask8)(U)))
13336
13337 #define _mm256_alignr_epi64(X, Y, C)                                        \
13338     ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X),          \
13339         (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)(X), (__mmask8)-1))
13340
13341 #define _mm256_mask_alignr_epi64(W, U, X, Y, C)                             \
13342     ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X),          \
13343         (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)(W), (__mmask8)(U)))
13344
13345 #define _mm256_maskz_alignr_epi64(U, X, Y, C)                               \
13346     ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X),          \
13347         (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)_mm256_setzero_si256 (),\
13348         (__mmask8)(U)))
13349
13350 #define _mm_alignr_epi32(X, Y, C)                                           \
13351     ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X),          \
13352         (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)(X), (__mmask8)-1))
13353
13354 #define _mm_mask_alignr_epi32(W, U, X, Y, C)                                \
13355     ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X),          \
13356         (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)(W), (__mmask8)(U)))
13357
13358 #define _mm_maskz_alignr_epi32(U, X, Y, C)                                  \
13359     ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X),          \
13360         (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)_mm_setzero_si128 (),\
13361         (__mmask8)(U)))
13362
13363 #define _mm_alignr_epi64(X, Y, C)                                           \
13364     ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X),          \
13365         (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)(X), (__mmask8)-1))
13366
13367 #define _mm_mask_alignr_epi64(W, U, X, Y, C)                                \
13368     ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X),          \
13369         (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)(X), (__mmask8)-1))
13370
13371 #define _mm_maskz_alignr_epi64(U, X, Y, C)                                  \
13372     ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X),          \
13373         (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)_mm_setzero_si128 (),\
13374         (__mmask8)(U)))
13375
13376 #define _mm_mask_cvtps_ph(W, U, A, I)                                           \
13377   ((__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf)(__m128) A, (int) (I),      \
13378       (__v8hi)(__m128i) (W), (__mmask8) (U)))
13379
13380 #define _mm_maskz_cvtps_ph(U, A, I)                                             \
13381   ((__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf)(__m128) A, (int) (I),      \
13382       (__v8hi)(__m128i) _mm_setzero_si128 (), (__mmask8) (U)))
13383
13384 #define _mm256_mask_cvtps_ph(W, U, A, I)                                        \
13385   ((__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf)(__m256) A, (int) (I),   \
13386       (__v8hi)(__m128i) (W), (__mmask8) (U)))
13387
13388 #define _mm256_maskz_cvtps_ph(U, A, I)                                          \
13389   ((__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf)(__m256) A, (int) (I),   \
13390       (__v8hi)(__m128i) _mm_setzero_si128 (), (__mmask8) (U)))
13391
13392 #define _mm256_mask_srai_epi32(W, U, A, B)                              \
13393   ((__m256i) __builtin_ia32_psradi256_mask ((__v8si)(__m256i)(A),       \
13394     (int)(B), (__v8si)(__m256i)(W), (__mmask8)(U)))
13395
13396 #define _mm256_maskz_srai_epi32(U, A, B)                                \
13397   ((__m256i) __builtin_ia32_psradi256_mask ((__v8si)(__m256i)(A),       \
13398     (int)(B), (__v8si)_mm256_setzero_si256 (), (__mmask8)(U)))
13399
13400 #define _mm_mask_srai_epi32(W, U, A, B)                                 \
13401   ((__m128i) __builtin_ia32_psradi128_mask ((__v4si)(__m128i)(A),       \
13402     (int)(B), (__v4si)(__m128i)(W), (__mmask8)(U)))
13403
13404 #define _mm_maskz_srai_epi32(U, A, B)                                   \
13405   ((__m128i) __builtin_ia32_psradi128_mask ((__v4si)(__m128i)(A),       \
13406     (int)(B), (__v4si)_mm_setzero_si128 (), (__mmask8)(U)))
13407
13408 #define _mm256_srai_epi64(A, B)                                         \
13409   ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A),       \
13410     (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)-1))
13411
13412 #define _mm256_mask_srai_epi64(W, U, A, B)                              \
13413   ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A),       \
13414     (int)(B), (__v4di)(__m256i)(W), (__mmask8)(U)))
13415
13416 #define _mm256_maskz_srai_epi64(U, A, B)                                \
13417   ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A),       \
13418     (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)(U)))
13419
13420 #define _mm_srai_epi64(A, B)                                            \
13421   ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A),       \
13422     (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)-1))
13423
13424 #define _mm_mask_srai_epi64(W, U, A, B)                                 \
13425   ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A),       \
13426     (int)(B), (__v2di)(__m128i)(W), (__mmask8)(U)))
13427
13428 #define _mm_maskz_srai_epi64(U, A, B)                                   \
13429   ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A),       \
13430     (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)(U)))
13431
13432 #define _mm256_mask_permutex_pd(W, U, A, B)                             \
13433   ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(A),       \
13434     (int)(B), (__v4df)(__m256d)(W), (__mmask8)(U)))
13435
13436 #define _mm256_maskz_permutex_pd(U, A, B)                               \
13437   ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(A),       \
13438     (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)(U)))
13439
13440 #define _mm256_mask_permute_pd(W, U, X, C)                                          \
13441   ((__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df)(__m256d)(X), (int)(C),      \
13442                                               (__v4df)(__m256d)(W),                 \
13443                                               (__mmask8)(U)))
13444
13445 #define _mm256_maskz_permute_pd(U, X, C)                                            \
13446   ((__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df)(__m256d)(X), (int)(C),      \
13447                                               (__v4df)(__m256d)_mm256_setzero_pd (),\
13448                                               (__mmask8)(U)))
13449
13450 #define _mm256_mask_permute_ps(W, U, X, C)                                          \
13451   ((__m256) __builtin_ia32_vpermilps256_mask ((__v8sf)(__m256)(X), (int)(C),        \
13452                                               (__v8sf)(__m256)(W), (__mmask8)(U)))
13453
13454 #define _mm256_maskz_permute_ps(U, X, C)                                            \
13455   ((__m256) __builtin_ia32_vpermilps256_mask ((__v8sf)(__m256)(X), (int)(C),        \
13456                                               (__v8sf)(__m256)_mm256_setzero_ps (), \
13457                                               (__mmask8)(U)))
13458
13459 #define _mm_mask_permute_pd(W, U, X, C)                                             \
13460   ((__m128d) __builtin_ia32_vpermilpd_mask ((__v2df)(__m128d)(X), (int)(C),         \
13461                                             (__v2df)(__m128d)(W), (__mmask8)(U)))
13462
13463 #define _mm_maskz_permute_pd(U, X, C)                                               \
13464   ((__m128d) __builtin_ia32_vpermilpd_mask ((__v2df)(__m128d)(X), (int)(C),         \
13465                                             (__v2df)(__m128d)_mm_setzero_pd (),     \
13466                                             (__mmask8)(U)))
13467
13468 #define _mm_mask_permute_ps(W, U, X, C)                                             \
13469   ((__m128) __builtin_ia32_vpermilps_mask ((__v4sf)(__m128)(X), (int)(C),           \
13470                                           (__v4sf)(__m128)(W), (__mmask8)(U)))
13471
13472 #define _mm_maskz_permute_ps(U, X, C)                                               \
13473   ((__m128) __builtin_ia32_vpermilps_mask ((__v4sf)(__m128)(X), (int)(C),           \
13474                                           (__v4sf)(__m128)_mm_setzero_ps (),        \
13475                                           (__mmask8)(U)))
13476
13477 #define _mm256_mask_blend_pd(__U, __A, __W)                           \
13478   ((__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) (__A),        \
13479                                                      (__v4df) (__W),  \
13480                                                      (__mmask8) (__U)))
13481
13482 #define _mm256_mask_blend_ps(__U, __A, __W)                           \
13483   ((__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) (__A),         \
13484                                                     (__v8sf) (__W),   \
13485                                                     (__mmask8) (__U)))
13486
13487 #define _mm256_mask_blend_epi64(__U, __A, __W)                        \
13488   ((__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) (__A),         \
13489                                                     (__v4di) (__W),   \
13490                                                     (__mmask8) (__U)))
13491
13492 #define _mm256_mask_blend_epi32(__U, __A, __W)                        \
13493   ((__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) (__A),         \
13494                                                     (__v8si) (__W),   \
13495                                                     (__mmask8) (__U)))
13496
13497 #define _mm_mask_blend_pd(__U, __A, __W)                              \
13498   ((__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) (__A),        \
13499                                                      (__v2df) (__W),  \
13500                                                      (__mmask8) (__U)))
13501
13502 #define _mm_mask_blend_ps(__U, __A, __W)                              \
13503   ((__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) (__A),         \
13504                                                     (__v4sf) (__W),   \
13505                                                     (__mmask8) (__U)))
13506
13507 #define _mm_mask_blend_epi64(__U, __A, __W)                           \
13508   ((__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) (__A),         \
13509                                                     (__v2di) (__W),   \
13510                                                     (__mmask8) (__U)))
13511
13512 #define _mm_mask_blend_epi32(__U, __A, __W)                           \
13513   ((__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) (__A),         \
13514                                                     (__v4si) (__W),   \
13515                                                     (__mmask8) (__U)))
13516
13517 #define _mm256_cmp_epu32_mask(X, Y, P)                                  \
13518   ((__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si)(__m256i)(X),       \
13519                                             (__v8si)(__m256i)(Y), (int)(P),\
13520                                             (__mmask8)-1))
13521
13522 #define _mm256_cmp_epi64_mask(X, Y, P)                                  \
13523   ((__mmask8) __builtin_ia32_cmpq256_mask ((__v4di)(__m256i)(X),        \
13524                                            (__v4di)(__m256i)(Y), (int)(P),\
13525                                            (__mmask8)-1))
13526
13527 #define _mm256_cmp_epi32_mask(X, Y, P)                                  \
13528   ((__mmask8) __builtin_ia32_cmpd256_mask ((__v8si)(__m256i)(X),        \
13529                                            (__v8si)(__m256i)(Y), (int)(P),\
13530                                            (__mmask8)-1))
13531
13532 #define _mm256_cmp_epu64_mask(X, Y, P)                                  \
13533   ((__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di)(__m256i)(X),       \
13534                                             (__v4di)(__m256i)(Y), (int)(P),\
13535                                             (__mmask8)-1))
13536
13537 #define _mm256_cmp_pd_mask(X, Y, P)                                     \
13538   ((__mmask8) __builtin_ia32_cmppd256_mask ((__v4df)(__m256d)(X),       \
13539                                             (__v4df)(__m256d)(Y), (int)(P),\
13540                                             (__mmask8)-1))
13541
13542 #define _mm256_cmp_ps_mask(X, Y, P)                                     \
13543   ((__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf)(__m256)(X),        \
13544                                              (__v8sf)(__m256)(Y), (int)(P),\
13545                                              (__mmask8)-1))
13546
13547 #define _mm256_mask_cmp_epi64_mask(M, X, Y, P)                          \
13548   ((__mmask8) __builtin_ia32_cmpq256_mask ((__v4di)(__m256i)(X),        \
13549                                            (__v4di)(__m256i)(Y), (int)(P),\
13550                                            (__mmask8)(M)))
13551
13552 #define _mm256_mask_cmp_epi32_mask(M, X, Y, P)                          \
13553   ((__mmask8) __builtin_ia32_cmpd256_mask ((__v8si)(__m256i)(X),        \
13554                                            (__v8si)(__m256i)(Y), (int)(P),\
13555                                            (__mmask8)(M)))
13556
13557 #define _mm256_mask_cmp_epu64_mask(M, X, Y, P)                          \
13558   ((__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di)(__m256i)(X),       \
13559                                             (__v4di)(__m256i)(Y), (int)(P),\
13560                                             (__mmask8)(M)))
13561
13562 #define _mm256_mask_cmp_epu32_mask(M, X, Y, P)                          \
13563   ((__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si)(__m256i)(X),       \
13564                                             (__v8si)(__m256i)(Y), (int)(P),\
13565                                             (__mmask8)(M)))
13566
13567 #define _mm256_mask_cmp_pd_mask(M, X, Y, P)                             \
13568   ((__mmask8) __builtin_ia32_cmppd256_mask ((__v4df)(__m256d)(X),       \
13569                                             (__v4df)(__m256d)(Y), (int)(P),\
13570                                             (__mmask8)(M)))
13571
13572 #define _mm256_mask_cmp_ps_mask(M, X, Y, P)                             \
13573   ((__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf)(__m256)(X),        \
13574                                              (__v8sf)(__m256)(Y), (int)(P),\
13575                                              (__mmask8)(M)))
13576
13577 #define _mm_cmp_epi64_mask(X, Y, P)                                     \
13578   ((__mmask8) __builtin_ia32_cmpq128_mask ((__v2di)(__m128i)(X),        \
13579                                            (__v2di)(__m128i)(Y), (int)(P),\
13580                                            (__mmask8)-1))
13581
13582 #define _mm_cmp_epi32_mask(X, Y, P)                                     \
13583   ((__mmask8) __builtin_ia32_cmpd128_mask ((__v4si)(__m128i)(X),        \
13584                                            (__v4si)(__m128i)(Y), (int)(P),\
13585                                            (__mmask8)-1))
13586
13587 #define _mm_cmp_epu64_mask(X, Y, P)                                     \
13588   ((__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di)(__m128i)(X),       \
13589                                             (__v2di)(__m128i)(Y), (int)(P),\
13590                                             (__mmask8)-1))
13591
13592 #define _mm_cmp_epu32_mask(X, Y, P)                                     \
13593   ((__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si)(__m128i)(X),       \
13594                                             (__v4si)(__m128i)(Y), (int)(P),\
13595                                             (__mmask8)-1))
13596
13597 #define _mm_cmp_pd_mask(X, Y, P)                                        \
13598   ((__mmask8) __builtin_ia32_cmppd128_mask ((__v2df)(__m128d)(X),       \
13599                                             (__v2df)(__m128d)(Y), (int)(P),\
13600                                             (__mmask8)-1))
13601
13602 #define _mm_cmp_ps_mask(X, Y, P)                                        \
13603   ((__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf)(__m128)(X),        \
13604                                              (__v4sf)(__m128)(Y), (int)(P),\
13605                                              (__mmask8)-1))
13606
13607 #define _mm_mask_cmp_epi64_mask(M, X, Y, P)                             \
13608   ((__mmask8) __builtin_ia32_cmpq128_mask ((__v2di)(__m128i)(X),        \
13609                                            (__v2di)(__m128i)(Y), (int)(P),\
13610                                            (__mmask8)(M)))
13611
13612 #define _mm_mask_cmp_epi32_mask(M, X, Y, P)                             \
13613   ((__mmask8) __builtin_ia32_cmpd128_mask ((__v4si)(__m128i)(X),        \
13614                                            (__v4si)(__m128i)(Y), (int)(P),\
13615                                            (__mmask8)(M)))
13616
13617 #define _mm_mask_cmp_epu64_mask(M, X, Y, P)                             \
13618   ((__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di)(__m128i)(X),       \
13619                                             (__v2di)(__m128i)(Y), (int)(P),\
13620                                             (__mmask8)(M)))
13621
13622 #define _mm_mask_cmp_epu32_mask(M, X, Y, P)                             \
13623   ((__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si)(__m128i)(X),       \
13624                                             (__v4si)(__m128i)(Y), (int)(P),\
13625                                             (__mmask8)(M)))
13626
13627 #define _mm_mask_cmp_pd_mask(M, X, Y, P)                                \
13628   ((__mmask8) __builtin_ia32_cmppd128_mask ((__v2df)(__m128d)(X),       \
13629                                             (__v2df)(__m128d)(Y), (int)(P),\
13630                                             (__mmask8)(M)))
13631
13632 #define _mm_mask_cmp_ps_mask(M, X, Y, P)                                \
13633   ((__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf)(__m128)(X),        \
13634                                              (__v4sf)(__m128)(Y), (int)(P),\
13635                                              (__mmask8)(M)))
13636
13637 #endif
13638
13639 #define _mm256_permutexvar_ps(A, B)     _mm256_permutevar8x32_ps ((B), (A))
13640
13641 #ifdef __DISABLE_AVX512VL__
13642 #undef __DISABLE_AVX512VL__
13643 #pragma GCC pop_options
13644 #endif /* __DISABLE_AVX512VL__ */
13645
13646 #endif /* _AVX512VLINTRIN_H_INCLUDED */