avx512vlintrin.h revision 1.1.1.8
1 /* Copyright (C) 2014-2022 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 typedef int __v4si_u __attribute__ ((__vector_size__ (16), \ 40 __may_alias__, __aligned__ (1))); 41 typedef int __v8si_u __attribute__ ((__vector_size__ (32), \ 42 __may_alias__, __aligned__ (1))); 43 typedef long long __v2di_u __attribute__ ((__vector_size__ (16), \ 44 __may_alias__, __aligned__ (1))); 45 typedef long long __v4di_u __attribute__ ((__vector_size__ (32), \ 46 __may_alias__, __aligned__ (1))); 47 48 extern __inline __m256d 49 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 50 _mm256_mask_mov_pd (__m256d __W, __mmask8 __U, __m256d __A) 51 { 52 return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A, 53 (__v4df) __W, 54 (__mmask8) __U); 55 } 56 57 extern __inline __m256d 58 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 59 _mm256_maskz_mov_pd (__mmask8 __U, __m256d __A) 60 { 61 return (__m256d) __builtin_ia32_movapd256_mask ((__v4df) __A, 62 (__v4df) 63 _mm256_setzero_pd (), 64 (__mmask8) __U); 65 } 66 67 extern __inline __m128d 68 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 69 _mm_mask_mov_pd (__m128d __W, __mmask8 __U, __m128d __A) 70 { 71 return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A, 72 (__v2df) __W, 73 (__mmask8) __U); 74 } 75 76 extern __inline __m128d 77 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 78 _mm_maskz_mov_pd (__mmask8 __U, __m128d __A) 79 { 80 return (__m128d) __builtin_ia32_movapd128_mask ((__v2df) __A, 81 (__v2df) 82 _mm_setzero_pd (), 83 (__mmask8) __U); 84 } 85 86 extern __inline __m256d 87 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 88 _mm256_mask_load_pd (__m256d __W, __mmask8 __U, void const *__P) 89 { 90 return (__m256d) __builtin_ia32_loadapd256_mask ((const __v4df *) __P, 91 (__v4df) __W, 92 (__mmask8) __U); 93 } 94 95 extern __inline __m256d 96 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 97 _mm256_maskz_load_pd (__mmask8 __U, void const *__P) 98 { 99 return (__m256d) __builtin_ia32_loadapd256_mask ((const __v4df *) __P, 100 (__v4df) 101 _mm256_setzero_pd (), 102 (__mmask8) __U); 103 } 104 105 extern __inline __m128d 106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 107 _mm_mask_load_pd (__m128d __W, __mmask8 __U, void const *__P) 108 { 109 return (__m128d) __builtin_ia32_loadapd128_mask ((const __v2df *) __P, 110 (__v2df) __W, 111 (__mmask8) __U); 112 } 113 114 extern __inline __m128d 115 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 116 _mm_maskz_load_pd (__mmask8 __U, void const *__P) 117 { 118 return (__m128d) __builtin_ia32_loadapd128_mask ((const __v2df *) __P, 119 (__v2df) 120 _mm_setzero_pd (), 121 (__mmask8) __U); 122 } 123 124 extern __inline void 125 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 126 _mm256_mask_store_pd (void *__P, __mmask8 __U, __m256d __A) 127 { 128 __builtin_ia32_storeapd256_mask ((__v4df *) __P, 129 (__v4df) __A, 130 (__mmask8) __U); 131 } 132 133 extern __inline void 134 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 135 _mm_mask_store_pd (void *__P, __mmask8 __U, __m128d __A) 136 { 137 __builtin_ia32_storeapd128_mask ((__v2df *) __P, 138 (__v2df) __A, 139 (__mmask8) __U); 140 } 141 142 extern __inline __m256 143 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 144 _mm256_mask_mov_ps (__m256 __W, __mmask8 __U, __m256 __A) 145 { 146 return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A, 147 (__v8sf) __W, 148 (__mmask8) __U); 149 } 150 151 extern __inline __m256 152 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 153 _mm256_maskz_mov_ps (__mmask8 __U, __m256 __A) 154 { 155 return (__m256) __builtin_ia32_movaps256_mask ((__v8sf) __A, 156 (__v8sf) 157 _mm256_setzero_ps (), 158 (__mmask8) __U); 159 } 160 161 extern __inline __m128 162 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 163 _mm_mask_mov_ps (__m128 __W, __mmask8 __U, __m128 __A) 164 { 165 return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A, 166 (__v4sf) __W, 167 (__mmask8) __U); 168 } 169 170 extern __inline __m128 171 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 172 _mm_maskz_mov_ps (__mmask8 __U, __m128 __A) 173 { 174 return (__m128) __builtin_ia32_movaps128_mask ((__v4sf) __A, 175 (__v4sf) 176 _mm_setzero_ps (), 177 (__mmask8) __U); 178 } 179 180 extern __inline __m256 181 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 182 _mm256_mask_load_ps (__m256 __W, __mmask8 __U, void const *__P) 183 { 184 return (__m256) __builtin_ia32_loadaps256_mask ((const __v8sf *) __P, 185 (__v8sf) __W, 186 (__mmask8) __U); 187 } 188 189 extern __inline __m256 190 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 191 _mm256_maskz_load_ps (__mmask8 __U, void const *__P) 192 { 193 return (__m256) __builtin_ia32_loadaps256_mask ((const __v8sf *) __P, 194 (__v8sf) 195 _mm256_setzero_ps (), 196 (__mmask8) __U); 197 } 198 199 extern __inline __m128 200 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 201 _mm_mask_load_ps (__m128 __W, __mmask8 __U, void const *__P) 202 { 203 return (__m128) __builtin_ia32_loadaps128_mask ((const __v4sf *) __P, 204 (__v4sf) __W, 205 (__mmask8) __U); 206 } 207 208 extern __inline __m128 209 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 210 _mm_maskz_load_ps (__mmask8 __U, void const *__P) 211 { 212 return (__m128) __builtin_ia32_loadaps128_mask ((const __v4sf *) __P, 213 (__v4sf) 214 _mm_setzero_ps (), 215 (__mmask8) __U); 216 } 217 218 extern __inline void 219 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 220 _mm256_mask_store_ps (void *__P, __mmask8 __U, __m256 __A) 221 { 222 __builtin_ia32_storeaps256_mask ((__v8sf *) __P, 223 (__v8sf) __A, 224 (__mmask8) __U); 225 } 226 227 extern __inline void 228 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 229 _mm_mask_store_ps (void *__P, __mmask8 __U, __m128 __A) 230 { 231 __builtin_ia32_storeaps128_mask ((__v4sf *) __P, 232 (__v4sf) __A, 233 (__mmask8) __U); 234 } 235 236 extern __inline __m256i 237 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 238 _mm256_mask_mov_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 239 { 240 return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A, 241 (__v4di) __W, 242 (__mmask8) __U); 243 } 244 245 extern __inline __m256i 246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 247 _mm256_maskz_mov_epi64 (__mmask8 __U, __m256i __A) 248 { 249 return (__m256i) __builtin_ia32_movdqa64_256_mask ((__v4di) __A, 250 (__v4di) 251 _mm256_setzero_si256 (), 252 (__mmask8) __U); 253 } 254 255 extern __inline __m128i 256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 257 _mm_mask_mov_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 258 { 259 return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A, 260 (__v2di) __W, 261 (__mmask8) __U); 262 } 263 264 extern __inline __m128i 265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 266 _mm_maskz_mov_epi64 (__mmask8 __U, __m128i __A) 267 { 268 return (__m128i) __builtin_ia32_movdqa64_128_mask ((__v2di) __A, 269 (__v2di) 270 _mm_setzero_si128 (), 271 (__mmask8) __U); 272 } 273 274 extern __inline __m256i 275 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 276 _mm256_load_epi64 (void const *__P) 277 { 278 return (__m256i) (*(const __v4di *) __P); 279 } 280 281 extern __inline __m256i 282 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 283 _mm256_mask_load_epi64 (__m256i __W, __mmask8 __U, void const *__P) 284 { 285 return (__m256i) __builtin_ia32_movdqa64load256_mask ((const __v4di *) __P, 286 (__v4di) __W, 287 (__mmask8) 288 __U); 289 } 290 291 extern __inline __m256i 292 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 293 _mm256_maskz_load_epi64 (__mmask8 __U, void const *__P) 294 { 295 return (__m256i) __builtin_ia32_movdqa64load256_mask ((const __v4di *) __P, 296 (__v4di) 297 _mm256_setzero_si256 (), 298 (__mmask8) 299 __U); 300 } 301 302 extern __inline __m128i 303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 304 _mm_load_epi64 (void const *__P) 305 { 306 return (__m128i) (*(const __v2di *) __P); 307 } 308 309 extern __inline __m128i 310 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 311 _mm_mask_load_epi64 (__m128i __W, __mmask8 __U, void const *__P) 312 { 313 return (__m128i) __builtin_ia32_movdqa64load128_mask ((const __v2di *) __P, 314 (__v2di) __W, 315 (__mmask8) 316 __U); 317 } 318 319 extern __inline __m128i 320 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 321 _mm_maskz_load_epi64 (__mmask8 __U, void const *__P) 322 { 323 return (__m128i) __builtin_ia32_movdqa64load128_mask ((const __v2di *) __P, 324 (__v2di) 325 _mm_setzero_si128 (), 326 (__mmask8) 327 __U); 328 } 329 330 extern __inline void 331 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 332 _mm256_mask_store_epi64 (void *__P, __mmask8 __U, __m256i __A) 333 { 334 __builtin_ia32_movdqa64store256_mask ((__v4di *) __P, 335 (__v4di) __A, 336 (__mmask8) __U); 337 } 338 339 extern __inline void 340 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 341 _mm_mask_store_epi64 (void *__P, __mmask8 __U, __m128i __A) 342 { 343 __builtin_ia32_movdqa64store128_mask ((__v2di *) __P, 344 (__v2di) __A, 345 (__mmask8) __U); 346 } 347 348 extern __inline __m256i 349 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 350 _mm256_mask_mov_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 351 { 352 return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A, 353 (__v8si) __W, 354 (__mmask8) __U); 355 } 356 357 extern __inline __m256i 358 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 359 _mm256_maskz_mov_epi32 (__mmask8 __U, __m256i __A) 360 { 361 return (__m256i) __builtin_ia32_movdqa32_256_mask ((__v8si) __A, 362 (__v8si) 363 _mm256_setzero_si256 (), 364 (__mmask8) __U); 365 } 366 367 extern __inline __m128i 368 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 369 _mm_mask_mov_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 370 { 371 return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A, 372 (__v4si) __W, 373 (__mmask8) __U); 374 } 375 376 extern __inline __m128i 377 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 378 _mm_maskz_mov_epi32 (__mmask8 __U, __m128i __A) 379 { 380 return (__m128i) __builtin_ia32_movdqa32_128_mask ((__v4si) __A, 381 (__v4si) 382 _mm_setzero_si128 (), 383 (__mmask8) __U); 384 } 385 386 extern __inline __m256i 387 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 388 _mm256_load_epi32 (void const *__P) 389 { 390 return (__m256i) (*(const __v8si *) __P); 391 } 392 393 extern __inline __m256i 394 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 395 _mm256_mask_load_epi32 (__m256i __W, __mmask8 __U, void const *__P) 396 { 397 return (__m256i) __builtin_ia32_movdqa32load256_mask ((const __v8si *) __P, 398 (__v8si) __W, 399 (__mmask8) 400 __U); 401 } 402 403 extern __inline __m256i 404 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 405 _mm256_maskz_load_epi32 (__mmask8 __U, void const *__P) 406 { 407 return (__m256i) __builtin_ia32_movdqa32load256_mask ((const __v8si *) __P, 408 (__v8si) 409 _mm256_setzero_si256 (), 410 (__mmask8) 411 __U); 412 } 413 414 extern __inline __m128i 415 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 416 _mm_load_epi32 (void const *__P) 417 { 418 return (__m128i) (*(const __v4si *) __P); 419 } 420 421 extern __inline __m128i 422 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 423 _mm_mask_load_epi32 (__m128i __W, __mmask8 __U, void const *__P) 424 { 425 return (__m128i) __builtin_ia32_movdqa32load128_mask ((const __v4si *) __P, 426 (__v4si) __W, 427 (__mmask8) 428 __U); 429 } 430 431 extern __inline __m128i 432 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 433 _mm_maskz_load_epi32 (__mmask8 __U, void const *__P) 434 { 435 return (__m128i) __builtin_ia32_movdqa32load128_mask ((const __v4si *) __P, 436 (__v4si) 437 _mm_setzero_si128 (), 438 (__mmask8) 439 __U); 440 } 441 442 extern __inline void 443 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 444 _mm256_store_epi32 (void *__P, __m256i __A) 445 { 446 *(__v8si *) __P = (__v8si) __A; 447 } 448 449 extern __inline void 450 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 451 _mm256_mask_store_epi32 (void *__P, __mmask8 __U, __m256i __A) 452 { 453 __builtin_ia32_movdqa32store256_mask ((__v8si *) __P, 454 (__v8si) __A, 455 (__mmask8) __U); 456 } 457 458 extern __inline void 459 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 460 _mm_store_epi32 (void *__P, __m128i __A) 461 { 462 *(__v4si *) __P = (__v4si) __A; 463 } 464 465 extern __inline void 466 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 467 _mm_mask_store_epi32 (void *__P, __mmask8 __U, __m128i __A) 468 { 469 __builtin_ia32_movdqa32store128_mask ((__v4si *) __P, 470 (__v4si) __A, 471 (__mmask8) __U); 472 } 473 474 extern __inline __m128d 475 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 476 _mm_mask_add_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 477 { 478 return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, 479 (__v2df) __B, 480 (__v2df) __W, 481 (__mmask8) __U); 482 } 483 484 extern __inline __m128d 485 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 486 _mm_maskz_add_pd (__mmask8 __U, __m128d __A, __m128d __B) 487 { 488 return (__m128d) __builtin_ia32_addpd128_mask ((__v2df) __A, 489 (__v2df) __B, 490 (__v2df) 491 _mm_setzero_pd (), 492 (__mmask8) __U); 493 } 494 495 extern __inline __m256d 496 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 497 _mm256_mask_add_pd (__m256d __W, __mmask8 __U, __m256d __A, 498 __m256d __B) 499 { 500 return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, 501 (__v4df) __B, 502 (__v4df) __W, 503 (__mmask8) __U); 504 } 505 506 extern __inline __m256d 507 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 508 _mm256_maskz_add_pd (__mmask8 __U, __m256d __A, __m256d __B) 509 { 510 return (__m256d) __builtin_ia32_addpd256_mask ((__v4df) __A, 511 (__v4df) __B, 512 (__v4df) 513 _mm256_setzero_pd (), 514 (__mmask8) __U); 515 } 516 517 extern __inline __m128 518 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 519 _mm_mask_add_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 520 { 521 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, 522 (__v4sf) __B, 523 (__v4sf) __W, 524 (__mmask8) __U); 525 } 526 527 extern __inline __m128 528 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 529 _mm_maskz_add_ps (__mmask8 __U, __m128 __A, __m128 __B) 530 { 531 return (__m128) __builtin_ia32_addps128_mask ((__v4sf) __A, 532 (__v4sf) __B, 533 (__v4sf) 534 _mm_setzero_ps (), 535 (__mmask8) __U); 536 } 537 538 extern __inline __m256 539 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 540 _mm256_mask_add_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 541 { 542 return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, 543 (__v8sf) __B, 544 (__v8sf) __W, 545 (__mmask8) __U); 546 } 547 548 extern __inline __m256 549 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 550 _mm256_maskz_add_ps (__mmask8 __U, __m256 __A, __m256 __B) 551 { 552 return (__m256) __builtin_ia32_addps256_mask ((__v8sf) __A, 553 (__v8sf) __B, 554 (__v8sf) 555 _mm256_setzero_ps (), 556 (__mmask8) __U); 557 } 558 559 extern __inline __m128d 560 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 561 _mm_mask_sub_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 562 { 563 return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, 564 (__v2df) __B, 565 (__v2df) __W, 566 (__mmask8) __U); 567 } 568 569 extern __inline __m128d 570 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 571 _mm_maskz_sub_pd (__mmask8 __U, __m128d __A, __m128d __B) 572 { 573 return (__m128d) __builtin_ia32_subpd128_mask ((__v2df) __A, 574 (__v2df) __B, 575 (__v2df) 576 _mm_setzero_pd (), 577 (__mmask8) __U); 578 } 579 580 extern __inline __m256d 581 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 582 _mm256_mask_sub_pd (__m256d __W, __mmask8 __U, __m256d __A, 583 __m256d __B) 584 { 585 return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, 586 (__v4df) __B, 587 (__v4df) __W, 588 (__mmask8) __U); 589 } 590 591 extern __inline __m256d 592 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 593 _mm256_maskz_sub_pd (__mmask8 __U, __m256d __A, __m256d __B) 594 { 595 return (__m256d) __builtin_ia32_subpd256_mask ((__v4df) __A, 596 (__v4df) __B, 597 (__v4df) 598 _mm256_setzero_pd (), 599 (__mmask8) __U); 600 } 601 602 extern __inline __m128 603 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 604 _mm_mask_sub_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 605 { 606 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, 607 (__v4sf) __B, 608 (__v4sf) __W, 609 (__mmask8) __U); 610 } 611 612 extern __inline __m128 613 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 614 _mm_maskz_sub_ps (__mmask8 __U, __m128 __A, __m128 __B) 615 { 616 return (__m128) __builtin_ia32_subps128_mask ((__v4sf) __A, 617 (__v4sf) __B, 618 (__v4sf) 619 _mm_setzero_ps (), 620 (__mmask8) __U); 621 } 622 623 extern __inline __m256 624 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 625 _mm256_mask_sub_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 626 { 627 return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, 628 (__v8sf) __B, 629 (__v8sf) __W, 630 (__mmask8) __U); 631 } 632 633 extern __inline __m256 634 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 635 _mm256_maskz_sub_ps (__mmask8 __U, __m256 __A, __m256 __B) 636 { 637 return (__m256) __builtin_ia32_subps256_mask ((__v8sf) __A, 638 (__v8sf) __B, 639 (__v8sf) 640 _mm256_setzero_ps (), 641 (__mmask8) __U); 642 } 643 644 extern __inline void 645 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 646 _mm256_store_epi64 (void *__P, __m256i __A) 647 { 648 *(__m256i *) __P = __A; 649 } 650 651 extern __inline void 652 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 653 _mm_store_epi64 (void *__P, __m128i __A) 654 { 655 *(__m128i *) __P = __A; 656 } 657 658 extern __inline __m256d 659 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 660 _mm256_mask_loadu_pd (__m256d __W, __mmask8 __U, void const *__P) 661 { 662 return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P, 663 (__v4df) __W, 664 (__mmask8) __U); 665 } 666 667 extern __inline __m256d 668 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 669 _mm256_maskz_loadu_pd (__mmask8 __U, void const *__P) 670 { 671 return (__m256d) __builtin_ia32_loadupd256_mask ((const double *) __P, 672 (__v4df) 673 _mm256_setzero_pd (), 674 (__mmask8) __U); 675 } 676 677 extern __inline __m128d 678 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 679 _mm_mask_loadu_pd (__m128d __W, __mmask8 __U, void const *__P) 680 { 681 return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P, 682 (__v2df) __W, 683 (__mmask8) __U); 684 } 685 686 extern __inline __m128d 687 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 688 _mm_maskz_loadu_pd (__mmask8 __U, void const *__P) 689 { 690 return (__m128d) __builtin_ia32_loadupd128_mask ((const double *) __P, 691 (__v2df) 692 _mm_setzero_pd (), 693 (__mmask8) __U); 694 } 695 696 extern __inline void 697 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 698 _mm256_mask_storeu_pd (void *__P, __mmask8 __U, __m256d __A) 699 { 700 __builtin_ia32_storeupd256_mask ((double *) __P, 701 (__v4df) __A, 702 (__mmask8) __U); 703 } 704 705 extern __inline void 706 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 707 _mm_mask_storeu_pd (void *__P, __mmask8 __U, __m128d __A) 708 { 709 __builtin_ia32_storeupd128_mask ((double *) __P, 710 (__v2df) __A, 711 (__mmask8) __U); 712 } 713 714 extern __inline __m256 715 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 716 _mm256_mask_loadu_ps (__m256 __W, __mmask8 __U, void const *__P) 717 { 718 return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P, 719 (__v8sf) __W, 720 (__mmask8) __U); 721 } 722 723 extern __inline __m256 724 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 725 _mm256_maskz_loadu_ps (__mmask8 __U, void const *__P) 726 { 727 return (__m256) __builtin_ia32_loadups256_mask ((const float *) __P, 728 (__v8sf) 729 _mm256_setzero_ps (), 730 (__mmask8) __U); 731 } 732 733 extern __inline __m128 734 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 735 _mm_mask_loadu_ps (__m128 __W, __mmask8 __U, void const *__P) 736 { 737 return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P, 738 (__v4sf) __W, 739 (__mmask8) __U); 740 } 741 742 extern __inline __m128 743 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 744 _mm_maskz_loadu_ps (__mmask8 __U, void const *__P) 745 { 746 return (__m128) __builtin_ia32_loadups128_mask ((const float *) __P, 747 (__v4sf) 748 _mm_setzero_ps (), 749 (__mmask8) __U); 750 } 751 752 extern __inline void 753 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 754 _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A) 755 { 756 __builtin_ia32_storeups256_mask ((float *) __P, 757 (__v8sf) __A, 758 (__mmask8) __U); 759 } 760 761 extern __inline void 762 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 763 _mm_mask_storeu_ps (void *__P, __mmask8 __U, __m128 __A) 764 { 765 __builtin_ia32_storeups128_mask ((float *) __P, 766 (__v4sf) __A, 767 (__mmask8) __U); 768 } 769 770 extern __inline __m256i 771 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 772 _mm256_loadu_epi64 (void const *__P) 773 { 774 return (__m256i) (*(const __v4di_u *) __P); 775 } 776 777 extern __inline __m256i 778 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 779 _mm256_mask_loadu_epi64 (__m256i __W, __mmask8 __U, void const *__P) 780 { 781 return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P, 782 (__v4di) __W, 783 (__mmask8) __U); 784 } 785 786 extern __inline __m256i 787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 788 _mm256_maskz_loadu_epi64 (__mmask8 __U, void const *__P) 789 { 790 return (__m256i) __builtin_ia32_loaddqudi256_mask ((const long long *) __P, 791 (__v4di) 792 _mm256_setzero_si256 (), 793 (__mmask8) __U); 794 } 795 796 extern __inline __m128i 797 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 798 _mm_loadu_epi64 (void const *__P) 799 { 800 return (__m128i) (*(const __v2di_u *) __P); 801 } 802 803 extern __inline __m128i 804 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 805 _mm_mask_loadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) 806 { 807 return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P, 808 (__v2di) __W, 809 (__mmask8) __U); 810 } 811 812 extern __inline __m128i 813 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 814 _mm_maskz_loadu_epi64 (__mmask8 __U, void const *__P) 815 { 816 return (__m128i) __builtin_ia32_loaddqudi128_mask ((const long long *) __P, 817 (__v2di) 818 _mm_setzero_si128 (), 819 (__mmask8) __U); 820 } 821 822 extern __inline void 823 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 824 _mm256_storeu_epi64 (void *__P, __m256i __A) 825 { 826 *(__m256i_u *) __P = (__m256i_u) __A; 827 } 828 829 extern __inline void 830 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 831 _mm256_mask_storeu_epi64 (void *__P, __mmask8 __U, __m256i __A) 832 { 833 __builtin_ia32_storedqudi256_mask ((long long *) __P, 834 (__v4di) __A, 835 (__mmask8) __U); 836 } 837 838 extern __inline void 839 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 840 _mm_storeu_epi64 (void *__P, __m128i __A) 841 { 842 *(__m128i_u *) __P = (__m128i_u) __A; 843 } 844 845 extern __inline void 846 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 847 _mm_mask_storeu_epi64 (void *__P, __mmask8 __U, __m128i __A) 848 { 849 __builtin_ia32_storedqudi128_mask ((long long *) __P, 850 (__v2di) __A, 851 (__mmask8) __U); 852 } 853 854 extern __inline __m256i 855 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 856 _mm256_loadu_epi32 (void const *__P) 857 { 858 return (__m256i) (*(const __v8si_u *) __P); 859 } 860 861 extern __inline __m256i 862 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 863 _mm256_mask_loadu_epi32 (__m256i __W, __mmask8 __U, void const *__P) 864 { 865 return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P, 866 (__v8si) __W, 867 (__mmask8) __U); 868 } 869 870 extern __inline __m256i 871 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 872 _mm256_maskz_loadu_epi32 (__mmask8 __U, void const *__P) 873 { 874 return (__m256i) __builtin_ia32_loaddqusi256_mask ((const int *) __P, 875 (__v8si) 876 _mm256_setzero_si256 (), 877 (__mmask8) __U); 878 } 879 880 extern __inline __m128i 881 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 882 _mm_loadu_epi32 (void const *__P) 883 { 884 return (__m128i) (*(const __v4si_u *) __P); 885 } 886 887 extern __inline __m128i 888 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 889 _mm_mask_loadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) 890 { 891 return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P, 892 (__v4si) __W, 893 (__mmask8) __U); 894 } 895 896 extern __inline __m128i 897 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 898 _mm_maskz_loadu_epi32 (__mmask8 __U, void const *__P) 899 { 900 return (__m128i) __builtin_ia32_loaddqusi128_mask ((const int *) __P, 901 (__v4si) 902 _mm_setzero_si128 (), 903 (__mmask8) __U); 904 } 905 906 extern __inline void 907 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 908 _mm256_storeu_epi32 (void *__P, __m256i __A) 909 { 910 *(__m256i_u *) __P = (__m256i_u) __A; 911 } 912 913 extern __inline void 914 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 915 _mm256_mask_storeu_epi32 (void *__P, __mmask8 __U, __m256i __A) 916 { 917 __builtin_ia32_storedqusi256_mask ((int *) __P, 918 (__v8si) __A, 919 (__mmask8) __U); 920 } 921 922 extern __inline void 923 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 924 _mm_storeu_epi32 (void *__P, __m128i __A) 925 { 926 *(__m128i_u *) __P = (__m128i_u) __A; 927 } 928 929 extern __inline void 930 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 931 _mm_mask_storeu_epi32 (void *__P, __mmask8 __U, __m128i __A) 932 { 933 __builtin_ia32_storedqusi128_mask ((int *) __P, 934 (__v4si) __A, 935 (__mmask8) __U); 936 } 937 938 extern __inline __m256i 939 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 940 _mm256_mask_abs_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 941 { 942 return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, 943 (__v8si) __W, 944 (__mmask8) __U); 945 } 946 947 extern __inline __m256i 948 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 949 _mm256_maskz_abs_epi32 (__mmask8 __U, __m256i __A) 950 { 951 return (__m256i) __builtin_ia32_pabsd256_mask ((__v8si) __A, 952 (__v8si) 953 _mm256_setzero_si256 (), 954 (__mmask8) __U); 955 } 956 957 extern __inline __m128i 958 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 959 _mm_mask_abs_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 960 { 961 return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, 962 (__v4si) __W, 963 (__mmask8) __U); 964 } 965 966 extern __inline __m128i 967 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 968 _mm_maskz_abs_epi32 (__mmask8 __U, __m128i __A) 969 { 970 return (__m128i) __builtin_ia32_pabsd128_mask ((__v4si) __A, 971 (__v4si) 972 _mm_setzero_si128 (), 973 (__mmask8) __U); 974 } 975 976 extern __inline __m256i 977 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 978 _mm256_abs_epi64 (__m256i __A) 979 { 980 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, 981 (__v4di) 982 _mm256_setzero_si256 (), 983 (__mmask8) -1); 984 } 985 986 extern __inline __m256i 987 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 988 _mm256_mask_abs_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 989 { 990 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, 991 (__v4di) __W, 992 (__mmask8) __U); 993 } 994 995 extern __inline __m256i 996 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 997 _mm256_maskz_abs_epi64 (__mmask8 __U, __m256i __A) 998 { 999 return (__m256i) __builtin_ia32_pabsq256_mask ((__v4di) __A, 1000 (__v4di) 1001 _mm256_setzero_si256 (), 1002 (__mmask8) __U); 1003 } 1004 1005 extern __inline __m128i 1006 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1007 _mm_abs_epi64 (__m128i __A) 1008 { 1009 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, 1010 (__v2di) 1011 _mm_setzero_si128 (), 1012 (__mmask8) -1); 1013 } 1014 1015 extern __inline __m128i 1016 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1017 _mm_mask_abs_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 1018 { 1019 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, 1020 (__v2di) __W, 1021 (__mmask8) __U); 1022 } 1023 1024 extern __inline __m128i 1025 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1026 _mm_maskz_abs_epi64 (__mmask8 __U, __m128i __A) 1027 { 1028 return (__m128i) __builtin_ia32_pabsq128_mask ((__v2di) __A, 1029 (__v2di) 1030 _mm_setzero_si128 (), 1031 (__mmask8) __U); 1032 } 1033 1034 extern __inline __m128i 1035 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1036 _mm256_cvtpd_epu32 (__m256d __A) 1037 { 1038 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, 1039 (__v4si) 1040 _mm_setzero_si128 (), 1041 (__mmask8) -1); 1042 } 1043 1044 extern __inline __m128i 1045 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1046 _mm256_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) 1047 { 1048 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, 1049 (__v4si) __W, 1050 (__mmask8) __U); 1051 } 1052 1053 extern __inline __m128i 1054 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1055 _mm256_maskz_cvtpd_epu32 (__mmask8 __U, __m256d __A) 1056 { 1057 return (__m128i) __builtin_ia32_cvtpd2udq256_mask ((__v4df) __A, 1058 (__v4si) 1059 _mm_setzero_si128 (), 1060 (__mmask8) __U); 1061 } 1062 1063 extern __inline __m128i 1064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1065 _mm_cvtpd_epu32 (__m128d __A) 1066 { 1067 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, 1068 (__v4si) 1069 _mm_setzero_si128 (), 1070 (__mmask8) -1); 1071 } 1072 1073 extern __inline __m128i 1074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1075 _mm_mask_cvtpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) 1076 { 1077 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, 1078 (__v4si) __W, 1079 (__mmask8) __U); 1080 } 1081 1082 extern __inline __m128i 1083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1084 _mm_maskz_cvtpd_epu32 (__mmask8 __U, __m128d __A) 1085 { 1086 return (__m128i) __builtin_ia32_cvtpd2udq128_mask ((__v2df) __A, 1087 (__v4si) 1088 _mm_setzero_si128 (), 1089 (__mmask8) __U); 1090 } 1091 1092 extern __inline __m256i 1093 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1094 _mm256_mask_cvttps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) 1095 { 1096 return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, 1097 (__v8si) __W, 1098 (__mmask8) __U); 1099 } 1100 1101 extern __inline __m256i 1102 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1103 _mm256_maskz_cvttps_epi32 (__mmask8 __U, __m256 __A) 1104 { 1105 return (__m256i) __builtin_ia32_cvttps2dq256_mask ((__v8sf) __A, 1106 (__v8si) 1107 _mm256_setzero_si256 (), 1108 (__mmask8) __U); 1109 } 1110 1111 extern __inline __m128i 1112 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1113 _mm_mask_cvttps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) 1114 { 1115 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, 1116 (__v4si) __W, 1117 (__mmask8) __U); 1118 } 1119 1120 extern __inline __m128i 1121 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1122 _mm_maskz_cvttps_epi32 (__mmask8 __U, __m128 __A) 1123 { 1124 return (__m128i) __builtin_ia32_cvttps2dq128_mask ((__v4sf) __A, 1125 (__v4si) 1126 _mm_setzero_si128 (), 1127 (__mmask8) __U); 1128 } 1129 1130 extern __inline __m256i 1131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1132 _mm256_cvttps_epu32 (__m256 __A) 1133 { 1134 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, 1135 (__v8si) 1136 _mm256_setzero_si256 (), 1137 (__mmask8) -1); 1138 } 1139 1140 extern __inline __m256i 1141 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1142 _mm256_mask_cvttps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) 1143 { 1144 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, 1145 (__v8si) __W, 1146 (__mmask8) __U); 1147 } 1148 1149 extern __inline __m256i 1150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1151 _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) 1152 { 1153 return (__m256i) __builtin_ia32_cvttps2udq256_mask ((__v8sf) __A, 1154 (__v8si) 1155 _mm256_setzero_si256 (), 1156 (__mmask8) __U); 1157 } 1158 1159 extern __inline __m128i 1160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1161 _mm_cvttps_epu32 (__m128 __A) 1162 { 1163 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, 1164 (__v4si) 1165 _mm_setzero_si128 (), 1166 (__mmask8) -1); 1167 } 1168 1169 extern __inline __m128i 1170 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1171 _mm_mask_cvttps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) 1172 { 1173 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, 1174 (__v4si) __W, 1175 (__mmask8) __U); 1176 } 1177 1178 extern __inline __m128i 1179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1180 _mm_maskz_cvttps_epu32 (__mmask8 __U, __m128 __A) 1181 { 1182 return (__m128i) __builtin_ia32_cvttps2udq128_mask ((__v4sf) __A, 1183 (__v4si) 1184 _mm_setzero_si128 (), 1185 (__mmask8) __U); 1186 } 1187 1188 extern __inline __m128i 1189 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1190 _mm256_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) 1191 { 1192 return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, 1193 (__v4si) __W, 1194 (__mmask8) __U); 1195 } 1196 1197 extern __inline __m128i 1198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1199 _mm256_maskz_cvttpd_epi32 (__mmask8 __U, __m256d __A) 1200 { 1201 return (__m128i) __builtin_ia32_cvttpd2dq256_mask ((__v4df) __A, 1202 (__v4si) 1203 _mm_setzero_si128 (), 1204 (__mmask8) __U); 1205 } 1206 1207 extern __inline __m128i 1208 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1209 _mm_mask_cvttpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) 1210 { 1211 return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, 1212 (__v4si) __W, 1213 (__mmask8) __U); 1214 } 1215 1216 extern __inline __m128i 1217 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1218 _mm_maskz_cvttpd_epi32 (__mmask8 __U, __m128d __A) 1219 { 1220 return (__m128i) __builtin_ia32_cvttpd2dq128_mask ((__v2df) __A, 1221 (__v4si) 1222 _mm_setzero_si128 (), 1223 (__mmask8) __U); 1224 } 1225 1226 extern __inline __m128i 1227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1228 _mm256_cvttpd_epu32 (__m256d __A) 1229 { 1230 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, 1231 (__v4si) 1232 _mm_setzero_si128 (), 1233 (__mmask8) -1); 1234 } 1235 1236 extern __inline __m128i 1237 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1238 _mm256_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m256d __A) 1239 { 1240 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, 1241 (__v4si) __W, 1242 (__mmask8) __U); 1243 } 1244 1245 extern __inline __m128i 1246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1247 _mm256_maskz_cvttpd_epu32 (__mmask8 __U, __m256d __A) 1248 { 1249 return (__m128i) __builtin_ia32_cvttpd2udq256_mask ((__v4df) __A, 1250 (__v4si) 1251 _mm_setzero_si128 (), 1252 (__mmask8) __U); 1253 } 1254 1255 extern __inline __m128i 1256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1257 _mm_cvttpd_epu32 (__m128d __A) 1258 { 1259 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, 1260 (__v4si) 1261 _mm_setzero_si128 (), 1262 (__mmask8) -1); 1263 } 1264 1265 extern __inline __m128i 1266 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1267 _mm_mask_cvttpd_epu32 (__m128i __W, __mmask8 __U, __m128d __A) 1268 { 1269 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, 1270 (__v4si) __W, 1271 (__mmask8) __U); 1272 } 1273 1274 extern __inline __m128i 1275 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1276 _mm_maskz_cvttpd_epu32 (__mmask8 __U, __m128d __A) 1277 { 1278 return (__m128i) __builtin_ia32_cvttpd2udq128_mask ((__v2df) __A, 1279 (__v4si) 1280 _mm_setzero_si128 (), 1281 (__mmask8) __U); 1282 } 1283 1284 extern __inline __m128i 1285 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1286 _mm256_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m256d __A) 1287 { 1288 return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, 1289 (__v4si) __W, 1290 (__mmask8) __U); 1291 } 1292 1293 extern __inline __m128i 1294 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1295 _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) 1296 { 1297 return (__m128i) __builtin_ia32_cvtpd2dq256_mask ((__v4df) __A, 1298 (__v4si) 1299 _mm_setzero_si128 (), 1300 (__mmask8) __U); 1301 } 1302 1303 extern __inline __m128i 1304 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1305 _mm_mask_cvtpd_epi32 (__m128i __W, __mmask8 __U, __m128d __A) 1306 { 1307 return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, 1308 (__v4si) __W, 1309 (__mmask8) __U); 1310 } 1311 1312 extern __inline __m128i 1313 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1314 _mm_maskz_cvtpd_epi32 (__mmask8 __U, __m128d __A) 1315 { 1316 return (__m128i) __builtin_ia32_cvtpd2dq128_mask ((__v2df) __A, 1317 (__v4si) 1318 _mm_setzero_si128 (), 1319 (__mmask8) __U); 1320 } 1321 1322 extern __inline __m256d 1323 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1324 _mm256_mask_cvtepi32_pd (__m256d __W, __mmask8 __U, __m128i __A) 1325 { 1326 return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, 1327 (__v4df) __W, 1328 (__mmask8) __U); 1329 } 1330 1331 extern __inline __m256d 1332 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1333 _mm256_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) 1334 { 1335 return (__m256d) __builtin_ia32_cvtdq2pd256_mask ((__v4si) __A, 1336 (__v4df) 1337 _mm256_setzero_pd (), 1338 (__mmask8) __U); 1339 } 1340 1341 extern __inline __m128d 1342 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1343 _mm_mask_cvtepi32_pd (__m128d __W, __mmask8 __U, __m128i __A) 1344 { 1345 return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, 1346 (__v2df) __W, 1347 (__mmask8) __U); 1348 } 1349 1350 extern __inline __m128d 1351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1352 _mm_maskz_cvtepi32_pd (__mmask8 __U, __m128i __A) 1353 { 1354 return (__m128d) __builtin_ia32_cvtdq2pd128_mask ((__v4si) __A, 1355 (__v2df) 1356 _mm_setzero_pd (), 1357 (__mmask8) __U); 1358 } 1359 1360 extern __inline __m256d 1361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1362 _mm256_cvtepu32_pd (__m128i __A) 1363 { 1364 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, 1365 (__v4df) 1366 _mm256_setzero_pd (), 1367 (__mmask8) -1); 1368 } 1369 1370 extern __inline __m256d 1371 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1372 _mm256_mask_cvtepu32_pd (__m256d __W, __mmask8 __U, __m128i __A) 1373 { 1374 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, 1375 (__v4df) __W, 1376 (__mmask8) __U); 1377 } 1378 1379 extern __inline __m256d 1380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1381 _mm256_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) 1382 { 1383 return (__m256d) __builtin_ia32_cvtudq2pd256_mask ((__v4si) __A, 1384 (__v4df) 1385 _mm256_setzero_pd (), 1386 (__mmask8) __U); 1387 } 1388 1389 extern __inline __m128d 1390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1391 _mm_cvtepu32_pd (__m128i __A) 1392 { 1393 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, 1394 (__v2df) 1395 _mm_setzero_pd (), 1396 (__mmask8) -1); 1397 } 1398 1399 extern __inline __m128d 1400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1401 _mm_mask_cvtepu32_pd (__m128d __W, __mmask8 __U, __m128i __A) 1402 { 1403 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, 1404 (__v2df) __W, 1405 (__mmask8) __U); 1406 } 1407 1408 extern __inline __m128d 1409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1410 _mm_maskz_cvtepu32_pd (__mmask8 __U, __m128i __A) 1411 { 1412 return (__m128d) __builtin_ia32_cvtudq2pd128_mask ((__v4si) __A, 1413 (__v2df) 1414 _mm_setzero_pd (), 1415 (__mmask8) __U); 1416 } 1417 1418 extern __inline __m256 1419 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1420 _mm256_mask_cvtepi32_ps (__m256 __W, __mmask8 __U, __m256i __A) 1421 { 1422 return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, 1423 (__v8sf) __W, 1424 (__mmask8) __U); 1425 } 1426 1427 extern __inline __m256 1428 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1429 _mm256_maskz_cvtepi32_ps (__mmask8 __U, __m256i __A) 1430 { 1431 return (__m256) __builtin_ia32_cvtdq2ps256_mask ((__v8si) __A, 1432 (__v8sf) 1433 _mm256_setzero_ps (), 1434 (__mmask8) __U); 1435 } 1436 1437 extern __inline __m128 1438 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1439 _mm_mask_cvtepi32_ps (__m128 __W, __mmask8 __U, __m128i __A) 1440 { 1441 return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, 1442 (__v4sf) __W, 1443 (__mmask8) __U); 1444 } 1445 1446 extern __inline __m128 1447 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1448 _mm_maskz_cvtepi32_ps (__mmask8 __U, __m128i __A) 1449 { 1450 return (__m128) __builtin_ia32_cvtdq2ps128_mask ((__v4si) __A, 1451 (__v4sf) 1452 _mm_setzero_ps (), 1453 (__mmask8) __U); 1454 } 1455 1456 extern __inline __m256 1457 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1458 _mm256_cvtepu32_ps (__m256i __A) 1459 { 1460 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, 1461 (__v8sf) 1462 _mm256_setzero_ps (), 1463 (__mmask8) -1); 1464 } 1465 1466 extern __inline __m256 1467 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1468 _mm256_mask_cvtepu32_ps (__m256 __W, __mmask8 __U, __m256i __A) 1469 { 1470 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, 1471 (__v8sf) __W, 1472 (__mmask8) __U); 1473 } 1474 1475 extern __inline __m256 1476 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1477 _mm256_maskz_cvtepu32_ps (__mmask8 __U, __m256i __A) 1478 { 1479 return (__m256) __builtin_ia32_cvtudq2ps256_mask ((__v8si) __A, 1480 (__v8sf) 1481 _mm256_setzero_ps (), 1482 (__mmask8) __U); 1483 } 1484 1485 extern __inline __m128 1486 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1487 _mm_cvtepu32_ps (__m128i __A) 1488 { 1489 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, 1490 (__v4sf) 1491 _mm_setzero_ps (), 1492 (__mmask8) -1); 1493 } 1494 1495 extern __inline __m128 1496 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1497 _mm_mask_cvtepu32_ps (__m128 __W, __mmask8 __U, __m128i __A) 1498 { 1499 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, 1500 (__v4sf) __W, 1501 (__mmask8) __U); 1502 } 1503 1504 extern __inline __m128 1505 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1506 _mm_maskz_cvtepu32_ps (__mmask8 __U, __m128i __A) 1507 { 1508 return (__m128) __builtin_ia32_cvtudq2ps128_mask ((__v4si) __A, 1509 (__v4sf) 1510 _mm_setzero_ps (), 1511 (__mmask8) __U); 1512 } 1513 1514 extern __inline __m256d 1515 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1516 _mm256_mask_cvtps_pd (__m256d __W, __mmask8 __U, __m128 __A) 1517 { 1518 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, 1519 (__v4df) __W, 1520 (__mmask8) __U); 1521 } 1522 1523 extern __inline __m256d 1524 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1525 _mm256_maskz_cvtps_pd (__mmask8 __U, __m128 __A) 1526 { 1527 return (__m256d) __builtin_ia32_cvtps2pd256_mask ((__v4sf) __A, 1528 (__v4df) 1529 _mm256_setzero_pd (), 1530 (__mmask8) __U); 1531 } 1532 1533 extern __inline __m128d 1534 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1535 _mm_mask_cvtps_pd (__m128d __W, __mmask8 __U, __m128 __A) 1536 { 1537 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, 1538 (__v2df) __W, 1539 (__mmask8) __U); 1540 } 1541 1542 extern __inline __m128d 1543 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1544 _mm_maskz_cvtps_pd (__mmask8 __U, __m128 __A) 1545 { 1546 return (__m128d) __builtin_ia32_cvtps2pd128_mask ((__v4sf) __A, 1547 (__v2df) 1548 _mm_setzero_pd (), 1549 (__mmask8) __U); 1550 } 1551 1552 extern __inline __m128i 1553 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1554 _mm_cvtepi32_epi8 (__m128i __A) 1555 { 1556 return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, 1557 (__v16qi) 1558 _mm_undefined_si128 (), 1559 (__mmask8) -1); 1560 } 1561 1562 extern __inline void 1563 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1564 _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 1565 { 1566 __builtin_ia32_pmovdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M); 1567 } 1568 1569 extern __inline __m128i 1570 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1571 _mm_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 1572 { 1573 return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, 1574 (__v16qi) __O, __M); 1575 } 1576 1577 extern __inline __m128i 1578 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1579 _mm_maskz_cvtepi32_epi8 (__mmask8 __M, __m128i __A) 1580 { 1581 return (__m128i) __builtin_ia32_pmovdb128_mask ((__v4si) __A, 1582 (__v16qi) 1583 _mm_setzero_si128 (), 1584 __M); 1585 } 1586 1587 extern __inline __m128i 1588 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1589 _mm256_cvtepi32_epi8 (__m256i __A) 1590 { 1591 return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, 1592 (__v16qi) 1593 _mm_undefined_si128 (), 1594 (__mmask8) -1); 1595 } 1596 1597 extern __inline __m128i 1598 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1599 _mm256_mask_cvtepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 1600 { 1601 return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, 1602 (__v16qi) __O, __M); 1603 } 1604 1605 extern __inline void 1606 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1607 _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 1608 { 1609 __builtin_ia32_pmovdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M); 1610 } 1611 1612 extern __inline __m128i 1613 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1614 _mm256_maskz_cvtepi32_epi8 (__mmask8 __M, __m256i __A) 1615 { 1616 return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A, 1617 (__v16qi) 1618 _mm_setzero_si128 (), 1619 __M); 1620 } 1621 1622 extern __inline __m128i 1623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1624 _mm_cvtsepi32_epi8 (__m128i __A) 1625 { 1626 return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, 1627 (__v16qi) 1628 _mm_undefined_si128 (), 1629 (__mmask8) -1); 1630 } 1631 1632 extern __inline void 1633 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1634 _mm_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 1635 { 1636 __builtin_ia32_pmovsdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M); 1637 } 1638 1639 extern __inline __m128i 1640 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1641 _mm_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 1642 { 1643 return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, 1644 (__v16qi) __O, __M); 1645 } 1646 1647 extern __inline __m128i 1648 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1649 _mm_maskz_cvtsepi32_epi8 (__mmask8 __M, __m128i __A) 1650 { 1651 return (__m128i) __builtin_ia32_pmovsdb128_mask ((__v4si) __A, 1652 (__v16qi) 1653 _mm_setzero_si128 (), 1654 __M); 1655 } 1656 1657 extern __inline __m128i 1658 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1659 _mm256_cvtsepi32_epi8 (__m256i __A) 1660 { 1661 return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, 1662 (__v16qi) 1663 _mm_undefined_si128 (), 1664 (__mmask8) -1); 1665 } 1666 1667 extern __inline void 1668 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1669 _mm256_mask_cvtsepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 1670 { 1671 __builtin_ia32_pmovsdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M); 1672 } 1673 1674 extern __inline __m128i 1675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1676 _mm256_mask_cvtsepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 1677 { 1678 return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, 1679 (__v16qi) __O, __M); 1680 } 1681 1682 extern __inline __m128i 1683 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1684 _mm256_maskz_cvtsepi32_epi8 (__mmask8 __M, __m256i __A) 1685 { 1686 return (__m128i) __builtin_ia32_pmovsdb256_mask ((__v8si) __A, 1687 (__v16qi) 1688 _mm_setzero_si128 (), 1689 __M); 1690 } 1691 1692 extern __inline __m128i 1693 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1694 _mm_cvtusepi32_epi8 (__m128i __A) 1695 { 1696 return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, 1697 (__v16qi) 1698 _mm_undefined_si128 (), 1699 (__mmask8) -1); 1700 } 1701 1702 extern __inline void 1703 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1704 _mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 1705 { 1706 __builtin_ia32_pmovusdb128mem_mask ((unsigned int *) __P, (__v4si) __A, __M); 1707 } 1708 1709 extern __inline __m128i 1710 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1711 _mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 1712 { 1713 return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, 1714 (__v16qi) __O, 1715 __M); 1716 } 1717 1718 extern __inline __m128i 1719 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1720 _mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A) 1721 { 1722 return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, 1723 (__v16qi) 1724 _mm_setzero_si128 (), 1725 __M); 1726 } 1727 1728 extern __inline __m128i 1729 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1730 _mm256_cvtusepi32_epi8 (__m256i __A) 1731 { 1732 return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, 1733 (__v16qi) 1734 _mm_undefined_si128 (), 1735 (__mmask8) -1); 1736 } 1737 1738 extern __inline void 1739 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1740 _mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 1741 { 1742 __builtin_ia32_pmovusdb256mem_mask ((unsigned long long *) __P, (__v8si) __A, __M); 1743 } 1744 1745 extern __inline __m128i 1746 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1747 _mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 1748 { 1749 return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, 1750 (__v16qi) __O, 1751 __M); 1752 } 1753 1754 extern __inline __m128i 1755 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1756 _mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A) 1757 { 1758 return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, 1759 (__v16qi) 1760 _mm_setzero_si128 (), 1761 __M); 1762 } 1763 1764 extern __inline __m128i 1765 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1766 _mm_cvtepi32_epi16 (__m128i __A) 1767 { 1768 return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, 1769 (__v8hi) 1770 _mm_setzero_si128 (), 1771 (__mmask8) -1); 1772 } 1773 1774 extern __inline void 1775 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1776 _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 1777 { 1778 __builtin_ia32_pmovdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M); 1779 } 1780 1781 extern __inline __m128i 1782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1783 _mm_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 1784 { 1785 return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, 1786 (__v8hi) __O, __M); 1787 } 1788 1789 extern __inline __m128i 1790 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1791 _mm_maskz_cvtepi32_epi16 (__mmask8 __M, __m128i __A) 1792 { 1793 return (__m128i) __builtin_ia32_pmovdw128_mask ((__v4si) __A, 1794 (__v8hi) 1795 _mm_setzero_si128 (), 1796 __M); 1797 } 1798 1799 extern __inline __m128i 1800 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1801 _mm256_cvtepi32_epi16 (__m256i __A) 1802 { 1803 return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, 1804 (__v8hi) 1805 _mm_setzero_si128 (), 1806 (__mmask8) -1); 1807 } 1808 1809 extern __inline void 1810 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1811 _mm256_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 1812 { 1813 __builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); 1814 } 1815 1816 extern __inline __m128i 1817 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1818 _mm256_mask_cvtepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 1819 { 1820 return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, 1821 (__v8hi) __O, __M); 1822 } 1823 1824 extern __inline __m128i 1825 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1826 _mm256_maskz_cvtepi32_epi16 (__mmask8 __M, __m256i __A) 1827 { 1828 return (__m128i) __builtin_ia32_pmovdw256_mask ((__v8si) __A, 1829 (__v8hi) 1830 _mm_setzero_si128 (), 1831 __M); 1832 } 1833 1834 extern __inline __m128i 1835 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1836 _mm_cvtsepi32_epi16 (__m128i __A) 1837 { 1838 return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, 1839 (__v8hi) 1840 _mm_setzero_si128 (), 1841 (__mmask8) -1); 1842 } 1843 1844 extern __inline void 1845 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1846 _mm_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 1847 { 1848 __builtin_ia32_pmovsdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M); 1849 } 1850 1851 extern __inline __m128i 1852 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1853 _mm_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 1854 { 1855 return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, 1856 (__v8hi)__O, 1857 __M); 1858 } 1859 1860 extern __inline __m128i 1861 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1862 _mm_maskz_cvtsepi32_epi16 (__mmask8 __M, __m128i __A) 1863 { 1864 return (__m128i) __builtin_ia32_pmovsdw128_mask ((__v4si) __A, 1865 (__v8hi) 1866 _mm_setzero_si128 (), 1867 __M); 1868 } 1869 1870 extern __inline __m128i 1871 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1872 _mm256_cvtsepi32_epi16 (__m256i __A) 1873 { 1874 return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, 1875 (__v8hi) 1876 _mm_undefined_si128 (), 1877 (__mmask8) -1); 1878 } 1879 1880 extern __inline void 1881 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1882 _mm256_mask_cvtsepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 1883 { 1884 __builtin_ia32_pmovsdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); 1885 } 1886 1887 extern __inline __m128i 1888 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1889 _mm256_mask_cvtsepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 1890 { 1891 return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, 1892 (__v8hi) __O, __M); 1893 } 1894 1895 extern __inline __m128i 1896 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1897 _mm256_maskz_cvtsepi32_epi16 (__mmask8 __M, __m256i __A) 1898 { 1899 return (__m128i) __builtin_ia32_pmovsdw256_mask ((__v8si) __A, 1900 (__v8hi) 1901 _mm_setzero_si128 (), 1902 __M); 1903 } 1904 1905 extern __inline __m128i 1906 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1907 _mm_cvtusepi32_epi16 (__m128i __A) 1908 { 1909 return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, 1910 (__v8hi) 1911 _mm_undefined_si128 (), 1912 (__mmask8) -1); 1913 } 1914 1915 extern __inline void 1916 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1917 _mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 1918 { 1919 __builtin_ia32_pmovusdw128mem_mask ((unsigned long long *) __P, (__v4si) __A, __M); 1920 } 1921 1922 extern __inline __m128i 1923 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1924 _mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 1925 { 1926 return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, 1927 (__v8hi) __O, __M); 1928 } 1929 1930 extern __inline __m128i 1931 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1932 _mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A) 1933 { 1934 return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, 1935 (__v8hi) 1936 _mm_setzero_si128 (), 1937 __M); 1938 } 1939 1940 extern __inline __m128i 1941 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1942 _mm256_cvtusepi32_epi16 (__m256i __A) 1943 { 1944 return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, 1945 (__v8hi) 1946 _mm_undefined_si128 (), 1947 (__mmask8) -1); 1948 } 1949 1950 extern __inline void 1951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1952 _mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 1953 { 1954 __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); 1955 } 1956 1957 extern __inline __m128i 1958 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1959 _mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 1960 { 1961 return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, 1962 (__v8hi) __O, __M); 1963 } 1964 1965 extern __inline __m128i 1966 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1967 _mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A) 1968 { 1969 return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, 1970 (__v8hi) 1971 _mm_setzero_si128 (), 1972 __M); 1973 } 1974 1975 extern __inline __m128i 1976 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1977 _mm_cvtepi64_epi8 (__m128i __A) 1978 { 1979 return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, 1980 (__v16qi) 1981 _mm_undefined_si128 (), 1982 (__mmask8) -1); 1983 } 1984 1985 extern __inline void 1986 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1987 _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 1988 { 1989 __builtin_ia32_pmovqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M); 1990 } 1991 1992 extern __inline __m128i 1993 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1994 _mm_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 1995 { 1996 return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, 1997 (__v16qi) __O, __M); 1998 } 1999 2000 extern __inline __m128i 2001 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2002 _mm_maskz_cvtepi64_epi8 (__mmask8 __M, __m128i __A) 2003 { 2004 return (__m128i) __builtin_ia32_pmovqb128_mask ((__v2di) __A, 2005 (__v16qi) 2006 _mm_setzero_si128 (), 2007 __M); 2008 } 2009 2010 extern __inline __m128i 2011 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2012 _mm256_cvtepi64_epi8 (__m256i __A) 2013 { 2014 return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, 2015 (__v16qi) 2016 _mm_undefined_si128 (), 2017 (__mmask8) -1); 2018 } 2019 2020 extern __inline void 2021 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2022 _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 2023 { 2024 __builtin_ia32_pmovqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M); 2025 } 2026 2027 extern __inline __m128i 2028 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2029 _mm256_mask_cvtepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 2030 { 2031 return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, 2032 (__v16qi) __O, __M); 2033 } 2034 2035 extern __inline __m128i 2036 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2037 _mm256_maskz_cvtepi64_epi8 (__mmask8 __M, __m256i __A) 2038 { 2039 return (__m128i) __builtin_ia32_pmovqb256_mask ((__v4di) __A, 2040 (__v16qi) 2041 _mm_setzero_si128 (), 2042 __M); 2043 } 2044 2045 extern __inline __m128i 2046 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2047 _mm_cvtsepi64_epi8 (__m128i __A) 2048 { 2049 return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, 2050 (__v16qi) 2051 _mm_undefined_si128 (), 2052 (__mmask8) -1); 2053 } 2054 2055 extern __inline void 2056 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2057 _mm_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 2058 { 2059 __builtin_ia32_pmovsqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M); 2060 } 2061 2062 extern __inline __m128i 2063 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2064 _mm_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 2065 { 2066 return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, 2067 (__v16qi) __O, __M); 2068 } 2069 2070 extern __inline __m128i 2071 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2072 _mm_maskz_cvtsepi64_epi8 (__mmask8 __M, __m128i __A) 2073 { 2074 return (__m128i) __builtin_ia32_pmovsqb128_mask ((__v2di) __A, 2075 (__v16qi) 2076 _mm_setzero_si128 (), 2077 __M); 2078 } 2079 2080 extern __inline __m128i 2081 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2082 _mm256_cvtsepi64_epi8 (__m256i __A) 2083 { 2084 return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, 2085 (__v16qi) 2086 _mm_undefined_si128 (), 2087 (__mmask8) -1); 2088 } 2089 2090 extern __inline void 2091 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2092 _mm256_mask_cvtsepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 2093 { 2094 __builtin_ia32_pmovsqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M); 2095 } 2096 2097 extern __inline __m128i 2098 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2099 _mm256_mask_cvtsepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 2100 { 2101 return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, 2102 (__v16qi) __O, __M); 2103 } 2104 2105 extern __inline __m128i 2106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2107 _mm256_maskz_cvtsepi64_epi8 (__mmask8 __M, __m256i __A) 2108 { 2109 return (__m128i) __builtin_ia32_pmovsqb256_mask ((__v4di) __A, 2110 (__v16qi) 2111 _mm_setzero_si128 (), 2112 __M); 2113 } 2114 2115 extern __inline __m128i 2116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2117 _mm_cvtusepi64_epi8 (__m128i __A) 2118 { 2119 return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, 2120 (__v16qi) 2121 _mm_undefined_si128 (), 2122 (__mmask8) -1); 2123 } 2124 2125 extern __inline void 2126 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2127 _mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) 2128 { 2129 __builtin_ia32_pmovusqb128mem_mask ((unsigned short *) __P, (__v2di) __A, __M); 2130 } 2131 2132 extern __inline __m128i 2133 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2134 _mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) 2135 { 2136 return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, 2137 (__v16qi) __O, 2138 __M); 2139 } 2140 2141 extern __inline __m128i 2142 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2143 _mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A) 2144 { 2145 return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, 2146 (__v16qi) 2147 _mm_setzero_si128 (), 2148 __M); 2149 } 2150 2151 extern __inline __m128i 2152 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2153 _mm256_cvtusepi64_epi8 (__m256i __A) 2154 { 2155 return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, 2156 (__v16qi) 2157 _mm_undefined_si128 (), 2158 (__mmask8) -1); 2159 } 2160 2161 extern __inline void 2162 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2163 _mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) 2164 { 2165 __builtin_ia32_pmovusqb256mem_mask ((unsigned int *) __P, (__v4di) __A, __M); 2166 } 2167 2168 extern __inline __m128i 2169 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2170 _mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) 2171 { 2172 return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, 2173 (__v16qi) __O, 2174 __M); 2175 } 2176 2177 extern __inline __m128i 2178 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2179 _mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A) 2180 { 2181 return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, 2182 (__v16qi) 2183 _mm_setzero_si128 (), 2184 __M); 2185 } 2186 2187 extern __inline __m128i 2188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2189 _mm_cvtepi64_epi16 (__m128i __A) 2190 { 2191 return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, 2192 (__v8hi) 2193 _mm_undefined_si128 (), 2194 (__mmask8) -1); 2195 } 2196 2197 extern __inline void 2198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2199 _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 2200 { 2201 __builtin_ia32_pmovqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M); 2202 } 2203 2204 extern __inline __m128i 2205 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2206 _mm_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 2207 { 2208 return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, 2209 (__v8hi)__O, 2210 __M); 2211 } 2212 2213 extern __inline __m128i 2214 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2215 _mm_maskz_cvtepi64_epi16 (__mmask8 __M, __m128i __A) 2216 { 2217 return (__m128i) __builtin_ia32_pmovqw128_mask ((__v2di) __A, 2218 (__v8hi) 2219 _mm_setzero_si128 (), 2220 __M); 2221 } 2222 2223 extern __inline __m128i 2224 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2225 _mm256_cvtepi64_epi16 (__m256i __A) 2226 { 2227 return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, 2228 (__v8hi) 2229 _mm_undefined_si128 (), 2230 (__mmask8) -1); 2231 } 2232 2233 extern __inline void 2234 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2235 _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 2236 { 2237 __builtin_ia32_pmovqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M); 2238 } 2239 2240 extern __inline __m128i 2241 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2242 _mm256_mask_cvtepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 2243 { 2244 return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, 2245 (__v8hi) __O, __M); 2246 } 2247 2248 extern __inline __m128i 2249 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2250 _mm256_maskz_cvtepi64_epi16 (__mmask8 __M, __m256i __A) 2251 { 2252 return (__m128i) __builtin_ia32_pmovqw256_mask ((__v4di) __A, 2253 (__v8hi) 2254 _mm_setzero_si128 (), 2255 __M); 2256 } 2257 2258 extern __inline __m128i 2259 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2260 _mm_cvtsepi64_epi16 (__m128i __A) 2261 { 2262 return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, 2263 (__v8hi) 2264 _mm_undefined_si128 (), 2265 (__mmask8) -1); 2266 } 2267 2268 extern __inline void 2269 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2270 _mm_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 2271 { 2272 __builtin_ia32_pmovsqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M); 2273 } 2274 2275 extern __inline __m128i 2276 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2277 _mm_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 2278 { 2279 return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, 2280 (__v8hi) __O, __M); 2281 } 2282 2283 extern __inline __m128i 2284 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2285 _mm_maskz_cvtsepi64_epi16 (__mmask8 __M, __m128i __A) 2286 { 2287 return (__m128i) __builtin_ia32_pmovsqw128_mask ((__v2di) __A, 2288 (__v8hi) 2289 _mm_setzero_si128 (), 2290 __M); 2291 } 2292 2293 extern __inline __m128i 2294 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2295 _mm256_cvtsepi64_epi16 (__m256i __A) 2296 { 2297 return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, 2298 (__v8hi) 2299 _mm_undefined_si128 (), 2300 (__mmask8) -1); 2301 } 2302 2303 extern __inline void 2304 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2305 _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 2306 { 2307 __builtin_ia32_pmovsqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M); 2308 } 2309 2310 extern __inline __m128i 2311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2312 _mm256_mask_cvtsepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 2313 { 2314 return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, 2315 (__v8hi) __O, __M); 2316 } 2317 2318 extern __inline __m128i 2319 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2320 _mm256_maskz_cvtsepi64_epi16 (__mmask8 __M, __m256i __A) 2321 { 2322 return (__m128i) __builtin_ia32_pmovsqw256_mask ((__v4di) __A, 2323 (__v8hi) 2324 _mm_setzero_si128 (), 2325 __M); 2326 } 2327 2328 extern __inline __m128i 2329 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2330 _mm_cvtusepi64_epi16 (__m128i __A) 2331 { 2332 return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, 2333 (__v8hi) 2334 _mm_undefined_si128 (), 2335 (__mmask8) -1); 2336 } 2337 2338 extern __inline void 2339 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2340 _mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) 2341 { 2342 __builtin_ia32_pmovusqw128mem_mask ((unsigned int *) __P, (__v2di) __A, __M); 2343 } 2344 2345 extern __inline __m128i 2346 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2347 _mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) 2348 { 2349 return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, 2350 (__v8hi) __O, __M); 2351 } 2352 2353 extern __inline __m128i 2354 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2355 _mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A) 2356 { 2357 return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, 2358 (__v8hi) 2359 _mm_setzero_si128 (), 2360 __M); 2361 } 2362 2363 extern __inline __m128i 2364 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2365 _mm256_cvtusepi64_epi16 (__m256i __A) 2366 { 2367 return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, 2368 (__v8hi) 2369 _mm_undefined_si128 (), 2370 (__mmask8) -1); 2371 } 2372 2373 extern __inline void 2374 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2375 _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) 2376 { 2377 __builtin_ia32_pmovusqw256mem_mask ((unsigned long long *) __P, (__v4di) __A, __M); 2378 } 2379 2380 extern __inline __m128i 2381 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2382 _mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) 2383 { 2384 return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, 2385 (__v8hi) __O, __M); 2386 } 2387 2388 extern __inline __m128i 2389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2390 _mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A) 2391 { 2392 return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, 2393 (__v8hi) 2394 _mm_setzero_si128 (), 2395 __M); 2396 } 2397 2398 extern __inline __m128i 2399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2400 _mm_cvtepi64_epi32 (__m128i __A) 2401 { 2402 return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, 2403 (__v4si) 2404 _mm_undefined_si128 (), 2405 (__mmask8) -1); 2406 } 2407 2408 extern __inline void 2409 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2410 _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) 2411 { 2412 __builtin_ia32_pmovqd128mem_mask ((unsigned long long *) __P, 2413 (__v2di) __A, __M); 2414 } 2415 2416 extern __inline __m128i 2417 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2418 _mm_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) 2419 { 2420 return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, 2421 (__v4si) __O, __M); 2422 } 2423 2424 extern __inline __m128i 2425 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2426 _mm_maskz_cvtepi64_epi32 (__mmask8 __M, __m128i __A) 2427 { 2428 return (__m128i) __builtin_ia32_pmovqd128_mask ((__v2di) __A, 2429 (__v4si) 2430 _mm_setzero_si128 (), 2431 __M); 2432 } 2433 2434 extern __inline __m128i 2435 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2436 _mm256_cvtepi64_epi32 (__m256i __A) 2437 { 2438 return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, 2439 (__v4si) 2440 _mm_undefined_si128 (), 2441 (__mmask8) -1); 2442 } 2443 2444 extern __inline void 2445 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2446 _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) 2447 { 2448 __builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); 2449 } 2450 2451 extern __inline __m128i 2452 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2453 _mm256_mask_cvtepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) 2454 { 2455 return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, 2456 (__v4si) __O, __M); 2457 } 2458 2459 extern __inline __m128i 2460 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2461 _mm256_maskz_cvtepi64_epi32 (__mmask8 __M, __m256i __A) 2462 { 2463 return (__m128i) __builtin_ia32_pmovqd256_mask ((__v4di) __A, 2464 (__v4si) 2465 _mm_setzero_si128 (), 2466 __M); 2467 } 2468 2469 extern __inline __m128i 2470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2471 _mm_cvtsepi64_epi32 (__m128i __A) 2472 { 2473 return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, 2474 (__v4si) 2475 _mm_undefined_si128 (), 2476 (__mmask8) -1); 2477 } 2478 2479 extern __inline void 2480 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2481 _mm_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) 2482 { 2483 __builtin_ia32_pmovsqd128mem_mask ((unsigned long long *) __P, (__v2di) __A, __M); 2484 } 2485 2486 extern __inline __m128i 2487 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2488 _mm_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) 2489 { 2490 return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, 2491 (__v4si) __O, __M); 2492 } 2493 2494 extern __inline __m128i 2495 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2496 _mm_maskz_cvtsepi64_epi32 (__mmask8 __M, __m128i __A) 2497 { 2498 return (__m128i) __builtin_ia32_pmovsqd128_mask ((__v2di) __A, 2499 (__v4si) 2500 _mm_setzero_si128 (), 2501 __M); 2502 } 2503 2504 extern __inline __m128i 2505 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2506 _mm256_cvtsepi64_epi32 (__m256i __A) 2507 { 2508 return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, 2509 (__v4si) 2510 _mm_undefined_si128 (), 2511 (__mmask8) -1); 2512 } 2513 2514 extern __inline void 2515 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2516 _mm256_mask_cvtsepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) 2517 { 2518 __builtin_ia32_pmovsqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); 2519 } 2520 2521 extern __inline __m128i 2522 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2523 _mm256_mask_cvtsepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) 2524 { 2525 return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, 2526 (__v4si)__O, 2527 __M); 2528 } 2529 2530 extern __inline __m128i 2531 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2532 _mm256_maskz_cvtsepi64_epi32 (__mmask8 __M, __m256i __A) 2533 { 2534 return (__m128i) __builtin_ia32_pmovsqd256_mask ((__v4di) __A, 2535 (__v4si) 2536 _mm_setzero_si128 (), 2537 __M); 2538 } 2539 2540 extern __inline __m128i 2541 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2542 _mm_cvtusepi64_epi32 (__m128i __A) 2543 { 2544 return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, 2545 (__v4si) 2546 _mm_undefined_si128 (), 2547 (__mmask8) -1); 2548 } 2549 2550 extern __inline void 2551 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2552 _mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) 2553 { 2554 __builtin_ia32_pmovusqd128mem_mask ((unsigned long long *) __P, (__v2di) __A, __M); 2555 } 2556 2557 extern __inline __m128i 2558 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2559 _mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) 2560 { 2561 return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, 2562 (__v4si) __O, __M); 2563 } 2564 2565 extern __inline __m128i 2566 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2567 _mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A) 2568 { 2569 return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, 2570 (__v4si) 2571 _mm_setzero_si128 (), 2572 __M); 2573 } 2574 2575 extern __inline __m128i 2576 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2577 _mm256_cvtusepi64_epi32 (__m256i __A) 2578 { 2579 return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, 2580 (__v4si) 2581 _mm_undefined_si128 (), 2582 (__mmask8) -1); 2583 } 2584 2585 extern __inline void 2586 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2587 _mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) 2588 { 2589 __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); 2590 } 2591 2592 extern __inline __m128i 2593 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2594 _mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) 2595 { 2596 return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, 2597 (__v4si) __O, __M); 2598 } 2599 2600 extern __inline __m128i 2601 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2602 _mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A) 2603 { 2604 return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, 2605 (__v4si) 2606 _mm_setzero_si128 (), 2607 __M); 2608 } 2609 2610 extern __inline __m256 2611 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2612 _mm256_mask_broadcastss_ps (__m256 __O, __mmask8 __M, __m128 __A) 2613 { 2614 return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A, 2615 (__v8sf) __O, 2616 __M); 2617 } 2618 2619 extern __inline __m256 2620 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2621 _mm256_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) 2622 { 2623 return (__m256) __builtin_ia32_broadcastss256_mask ((__v4sf) __A, 2624 (__v8sf) 2625 _mm256_setzero_ps (), 2626 __M); 2627 } 2628 2629 extern __inline __m128 2630 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2631 _mm_mask_broadcastss_ps (__m128 __O, __mmask8 __M, __m128 __A) 2632 { 2633 return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A, 2634 (__v4sf) __O, 2635 __M); 2636 } 2637 2638 extern __inline __m128 2639 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2640 _mm_maskz_broadcastss_ps (__mmask8 __M, __m128 __A) 2641 { 2642 return (__m128) __builtin_ia32_broadcastss128_mask ((__v4sf) __A, 2643 (__v4sf) 2644 _mm_setzero_ps (), 2645 __M); 2646 } 2647 2648 extern __inline __m256d 2649 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2650 _mm256_mask_broadcastsd_pd (__m256d __O, __mmask8 __M, __m128d __A) 2651 { 2652 return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A, 2653 (__v4df) __O, 2654 __M); 2655 } 2656 2657 extern __inline __m256d 2658 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2659 _mm256_maskz_broadcastsd_pd (__mmask8 __M, __m128d __A) 2660 { 2661 return (__m256d) __builtin_ia32_broadcastsd256_mask ((__v2df) __A, 2662 (__v4df) 2663 _mm256_setzero_pd (), 2664 __M); 2665 } 2666 2667 extern __inline __m256i 2668 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2669 _mm256_mask_broadcastd_epi32 (__m256i __O, __mmask8 __M, __m128i __A) 2670 { 2671 return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A, 2672 (__v8si) __O, 2673 __M); 2674 } 2675 2676 extern __inline __m256i 2677 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2678 _mm256_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) 2679 { 2680 return (__m256i) __builtin_ia32_pbroadcastd256_mask ((__v4si) __A, 2681 (__v8si) 2682 _mm256_setzero_si256 (), 2683 __M); 2684 } 2685 2686 extern __inline __m256i 2687 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2688 _mm256_mask_set1_epi32 (__m256i __O, __mmask8 __M, int __A) 2689 { 2690 return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A, (__v8si) __O, 2691 __M); 2692 } 2693 2694 extern __inline __m256i 2695 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2696 _mm256_maskz_set1_epi32 (__mmask8 __M, int __A) 2697 { 2698 return (__m256i) __builtin_ia32_pbroadcastd256_gpr_mask (__A, 2699 (__v8si) 2700 _mm256_setzero_si256 (), 2701 __M); 2702 } 2703 2704 extern __inline __m128i 2705 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2706 _mm_mask_broadcastd_epi32 (__m128i __O, __mmask8 __M, __m128i __A) 2707 { 2708 return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A, 2709 (__v4si) __O, 2710 __M); 2711 } 2712 2713 extern __inline __m128i 2714 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2715 _mm_maskz_broadcastd_epi32 (__mmask8 __M, __m128i __A) 2716 { 2717 return (__m128i) __builtin_ia32_pbroadcastd128_mask ((__v4si) __A, 2718 (__v4si) 2719 _mm_setzero_si128 (), 2720 __M); 2721 } 2722 2723 extern __inline __m128i 2724 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2725 _mm_mask_set1_epi32 (__m128i __O, __mmask8 __M, int __A) 2726 { 2727 return (__m128i) __builtin_ia32_pbroadcastd128_gpr_mask (__A, (__v4si) __O, 2728 __M); 2729 } 2730 2731 extern __inline __m128i 2732 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2733 _mm_maskz_set1_epi32 (__mmask8 __M, int __A) 2734 { 2735 return (__m128i) 2736 __builtin_ia32_pbroadcastd128_gpr_mask (__A, 2737 (__v4si) _mm_setzero_si128 (), 2738 __M); 2739 } 2740 2741 extern __inline __m256i 2742 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2743 _mm256_mask_broadcastq_epi64 (__m256i __O, __mmask8 __M, __m128i __A) 2744 { 2745 return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A, 2746 (__v4di) __O, 2747 __M); 2748 } 2749 2750 extern __inline __m256i 2751 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2752 _mm256_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) 2753 { 2754 return (__m256i) __builtin_ia32_pbroadcastq256_mask ((__v2di) __A, 2755 (__v4di) 2756 _mm256_setzero_si256 (), 2757 __M); 2758 } 2759 2760 extern __inline __m256i 2761 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2762 _mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A) 2763 { 2764 return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O, 2765 __M); 2766 } 2767 2768 extern __inline __m256i 2769 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2770 _mm256_maskz_set1_epi64 (__mmask8 __M, long long __A) 2771 { 2772 return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, 2773 (__v4di) 2774 _mm256_setzero_si256 (), 2775 __M); 2776 } 2777 2778 extern __inline __m128i 2779 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2780 _mm_mask_broadcastq_epi64 (__m128i __O, __mmask8 __M, __m128i __A) 2781 { 2782 return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A, 2783 (__v2di) __O, 2784 __M); 2785 } 2786 2787 extern __inline __m128i 2788 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2789 _mm_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A) 2790 { 2791 return (__m128i) __builtin_ia32_pbroadcastq128_mask ((__v2di) __A, 2792 (__v2di) 2793 _mm_setzero_si128 (), 2794 __M); 2795 } 2796 2797 extern __inline __m128i 2798 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2799 _mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A) 2800 { 2801 return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O, 2802 __M); 2803 } 2804 2805 extern __inline __m128i 2806 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2807 _mm_maskz_set1_epi64 (__mmask8 __M, long long __A) 2808 { 2809 return (__m128i) 2810 __builtin_ia32_pbroadcastq128_gpr_mask (__A, 2811 (__v2di) _mm_setzero_si128 (), 2812 __M); 2813 } 2814 2815 extern __inline __m256 2816 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2817 _mm256_broadcast_f32x4 (__m128 __A) 2818 { 2819 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, 2820 (__v8sf)_mm256_undefined_pd (), 2821 (__mmask8) -1); 2822 } 2823 2824 extern __inline __m256 2825 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2826 _mm256_mask_broadcast_f32x4 (__m256 __O, __mmask8 __M, __m128 __A) 2827 { 2828 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, 2829 (__v8sf) __O, 2830 __M); 2831 } 2832 2833 extern __inline __m256 2834 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2835 _mm256_maskz_broadcast_f32x4 (__mmask8 __M, __m128 __A) 2836 { 2837 return (__m256) __builtin_ia32_broadcastf32x4_256_mask ((__v4sf) __A, 2838 (__v8sf) 2839 _mm256_setzero_ps (), 2840 __M); 2841 } 2842 2843 extern __inline __m256i 2844 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2845 _mm256_broadcast_i32x4 (__m128i __A) 2846 { 2847 return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) 2848 __A, 2849 (__v8si)_mm256_undefined_si256 (), 2850 (__mmask8) -1); 2851 } 2852 2853 extern __inline __m256i 2854 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2855 _mm256_mask_broadcast_i32x4 (__m256i __O, __mmask8 __M, __m128i __A) 2856 { 2857 return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) 2858 __A, 2859 (__v8si) 2860 __O, __M); 2861 } 2862 2863 extern __inline __m256i 2864 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2865 _mm256_maskz_broadcast_i32x4 (__mmask8 __M, __m128i __A) 2866 { 2867 return (__m256i) __builtin_ia32_broadcasti32x4_256_mask ((__v4si) 2868 __A, 2869 (__v8si) 2870 _mm256_setzero_si256 (), 2871 __M); 2872 } 2873 2874 extern __inline __m256i 2875 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2876 _mm256_mask_cvtepi8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) 2877 { 2878 return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, 2879 (__v8si) __W, 2880 (__mmask8) __U); 2881 } 2882 2883 extern __inline __m256i 2884 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2885 _mm256_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) 2886 { 2887 return (__m256i) __builtin_ia32_pmovsxbd256_mask ((__v16qi) __A, 2888 (__v8si) 2889 _mm256_setzero_si256 (), 2890 (__mmask8) __U); 2891 } 2892 2893 extern __inline __m128i 2894 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2895 _mm_mask_cvtepi8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 2896 { 2897 return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, 2898 (__v4si) __W, 2899 (__mmask8) __U); 2900 } 2901 2902 extern __inline __m128i 2903 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2904 _mm_maskz_cvtepi8_epi32 (__mmask8 __U, __m128i __A) 2905 { 2906 return (__m128i) __builtin_ia32_pmovsxbd128_mask ((__v16qi) __A, 2907 (__v4si) 2908 _mm_setzero_si128 (), 2909 (__mmask8) __U); 2910 } 2911 2912 extern __inline __m256i 2913 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2914 _mm256_mask_cvtepi8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) 2915 { 2916 return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, 2917 (__v4di) __W, 2918 (__mmask8) __U); 2919 } 2920 2921 extern __inline __m256i 2922 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2923 _mm256_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) 2924 { 2925 return (__m256i) __builtin_ia32_pmovsxbq256_mask ((__v16qi) __A, 2926 (__v4di) 2927 _mm256_setzero_si256 (), 2928 (__mmask8) __U); 2929 } 2930 2931 extern __inline __m128i 2932 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2933 _mm_mask_cvtepi8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 2934 { 2935 return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, 2936 (__v2di) __W, 2937 (__mmask8) __U); 2938 } 2939 2940 extern __inline __m128i 2941 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2942 _mm_maskz_cvtepi8_epi64 (__mmask8 __U, __m128i __A) 2943 { 2944 return (__m128i) __builtin_ia32_pmovsxbq128_mask ((__v16qi) __A, 2945 (__v2di) 2946 _mm_setzero_si128 (), 2947 (__mmask8) __U); 2948 } 2949 2950 extern __inline __m256i 2951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2952 _mm256_mask_cvtepi16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) 2953 { 2954 return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, 2955 (__v8si) __W, 2956 (__mmask8) __U); 2957 } 2958 2959 extern __inline __m256i 2960 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2961 _mm256_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) 2962 { 2963 return (__m256i) __builtin_ia32_pmovsxwd256_mask ((__v8hi) __A, 2964 (__v8si) 2965 _mm256_setzero_si256 (), 2966 (__mmask8) __U); 2967 } 2968 2969 extern __inline __m128i 2970 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2971 _mm_mask_cvtepi16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 2972 { 2973 return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, 2974 (__v4si) __W, 2975 (__mmask8) __U); 2976 } 2977 2978 extern __inline __m128i 2979 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2980 _mm_maskz_cvtepi16_epi32 (__mmask8 __U, __m128i __A) 2981 { 2982 return (__m128i) __builtin_ia32_pmovsxwd128_mask ((__v8hi) __A, 2983 (__v4si) 2984 _mm_setzero_si128 (), 2985 (__mmask8) __U); 2986 } 2987 2988 extern __inline __m256i 2989 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2990 _mm256_mask_cvtepi16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) 2991 { 2992 return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, 2993 (__v4di) __W, 2994 (__mmask8) __U); 2995 } 2996 2997 extern __inline __m256i 2998 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2999 _mm256_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) 3000 { 3001 return (__m256i) __builtin_ia32_pmovsxwq256_mask ((__v8hi) __A, 3002 (__v4di) 3003 _mm256_setzero_si256 (), 3004 (__mmask8) __U); 3005 } 3006 3007 extern __inline __m128i 3008 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3009 _mm_mask_cvtepi16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 3010 { 3011 return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, 3012 (__v2di) __W, 3013 (__mmask8) __U); 3014 } 3015 3016 extern __inline __m128i 3017 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3018 _mm_maskz_cvtepi16_epi64 (__mmask8 __U, __m128i __A) 3019 { 3020 return (__m128i) __builtin_ia32_pmovsxwq128_mask ((__v8hi) __A, 3021 (__v2di) 3022 _mm_setzero_si128 (), 3023 (__mmask8) __U); 3024 } 3025 3026 extern __inline __m256i 3027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3028 _mm256_mask_cvtepi32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) 3029 { 3030 return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, 3031 (__v4di) __W, 3032 (__mmask8) __U); 3033 } 3034 3035 extern __inline __m256i 3036 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3037 _mm256_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) 3038 { 3039 return (__m256i) __builtin_ia32_pmovsxdq256_mask ((__v4si) __X, 3040 (__v4di) 3041 _mm256_setzero_si256 (), 3042 (__mmask8) __U); 3043 } 3044 3045 extern __inline __m128i 3046 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3047 _mm_mask_cvtepi32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) 3048 { 3049 return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, 3050 (__v2di) __W, 3051 (__mmask8) __U); 3052 } 3053 3054 extern __inline __m128i 3055 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3056 _mm_maskz_cvtepi32_epi64 (__mmask8 __U, __m128i __X) 3057 { 3058 return (__m128i) __builtin_ia32_pmovsxdq128_mask ((__v4si) __X, 3059 (__v2di) 3060 _mm_setzero_si128 (), 3061 (__mmask8) __U); 3062 } 3063 3064 extern __inline __m256i 3065 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3066 _mm256_mask_cvtepu8_epi32 (__m256i __W, __mmask8 __U, __m128i __A) 3067 { 3068 return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, 3069 (__v8si) __W, 3070 (__mmask8) __U); 3071 } 3072 3073 extern __inline __m256i 3074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3075 _mm256_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) 3076 { 3077 return (__m256i) __builtin_ia32_pmovzxbd256_mask ((__v16qi) __A, 3078 (__v8si) 3079 _mm256_setzero_si256 (), 3080 (__mmask8) __U); 3081 } 3082 3083 extern __inline __m128i 3084 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3085 _mm_mask_cvtepu8_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 3086 { 3087 return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, 3088 (__v4si) __W, 3089 (__mmask8) __U); 3090 } 3091 3092 extern __inline __m128i 3093 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3094 _mm_maskz_cvtepu8_epi32 (__mmask8 __U, __m128i __A) 3095 { 3096 return (__m128i) __builtin_ia32_pmovzxbd128_mask ((__v16qi) __A, 3097 (__v4si) 3098 _mm_setzero_si128 (), 3099 (__mmask8) __U); 3100 } 3101 3102 extern __inline __m256i 3103 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3104 _mm256_mask_cvtepu8_epi64 (__m256i __W, __mmask8 __U, __m128i __A) 3105 { 3106 return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, 3107 (__v4di) __W, 3108 (__mmask8) __U); 3109 } 3110 3111 extern __inline __m256i 3112 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3113 _mm256_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) 3114 { 3115 return (__m256i) __builtin_ia32_pmovzxbq256_mask ((__v16qi) __A, 3116 (__v4di) 3117 _mm256_setzero_si256 (), 3118 (__mmask8) __U); 3119 } 3120 3121 extern __inline __m128i 3122 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3123 _mm_mask_cvtepu8_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 3124 { 3125 return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, 3126 (__v2di) __W, 3127 (__mmask8) __U); 3128 } 3129 3130 extern __inline __m128i 3131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3132 _mm_maskz_cvtepu8_epi64 (__mmask8 __U, __m128i __A) 3133 { 3134 return (__m128i) __builtin_ia32_pmovzxbq128_mask ((__v16qi) __A, 3135 (__v2di) 3136 _mm_setzero_si128 (), 3137 (__mmask8) __U); 3138 } 3139 3140 extern __inline __m256i 3141 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3142 _mm256_mask_cvtepu16_epi32 (__m256i __W, __mmask8 __U, __m128i __A) 3143 { 3144 return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, 3145 (__v8si) __W, 3146 (__mmask8) __U); 3147 } 3148 3149 extern __inline __m256i 3150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3151 _mm256_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) 3152 { 3153 return (__m256i) __builtin_ia32_pmovzxwd256_mask ((__v8hi) __A, 3154 (__v8si) 3155 _mm256_setzero_si256 (), 3156 (__mmask8) __U); 3157 } 3158 3159 extern __inline __m128i 3160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3161 _mm_mask_cvtepu16_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 3162 { 3163 return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, 3164 (__v4si) __W, 3165 (__mmask8) __U); 3166 } 3167 3168 extern __inline __m128i 3169 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3170 _mm_maskz_cvtepu16_epi32 (__mmask8 __U, __m128i __A) 3171 { 3172 return (__m128i) __builtin_ia32_pmovzxwd128_mask ((__v8hi) __A, 3173 (__v4si) 3174 _mm_setzero_si128 (), 3175 (__mmask8) __U); 3176 } 3177 3178 extern __inline __m256i 3179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3180 _mm256_mask_cvtepu16_epi64 (__m256i __W, __mmask8 __U, __m128i __A) 3181 { 3182 return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, 3183 (__v4di) __W, 3184 (__mmask8) __U); 3185 } 3186 3187 extern __inline __m256i 3188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3189 _mm256_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) 3190 { 3191 return (__m256i) __builtin_ia32_pmovzxwq256_mask ((__v8hi) __A, 3192 (__v4di) 3193 _mm256_setzero_si256 (), 3194 (__mmask8) __U); 3195 } 3196 3197 extern __inline __m128i 3198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3199 _mm_mask_cvtepu16_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 3200 { 3201 return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, 3202 (__v2di) __W, 3203 (__mmask8) __U); 3204 } 3205 3206 extern __inline __m128i 3207 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3208 _mm_maskz_cvtepu16_epi64 (__mmask8 __U, __m128i __A) 3209 { 3210 return (__m128i) __builtin_ia32_pmovzxwq128_mask ((__v8hi) __A, 3211 (__v2di) 3212 _mm_setzero_si128 (), 3213 (__mmask8) __U); 3214 } 3215 3216 extern __inline __m256i 3217 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3218 _mm256_mask_cvtepu32_epi64 (__m256i __W, __mmask8 __U, __m128i __X) 3219 { 3220 return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, 3221 (__v4di) __W, 3222 (__mmask8) __U); 3223 } 3224 3225 extern __inline __m256i 3226 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3227 _mm256_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) 3228 { 3229 return (__m256i) __builtin_ia32_pmovzxdq256_mask ((__v4si) __X, 3230 (__v4di) 3231 _mm256_setzero_si256 (), 3232 (__mmask8) __U); 3233 } 3234 3235 extern __inline __m128i 3236 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3237 _mm_mask_cvtepu32_epi64 (__m128i __W, __mmask8 __U, __m128i __X) 3238 { 3239 return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, 3240 (__v2di) __W, 3241 (__mmask8) __U); 3242 } 3243 3244 extern __inline __m128i 3245 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3246 _mm_maskz_cvtepu32_epi64 (__mmask8 __U, __m128i __X) 3247 { 3248 return (__m128i) __builtin_ia32_pmovzxdq128_mask ((__v4si) __X, 3249 (__v2di) 3250 _mm_setzero_si128 (), 3251 (__mmask8) __U); 3252 } 3253 3254 extern __inline __m256d 3255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3256 _mm256_rcp14_pd (__m256d __A) 3257 { 3258 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, 3259 (__v4df) 3260 _mm256_setzero_pd (), 3261 (__mmask8) -1); 3262 } 3263 3264 extern __inline __m256d 3265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3266 _mm256_mask_rcp14_pd (__m256d __W, __mmask8 __U, __m256d __A) 3267 { 3268 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, 3269 (__v4df) __W, 3270 (__mmask8) __U); 3271 } 3272 3273 extern __inline __m256d 3274 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3275 _mm256_maskz_rcp14_pd (__mmask8 __U, __m256d __A) 3276 { 3277 return (__m256d) __builtin_ia32_rcp14pd256_mask ((__v4df) __A, 3278 (__v4df) 3279 _mm256_setzero_pd (), 3280 (__mmask8) __U); 3281 } 3282 3283 extern __inline __m128d 3284 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3285 _mm_rcp14_pd (__m128d __A) 3286 { 3287 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, 3288 (__v2df) 3289 _mm_setzero_pd (), 3290 (__mmask8) -1); 3291 } 3292 3293 extern __inline __m128d 3294 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3295 _mm_mask_rcp14_pd (__m128d __W, __mmask8 __U, __m128d __A) 3296 { 3297 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, 3298 (__v2df) __W, 3299 (__mmask8) __U); 3300 } 3301 3302 extern __inline __m128d 3303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3304 _mm_maskz_rcp14_pd (__mmask8 __U, __m128d __A) 3305 { 3306 return (__m128d) __builtin_ia32_rcp14pd128_mask ((__v2df) __A, 3307 (__v2df) 3308 _mm_setzero_pd (), 3309 (__mmask8) __U); 3310 } 3311 3312 extern __inline __m256 3313 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3314 _mm256_rcp14_ps (__m256 __A) 3315 { 3316 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, 3317 (__v8sf) 3318 _mm256_setzero_ps (), 3319 (__mmask8) -1); 3320 } 3321 3322 extern __inline __m256 3323 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3324 _mm256_mask_rcp14_ps (__m256 __W, __mmask8 __U, __m256 __A) 3325 { 3326 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, 3327 (__v8sf) __W, 3328 (__mmask8) __U); 3329 } 3330 3331 extern __inline __m256 3332 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3333 _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A) 3334 { 3335 return (__m256) __builtin_ia32_rcp14ps256_mask ((__v8sf) __A, 3336 (__v8sf) 3337 _mm256_setzero_ps (), 3338 (__mmask8) __U); 3339 } 3340 3341 extern __inline __m128 3342 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3343 _mm_rcp14_ps (__m128 __A) 3344 { 3345 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, 3346 (__v4sf) 3347 _mm_setzero_ps (), 3348 (__mmask8) -1); 3349 } 3350 3351 extern __inline __m128 3352 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3353 _mm_mask_rcp14_ps (__m128 __W, __mmask8 __U, __m128 __A) 3354 { 3355 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, 3356 (__v4sf) __W, 3357 (__mmask8) __U); 3358 } 3359 3360 extern __inline __m128 3361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3362 _mm_maskz_rcp14_ps (__mmask8 __U, __m128 __A) 3363 { 3364 return (__m128) __builtin_ia32_rcp14ps128_mask ((__v4sf) __A, 3365 (__v4sf) 3366 _mm_setzero_ps (), 3367 (__mmask8) __U); 3368 } 3369 3370 extern __inline __m256d 3371 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3372 _mm256_rsqrt14_pd (__m256d __A) 3373 { 3374 return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, 3375 (__v4df) 3376 _mm256_setzero_pd (), 3377 (__mmask8) -1); 3378 } 3379 3380 extern __inline __m256d 3381 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3382 _mm256_mask_rsqrt14_pd (__m256d __W, __mmask8 __U, __m256d __A) 3383 { 3384 return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, 3385 (__v4df) __W, 3386 (__mmask8) __U); 3387 } 3388 3389 extern __inline __m256d 3390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3391 _mm256_maskz_rsqrt14_pd (__mmask8 __U, __m256d __A) 3392 { 3393 return (__m256d) __builtin_ia32_rsqrt14pd256_mask ((__v4df) __A, 3394 (__v4df) 3395 _mm256_setzero_pd (), 3396 (__mmask8) __U); 3397 } 3398 3399 extern __inline __m128d 3400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3401 _mm_rsqrt14_pd (__m128d __A) 3402 { 3403 return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, 3404 (__v2df) 3405 _mm_setzero_pd (), 3406 (__mmask8) -1); 3407 } 3408 3409 extern __inline __m128d 3410 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3411 _mm_mask_rsqrt14_pd (__m128d __W, __mmask8 __U, __m128d __A) 3412 { 3413 return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, 3414 (__v2df) __W, 3415 (__mmask8) __U); 3416 } 3417 3418 extern __inline __m128d 3419 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3420 _mm_maskz_rsqrt14_pd (__mmask8 __U, __m128d __A) 3421 { 3422 return (__m128d) __builtin_ia32_rsqrt14pd128_mask ((__v2df) __A, 3423 (__v2df) 3424 _mm_setzero_pd (), 3425 (__mmask8) __U); 3426 } 3427 3428 extern __inline __m256 3429 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3430 _mm256_rsqrt14_ps (__m256 __A) 3431 { 3432 return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, 3433 (__v8sf) 3434 _mm256_setzero_ps (), 3435 (__mmask8) -1); 3436 } 3437 3438 extern __inline __m256 3439 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3440 _mm256_mask_rsqrt14_ps (__m256 __W, __mmask8 __U, __m256 __A) 3441 { 3442 return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, 3443 (__v8sf) __W, 3444 (__mmask8) __U); 3445 } 3446 3447 extern __inline __m256 3448 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3449 _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A) 3450 { 3451 return (__m256) __builtin_ia32_rsqrt14ps256_mask ((__v8sf) __A, 3452 (__v8sf) 3453 _mm256_setzero_ps (), 3454 (__mmask8) __U); 3455 } 3456 3457 extern __inline __m128 3458 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3459 _mm_rsqrt14_ps (__m128 __A) 3460 { 3461 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, 3462 (__v4sf) 3463 _mm_setzero_ps (), 3464 (__mmask8) -1); 3465 } 3466 3467 extern __inline __m128 3468 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3469 _mm_mask_rsqrt14_ps (__m128 __W, __mmask8 __U, __m128 __A) 3470 { 3471 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, 3472 (__v4sf) __W, 3473 (__mmask8) __U); 3474 } 3475 3476 extern __inline __m128 3477 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3478 _mm_maskz_rsqrt14_ps (__mmask8 __U, __m128 __A) 3479 { 3480 return (__m128) __builtin_ia32_rsqrt14ps128_mask ((__v4sf) __A, 3481 (__v4sf) 3482 _mm_setzero_ps (), 3483 (__mmask8) __U); 3484 } 3485 3486 extern __inline __m256d 3487 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3488 _mm256_mask_sqrt_pd (__m256d __W, __mmask8 __U, __m256d __A) 3489 { 3490 return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, 3491 (__v4df) __W, 3492 (__mmask8) __U); 3493 } 3494 3495 extern __inline __m256d 3496 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3497 _mm256_maskz_sqrt_pd (__mmask8 __U, __m256d __A) 3498 { 3499 return (__m256d) __builtin_ia32_sqrtpd256_mask ((__v4df) __A, 3500 (__v4df) 3501 _mm256_setzero_pd (), 3502 (__mmask8) __U); 3503 } 3504 3505 extern __inline __m128d 3506 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3507 _mm_mask_sqrt_pd (__m128d __W, __mmask8 __U, __m128d __A) 3508 { 3509 return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, 3510 (__v2df) __W, 3511 (__mmask8) __U); 3512 } 3513 3514 extern __inline __m128d 3515 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3516 _mm_maskz_sqrt_pd (__mmask8 __U, __m128d __A) 3517 { 3518 return (__m128d) __builtin_ia32_sqrtpd128_mask ((__v2df) __A, 3519 (__v2df) 3520 _mm_setzero_pd (), 3521 (__mmask8) __U); 3522 } 3523 3524 extern __inline __m256 3525 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3526 _mm256_mask_sqrt_ps (__m256 __W, __mmask8 __U, __m256 __A) 3527 { 3528 return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, 3529 (__v8sf) __W, 3530 (__mmask8) __U); 3531 } 3532 3533 extern __inline __m256 3534 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3535 _mm256_maskz_sqrt_ps (__mmask8 __U, __m256 __A) 3536 { 3537 return (__m256) __builtin_ia32_sqrtps256_mask ((__v8sf) __A, 3538 (__v8sf) 3539 _mm256_setzero_ps (), 3540 (__mmask8) __U); 3541 } 3542 3543 extern __inline __m128 3544 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3545 _mm_mask_sqrt_ps (__m128 __W, __mmask8 __U, __m128 __A) 3546 { 3547 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, 3548 (__v4sf) __W, 3549 (__mmask8) __U); 3550 } 3551 3552 extern __inline __m128 3553 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3554 _mm_maskz_sqrt_ps (__mmask8 __U, __m128 __A) 3555 { 3556 return (__m128) __builtin_ia32_sqrtps128_mask ((__v4sf) __A, 3557 (__v4sf) 3558 _mm_setzero_ps (), 3559 (__mmask8) __U); 3560 } 3561 3562 extern __inline __m256i 3563 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3564 _mm256_mask_add_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 3565 __m256i __B) 3566 { 3567 return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A, 3568 (__v8si) __B, 3569 (__v8si) __W, 3570 (__mmask8) __U); 3571 } 3572 3573 extern __inline __m256i 3574 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3575 _mm256_maskz_add_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 3576 { 3577 return (__m256i) __builtin_ia32_paddd256_mask ((__v8si) __A, 3578 (__v8si) __B, 3579 (__v8si) 3580 _mm256_setzero_si256 (), 3581 (__mmask8) __U); 3582 } 3583 3584 extern __inline __m256i 3585 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3586 _mm256_mask_add_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 3587 __m256i __B) 3588 { 3589 return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A, 3590 (__v4di) __B, 3591 (__v4di) __W, 3592 (__mmask8) __U); 3593 } 3594 3595 extern __inline __m256i 3596 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3597 _mm256_maskz_add_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 3598 { 3599 return (__m256i) __builtin_ia32_paddq256_mask ((__v4di) __A, 3600 (__v4di) __B, 3601 (__v4di) 3602 _mm256_setzero_si256 (), 3603 (__mmask8) __U); 3604 } 3605 3606 extern __inline __m256i 3607 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3608 _mm256_mask_sub_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 3609 __m256i __B) 3610 { 3611 return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A, 3612 (__v8si) __B, 3613 (__v8si) __W, 3614 (__mmask8) __U); 3615 } 3616 3617 extern __inline __m256i 3618 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3619 _mm256_maskz_sub_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 3620 { 3621 return (__m256i) __builtin_ia32_psubd256_mask ((__v8si) __A, 3622 (__v8si) __B, 3623 (__v8si) 3624 _mm256_setzero_si256 (), 3625 (__mmask8) __U); 3626 } 3627 3628 extern __inline __m256i 3629 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3630 _mm256_mask_sub_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 3631 __m256i __B) 3632 { 3633 return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A, 3634 (__v4di) __B, 3635 (__v4di) __W, 3636 (__mmask8) __U); 3637 } 3638 3639 extern __inline __m256i 3640 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3641 _mm256_maskz_sub_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 3642 { 3643 return (__m256i) __builtin_ia32_psubq256_mask ((__v4di) __A, 3644 (__v4di) __B, 3645 (__v4di) 3646 _mm256_setzero_si256 (), 3647 (__mmask8) __U); 3648 } 3649 3650 extern __inline __m128i 3651 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3652 _mm_mask_add_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 3653 __m128i __B) 3654 { 3655 return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A, 3656 (__v4si) __B, 3657 (__v4si) __W, 3658 (__mmask8) __U); 3659 } 3660 3661 extern __inline __m128i 3662 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3663 _mm_maskz_add_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 3664 { 3665 return (__m128i) __builtin_ia32_paddd128_mask ((__v4si) __A, 3666 (__v4si) __B, 3667 (__v4si) 3668 _mm_setzero_si128 (), 3669 (__mmask8) __U); 3670 } 3671 3672 extern __inline __m128i 3673 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3674 _mm_mask_add_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 3675 __m128i __B) 3676 { 3677 return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A, 3678 (__v2di) __B, 3679 (__v2di) __W, 3680 (__mmask8) __U); 3681 } 3682 3683 extern __inline __m128i 3684 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3685 _mm_maskz_add_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 3686 { 3687 return (__m128i) __builtin_ia32_paddq128_mask ((__v2di) __A, 3688 (__v2di) __B, 3689 (__v2di) 3690 _mm_setzero_si128 (), 3691 (__mmask8) __U); 3692 } 3693 3694 extern __inline __m128i 3695 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3696 _mm_mask_sub_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 3697 __m128i __B) 3698 { 3699 return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A, 3700 (__v4si) __B, 3701 (__v4si) __W, 3702 (__mmask8) __U); 3703 } 3704 3705 extern __inline __m128i 3706 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3707 _mm_maskz_sub_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 3708 { 3709 return (__m128i) __builtin_ia32_psubd128_mask ((__v4si) __A, 3710 (__v4si) __B, 3711 (__v4si) 3712 _mm_setzero_si128 (), 3713 (__mmask8) __U); 3714 } 3715 3716 extern __inline __m128i 3717 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3718 _mm_mask_sub_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 3719 __m128i __B) 3720 { 3721 return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A, 3722 (__v2di) __B, 3723 (__v2di) __W, 3724 (__mmask8) __U); 3725 } 3726 3727 extern __inline __m128i 3728 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3729 _mm_maskz_sub_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 3730 { 3731 return (__m128i) __builtin_ia32_psubq128_mask ((__v2di) __A, 3732 (__v2di) __B, 3733 (__v2di) 3734 _mm_setzero_si128 (), 3735 (__mmask8) __U); 3736 } 3737 3738 extern __inline __m256 3739 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3740 _mm256_getexp_ps (__m256 __A) 3741 { 3742 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, 3743 (__v8sf) 3744 _mm256_setzero_ps (), 3745 (__mmask8) -1); 3746 } 3747 3748 extern __inline __m256 3749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3750 _mm256_mask_getexp_ps (__m256 __W, __mmask8 __U, __m256 __A) 3751 { 3752 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, 3753 (__v8sf) __W, 3754 (__mmask8) __U); 3755 } 3756 3757 extern __inline __m256 3758 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3759 _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) 3760 { 3761 return (__m256) __builtin_ia32_getexpps256_mask ((__v8sf) __A, 3762 (__v8sf) 3763 _mm256_setzero_ps (), 3764 (__mmask8) __U); 3765 } 3766 3767 extern __inline __m256d 3768 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3769 _mm256_getexp_pd (__m256d __A) 3770 { 3771 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, 3772 (__v4df) 3773 _mm256_setzero_pd (), 3774 (__mmask8) -1); 3775 } 3776 3777 extern __inline __m256d 3778 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3779 _mm256_mask_getexp_pd (__m256d __W, __mmask8 __U, __m256d __A) 3780 { 3781 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, 3782 (__v4df) __W, 3783 (__mmask8) __U); 3784 } 3785 3786 extern __inline __m256d 3787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3788 _mm256_maskz_getexp_pd (__mmask8 __U, __m256d __A) 3789 { 3790 return (__m256d) __builtin_ia32_getexppd256_mask ((__v4df) __A, 3791 (__v4df) 3792 _mm256_setzero_pd (), 3793 (__mmask8) __U); 3794 } 3795 3796 extern __inline __m128 3797 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3798 _mm_getexp_ps (__m128 __A) 3799 { 3800 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, 3801 (__v4sf) 3802 _mm_setzero_ps (), 3803 (__mmask8) -1); 3804 } 3805 3806 extern __inline __m128 3807 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3808 _mm_mask_getexp_ps (__m128 __W, __mmask8 __U, __m128 __A) 3809 { 3810 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, 3811 (__v4sf) __W, 3812 (__mmask8) __U); 3813 } 3814 3815 extern __inline __m128 3816 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3817 _mm_maskz_getexp_ps (__mmask8 __U, __m128 __A) 3818 { 3819 return (__m128) __builtin_ia32_getexpps128_mask ((__v4sf) __A, 3820 (__v4sf) 3821 _mm_setzero_ps (), 3822 (__mmask8) __U); 3823 } 3824 3825 extern __inline __m128d 3826 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3827 _mm_getexp_pd (__m128d __A) 3828 { 3829 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, 3830 (__v2df) 3831 _mm_setzero_pd (), 3832 (__mmask8) -1); 3833 } 3834 3835 extern __inline __m128d 3836 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3837 _mm_mask_getexp_pd (__m128d __W, __mmask8 __U, __m128d __A) 3838 { 3839 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, 3840 (__v2df) __W, 3841 (__mmask8) __U); 3842 } 3843 3844 extern __inline __m128d 3845 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3846 _mm_maskz_getexp_pd (__mmask8 __U, __m128d __A) 3847 { 3848 return (__m128d) __builtin_ia32_getexppd128_mask ((__v2df) __A, 3849 (__v2df) 3850 _mm_setzero_pd (), 3851 (__mmask8) __U); 3852 } 3853 3854 extern __inline __m256i 3855 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3856 _mm256_mask_srl_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 3857 __m128i __B) 3858 { 3859 return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, 3860 (__v4si) __B, 3861 (__v8si) __W, 3862 (__mmask8) __U); 3863 } 3864 3865 extern __inline __m256i 3866 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3867 _mm256_maskz_srl_epi32 (__mmask8 __U, __m256i __A, __m128i __B) 3868 { 3869 return (__m256i) __builtin_ia32_psrld256_mask ((__v8si) __A, 3870 (__v4si) __B, 3871 (__v8si) 3872 _mm256_setzero_si256 (), 3873 (__mmask8) __U); 3874 } 3875 3876 extern __inline __m128i 3877 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3878 _mm_mask_srl_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 3879 __m128i __B) 3880 { 3881 return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, 3882 (__v4si) __B, 3883 (__v4si) __W, 3884 (__mmask8) __U); 3885 } 3886 3887 extern __inline __m128i 3888 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3889 _mm_maskz_srl_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 3890 { 3891 return (__m128i) __builtin_ia32_psrld128_mask ((__v4si) __A, 3892 (__v4si) __B, 3893 (__v4si) 3894 _mm_setzero_si128 (), 3895 (__mmask8) __U); 3896 } 3897 3898 extern __inline __m256i 3899 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3900 _mm256_mask_srl_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 3901 __m128i __B) 3902 { 3903 return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, 3904 (__v2di) __B, 3905 (__v4di) __W, 3906 (__mmask8) __U); 3907 } 3908 3909 extern __inline __m256i 3910 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3911 _mm256_maskz_srl_epi64 (__mmask8 __U, __m256i __A, __m128i __B) 3912 { 3913 return (__m256i) __builtin_ia32_psrlq256_mask ((__v4di) __A, 3914 (__v2di) __B, 3915 (__v4di) 3916 _mm256_setzero_si256 (), 3917 (__mmask8) __U); 3918 } 3919 3920 extern __inline __m128i 3921 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3922 _mm_mask_srl_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 3923 __m128i __B) 3924 { 3925 return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, 3926 (__v2di) __B, 3927 (__v2di) __W, 3928 (__mmask8) __U); 3929 } 3930 3931 extern __inline __m128i 3932 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3933 _mm_maskz_srl_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 3934 { 3935 return (__m128i) __builtin_ia32_psrlq128_mask ((__v2di) __A, 3936 (__v2di) __B, 3937 (__v2di) 3938 _mm_setzero_si128 (), 3939 (__mmask8) __U); 3940 } 3941 3942 extern __inline __m256i 3943 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3944 _mm256_mask_and_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 3945 __m256i __B) 3946 { 3947 return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A, 3948 (__v8si) __B, 3949 (__v8si) __W, 3950 (__mmask8) __U); 3951 } 3952 3953 extern __inline __m256i 3954 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3955 _mm256_maskz_and_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 3956 { 3957 return (__m256i) __builtin_ia32_pandd256_mask ((__v8si) __A, 3958 (__v8si) __B, 3959 (__v8si) 3960 _mm256_setzero_si256 (), 3961 (__mmask8) __U); 3962 } 3963 3964 extern __inline __m256d 3965 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3966 _mm256_scalef_pd (__m256d __A, __m256d __B) 3967 { 3968 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, 3969 (__v4df) __B, 3970 (__v4df) 3971 _mm256_setzero_pd (), 3972 (__mmask8) -1); 3973 } 3974 3975 extern __inline __m256d 3976 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3977 _mm256_mask_scalef_pd (__m256d __W, __mmask8 __U, __m256d __A, 3978 __m256d __B) 3979 { 3980 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, 3981 (__v4df) __B, 3982 (__v4df) __W, 3983 (__mmask8) __U); 3984 } 3985 3986 extern __inline __m256d 3987 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3988 _mm256_maskz_scalef_pd (__mmask8 __U, __m256d __A, __m256d __B) 3989 { 3990 return (__m256d) __builtin_ia32_scalefpd256_mask ((__v4df) __A, 3991 (__v4df) __B, 3992 (__v4df) 3993 _mm256_setzero_pd (), 3994 (__mmask8) __U); 3995 } 3996 3997 extern __inline __m256 3998 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3999 _mm256_scalef_ps (__m256 __A, __m256 __B) 4000 { 4001 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, 4002 (__v8sf) __B, 4003 (__v8sf) 4004 _mm256_setzero_ps (), 4005 (__mmask8) -1); 4006 } 4007 4008 extern __inline __m256 4009 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4010 _mm256_mask_scalef_ps (__m256 __W, __mmask8 __U, __m256 __A, 4011 __m256 __B) 4012 { 4013 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, 4014 (__v8sf) __B, 4015 (__v8sf) __W, 4016 (__mmask8) __U); 4017 } 4018 4019 extern __inline __m256 4020 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4021 _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) 4022 { 4023 return (__m256) __builtin_ia32_scalefps256_mask ((__v8sf) __A, 4024 (__v8sf) __B, 4025 (__v8sf) 4026 _mm256_setzero_ps (), 4027 (__mmask8) __U); 4028 } 4029 4030 extern __inline __m128d 4031 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4032 _mm_scalef_pd (__m128d __A, __m128d __B) 4033 { 4034 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, 4035 (__v2df) __B, 4036 (__v2df) 4037 _mm_setzero_pd (), 4038 (__mmask8) -1); 4039 } 4040 4041 extern __inline __m128d 4042 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4043 _mm_mask_scalef_pd (__m128d __W, __mmask8 __U, __m128d __A, 4044 __m128d __B) 4045 { 4046 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, 4047 (__v2df) __B, 4048 (__v2df) __W, 4049 (__mmask8) __U); 4050 } 4051 4052 extern __inline __m128d 4053 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4054 _mm_maskz_scalef_pd (__mmask8 __U, __m128d __A, __m128d __B) 4055 { 4056 return (__m128d) __builtin_ia32_scalefpd128_mask ((__v2df) __A, 4057 (__v2df) __B, 4058 (__v2df) 4059 _mm_setzero_pd (), 4060 (__mmask8) __U); 4061 } 4062 4063 extern __inline __m128 4064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4065 _mm_scalef_ps (__m128 __A, __m128 __B) 4066 { 4067 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, 4068 (__v4sf) __B, 4069 (__v4sf) 4070 _mm_setzero_ps (), 4071 (__mmask8) -1); 4072 } 4073 4074 extern __inline __m128 4075 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4076 _mm_mask_scalef_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 4077 { 4078 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, 4079 (__v4sf) __B, 4080 (__v4sf) __W, 4081 (__mmask8) __U); 4082 } 4083 4084 extern __inline __m128 4085 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4086 _mm_maskz_scalef_ps (__mmask8 __U, __m128 __A, __m128 __B) 4087 { 4088 return (__m128) __builtin_ia32_scalefps128_mask ((__v4sf) __A, 4089 (__v4sf) __B, 4090 (__v4sf) 4091 _mm_setzero_ps (), 4092 (__mmask8) __U); 4093 } 4094 4095 extern __inline __m256d 4096 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4097 _mm256_mask_fmadd_pd (__m256d __A, __mmask8 __U, __m256d __B, 4098 __m256d __C) 4099 { 4100 return (__m256d) __builtin_ia32_vfmaddpd256_mask ((__v4df) __A, 4101 (__v4df) __B, 4102 (__v4df) __C, 4103 (__mmask8) __U); 4104 } 4105 4106 extern __inline __m256d 4107 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4108 _mm256_mask3_fmadd_pd (__m256d __A, __m256d __B, __m256d __C, 4109 __mmask8 __U) 4110 { 4111 return (__m256d) __builtin_ia32_vfmaddpd256_mask3 ((__v4df) __A, 4112 (__v4df) __B, 4113 (__v4df) __C, 4114 (__mmask8) __U); 4115 } 4116 4117 extern __inline __m256d 4118 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4119 _mm256_maskz_fmadd_pd (__mmask8 __U, __m256d __A, __m256d __B, 4120 __m256d __C) 4121 { 4122 return (__m256d) __builtin_ia32_vfmaddpd256_maskz ((__v4df) __A, 4123 (__v4df) __B, 4124 (__v4df) __C, 4125 (__mmask8) __U); 4126 } 4127 4128 extern __inline __m128d 4129 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4130 _mm_mask_fmadd_pd (__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) 4131 { 4132 return (__m128d) __builtin_ia32_vfmaddpd128_mask ((__v2df) __A, 4133 (__v2df) __B, 4134 (__v2df) __C, 4135 (__mmask8) __U); 4136 } 4137 4138 extern __inline __m128d 4139 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4140 _mm_mask3_fmadd_pd (__m128d __A, __m128d __B, __m128d __C, 4141 __mmask8 __U) 4142 { 4143 return (__m128d) __builtin_ia32_vfmaddpd128_mask3 ((__v2df) __A, 4144 (__v2df) __B, 4145 (__v2df) __C, 4146 (__mmask8) __U); 4147 } 4148 4149 extern __inline __m128d 4150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4151 _mm_maskz_fmadd_pd (__mmask8 __U, __m128d __A, __m128d __B, 4152 __m128d __C) 4153 { 4154 return (__m128d) __builtin_ia32_vfmaddpd128_maskz ((__v2df) __A, 4155 (__v2df) __B, 4156 (__v2df) __C, 4157 (__mmask8) __U); 4158 } 4159 4160 extern __inline __m256 4161 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4162 _mm256_mask_fmadd_ps (__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) 4163 { 4164 return (__m256) __builtin_ia32_vfmaddps256_mask ((__v8sf) __A, 4165 (__v8sf) __B, 4166 (__v8sf) __C, 4167 (__mmask8) __U); 4168 } 4169 4170 extern __inline __m256 4171 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4172 _mm256_mask3_fmadd_ps (__m256 __A, __m256 __B, __m256 __C, 4173 __mmask8 __U) 4174 { 4175 return (__m256) __builtin_ia32_vfmaddps256_mask3 ((__v8sf) __A, 4176 (__v8sf) __B, 4177 (__v8sf) __C, 4178 (__mmask8) __U); 4179 } 4180 4181 extern __inline __m256 4182 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4183 _mm256_maskz_fmadd_ps (__mmask8 __U, __m256 __A, __m256 __B, 4184 __m256 __C) 4185 { 4186 return (__m256) __builtin_ia32_vfmaddps256_maskz ((__v8sf) __A, 4187 (__v8sf) __B, 4188 (__v8sf) __C, 4189 (__mmask8) __U); 4190 } 4191 4192 extern __inline __m128 4193 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4194 _mm_mask_fmadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4195 { 4196 return (__m128) __builtin_ia32_vfmaddps128_mask ((__v4sf) __A, 4197 (__v4sf) __B, 4198 (__v4sf) __C, 4199 (__mmask8) __U); 4200 } 4201 4202 extern __inline __m128 4203 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4204 _mm_mask3_fmadd_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) 4205 { 4206 return (__m128) __builtin_ia32_vfmaddps128_mask3 ((__v4sf) __A, 4207 (__v4sf) __B, 4208 (__v4sf) __C, 4209 (__mmask8) __U); 4210 } 4211 4212 extern __inline __m128 4213 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4214 _mm_maskz_fmadd_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) 4215 { 4216 return (__m128) __builtin_ia32_vfmaddps128_maskz ((__v4sf) __A, 4217 (__v4sf) __B, 4218 (__v4sf) __C, 4219 (__mmask8) __U); 4220 } 4221 4222 extern __inline __m256d 4223 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4224 _mm256_mask_fmsub_pd (__m256d __A, __mmask8 __U, __m256d __B, 4225 __m256d __C) 4226 { 4227 return (__m256d) __builtin_ia32_vfmsubpd256_mask ((__v4df) __A, 4228 (__v4df) __B, 4229 (__v4df) __C, 4230 (__mmask8) __U); 4231 } 4232 4233 extern __inline __m256d 4234 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4235 _mm256_mask3_fmsub_pd (__m256d __A, __m256d __B, __m256d __C, 4236 __mmask8 __U) 4237 { 4238 return (__m256d) __builtin_ia32_vfmsubpd256_mask3 ((__v4df) __A, 4239 (__v4df) __B, 4240 (__v4df) __C, 4241 (__mmask8) __U); 4242 } 4243 4244 extern __inline __m256d 4245 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4246 _mm256_maskz_fmsub_pd (__mmask8 __U, __m256d __A, __m256d __B, 4247 __m256d __C) 4248 { 4249 return (__m256d) __builtin_ia32_vfmsubpd256_maskz ((__v4df) __A, 4250 (__v4df) __B, 4251 (__v4df) __C, 4252 (__mmask8) __U); 4253 } 4254 4255 extern __inline __m128d 4256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4257 _mm_mask_fmsub_pd (__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) 4258 { 4259 return (__m128d) __builtin_ia32_vfmsubpd128_mask ((__v2df) __A, 4260 (__v2df) __B, 4261 (__v2df) __C, 4262 (__mmask8) __U); 4263 } 4264 4265 extern __inline __m128d 4266 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4267 _mm_mask3_fmsub_pd (__m128d __A, __m128d __B, __m128d __C, 4268 __mmask8 __U) 4269 { 4270 return (__m128d) __builtin_ia32_vfmsubpd128_mask3 ((__v2df) __A, 4271 (__v2df) __B, 4272 (__v2df) __C, 4273 (__mmask8) __U); 4274 } 4275 4276 extern __inline __m128d 4277 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4278 _mm_maskz_fmsub_pd (__mmask8 __U, __m128d __A, __m128d __B, 4279 __m128d __C) 4280 { 4281 return (__m128d) __builtin_ia32_vfmsubpd128_maskz ((__v2df) __A, 4282 (__v2df) __B, 4283 (__v2df) __C, 4284 (__mmask8) __U); 4285 } 4286 4287 extern __inline __m256 4288 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4289 _mm256_mask_fmsub_ps (__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) 4290 { 4291 return (__m256) __builtin_ia32_vfmsubps256_mask ((__v8sf) __A, 4292 (__v8sf) __B, 4293 (__v8sf) __C, 4294 (__mmask8) __U); 4295 } 4296 4297 extern __inline __m256 4298 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4299 _mm256_mask3_fmsub_ps (__m256 __A, __m256 __B, __m256 __C, 4300 __mmask8 __U) 4301 { 4302 return (__m256) __builtin_ia32_vfmsubps256_mask3 ((__v8sf) __A, 4303 (__v8sf) __B, 4304 (__v8sf) __C, 4305 (__mmask8) __U); 4306 } 4307 4308 extern __inline __m256 4309 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4310 _mm256_maskz_fmsub_ps (__mmask8 __U, __m256 __A, __m256 __B, 4311 __m256 __C) 4312 { 4313 return (__m256) __builtin_ia32_vfmsubps256_maskz ((__v8sf) __A, 4314 (__v8sf) __B, 4315 (__v8sf) __C, 4316 (__mmask8) __U); 4317 } 4318 4319 extern __inline __m128 4320 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4321 _mm_mask_fmsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4322 { 4323 return (__m128) __builtin_ia32_vfmsubps128_mask ((__v4sf) __A, 4324 (__v4sf) __B, 4325 (__v4sf) __C, 4326 (__mmask8) __U); 4327 } 4328 4329 extern __inline __m128 4330 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4331 _mm_mask3_fmsub_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) 4332 { 4333 return (__m128) __builtin_ia32_vfmsubps128_mask3 ((__v4sf) __A, 4334 (__v4sf) __B, 4335 (__v4sf) __C, 4336 (__mmask8) __U); 4337 } 4338 4339 extern __inline __m128 4340 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4341 _mm_maskz_fmsub_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) 4342 { 4343 return (__m128) __builtin_ia32_vfmsubps128_maskz ((__v4sf) __A, 4344 (__v4sf) __B, 4345 (__v4sf) __C, 4346 (__mmask8) __U); 4347 } 4348 4349 extern __inline __m256d 4350 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4351 _mm256_mask_fmaddsub_pd (__m256d __A, __mmask8 __U, __m256d __B, 4352 __m256d __C) 4353 { 4354 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, 4355 (__v4df) __B, 4356 (__v4df) __C, 4357 (__mmask8) __U); 4358 } 4359 4360 extern __inline __m256d 4361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4362 _mm256_mask3_fmaddsub_pd (__m256d __A, __m256d __B, __m256d __C, 4363 __mmask8 __U) 4364 { 4365 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask3 ((__v4df) __A, 4366 (__v4df) __B, 4367 (__v4df) __C, 4368 (__mmask8) 4369 __U); 4370 } 4371 4372 extern __inline __m256d 4373 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4374 _mm256_maskz_fmaddsub_pd (__mmask8 __U, __m256d __A, __m256d __B, 4375 __m256d __C) 4376 { 4377 return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, 4378 (__v4df) __B, 4379 (__v4df) __C, 4380 (__mmask8) 4381 __U); 4382 } 4383 4384 extern __inline __m128d 4385 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4386 _mm_mask_fmaddsub_pd (__m128d __A, __mmask8 __U, __m128d __B, 4387 __m128d __C) 4388 { 4389 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, 4390 (__v2df) __B, 4391 (__v2df) __C, 4392 (__mmask8) __U); 4393 } 4394 4395 extern __inline __m128d 4396 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4397 _mm_mask3_fmaddsub_pd (__m128d __A, __m128d __B, __m128d __C, 4398 __mmask8 __U) 4399 { 4400 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask3 ((__v2df) __A, 4401 (__v2df) __B, 4402 (__v2df) __C, 4403 (__mmask8) 4404 __U); 4405 } 4406 4407 extern __inline __m128d 4408 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4409 _mm_maskz_fmaddsub_pd (__mmask8 __U, __m128d __A, __m128d __B, 4410 __m128d __C) 4411 { 4412 return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, 4413 (__v2df) __B, 4414 (__v2df) __C, 4415 (__mmask8) 4416 __U); 4417 } 4418 4419 extern __inline __m256 4420 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4421 _mm256_mask_fmaddsub_ps (__m256 __A, __mmask8 __U, __m256 __B, 4422 __m256 __C) 4423 { 4424 return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, 4425 (__v8sf) __B, 4426 (__v8sf) __C, 4427 (__mmask8) __U); 4428 } 4429 4430 extern __inline __m256 4431 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4432 _mm256_mask3_fmaddsub_ps (__m256 __A, __m256 __B, __m256 __C, 4433 __mmask8 __U) 4434 { 4435 return (__m256) __builtin_ia32_vfmaddsubps256_mask3 ((__v8sf) __A, 4436 (__v8sf) __B, 4437 (__v8sf) __C, 4438 (__mmask8) __U); 4439 } 4440 4441 extern __inline __m256 4442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4443 _mm256_maskz_fmaddsub_ps (__mmask8 __U, __m256 __A, __m256 __B, 4444 __m256 __C) 4445 { 4446 return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, 4447 (__v8sf) __B, 4448 (__v8sf) __C, 4449 (__mmask8) __U); 4450 } 4451 4452 extern __inline __m128 4453 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4454 _mm_mask_fmaddsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4455 { 4456 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, 4457 (__v4sf) __B, 4458 (__v4sf) __C, 4459 (__mmask8) __U); 4460 } 4461 4462 extern __inline __m128 4463 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4464 _mm_mask3_fmaddsub_ps (__m128 __A, __m128 __B, __m128 __C, 4465 __mmask8 __U) 4466 { 4467 return (__m128) __builtin_ia32_vfmaddsubps128_mask3 ((__v4sf) __A, 4468 (__v4sf) __B, 4469 (__v4sf) __C, 4470 (__mmask8) __U); 4471 } 4472 4473 extern __inline __m128 4474 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4475 _mm_maskz_fmaddsub_ps (__mmask8 __U, __m128 __A, __m128 __B, 4476 __m128 __C) 4477 { 4478 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, 4479 (__v4sf) __B, 4480 (__v4sf) __C, 4481 (__mmask8) __U); 4482 } 4483 4484 extern __inline __m256d 4485 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4486 _mm256_mask_fmsubadd_pd (__m256d __A, __mmask8 __U, __m256d __B, 4487 __m256d __C) 4488 { 4489 return (__m256d) __builtin_ia32_vfmaddsubpd256_mask ((__v4df) __A, 4490 (__v4df) __B, 4491 -(__v4df) __C, 4492 (__mmask8) __U); 4493 } 4494 4495 extern __inline __m256d 4496 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4497 _mm256_mask3_fmsubadd_pd (__m256d __A, __m256d __B, __m256d __C, 4498 __mmask8 __U) 4499 { 4500 return (__m256d) __builtin_ia32_vfmsubaddpd256_mask3 ((__v4df) __A, 4501 (__v4df) __B, 4502 (__v4df) __C, 4503 (__mmask8) 4504 __U); 4505 } 4506 4507 extern __inline __m256d 4508 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4509 _mm256_maskz_fmsubadd_pd (__mmask8 __U, __m256d __A, __m256d __B, 4510 __m256d __C) 4511 { 4512 return (__m256d) __builtin_ia32_vfmaddsubpd256_maskz ((__v4df) __A, 4513 (__v4df) __B, 4514 -(__v4df) __C, 4515 (__mmask8) 4516 __U); 4517 } 4518 4519 extern __inline __m128d 4520 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4521 _mm_mask_fmsubadd_pd (__m128d __A, __mmask8 __U, __m128d __B, 4522 __m128d __C) 4523 { 4524 return (__m128d) __builtin_ia32_vfmaddsubpd128_mask ((__v2df) __A, 4525 (__v2df) __B, 4526 -(__v2df) __C, 4527 (__mmask8) __U); 4528 } 4529 4530 extern __inline __m128d 4531 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4532 _mm_mask3_fmsubadd_pd (__m128d __A, __m128d __B, __m128d __C, 4533 __mmask8 __U) 4534 { 4535 return (__m128d) __builtin_ia32_vfmsubaddpd128_mask3 ((__v2df) __A, 4536 (__v2df) __B, 4537 (__v2df) __C, 4538 (__mmask8) 4539 __U); 4540 } 4541 4542 extern __inline __m128d 4543 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4544 _mm_maskz_fmsubadd_pd (__mmask8 __U, __m128d __A, __m128d __B, 4545 __m128d __C) 4546 { 4547 return (__m128d) __builtin_ia32_vfmaddsubpd128_maskz ((__v2df) __A, 4548 (__v2df) __B, 4549 -(__v2df) __C, 4550 (__mmask8) 4551 __U); 4552 } 4553 4554 extern __inline __m256 4555 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4556 _mm256_mask_fmsubadd_ps (__m256 __A, __mmask8 __U, __m256 __B, 4557 __m256 __C) 4558 { 4559 return (__m256) __builtin_ia32_vfmaddsubps256_mask ((__v8sf) __A, 4560 (__v8sf) __B, 4561 -(__v8sf) __C, 4562 (__mmask8) __U); 4563 } 4564 4565 extern __inline __m256 4566 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4567 _mm256_mask3_fmsubadd_ps (__m256 __A, __m256 __B, __m256 __C, 4568 __mmask8 __U) 4569 { 4570 return (__m256) __builtin_ia32_vfmsubaddps256_mask3 ((__v8sf) __A, 4571 (__v8sf) __B, 4572 (__v8sf) __C, 4573 (__mmask8) __U); 4574 } 4575 4576 extern __inline __m256 4577 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4578 _mm256_maskz_fmsubadd_ps (__mmask8 __U, __m256 __A, __m256 __B, 4579 __m256 __C) 4580 { 4581 return (__m256) __builtin_ia32_vfmaddsubps256_maskz ((__v8sf) __A, 4582 (__v8sf) __B, 4583 -(__v8sf) __C, 4584 (__mmask8) __U); 4585 } 4586 4587 extern __inline __m128 4588 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4589 _mm_mask_fmsubadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4590 { 4591 return (__m128) __builtin_ia32_vfmaddsubps128_mask ((__v4sf) __A, 4592 (__v4sf) __B, 4593 -(__v4sf) __C, 4594 (__mmask8) __U); 4595 } 4596 4597 extern __inline __m128 4598 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4599 _mm_mask3_fmsubadd_ps (__m128 __A, __m128 __B, __m128 __C, 4600 __mmask8 __U) 4601 { 4602 return (__m128) __builtin_ia32_vfmsubaddps128_mask3 ((__v4sf) __A, 4603 (__v4sf) __B, 4604 (__v4sf) __C, 4605 (__mmask8) __U); 4606 } 4607 4608 extern __inline __m128 4609 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4610 _mm_maskz_fmsubadd_ps (__mmask8 __U, __m128 __A, __m128 __B, 4611 __m128 __C) 4612 { 4613 return (__m128) __builtin_ia32_vfmaddsubps128_maskz ((__v4sf) __A, 4614 (__v4sf) __B, 4615 -(__v4sf) __C, 4616 (__mmask8) __U); 4617 } 4618 4619 extern __inline __m256d 4620 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4621 _mm256_mask_fnmadd_pd (__m256d __A, __mmask8 __U, __m256d __B, 4622 __m256d __C) 4623 { 4624 return (__m256d) __builtin_ia32_vfnmaddpd256_mask ((__v4df) __A, 4625 (__v4df) __B, 4626 (__v4df) __C, 4627 (__mmask8) __U); 4628 } 4629 4630 extern __inline __m256d 4631 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4632 _mm256_mask3_fnmadd_pd (__m256d __A, __m256d __B, __m256d __C, 4633 __mmask8 __U) 4634 { 4635 return (__m256d) __builtin_ia32_vfnmaddpd256_mask3 ((__v4df) __A, 4636 (__v4df) __B, 4637 (__v4df) __C, 4638 (__mmask8) __U); 4639 } 4640 4641 extern __inline __m256d 4642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4643 _mm256_maskz_fnmadd_pd (__mmask8 __U, __m256d __A, __m256d __B, 4644 __m256d __C) 4645 { 4646 return (__m256d) __builtin_ia32_vfnmaddpd256_maskz ((__v4df) __A, 4647 (__v4df) __B, 4648 (__v4df) __C, 4649 (__mmask8) __U); 4650 } 4651 4652 extern __inline __m128d 4653 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4654 _mm_mask_fnmadd_pd (__m128d __A, __mmask8 __U, __m128d __B, 4655 __m128d __C) 4656 { 4657 return (__m128d) __builtin_ia32_vfnmaddpd128_mask ((__v2df) __A, 4658 (__v2df) __B, 4659 (__v2df) __C, 4660 (__mmask8) __U); 4661 } 4662 4663 extern __inline __m128d 4664 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4665 _mm_mask3_fnmadd_pd (__m128d __A, __m128d __B, __m128d __C, 4666 __mmask8 __U) 4667 { 4668 return (__m128d) __builtin_ia32_vfnmaddpd128_mask3 ((__v2df) __A, 4669 (__v2df) __B, 4670 (__v2df) __C, 4671 (__mmask8) __U); 4672 } 4673 4674 extern __inline __m128d 4675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4676 _mm_maskz_fnmadd_pd (__mmask8 __U, __m128d __A, __m128d __B, 4677 __m128d __C) 4678 { 4679 return (__m128d) __builtin_ia32_vfnmaddpd128_maskz ((__v2df) __A, 4680 (__v2df) __B, 4681 (__v2df) __C, 4682 (__mmask8) __U); 4683 } 4684 4685 extern __inline __m256 4686 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4687 _mm256_mask_fnmadd_ps (__m256 __A, __mmask8 __U, __m256 __B, 4688 __m256 __C) 4689 { 4690 return (__m256) __builtin_ia32_vfnmaddps256_mask ((__v8sf) __A, 4691 (__v8sf) __B, 4692 (__v8sf) __C, 4693 (__mmask8) __U); 4694 } 4695 4696 extern __inline __m256 4697 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4698 _mm256_mask3_fnmadd_ps (__m256 __A, __m256 __B, __m256 __C, 4699 __mmask8 __U) 4700 { 4701 return (__m256) __builtin_ia32_vfnmaddps256_mask3 ((__v8sf) __A, 4702 (__v8sf) __B, 4703 (__v8sf) __C, 4704 (__mmask8) __U); 4705 } 4706 4707 extern __inline __m256 4708 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4709 _mm256_maskz_fnmadd_ps (__mmask8 __U, __m256 __A, __m256 __B, 4710 __m256 __C) 4711 { 4712 return (__m256) __builtin_ia32_vfnmaddps256_maskz ((__v8sf) __A, 4713 (__v8sf) __B, 4714 (__v8sf) __C, 4715 (__mmask8) __U); 4716 } 4717 4718 extern __inline __m128 4719 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4720 _mm_mask_fnmadd_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4721 { 4722 return (__m128) __builtin_ia32_vfnmaddps128_mask ((__v4sf) __A, 4723 (__v4sf) __B, 4724 (__v4sf) __C, 4725 (__mmask8) __U); 4726 } 4727 4728 extern __inline __m128 4729 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4730 _mm_mask3_fnmadd_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) 4731 { 4732 return (__m128) __builtin_ia32_vfnmaddps128_mask3 ((__v4sf) __A, 4733 (__v4sf) __B, 4734 (__v4sf) __C, 4735 (__mmask8) __U); 4736 } 4737 4738 extern __inline __m128 4739 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4740 _mm_maskz_fnmadd_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) 4741 { 4742 return (__m128) __builtin_ia32_vfnmaddps128_maskz ((__v4sf) __A, 4743 (__v4sf) __B, 4744 (__v4sf) __C, 4745 (__mmask8) __U); 4746 } 4747 4748 extern __inline __m256d 4749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4750 _mm256_mask_fnmsub_pd (__m256d __A, __mmask8 __U, __m256d __B, 4751 __m256d __C) 4752 { 4753 return (__m256d) __builtin_ia32_vfnmsubpd256_mask ((__v4df) __A, 4754 (__v4df) __B, 4755 (__v4df) __C, 4756 (__mmask8) __U); 4757 } 4758 4759 extern __inline __m256d 4760 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4761 _mm256_mask3_fnmsub_pd (__m256d __A, __m256d __B, __m256d __C, 4762 __mmask8 __U) 4763 { 4764 return (__m256d) __builtin_ia32_vfnmsubpd256_mask3 ((__v4df) __A, 4765 (__v4df) __B, 4766 (__v4df) __C, 4767 (__mmask8) __U); 4768 } 4769 4770 extern __inline __m256d 4771 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4772 _mm256_maskz_fnmsub_pd (__mmask8 __U, __m256d __A, __m256d __B, 4773 __m256d __C) 4774 { 4775 return (__m256d) __builtin_ia32_vfnmsubpd256_maskz ((__v4df) __A, 4776 (__v4df) __B, 4777 (__v4df) __C, 4778 (__mmask8) __U); 4779 } 4780 4781 extern __inline __m128d 4782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4783 _mm_mask_fnmsub_pd (__m128d __A, __mmask8 __U, __m128d __B, 4784 __m128d __C) 4785 { 4786 return (__m128d) __builtin_ia32_vfnmsubpd128_mask ((__v2df) __A, 4787 (__v2df) __B, 4788 (__v2df) __C, 4789 (__mmask8) __U); 4790 } 4791 4792 extern __inline __m128d 4793 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4794 _mm_mask3_fnmsub_pd (__m128d __A, __m128d __B, __m128d __C, 4795 __mmask8 __U) 4796 { 4797 return (__m128d) __builtin_ia32_vfnmsubpd128_mask3 ((__v2df) __A, 4798 (__v2df) __B, 4799 (__v2df) __C, 4800 (__mmask8) __U); 4801 } 4802 4803 extern __inline __m128d 4804 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4805 _mm_maskz_fnmsub_pd (__mmask8 __U, __m128d __A, __m128d __B, 4806 __m128d __C) 4807 { 4808 return (__m128d) __builtin_ia32_vfnmsubpd128_maskz ((__v2df) __A, 4809 (__v2df) __B, 4810 (__v2df) __C, 4811 (__mmask8) __U); 4812 } 4813 4814 extern __inline __m256 4815 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4816 _mm256_mask_fnmsub_ps (__m256 __A, __mmask8 __U, __m256 __B, 4817 __m256 __C) 4818 { 4819 return (__m256) __builtin_ia32_vfnmsubps256_mask ((__v8sf) __A, 4820 (__v8sf) __B, 4821 (__v8sf) __C, 4822 (__mmask8) __U); 4823 } 4824 4825 extern __inline __m256 4826 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4827 _mm256_mask3_fnmsub_ps (__m256 __A, __m256 __B, __m256 __C, 4828 __mmask8 __U) 4829 { 4830 return (__m256) __builtin_ia32_vfnmsubps256_mask3 ((__v8sf) __A, 4831 (__v8sf) __B, 4832 (__v8sf) __C, 4833 (__mmask8) __U); 4834 } 4835 4836 extern __inline __m256 4837 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4838 _mm256_maskz_fnmsub_ps (__mmask8 __U, __m256 __A, __m256 __B, 4839 __m256 __C) 4840 { 4841 return (__m256) __builtin_ia32_vfnmsubps256_maskz ((__v8sf) __A, 4842 (__v8sf) __B, 4843 (__v8sf) __C, 4844 (__mmask8) __U); 4845 } 4846 4847 extern __inline __m128 4848 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4849 _mm_mask_fnmsub_ps (__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) 4850 { 4851 return (__m128) __builtin_ia32_vfnmsubps128_mask ((__v4sf) __A, 4852 (__v4sf) __B, 4853 (__v4sf) __C, 4854 (__mmask8) __U); 4855 } 4856 4857 extern __inline __m128 4858 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4859 _mm_mask3_fnmsub_ps (__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) 4860 { 4861 return (__m128) __builtin_ia32_vfnmsubps128_mask3 ((__v4sf) __A, 4862 (__v4sf) __B, 4863 (__v4sf) __C, 4864 (__mmask8) __U); 4865 } 4866 4867 extern __inline __m128 4868 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4869 _mm_maskz_fnmsub_ps (__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) 4870 { 4871 return (__m128) __builtin_ia32_vfnmsubps128_maskz ((__v4sf) __A, 4872 (__v4sf) __B, 4873 (__v4sf) __C, 4874 (__mmask8) __U); 4875 } 4876 4877 extern __inline __m128i 4878 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4879 _mm_mask_and_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 4880 __m128i __B) 4881 { 4882 return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A, 4883 (__v4si) __B, 4884 (__v4si) __W, 4885 (__mmask8) __U); 4886 } 4887 4888 extern __inline __m128i 4889 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4890 _mm_maskz_and_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 4891 { 4892 return (__m128i) __builtin_ia32_pandd128_mask ((__v4si) __A, 4893 (__v4si) __B, 4894 (__v4si) 4895 _mm_setzero_si128 (), 4896 (__mmask8) __U); 4897 } 4898 4899 extern __inline __m256i 4900 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4901 _mm256_mask_andnot_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 4902 __m256i __B) 4903 { 4904 return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A, 4905 (__v8si) __B, 4906 (__v8si) __W, 4907 (__mmask8) __U); 4908 } 4909 4910 extern __inline __m256i 4911 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4912 _mm256_maskz_andnot_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 4913 { 4914 return (__m256i) __builtin_ia32_pandnd256_mask ((__v8si) __A, 4915 (__v8si) __B, 4916 (__v8si) 4917 _mm256_setzero_si256 (), 4918 (__mmask8) __U); 4919 } 4920 4921 extern __inline __m128i 4922 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4923 _mm_mask_andnot_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 4924 __m128i __B) 4925 { 4926 return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A, 4927 (__v4si) __B, 4928 (__v4si) __W, 4929 (__mmask8) __U); 4930 } 4931 4932 extern __inline __m128i 4933 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4934 _mm_maskz_andnot_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 4935 { 4936 return (__m128i) __builtin_ia32_pandnd128_mask ((__v4si) __A, 4937 (__v4si) __B, 4938 (__v4si) 4939 _mm_setzero_si128 (), 4940 (__mmask8) __U); 4941 } 4942 4943 extern __inline __m256i 4944 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4945 _mm256_mask_or_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 4946 __m256i __B) 4947 { 4948 return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A, 4949 (__v8si) __B, 4950 (__v8si) __W, 4951 (__mmask8) __U); 4952 } 4953 4954 extern __inline __m256i 4955 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4956 _mm256_maskz_or_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 4957 { 4958 return (__m256i) __builtin_ia32_pord256_mask ((__v8si) __A, 4959 (__v8si) __B, 4960 (__v8si) 4961 _mm256_setzero_si256 (), 4962 (__mmask8) __U); 4963 } 4964 4965 extern __inline __m256i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 4966 _mm256_or_epi32 (__m256i __A, __m256i __B) 4967 { 4968 return (__m256i) ((__v8su)__A | (__v8su)__B); 4969 } 4970 4971 extern __inline __m128i 4972 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4973 _mm_mask_or_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) 4974 { 4975 return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A, 4976 (__v4si) __B, 4977 (__v4si) __W, 4978 (__mmask8) __U); 4979 } 4980 4981 extern __inline __m128i 4982 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 4983 _mm_maskz_or_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 4984 { 4985 return (__m128i) __builtin_ia32_pord128_mask ((__v4si) __A, 4986 (__v4si) __B, 4987 (__v4si) 4988 _mm_setzero_si128 (), 4989 (__mmask8) __U); 4990 } 4991 4992 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 4993 _mm_or_epi32 (__m128i __A, __m128i __B) 4994 { 4995 return (__m128i) ((__v4su)__A | (__v4su)__B); 4996 } 4997 4998 extern __inline __m256i 4999 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5000 _mm256_mask_xor_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 5001 __m256i __B) 5002 { 5003 return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A, 5004 (__v8si) __B, 5005 (__v8si) __W, 5006 (__mmask8) __U); 5007 } 5008 5009 extern __inline __m256i 5010 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5011 _mm256_maskz_xor_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 5012 { 5013 return (__m256i) __builtin_ia32_pxord256_mask ((__v8si) __A, 5014 (__v8si) __B, 5015 (__v8si) 5016 _mm256_setzero_si256 (), 5017 (__mmask8) __U); 5018 } 5019 5020 extern __inline __m256i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 5021 _mm256_xor_epi32 (__m256i __A, __m256i __B) 5022 { 5023 return (__m256i) ((__v8su)__A ^ (__v8su)__B); 5024 } 5025 5026 extern __inline __m128i 5027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5028 _mm_mask_xor_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 5029 __m128i __B) 5030 { 5031 return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A, 5032 (__v4si) __B, 5033 (__v4si) __W, 5034 (__mmask8) __U); 5035 } 5036 5037 extern __inline __m128i 5038 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5039 _mm_maskz_xor_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 5040 { 5041 return (__m128i) __builtin_ia32_pxord128_mask ((__v4si) __A, 5042 (__v4si) __B, 5043 (__v4si) 5044 _mm_setzero_si128 (), 5045 (__mmask8) __U); 5046 } 5047 5048 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 5049 _mm_xor_epi32 (__m128i __A, __m128i __B) 5050 { 5051 return (__m128i) ((__v4su)__A ^ (__v4su)__B); 5052 } 5053 5054 extern __inline __m128 5055 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5056 _mm_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m128d __A) 5057 { 5058 return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, 5059 (__v4sf) __W, 5060 (__mmask8) __U); 5061 } 5062 5063 extern __inline __m128 5064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5065 _mm_maskz_cvtpd_ps (__mmask8 __U, __m128d __A) 5066 { 5067 return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A, 5068 (__v4sf) 5069 _mm_setzero_ps (), 5070 (__mmask8) __U); 5071 } 5072 5073 extern __inline __m128 5074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5075 _mm256_mask_cvtpd_ps (__m128 __W, __mmask8 __U, __m256d __A) 5076 { 5077 return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, 5078 (__v4sf) __W, 5079 (__mmask8) __U); 5080 } 5081 5082 extern __inline __m128 5083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5084 _mm256_maskz_cvtpd_ps (__mmask8 __U, __m256d __A) 5085 { 5086 return (__m128) __builtin_ia32_cvtpd2ps256_mask ((__v4df) __A, 5087 (__v4sf) 5088 _mm_setzero_ps (), 5089 (__mmask8) __U); 5090 } 5091 5092 extern __inline __m256i 5093 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5094 _mm256_mask_cvtps_epi32 (__m256i __W, __mmask8 __U, __m256 __A) 5095 { 5096 return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, 5097 (__v8si) __W, 5098 (__mmask8) __U); 5099 } 5100 5101 extern __inline __m256i 5102 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5103 _mm256_maskz_cvtps_epi32 (__mmask8 __U, __m256 __A) 5104 { 5105 return (__m256i) __builtin_ia32_cvtps2dq256_mask ((__v8sf) __A, 5106 (__v8si) 5107 _mm256_setzero_si256 (), 5108 (__mmask8) __U); 5109 } 5110 5111 extern __inline __m128i 5112 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5113 _mm_mask_cvtps_epi32 (__m128i __W, __mmask8 __U, __m128 __A) 5114 { 5115 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, 5116 (__v4si) __W, 5117 (__mmask8) __U); 5118 } 5119 5120 extern __inline __m128i 5121 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5122 _mm_maskz_cvtps_epi32 (__mmask8 __U, __m128 __A) 5123 { 5124 return (__m128i) __builtin_ia32_cvtps2dq128_mask ((__v4sf) __A, 5125 (__v4si) 5126 _mm_setzero_si128 (), 5127 (__mmask8) __U); 5128 } 5129 5130 extern __inline __m256i 5131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5132 _mm256_cvtps_epu32 (__m256 __A) 5133 { 5134 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, 5135 (__v8si) 5136 _mm256_setzero_si256 (), 5137 (__mmask8) -1); 5138 } 5139 5140 extern __inline __m256i 5141 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5142 _mm256_mask_cvtps_epu32 (__m256i __W, __mmask8 __U, __m256 __A) 5143 { 5144 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, 5145 (__v8si) __W, 5146 (__mmask8) __U); 5147 } 5148 5149 extern __inline __m256i 5150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5151 _mm256_maskz_cvtps_epu32 (__mmask8 __U, __m256 __A) 5152 { 5153 return (__m256i) __builtin_ia32_cvtps2udq256_mask ((__v8sf) __A, 5154 (__v8si) 5155 _mm256_setzero_si256 (), 5156 (__mmask8) __U); 5157 } 5158 5159 extern __inline __m128i 5160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5161 _mm_cvtps_epu32 (__m128 __A) 5162 { 5163 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, 5164 (__v4si) 5165 _mm_setzero_si128 (), 5166 (__mmask8) -1); 5167 } 5168 5169 extern __inline __m128i 5170 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5171 _mm_mask_cvtps_epu32 (__m128i __W, __mmask8 __U, __m128 __A) 5172 { 5173 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, 5174 (__v4si) __W, 5175 (__mmask8) __U); 5176 } 5177 5178 extern __inline __m128i 5179 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5180 _mm_maskz_cvtps_epu32 (__mmask8 __U, __m128 __A) 5181 { 5182 return (__m128i) __builtin_ia32_cvtps2udq128_mask ((__v4sf) __A, 5183 (__v4si) 5184 _mm_setzero_si128 (), 5185 (__mmask8) __U); 5186 } 5187 5188 extern __inline __m256d 5189 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5190 _mm256_mask_movedup_pd (__m256d __W, __mmask8 __U, __m256d __A) 5191 { 5192 return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A, 5193 (__v4df) __W, 5194 (__mmask8) __U); 5195 } 5196 5197 extern __inline __m256d 5198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5199 _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A) 5200 { 5201 return (__m256d) __builtin_ia32_movddup256_mask ((__v4df) __A, 5202 (__v4df) 5203 _mm256_setzero_pd (), 5204 (__mmask8) __U); 5205 } 5206 5207 extern __inline __m128d 5208 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5209 _mm_mask_movedup_pd (__m128d __W, __mmask8 __U, __m128d __A) 5210 { 5211 return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A, 5212 (__v2df) __W, 5213 (__mmask8) __U); 5214 } 5215 5216 extern __inline __m128d 5217 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5218 _mm_maskz_movedup_pd (__mmask8 __U, __m128d __A) 5219 { 5220 return (__m128d) __builtin_ia32_movddup128_mask ((__v2df) __A, 5221 (__v2df) 5222 _mm_setzero_pd (), 5223 (__mmask8) __U); 5224 } 5225 5226 extern __inline __m256 5227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5228 _mm256_mask_movehdup_ps (__m256 __W, __mmask8 __U, __m256 __A) 5229 { 5230 return (__m256) __builtin_ia32_movshdup256_mask ((__v8sf) __A, 5231 (__v8sf) __W, 5232 (__mmask8) __U); 5233 } 5234 5235 extern __inline __m256 5236 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5237 _mm256_maskz_movehdup_ps (__mmask8 __U, __m256 __A) 5238 { 5239 return (__m256) __builtin_ia32_movshdup256_mask ((__v8sf) __A, 5240 (__v8sf) 5241 _mm256_setzero_ps (), 5242 (__mmask8) __U); 5243 } 5244 5245 extern __inline __m128 5246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5247 _mm_mask_movehdup_ps (__m128 __W, __mmask8 __U, __m128 __A) 5248 { 5249 return (__m128) __builtin_ia32_movshdup128_mask ((__v4sf) __A, 5250 (__v4sf) __W, 5251 (__mmask8) __U); 5252 } 5253 5254 extern __inline __m128 5255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5256 _mm_maskz_movehdup_ps (__mmask8 __U, __m128 __A) 5257 { 5258 return (__m128) __builtin_ia32_movshdup128_mask ((__v4sf) __A, 5259 (__v4sf) 5260 _mm_setzero_ps (), 5261 (__mmask8) __U); 5262 } 5263 5264 extern __inline __m256 5265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5266 _mm256_mask_moveldup_ps (__m256 __W, __mmask8 __U, __m256 __A) 5267 { 5268 return (__m256) __builtin_ia32_movsldup256_mask ((__v8sf) __A, 5269 (__v8sf) __W, 5270 (__mmask8) __U); 5271 } 5272 5273 extern __inline __m256 5274 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5275 _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A) 5276 { 5277 return (__m256) __builtin_ia32_movsldup256_mask ((__v8sf) __A, 5278 (__v8sf) 5279 _mm256_setzero_ps (), 5280 (__mmask8) __U); 5281 } 5282 5283 extern __inline __m128 5284 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5285 _mm_mask_moveldup_ps (__m128 __W, __mmask8 __U, __m128 __A) 5286 { 5287 return (__m128) __builtin_ia32_movsldup128_mask ((__v4sf) __A, 5288 (__v4sf) __W, 5289 (__mmask8) __U); 5290 } 5291 5292 extern __inline __m128 5293 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5294 _mm_maskz_moveldup_ps (__mmask8 __U, __m128 __A) 5295 { 5296 return (__m128) __builtin_ia32_movsldup128_mask ((__v4sf) __A, 5297 (__v4sf) 5298 _mm_setzero_ps (), 5299 (__mmask8) __U); 5300 } 5301 5302 extern __inline __m128i 5303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5304 _mm_mask_unpackhi_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 5305 __m128i __B) 5306 { 5307 return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A, 5308 (__v4si) __B, 5309 (__v4si) __W, 5310 (__mmask8) __U); 5311 } 5312 5313 extern __inline __m128i 5314 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5315 _mm_maskz_unpackhi_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 5316 { 5317 return (__m128i) __builtin_ia32_punpckhdq128_mask ((__v4si) __A, 5318 (__v4si) __B, 5319 (__v4si) 5320 _mm_setzero_si128 (), 5321 (__mmask8) __U); 5322 } 5323 5324 extern __inline __m256i 5325 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5326 _mm256_mask_unpackhi_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 5327 __m256i __B) 5328 { 5329 return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A, 5330 (__v8si) __B, 5331 (__v8si) __W, 5332 (__mmask8) __U); 5333 } 5334 5335 extern __inline __m256i 5336 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5337 _mm256_maskz_unpackhi_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 5338 { 5339 return (__m256i) __builtin_ia32_punpckhdq256_mask ((__v8si) __A, 5340 (__v8si) __B, 5341 (__v8si) 5342 _mm256_setzero_si256 (), 5343 (__mmask8) __U); 5344 } 5345 5346 extern __inline __m128i 5347 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5348 _mm_mask_unpackhi_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 5349 __m128i __B) 5350 { 5351 return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A, 5352 (__v2di) __B, 5353 (__v2di) __W, 5354 (__mmask8) __U); 5355 } 5356 5357 extern __inline __m128i 5358 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5359 _mm_maskz_unpackhi_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 5360 { 5361 return (__m128i) __builtin_ia32_punpckhqdq128_mask ((__v2di) __A, 5362 (__v2di) __B, 5363 (__v2di) 5364 _mm_setzero_si128 (), 5365 (__mmask8) __U); 5366 } 5367 5368 extern __inline __m256i 5369 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5370 _mm256_mask_unpackhi_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 5371 __m256i __B) 5372 { 5373 return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A, 5374 (__v4di) __B, 5375 (__v4di) __W, 5376 (__mmask8) __U); 5377 } 5378 5379 extern __inline __m256i 5380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5381 _mm256_maskz_unpackhi_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 5382 { 5383 return (__m256i) __builtin_ia32_punpckhqdq256_mask ((__v4di) __A, 5384 (__v4di) __B, 5385 (__v4di) 5386 _mm256_setzero_si256 (), 5387 (__mmask8) __U); 5388 } 5389 5390 extern __inline __m128i 5391 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5392 _mm_mask_unpacklo_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 5393 __m128i __B) 5394 { 5395 return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A, 5396 (__v4si) __B, 5397 (__v4si) __W, 5398 (__mmask8) __U); 5399 } 5400 5401 extern __inline __m128i 5402 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5403 _mm_maskz_unpacklo_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 5404 { 5405 return (__m128i) __builtin_ia32_punpckldq128_mask ((__v4si) __A, 5406 (__v4si) __B, 5407 (__v4si) 5408 _mm_setzero_si128 (), 5409 (__mmask8) __U); 5410 } 5411 5412 extern __inline __m256i 5413 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5414 _mm256_mask_unpacklo_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 5415 __m256i __B) 5416 { 5417 return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A, 5418 (__v8si) __B, 5419 (__v8si) __W, 5420 (__mmask8) __U); 5421 } 5422 5423 extern __inline __m256i 5424 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5425 _mm256_maskz_unpacklo_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 5426 { 5427 return (__m256i) __builtin_ia32_punpckldq256_mask ((__v8si) __A, 5428 (__v8si) __B, 5429 (__v8si) 5430 _mm256_setzero_si256 (), 5431 (__mmask8) __U); 5432 } 5433 5434 extern __inline __m128i 5435 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5436 _mm_mask_unpacklo_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 5437 __m128i __B) 5438 { 5439 return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A, 5440 (__v2di) __B, 5441 (__v2di) __W, 5442 (__mmask8) __U); 5443 } 5444 5445 extern __inline __m128i 5446 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5447 _mm_maskz_unpacklo_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 5448 { 5449 return (__m128i) __builtin_ia32_punpcklqdq128_mask ((__v2di) __A, 5450 (__v2di) __B, 5451 (__v2di) 5452 _mm_setzero_si128 (), 5453 (__mmask8) __U); 5454 } 5455 5456 extern __inline __m256i 5457 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5458 _mm256_mask_unpacklo_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 5459 __m256i __B) 5460 { 5461 return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A, 5462 (__v4di) __B, 5463 (__v4di) __W, 5464 (__mmask8) __U); 5465 } 5466 5467 extern __inline __m256i 5468 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5469 _mm256_maskz_unpacklo_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 5470 { 5471 return (__m256i) __builtin_ia32_punpcklqdq256_mask ((__v4di) __A, 5472 (__v4di) __B, 5473 (__v4di) 5474 _mm256_setzero_si256 (), 5475 (__mmask8) __U); 5476 } 5477 5478 extern __inline __mmask8 5479 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5480 _mm_cmpeq_epu32_mask (__m128i __A, __m128i __B) 5481 { 5482 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A, 5483 (__v4si) __B, 0, 5484 (__mmask8) -1); 5485 } 5486 5487 extern __inline __mmask8 5488 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5489 _mm_cmpeq_epi32_mask (__m128i __A, __m128i __B) 5490 { 5491 return (__mmask8) __builtin_ia32_pcmpeqd128_mask ((__v4si) __A, 5492 (__v4si) __B, 5493 (__mmask8) -1); 5494 } 5495 5496 extern __inline __mmask8 5497 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5498 _mm_mask_cmpeq_epu32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5499 { 5500 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A, 5501 (__v4si) __B, 0, __U); 5502 } 5503 5504 extern __inline __mmask8 5505 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5506 _mm_mask_cmpeq_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5507 { 5508 return (__mmask8) __builtin_ia32_pcmpeqd128_mask ((__v4si) __A, 5509 (__v4si) __B, __U); 5510 } 5511 5512 extern __inline __mmask8 5513 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5514 _mm256_cmpeq_epu32_mask (__m256i __A, __m256i __B) 5515 { 5516 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A, 5517 (__v8si) __B, 0, 5518 (__mmask8) -1); 5519 } 5520 5521 extern __inline __mmask8 5522 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5523 _mm256_cmpeq_epi32_mask (__m256i __A, __m256i __B) 5524 { 5525 return (__mmask8) __builtin_ia32_pcmpeqd256_mask ((__v8si) __A, 5526 (__v8si) __B, 5527 (__mmask8) -1); 5528 } 5529 5530 extern __inline __mmask8 5531 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5532 _mm256_mask_cmpeq_epu32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5533 { 5534 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A, 5535 (__v8si) __B, 0, __U); 5536 } 5537 5538 extern __inline __mmask8 5539 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5540 _mm256_mask_cmpeq_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5541 { 5542 return (__mmask8) __builtin_ia32_pcmpeqd256_mask ((__v8si) __A, 5543 (__v8si) __B, __U); 5544 } 5545 5546 extern __inline __mmask8 5547 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5548 _mm_cmpeq_epu64_mask (__m128i __A, __m128i __B) 5549 { 5550 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A, 5551 (__v2di) __B, 0, 5552 (__mmask8) -1); 5553 } 5554 5555 extern __inline __mmask8 5556 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5557 _mm_cmpeq_epi64_mask (__m128i __A, __m128i __B) 5558 { 5559 return (__mmask8) __builtin_ia32_pcmpeqq128_mask ((__v2di) __A, 5560 (__v2di) __B, 5561 (__mmask8) -1); 5562 } 5563 5564 extern __inline __mmask8 5565 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5566 _mm_mask_cmpeq_epu64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5567 { 5568 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A, 5569 (__v2di) __B, 0, __U); 5570 } 5571 5572 extern __inline __mmask8 5573 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5574 _mm_mask_cmpeq_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5575 { 5576 return (__mmask8) __builtin_ia32_pcmpeqq128_mask ((__v2di) __A, 5577 (__v2di) __B, __U); 5578 } 5579 5580 extern __inline __mmask8 5581 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5582 _mm256_cmpeq_epu64_mask (__m256i __A, __m256i __B) 5583 { 5584 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A, 5585 (__v4di) __B, 0, 5586 (__mmask8) -1); 5587 } 5588 5589 extern __inline __mmask8 5590 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5591 _mm256_cmpeq_epi64_mask (__m256i __A, __m256i __B) 5592 { 5593 return (__mmask8) __builtin_ia32_pcmpeqq256_mask ((__v4di) __A, 5594 (__v4di) __B, 5595 (__mmask8) -1); 5596 } 5597 5598 extern __inline __mmask8 5599 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5600 _mm256_mask_cmpeq_epu64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5601 { 5602 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A, 5603 (__v4di) __B, 0, __U); 5604 } 5605 5606 extern __inline __mmask8 5607 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5608 _mm256_mask_cmpeq_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5609 { 5610 return (__mmask8) __builtin_ia32_pcmpeqq256_mask ((__v4di) __A, 5611 (__v4di) __B, __U); 5612 } 5613 5614 extern __inline __mmask8 5615 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5616 _mm_cmpgt_epu32_mask (__m128i __A, __m128i __B) 5617 { 5618 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A, 5619 (__v4si) __B, 6, 5620 (__mmask8) -1); 5621 } 5622 5623 extern __inline __mmask8 5624 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5625 _mm_cmpgt_epi32_mask (__m128i __A, __m128i __B) 5626 { 5627 return (__mmask8) __builtin_ia32_pcmpgtd128_mask ((__v4si) __A, 5628 (__v4si) __B, 5629 (__mmask8) -1); 5630 } 5631 5632 extern __inline __mmask8 5633 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5634 _mm_mask_cmpgt_epu32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5635 { 5636 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __A, 5637 (__v4si) __B, 6, __U); 5638 } 5639 5640 extern __inline __mmask8 5641 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5642 _mm_mask_cmpgt_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5643 { 5644 return (__mmask8) __builtin_ia32_pcmpgtd128_mask ((__v4si) __A, 5645 (__v4si) __B, __U); 5646 } 5647 5648 extern __inline __mmask8 5649 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5650 _mm256_cmpgt_epu32_mask (__m256i __A, __m256i __B) 5651 { 5652 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A, 5653 (__v8si) __B, 6, 5654 (__mmask8) -1); 5655 } 5656 5657 extern __inline __mmask8 5658 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5659 _mm256_cmpgt_epi32_mask (__m256i __A, __m256i __B) 5660 { 5661 return (__mmask8) __builtin_ia32_pcmpgtd256_mask ((__v8si) __A, 5662 (__v8si) __B, 5663 (__mmask8) -1); 5664 } 5665 5666 extern __inline __mmask8 5667 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5668 _mm256_mask_cmpgt_epu32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5669 { 5670 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __A, 5671 (__v8si) __B, 6, __U); 5672 } 5673 5674 extern __inline __mmask8 5675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5676 _mm256_mask_cmpgt_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5677 { 5678 return (__mmask8) __builtin_ia32_pcmpgtd256_mask ((__v8si) __A, 5679 (__v8si) __B, __U); 5680 } 5681 5682 extern __inline __mmask8 5683 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5684 _mm_cmpgt_epu64_mask (__m128i __A, __m128i __B) 5685 { 5686 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A, 5687 (__v2di) __B, 6, 5688 (__mmask8) -1); 5689 } 5690 5691 extern __inline __mmask8 5692 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5693 _mm_cmpgt_epi64_mask (__m128i __A, __m128i __B) 5694 { 5695 return (__mmask8) __builtin_ia32_pcmpgtq128_mask ((__v2di) __A, 5696 (__v2di) __B, 5697 (__mmask8) -1); 5698 } 5699 5700 extern __inline __mmask8 5701 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5702 _mm_mask_cmpgt_epu64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5703 { 5704 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __A, 5705 (__v2di) __B, 6, __U); 5706 } 5707 5708 extern __inline __mmask8 5709 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5710 _mm_mask_cmpgt_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5711 { 5712 return (__mmask8) __builtin_ia32_pcmpgtq128_mask ((__v2di) __A, 5713 (__v2di) __B, __U); 5714 } 5715 5716 extern __inline __mmask8 5717 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5718 _mm256_cmpgt_epu64_mask (__m256i __A, __m256i __B) 5719 { 5720 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A, 5721 (__v4di) __B, 6, 5722 (__mmask8) -1); 5723 } 5724 5725 extern __inline __mmask8 5726 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5727 _mm256_cmpgt_epi64_mask (__m256i __A, __m256i __B) 5728 { 5729 return (__mmask8) __builtin_ia32_pcmpgtq256_mask ((__v4di) __A, 5730 (__v4di) __B, 5731 (__mmask8) -1); 5732 } 5733 5734 extern __inline __mmask8 5735 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5736 _mm256_mask_cmpgt_epu64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5737 { 5738 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __A, 5739 (__v4di) __B, 6, __U); 5740 } 5741 5742 extern __inline __mmask8 5743 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5744 _mm256_mask_cmpgt_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5745 { 5746 return (__mmask8) __builtin_ia32_pcmpgtq256_mask ((__v4di) __A, 5747 (__v4di) __B, __U); 5748 } 5749 5750 extern __inline __mmask8 5751 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5752 _mm_test_epi32_mask (__m128i __A, __m128i __B) 5753 { 5754 return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, 5755 (__v4si) __B, 5756 (__mmask8) -1); 5757 } 5758 5759 extern __inline __mmask8 5760 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5761 _mm_mask_test_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5762 { 5763 return (__mmask8) __builtin_ia32_ptestmd128 ((__v4si) __A, 5764 (__v4si) __B, __U); 5765 } 5766 5767 extern __inline __mmask8 5768 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5769 _mm256_test_epi32_mask (__m256i __A, __m256i __B) 5770 { 5771 return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, 5772 (__v8si) __B, 5773 (__mmask8) -1); 5774 } 5775 5776 extern __inline __mmask8 5777 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5778 _mm256_mask_test_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5779 { 5780 return (__mmask8) __builtin_ia32_ptestmd256 ((__v8si) __A, 5781 (__v8si) __B, __U); 5782 } 5783 5784 extern __inline __mmask8 5785 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5786 _mm_test_epi64_mask (__m128i __A, __m128i __B) 5787 { 5788 return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, 5789 (__v2di) __B, 5790 (__mmask8) -1); 5791 } 5792 5793 extern __inline __mmask8 5794 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5795 _mm_mask_test_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5796 { 5797 return (__mmask8) __builtin_ia32_ptestmq128 ((__v2di) __A, 5798 (__v2di) __B, __U); 5799 } 5800 5801 extern __inline __mmask8 5802 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5803 _mm256_test_epi64_mask (__m256i __A, __m256i __B) 5804 { 5805 return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, 5806 (__v4di) __B, 5807 (__mmask8) -1); 5808 } 5809 5810 extern __inline __mmask8 5811 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5812 _mm256_mask_test_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5813 { 5814 return (__mmask8) __builtin_ia32_ptestmq256 ((__v4di) __A, 5815 (__v4di) __B, __U); 5816 } 5817 5818 extern __inline __mmask8 5819 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5820 _mm_testn_epi32_mask (__m128i __A, __m128i __B) 5821 { 5822 return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, 5823 (__v4si) __B, 5824 (__mmask8) -1); 5825 } 5826 5827 extern __inline __mmask8 5828 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5829 _mm_mask_testn_epi32_mask (__mmask8 __U, __m128i __A, __m128i __B) 5830 { 5831 return (__mmask8) __builtin_ia32_ptestnmd128 ((__v4si) __A, 5832 (__v4si) __B, __U); 5833 } 5834 5835 extern __inline __mmask8 5836 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5837 _mm256_testn_epi32_mask (__m256i __A, __m256i __B) 5838 { 5839 return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, 5840 (__v8si) __B, 5841 (__mmask8) -1); 5842 } 5843 5844 extern __inline __mmask8 5845 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5846 _mm256_mask_testn_epi32_mask (__mmask8 __U, __m256i __A, __m256i __B) 5847 { 5848 return (__mmask8) __builtin_ia32_ptestnmd256 ((__v8si) __A, 5849 (__v8si) __B, __U); 5850 } 5851 5852 extern __inline __mmask8 5853 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5854 _mm_testn_epi64_mask (__m128i __A, __m128i __B) 5855 { 5856 return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, 5857 (__v2di) __B, 5858 (__mmask8) -1); 5859 } 5860 5861 extern __inline __mmask8 5862 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5863 _mm_mask_testn_epi64_mask (__mmask8 __U, __m128i __A, __m128i __B) 5864 { 5865 return (__mmask8) __builtin_ia32_ptestnmq128 ((__v2di) __A, 5866 (__v2di) __B, __U); 5867 } 5868 5869 extern __inline __mmask8 5870 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5871 _mm256_testn_epi64_mask (__m256i __A, __m256i __B) 5872 { 5873 return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, 5874 (__v4di) __B, 5875 (__mmask8) -1); 5876 } 5877 5878 extern __inline __mmask8 5879 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5880 _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B) 5881 { 5882 return (__mmask8) __builtin_ia32_ptestnmq256 ((__v4di) __A, 5883 (__v4di) __B, __U); 5884 } 5885 5886 extern __inline __m256d 5887 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5888 _mm256_mask_compress_pd (__m256d __W, __mmask8 __U, __m256d __A) 5889 { 5890 return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, 5891 (__v4df) __W, 5892 (__mmask8) __U); 5893 } 5894 5895 extern __inline __m256d 5896 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5897 _mm256_maskz_compress_pd (__mmask8 __U, __m256d __A) 5898 { 5899 return (__m256d) __builtin_ia32_compressdf256_mask ((__v4df) __A, 5900 (__v4df) 5901 _mm256_setzero_pd (), 5902 (__mmask8) __U); 5903 } 5904 5905 extern __inline void 5906 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5907 _mm256_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m256d __A) 5908 { 5909 __builtin_ia32_compressstoredf256_mask ((__v4df *) __P, 5910 (__v4df) __A, 5911 (__mmask8) __U); 5912 } 5913 5914 extern __inline __m128d 5915 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5916 _mm_mask_compress_pd (__m128d __W, __mmask8 __U, __m128d __A) 5917 { 5918 return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, 5919 (__v2df) __W, 5920 (__mmask8) __U); 5921 } 5922 5923 extern __inline __m128d 5924 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5925 _mm_maskz_compress_pd (__mmask8 __U, __m128d __A) 5926 { 5927 return (__m128d) __builtin_ia32_compressdf128_mask ((__v2df) __A, 5928 (__v2df) 5929 _mm_setzero_pd (), 5930 (__mmask8) __U); 5931 } 5932 5933 extern __inline void 5934 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5935 _mm_mask_compressstoreu_pd (void *__P, __mmask8 __U, __m128d __A) 5936 { 5937 __builtin_ia32_compressstoredf128_mask ((__v2df *) __P, 5938 (__v2df) __A, 5939 (__mmask8) __U); 5940 } 5941 5942 extern __inline __m256 5943 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5944 _mm256_mask_compress_ps (__m256 __W, __mmask8 __U, __m256 __A) 5945 { 5946 return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, 5947 (__v8sf) __W, 5948 (__mmask8) __U); 5949 } 5950 5951 extern __inline __m256 5952 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5953 _mm256_maskz_compress_ps (__mmask8 __U, __m256 __A) 5954 { 5955 return (__m256) __builtin_ia32_compresssf256_mask ((__v8sf) __A, 5956 (__v8sf) 5957 _mm256_setzero_ps (), 5958 (__mmask8) __U); 5959 } 5960 5961 extern __inline void 5962 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5963 _mm256_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m256 __A) 5964 { 5965 __builtin_ia32_compressstoresf256_mask ((__v8sf *) __P, 5966 (__v8sf) __A, 5967 (__mmask8) __U); 5968 } 5969 5970 extern __inline __m128 5971 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5972 _mm_mask_compress_ps (__m128 __W, __mmask8 __U, __m128 __A) 5973 { 5974 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, 5975 (__v4sf) __W, 5976 (__mmask8) __U); 5977 } 5978 5979 extern __inline __m128 5980 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5981 _mm_maskz_compress_ps (__mmask8 __U, __m128 __A) 5982 { 5983 return (__m128) __builtin_ia32_compresssf128_mask ((__v4sf) __A, 5984 (__v4sf) 5985 _mm_setzero_ps (), 5986 (__mmask8) __U); 5987 } 5988 5989 extern __inline void 5990 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 5991 _mm_mask_compressstoreu_ps (void *__P, __mmask8 __U, __m128 __A) 5992 { 5993 __builtin_ia32_compressstoresf128_mask ((__v4sf *) __P, 5994 (__v4sf) __A, 5995 (__mmask8) __U); 5996 } 5997 5998 extern __inline __m256i 5999 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6000 _mm256_mask_compress_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 6001 { 6002 return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, 6003 (__v4di) __W, 6004 (__mmask8) __U); 6005 } 6006 6007 extern __inline __m256i 6008 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6009 _mm256_maskz_compress_epi64 (__mmask8 __U, __m256i __A) 6010 { 6011 return (__m256i) __builtin_ia32_compressdi256_mask ((__v4di) __A, 6012 (__v4di) 6013 _mm256_setzero_si256 (), 6014 (__mmask8) __U); 6015 } 6016 6017 extern __inline void 6018 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6019 _mm256_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m256i __A) 6020 { 6021 __builtin_ia32_compressstoredi256_mask ((__v4di *) __P, 6022 (__v4di) __A, 6023 (__mmask8) __U); 6024 } 6025 6026 extern __inline __m128i 6027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6028 _mm_mask_compress_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 6029 { 6030 return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, 6031 (__v2di) __W, 6032 (__mmask8) __U); 6033 } 6034 6035 extern __inline __m128i 6036 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6037 _mm_maskz_compress_epi64 (__mmask8 __U, __m128i __A) 6038 { 6039 return (__m128i) __builtin_ia32_compressdi128_mask ((__v2di) __A, 6040 (__v2di) 6041 _mm_setzero_si128 (), 6042 (__mmask8) __U); 6043 } 6044 6045 extern __inline void 6046 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6047 _mm_mask_compressstoreu_epi64 (void *__P, __mmask8 __U, __m128i __A) 6048 { 6049 __builtin_ia32_compressstoredi128_mask ((__v2di *) __P, 6050 (__v2di) __A, 6051 (__mmask8) __U); 6052 } 6053 6054 extern __inline __m256i 6055 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6056 _mm256_mask_compress_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 6057 { 6058 return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, 6059 (__v8si) __W, 6060 (__mmask8) __U); 6061 } 6062 6063 extern __inline __m256i 6064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6065 _mm256_maskz_compress_epi32 (__mmask8 __U, __m256i __A) 6066 { 6067 return (__m256i) __builtin_ia32_compresssi256_mask ((__v8si) __A, 6068 (__v8si) 6069 _mm256_setzero_si256 (), 6070 (__mmask8) __U); 6071 } 6072 6073 extern __inline void 6074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6075 _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) 6076 { 6077 __builtin_ia32_compressstoresi256_mask ((__v8si *) __P, 6078 (__v8si) __A, 6079 (__mmask8) __U); 6080 } 6081 6082 extern __inline __m128i 6083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6084 _mm_mask_compress_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 6085 { 6086 return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, 6087 (__v4si) __W, 6088 (__mmask8) __U); 6089 } 6090 6091 extern __inline __m128i 6092 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6093 _mm_maskz_compress_epi32 (__mmask8 __U, __m128i __A) 6094 { 6095 return (__m128i) __builtin_ia32_compresssi128_mask ((__v4si) __A, 6096 (__v4si) 6097 _mm_setzero_si128 (), 6098 (__mmask8) __U); 6099 } 6100 6101 extern __inline void 6102 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6103 _mm_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m128i __A) 6104 { 6105 __builtin_ia32_compressstoresi128_mask ((__v4si *) __P, 6106 (__v4si) __A, 6107 (__mmask8) __U); 6108 } 6109 6110 extern __inline __m256d 6111 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6112 _mm256_mask_expand_pd (__m256d __W, __mmask8 __U, __m256d __A) 6113 { 6114 return (__m256d) __builtin_ia32_expanddf256_mask ((__v4df) __A, 6115 (__v4df) __W, 6116 (__mmask8) __U); 6117 } 6118 6119 extern __inline __m256d 6120 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6121 _mm256_maskz_expand_pd (__mmask8 __U, __m256d __A) 6122 { 6123 return (__m256d) __builtin_ia32_expanddf256_maskz ((__v4df) __A, 6124 (__v4df) 6125 _mm256_setzero_pd (), 6126 (__mmask8) __U); 6127 } 6128 6129 extern __inline __m256d 6130 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6131 _mm256_mask_expandloadu_pd (__m256d __W, __mmask8 __U, void const *__P) 6132 { 6133 return (__m256d) __builtin_ia32_expandloaddf256_mask ((const __v4df *) __P, 6134 (__v4df) __W, 6135 (__mmask8) 6136 __U); 6137 } 6138 6139 extern __inline __m256d 6140 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6141 _mm256_maskz_expandloadu_pd (__mmask8 __U, void const *__P) 6142 { 6143 return (__m256d) __builtin_ia32_expandloaddf256_maskz ((const __v4df *) __P, 6144 (__v4df) 6145 _mm256_setzero_pd (), 6146 (__mmask8) 6147 __U); 6148 } 6149 6150 extern __inline __m128d 6151 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6152 _mm_mask_expand_pd (__m128d __W, __mmask8 __U, __m128d __A) 6153 { 6154 return (__m128d) __builtin_ia32_expanddf128_mask ((__v2df) __A, 6155 (__v2df) __W, 6156 (__mmask8) __U); 6157 } 6158 6159 extern __inline __m128d 6160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6161 _mm_maskz_expand_pd (__mmask8 __U, __m128d __A) 6162 { 6163 return (__m128d) __builtin_ia32_expanddf128_maskz ((__v2df) __A, 6164 (__v2df) 6165 _mm_setzero_pd (), 6166 (__mmask8) __U); 6167 } 6168 6169 extern __inline __m128d 6170 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6171 _mm_mask_expandloadu_pd (__m128d __W, __mmask8 __U, void const *__P) 6172 { 6173 return (__m128d) __builtin_ia32_expandloaddf128_mask ((const __v2df *) __P, 6174 (__v2df) __W, 6175 (__mmask8) 6176 __U); 6177 } 6178 6179 extern __inline __m128d 6180 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6181 _mm_maskz_expandloadu_pd (__mmask8 __U, void const *__P) 6182 { 6183 return (__m128d) __builtin_ia32_expandloaddf128_maskz ((const __v2df *) __P, 6184 (__v2df) 6185 _mm_setzero_pd (), 6186 (__mmask8) 6187 __U); 6188 } 6189 6190 extern __inline __m256 6191 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6192 _mm256_mask_expand_ps (__m256 __W, __mmask8 __U, __m256 __A) 6193 { 6194 return (__m256) __builtin_ia32_expandsf256_mask ((__v8sf) __A, 6195 (__v8sf) __W, 6196 (__mmask8) __U); 6197 } 6198 6199 extern __inline __m256 6200 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6201 _mm256_maskz_expand_ps (__mmask8 __U, __m256 __A) 6202 { 6203 return (__m256) __builtin_ia32_expandsf256_maskz ((__v8sf) __A, 6204 (__v8sf) 6205 _mm256_setzero_ps (), 6206 (__mmask8) __U); 6207 } 6208 6209 extern __inline __m256 6210 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6211 _mm256_mask_expandloadu_ps (__m256 __W, __mmask8 __U, void const *__P) 6212 { 6213 return (__m256) __builtin_ia32_expandloadsf256_mask ((const __v8sf *) __P, 6214 (__v8sf) __W, 6215 (__mmask8) __U); 6216 } 6217 6218 extern __inline __m256 6219 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6220 _mm256_maskz_expandloadu_ps (__mmask8 __U, void const *__P) 6221 { 6222 return (__m256) __builtin_ia32_expandloadsf256_maskz ((const __v8sf *) __P, 6223 (__v8sf) 6224 _mm256_setzero_ps (), 6225 (__mmask8) 6226 __U); 6227 } 6228 6229 extern __inline __m128 6230 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6231 _mm_mask_expand_ps (__m128 __W, __mmask8 __U, __m128 __A) 6232 { 6233 return (__m128) __builtin_ia32_expandsf128_mask ((__v4sf) __A, 6234 (__v4sf) __W, 6235 (__mmask8) __U); 6236 } 6237 6238 extern __inline __m128 6239 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6240 _mm_maskz_expand_ps (__mmask8 __U, __m128 __A) 6241 { 6242 return (__m128) __builtin_ia32_expandsf128_maskz ((__v4sf) __A, 6243 (__v4sf) 6244 _mm_setzero_ps (), 6245 (__mmask8) __U); 6246 } 6247 6248 extern __inline __m128 6249 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6250 _mm_mask_expandloadu_ps (__m128 __W, __mmask8 __U, void const *__P) 6251 { 6252 return (__m128) __builtin_ia32_expandloadsf128_mask ((const __v4sf *) __P, 6253 (__v4sf) __W, 6254 (__mmask8) __U); 6255 } 6256 6257 extern __inline __m128 6258 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6259 _mm_maskz_expandloadu_ps (__mmask8 __U, void const *__P) 6260 { 6261 return (__m128) __builtin_ia32_expandloadsf128_maskz ((const __v4sf *) __P, 6262 (__v4sf) 6263 _mm_setzero_ps (), 6264 (__mmask8) 6265 __U); 6266 } 6267 6268 extern __inline __m256i 6269 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6270 _mm256_mask_expand_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 6271 { 6272 return (__m256i) __builtin_ia32_expanddi256_mask ((__v4di) __A, 6273 (__v4di) __W, 6274 (__mmask8) __U); 6275 } 6276 6277 extern __inline __m256i 6278 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6279 _mm256_maskz_expand_epi64 (__mmask8 __U, __m256i __A) 6280 { 6281 return (__m256i) __builtin_ia32_expanddi256_maskz ((__v4di) __A, 6282 (__v4di) 6283 _mm256_setzero_si256 (), 6284 (__mmask8) __U); 6285 } 6286 6287 extern __inline __m256i 6288 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6289 _mm256_mask_expandloadu_epi64 (__m256i __W, __mmask8 __U, 6290 void const *__P) 6291 { 6292 return (__m256i) __builtin_ia32_expandloaddi256_mask ((const __v4di *) __P, 6293 (__v4di) __W, 6294 (__mmask8) 6295 __U); 6296 } 6297 6298 extern __inline __m256i 6299 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6300 _mm256_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) 6301 { 6302 return (__m256i) __builtin_ia32_expandloaddi256_maskz ((const __v4di *) __P, 6303 (__v4di) 6304 _mm256_setzero_si256 (), 6305 (__mmask8) 6306 __U); 6307 } 6308 6309 extern __inline __m128i 6310 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6311 _mm_mask_expand_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 6312 { 6313 return (__m128i) __builtin_ia32_expanddi128_mask ((__v2di) __A, 6314 (__v2di) __W, 6315 (__mmask8) __U); 6316 } 6317 6318 extern __inline __m128i 6319 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6320 _mm_maskz_expand_epi64 (__mmask8 __U, __m128i __A) 6321 { 6322 return (__m128i) __builtin_ia32_expanddi128_maskz ((__v2di) __A, 6323 (__v2di) 6324 _mm_setzero_si128 (), 6325 (__mmask8) __U); 6326 } 6327 6328 extern __inline __m128i 6329 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6330 _mm_mask_expandloadu_epi64 (__m128i __W, __mmask8 __U, void const *__P) 6331 { 6332 return (__m128i) __builtin_ia32_expandloaddi128_mask ((const __v2di *) __P, 6333 (__v2di) __W, 6334 (__mmask8) 6335 __U); 6336 } 6337 6338 extern __inline __m128i 6339 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6340 _mm_maskz_expandloadu_epi64 (__mmask8 __U, void const *__P) 6341 { 6342 return (__m128i) __builtin_ia32_expandloaddi128_maskz ((const __v2di *) __P, 6343 (__v2di) 6344 _mm_setzero_si128 (), 6345 (__mmask8) 6346 __U); 6347 } 6348 6349 extern __inline __m256i 6350 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6351 _mm256_mask_expand_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 6352 { 6353 return (__m256i) __builtin_ia32_expandsi256_mask ((__v8si) __A, 6354 (__v8si) __W, 6355 (__mmask8) __U); 6356 } 6357 6358 extern __inline __m256i 6359 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6360 _mm256_maskz_expand_epi32 (__mmask8 __U, __m256i __A) 6361 { 6362 return (__m256i) __builtin_ia32_expandsi256_maskz ((__v8si) __A, 6363 (__v8si) 6364 _mm256_setzero_si256 (), 6365 (__mmask8) __U); 6366 } 6367 6368 extern __inline __m256i 6369 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6370 _mm256_mask_expandloadu_epi32 (__m256i __W, __mmask8 __U, 6371 void const *__P) 6372 { 6373 return (__m256i) __builtin_ia32_expandloadsi256_mask ((const __v8si *) __P, 6374 (__v8si) __W, 6375 (__mmask8) 6376 __U); 6377 } 6378 6379 extern __inline __m256i 6380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6381 _mm256_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) 6382 { 6383 return (__m256i) __builtin_ia32_expandloadsi256_maskz ((const __v8si *) __P, 6384 (__v8si) 6385 _mm256_setzero_si256 (), 6386 (__mmask8) 6387 __U); 6388 } 6389 6390 extern __inline __m128i 6391 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6392 _mm_mask_expand_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 6393 { 6394 return (__m128i) __builtin_ia32_expandsi128_mask ((__v4si) __A, 6395 (__v4si) __W, 6396 (__mmask8) __U); 6397 } 6398 6399 extern __inline __m128i 6400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6401 _mm_maskz_expand_epi32 (__mmask8 __U, __m128i __A) 6402 { 6403 return (__m128i) __builtin_ia32_expandsi128_maskz ((__v4si) __A, 6404 (__v4si) 6405 _mm_setzero_si128 (), 6406 (__mmask8) __U); 6407 } 6408 6409 extern __inline __m128i 6410 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6411 _mm_mask_expandloadu_epi32 (__m128i __W, __mmask8 __U, void const *__P) 6412 { 6413 return (__m128i) __builtin_ia32_expandloadsi128_mask ((const __v4si *) __P, 6414 (__v4si) __W, 6415 (__mmask8) 6416 __U); 6417 } 6418 6419 extern __inline __m128i 6420 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6421 _mm_maskz_expandloadu_epi32 (__mmask8 __U, void const *__P) 6422 { 6423 return (__m128i) __builtin_ia32_expandloadsi128_maskz ((const __v4si *) __P, 6424 (__v4si) 6425 _mm_setzero_si128 (), 6426 (__mmask8) 6427 __U); 6428 } 6429 6430 extern __inline __m256d 6431 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6432 _mm256_permutex2var_pd (__m256d __A, __m256i __I, __m256d __B) 6433 { 6434 return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I 6435 /* idx */ , 6436 (__v4df) __A, 6437 (__v4df) __B, 6438 (__mmask8) -1); 6439 } 6440 6441 extern __inline __m256d 6442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6443 _mm256_mask_permutex2var_pd (__m256d __A, __mmask8 __U, __m256i __I, 6444 __m256d __B) 6445 { 6446 return (__m256d) __builtin_ia32_vpermt2varpd256_mask ((__v4di) __I 6447 /* idx */ , 6448 (__v4df) __A, 6449 (__v4df) __B, 6450 (__mmask8) 6451 __U); 6452 } 6453 6454 extern __inline __m256d 6455 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6456 _mm256_mask2_permutex2var_pd (__m256d __A, __m256i __I, __mmask8 __U, 6457 __m256d __B) 6458 { 6459 return (__m256d) __builtin_ia32_vpermi2varpd256_mask ((__v4df) __A, 6460 (__v4di) __I 6461 /* idx */ , 6462 (__v4df) __B, 6463 (__mmask8) 6464 __U); 6465 } 6466 6467 extern __inline __m256d 6468 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6469 _mm256_maskz_permutex2var_pd (__mmask8 __U, __m256d __A, __m256i __I, 6470 __m256d __B) 6471 { 6472 return (__m256d) __builtin_ia32_vpermt2varpd256_maskz ((__v4di) __I 6473 /* idx */ , 6474 (__v4df) __A, 6475 (__v4df) __B, 6476 (__mmask8) 6477 __U); 6478 } 6479 6480 extern __inline __m256 6481 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6482 _mm256_permutex2var_ps (__m256 __A, __m256i __I, __m256 __B) 6483 { 6484 return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I 6485 /* idx */ , 6486 (__v8sf) __A, 6487 (__v8sf) __B, 6488 (__mmask8) -1); 6489 } 6490 6491 extern __inline __m256 6492 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6493 _mm256_mask_permutex2var_ps (__m256 __A, __mmask8 __U, __m256i __I, 6494 __m256 __B) 6495 { 6496 return (__m256) __builtin_ia32_vpermt2varps256_mask ((__v8si) __I 6497 /* idx */ , 6498 (__v8sf) __A, 6499 (__v8sf) __B, 6500 (__mmask8) __U); 6501 } 6502 6503 extern __inline __m256 6504 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6505 _mm256_mask2_permutex2var_ps (__m256 __A, __m256i __I, __mmask8 __U, 6506 __m256 __B) 6507 { 6508 return (__m256) __builtin_ia32_vpermi2varps256_mask ((__v8sf) __A, 6509 (__v8si) __I 6510 /* idx */ , 6511 (__v8sf) __B, 6512 (__mmask8) __U); 6513 } 6514 6515 extern __inline __m256 6516 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6517 _mm256_maskz_permutex2var_ps (__mmask8 __U, __m256 __A, __m256i __I, 6518 __m256 __B) 6519 { 6520 return (__m256) __builtin_ia32_vpermt2varps256_maskz ((__v8si) __I 6521 /* idx */ , 6522 (__v8sf) __A, 6523 (__v8sf) __B, 6524 (__mmask8) 6525 __U); 6526 } 6527 6528 extern __inline __m128i 6529 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6530 _mm_permutex2var_epi64 (__m128i __A, __m128i __I, __m128i __B) 6531 { 6532 return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I 6533 /* idx */ , 6534 (__v2di) __A, 6535 (__v2di) __B, 6536 (__mmask8) -1); 6537 } 6538 6539 extern __inline __m128i 6540 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6541 _mm_mask_permutex2var_epi64 (__m128i __A, __mmask8 __U, __m128i __I, 6542 __m128i __B) 6543 { 6544 return (__m128i) __builtin_ia32_vpermt2varq128_mask ((__v2di) __I 6545 /* idx */ , 6546 (__v2di) __A, 6547 (__v2di) __B, 6548 (__mmask8) __U); 6549 } 6550 6551 extern __inline __m128i 6552 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6553 _mm_mask2_permutex2var_epi64 (__m128i __A, __m128i __I, __mmask8 __U, 6554 __m128i __B) 6555 { 6556 return (__m128i) __builtin_ia32_vpermi2varq128_mask ((__v2di) __A, 6557 (__v2di) __I 6558 /* idx */ , 6559 (__v2di) __B, 6560 (__mmask8) __U); 6561 } 6562 6563 extern __inline __m128i 6564 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6565 _mm_maskz_permutex2var_epi64 (__mmask8 __U, __m128i __A, __m128i __I, 6566 __m128i __B) 6567 { 6568 return (__m128i) __builtin_ia32_vpermt2varq128_maskz ((__v2di) __I 6569 /* idx */ , 6570 (__v2di) __A, 6571 (__v2di) __B, 6572 (__mmask8) 6573 __U); 6574 } 6575 6576 extern __inline __m128i 6577 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6578 _mm_permutex2var_epi32 (__m128i __A, __m128i __I, __m128i __B) 6579 { 6580 return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I 6581 /* idx */ , 6582 (__v4si) __A, 6583 (__v4si) __B, 6584 (__mmask8) -1); 6585 } 6586 6587 extern __inline __m128i 6588 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6589 _mm_mask_permutex2var_epi32 (__m128i __A, __mmask8 __U, __m128i __I, 6590 __m128i __B) 6591 { 6592 return (__m128i) __builtin_ia32_vpermt2vard128_mask ((__v4si) __I 6593 /* idx */ , 6594 (__v4si) __A, 6595 (__v4si) __B, 6596 (__mmask8) __U); 6597 } 6598 6599 extern __inline __m128i 6600 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6601 _mm_mask2_permutex2var_epi32 (__m128i __A, __m128i __I, __mmask8 __U, 6602 __m128i __B) 6603 { 6604 return (__m128i) __builtin_ia32_vpermi2vard128_mask ((__v4si) __A, 6605 (__v4si) __I 6606 /* idx */ , 6607 (__v4si) __B, 6608 (__mmask8) __U); 6609 } 6610 6611 extern __inline __m128i 6612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6613 _mm_maskz_permutex2var_epi32 (__mmask8 __U, __m128i __A, __m128i __I, 6614 __m128i __B) 6615 { 6616 return (__m128i) __builtin_ia32_vpermt2vard128_maskz ((__v4si) __I 6617 /* idx */ , 6618 (__v4si) __A, 6619 (__v4si) __B, 6620 (__mmask8) 6621 __U); 6622 } 6623 6624 extern __inline __m256i 6625 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6626 _mm256_permutex2var_epi64 (__m256i __A, __m256i __I, __m256i __B) 6627 { 6628 return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I 6629 /* idx */ , 6630 (__v4di) __A, 6631 (__v4di) __B, 6632 (__mmask8) -1); 6633 } 6634 6635 extern __inline __m256i 6636 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6637 _mm256_mask_permutex2var_epi64 (__m256i __A, __mmask8 __U, __m256i __I, 6638 __m256i __B) 6639 { 6640 return (__m256i) __builtin_ia32_vpermt2varq256_mask ((__v4di) __I 6641 /* idx */ , 6642 (__v4di) __A, 6643 (__v4di) __B, 6644 (__mmask8) __U); 6645 } 6646 6647 extern __inline __m256i 6648 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6649 _mm256_mask2_permutex2var_epi64 (__m256i __A, __m256i __I, 6650 __mmask8 __U, __m256i __B) 6651 { 6652 return (__m256i) __builtin_ia32_vpermi2varq256_mask ((__v4di) __A, 6653 (__v4di) __I 6654 /* idx */ , 6655 (__v4di) __B, 6656 (__mmask8) __U); 6657 } 6658 6659 extern __inline __m256i 6660 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6661 _mm256_maskz_permutex2var_epi64 (__mmask8 __U, __m256i __A, 6662 __m256i __I, __m256i __B) 6663 { 6664 return (__m256i) __builtin_ia32_vpermt2varq256_maskz ((__v4di) __I 6665 /* idx */ , 6666 (__v4di) __A, 6667 (__v4di) __B, 6668 (__mmask8) 6669 __U); 6670 } 6671 6672 extern __inline __m256i 6673 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6674 _mm256_permutex2var_epi32 (__m256i __A, __m256i __I, __m256i __B) 6675 { 6676 return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I 6677 /* idx */ , 6678 (__v8si) __A, 6679 (__v8si) __B, 6680 (__mmask8) -1); 6681 } 6682 6683 extern __inline __m256i 6684 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6685 _mm256_mask_permutex2var_epi32 (__m256i __A, __mmask8 __U, __m256i __I, 6686 __m256i __B) 6687 { 6688 return (__m256i) __builtin_ia32_vpermt2vard256_mask ((__v8si) __I 6689 /* idx */ , 6690 (__v8si) __A, 6691 (__v8si) __B, 6692 (__mmask8) __U); 6693 } 6694 6695 extern __inline __m256i 6696 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6697 _mm256_mask2_permutex2var_epi32 (__m256i __A, __m256i __I, 6698 __mmask8 __U, __m256i __B) 6699 { 6700 return (__m256i) __builtin_ia32_vpermi2vard256_mask ((__v8si) __A, 6701 (__v8si) __I 6702 /* idx */ , 6703 (__v8si) __B, 6704 (__mmask8) __U); 6705 } 6706 6707 extern __inline __m256i 6708 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6709 _mm256_maskz_permutex2var_epi32 (__mmask8 __U, __m256i __A, 6710 __m256i __I, __m256i __B) 6711 { 6712 return (__m256i) __builtin_ia32_vpermt2vard256_maskz ((__v8si) __I 6713 /* idx */ , 6714 (__v8si) __A, 6715 (__v8si) __B, 6716 (__mmask8) 6717 __U); 6718 } 6719 6720 extern __inline __m128d 6721 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6722 _mm_permutex2var_pd (__m128d __A, __m128i __I, __m128d __B) 6723 { 6724 return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I 6725 /* idx */ , 6726 (__v2df) __A, 6727 (__v2df) __B, 6728 (__mmask8) -1); 6729 } 6730 6731 extern __inline __m128d 6732 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6733 _mm_mask_permutex2var_pd (__m128d __A, __mmask8 __U, __m128i __I, 6734 __m128d __B) 6735 { 6736 return (__m128d) __builtin_ia32_vpermt2varpd128_mask ((__v2di) __I 6737 /* idx */ , 6738 (__v2df) __A, 6739 (__v2df) __B, 6740 (__mmask8) 6741 __U); 6742 } 6743 6744 extern __inline __m128d 6745 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6746 _mm_mask2_permutex2var_pd (__m128d __A, __m128i __I, __mmask8 __U, 6747 __m128d __B) 6748 { 6749 return (__m128d) __builtin_ia32_vpermi2varpd128_mask ((__v2df) __A, 6750 (__v2di) __I 6751 /* idx */ , 6752 (__v2df) __B, 6753 (__mmask8) 6754 __U); 6755 } 6756 6757 extern __inline __m128d 6758 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6759 _mm_maskz_permutex2var_pd (__mmask8 __U, __m128d __A, __m128i __I, 6760 __m128d __B) 6761 { 6762 return (__m128d) __builtin_ia32_vpermt2varpd128_maskz ((__v2di) __I 6763 /* idx */ , 6764 (__v2df) __A, 6765 (__v2df) __B, 6766 (__mmask8) 6767 __U); 6768 } 6769 6770 extern __inline __m128 6771 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6772 _mm_permutex2var_ps (__m128 __A, __m128i __I, __m128 __B) 6773 { 6774 return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I 6775 /* idx */ , 6776 (__v4sf) __A, 6777 (__v4sf) __B, 6778 (__mmask8) -1); 6779 } 6780 6781 extern __inline __m128 6782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6783 _mm_mask_permutex2var_ps (__m128 __A, __mmask8 __U, __m128i __I, 6784 __m128 __B) 6785 { 6786 return (__m128) __builtin_ia32_vpermt2varps128_mask ((__v4si) __I 6787 /* idx */ , 6788 (__v4sf) __A, 6789 (__v4sf) __B, 6790 (__mmask8) __U); 6791 } 6792 6793 extern __inline __m128 6794 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6795 _mm_mask2_permutex2var_ps (__m128 __A, __m128i __I, __mmask8 __U, 6796 __m128 __B) 6797 { 6798 return (__m128) __builtin_ia32_vpermi2varps128_mask ((__v4sf) __A, 6799 (__v4si) __I 6800 /* idx */ , 6801 (__v4sf) __B, 6802 (__mmask8) __U); 6803 } 6804 6805 extern __inline __m128 6806 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6807 _mm_maskz_permutex2var_ps (__mmask8 __U, __m128 __A, __m128i __I, 6808 __m128 __B) 6809 { 6810 return (__m128) __builtin_ia32_vpermt2varps128_maskz ((__v4si) __I 6811 /* idx */ , 6812 (__v4sf) __A, 6813 (__v4sf) __B, 6814 (__mmask8) 6815 __U); 6816 } 6817 6818 extern __inline __m128i 6819 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6820 _mm_srav_epi64 (__m128i __X, __m128i __Y) 6821 { 6822 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, 6823 (__v2di) __Y, 6824 (__v2di) 6825 _mm_setzero_si128 (), 6826 (__mmask8) -1); 6827 } 6828 6829 extern __inline __m128i 6830 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6831 _mm_mask_srav_epi64 (__m128i __W, __mmask8 __U, __m128i __X, 6832 __m128i __Y) 6833 { 6834 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, 6835 (__v2di) __Y, 6836 (__v2di) __W, 6837 (__mmask8) __U); 6838 } 6839 6840 extern __inline __m128i 6841 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6842 _mm_maskz_srav_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) 6843 { 6844 return (__m128i) __builtin_ia32_psravq128_mask ((__v2di) __X, 6845 (__v2di) __Y, 6846 (__v2di) 6847 _mm_setzero_si128 (), 6848 (__mmask8) __U); 6849 } 6850 6851 extern __inline __m256i 6852 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6853 _mm256_mask_sllv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, 6854 __m256i __Y) 6855 { 6856 return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, 6857 (__v8si) __Y, 6858 (__v8si) __W, 6859 (__mmask8) __U); 6860 } 6861 6862 extern __inline __m256i 6863 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6864 _mm256_maskz_sllv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) 6865 { 6866 return (__m256i) __builtin_ia32_psllv8si_mask ((__v8si) __X, 6867 (__v8si) __Y, 6868 (__v8si) 6869 _mm256_setzero_si256 (), 6870 (__mmask8) __U); 6871 } 6872 6873 extern __inline __m128i 6874 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6875 _mm_mask_sllv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, 6876 __m128i __Y) 6877 { 6878 return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, 6879 (__v4si) __Y, 6880 (__v4si) __W, 6881 (__mmask8) __U); 6882 } 6883 6884 extern __inline __m128i 6885 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6886 _mm_maskz_sllv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) 6887 { 6888 return (__m128i) __builtin_ia32_psllv4si_mask ((__v4si) __X, 6889 (__v4si) __Y, 6890 (__v4si) 6891 _mm_setzero_si128 (), 6892 (__mmask8) __U); 6893 } 6894 6895 extern __inline __m256i 6896 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6897 _mm256_mask_sllv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, 6898 __m256i __Y) 6899 { 6900 return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, 6901 (__v4di) __Y, 6902 (__v4di) __W, 6903 (__mmask8) __U); 6904 } 6905 6906 extern __inline __m256i 6907 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6908 _mm256_maskz_sllv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) 6909 { 6910 return (__m256i) __builtin_ia32_psllv4di_mask ((__v4di) __X, 6911 (__v4di) __Y, 6912 (__v4di) 6913 _mm256_setzero_si256 (), 6914 (__mmask8) __U); 6915 } 6916 6917 extern __inline __m128i 6918 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6919 _mm_mask_sllv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, 6920 __m128i __Y) 6921 { 6922 return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, 6923 (__v2di) __Y, 6924 (__v2di) __W, 6925 (__mmask8) __U); 6926 } 6927 6928 extern __inline __m128i 6929 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6930 _mm_maskz_sllv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) 6931 { 6932 return (__m128i) __builtin_ia32_psllv2di_mask ((__v2di) __X, 6933 (__v2di) __Y, 6934 (__v2di) 6935 _mm_setzero_si128 (), 6936 (__mmask8) __U); 6937 } 6938 6939 extern __inline __m256i 6940 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6941 _mm256_mask_srav_epi32 (__m256i __W, __mmask8 __U, __m256i __X, 6942 __m256i __Y) 6943 { 6944 return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, 6945 (__v8si) __Y, 6946 (__v8si) __W, 6947 (__mmask8) __U); 6948 } 6949 6950 extern __inline __m256i 6951 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6952 _mm256_maskz_srav_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) 6953 { 6954 return (__m256i) __builtin_ia32_psrav8si_mask ((__v8si) __X, 6955 (__v8si) __Y, 6956 (__v8si) 6957 _mm256_setzero_si256 (), 6958 (__mmask8) __U); 6959 } 6960 6961 extern __inline __m128i 6962 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6963 _mm_mask_srav_epi32 (__m128i __W, __mmask8 __U, __m128i __X, 6964 __m128i __Y) 6965 { 6966 return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, 6967 (__v4si) __Y, 6968 (__v4si) __W, 6969 (__mmask8) __U); 6970 } 6971 6972 extern __inline __m128i 6973 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6974 _mm_maskz_srav_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) 6975 { 6976 return (__m128i) __builtin_ia32_psrav4si_mask ((__v4si) __X, 6977 (__v4si) __Y, 6978 (__v4si) 6979 _mm_setzero_si128 (), 6980 (__mmask8) __U); 6981 } 6982 6983 extern __inline __m256i 6984 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6985 _mm256_mask_srlv_epi32 (__m256i __W, __mmask8 __U, __m256i __X, 6986 __m256i __Y) 6987 { 6988 return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, 6989 (__v8si) __Y, 6990 (__v8si) __W, 6991 (__mmask8) __U); 6992 } 6993 6994 extern __inline __m256i 6995 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 6996 _mm256_maskz_srlv_epi32 (__mmask8 __U, __m256i __X, __m256i __Y) 6997 { 6998 return (__m256i) __builtin_ia32_psrlv8si_mask ((__v8si) __X, 6999 (__v8si) __Y, 7000 (__v8si) 7001 _mm256_setzero_si256 (), 7002 (__mmask8) __U); 7003 } 7004 7005 extern __inline __m128i 7006 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7007 _mm_mask_srlv_epi32 (__m128i __W, __mmask8 __U, __m128i __X, 7008 __m128i __Y) 7009 { 7010 return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, 7011 (__v4si) __Y, 7012 (__v4si) __W, 7013 (__mmask8) __U); 7014 } 7015 7016 extern __inline __m128i 7017 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7018 _mm_maskz_srlv_epi32 (__mmask8 __U, __m128i __X, __m128i __Y) 7019 { 7020 return (__m128i) __builtin_ia32_psrlv4si_mask ((__v4si) __X, 7021 (__v4si) __Y, 7022 (__v4si) 7023 _mm_setzero_si128 (), 7024 (__mmask8) __U); 7025 } 7026 7027 extern __inline __m256i 7028 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7029 _mm256_mask_srlv_epi64 (__m256i __W, __mmask8 __U, __m256i __X, 7030 __m256i __Y) 7031 { 7032 return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, 7033 (__v4di) __Y, 7034 (__v4di) __W, 7035 (__mmask8) __U); 7036 } 7037 7038 extern __inline __m256i 7039 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7040 _mm256_maskz_srlv_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) 7041 { 7042 return (__m256i) __builtin_ia32_psrlv4di_mask ((__v4di) __X, 7043 (__v4di) __Y, 7044 (__v4di) 7045 _mm256_setzero_si256 (), 7046 (__mmask8) __U); 7047 } 7048 7049 extern __inline __m128i 7050 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7051 _mm_mask_srlv_epi64 (__m128i __W, __mmask8 __U, __m128i __X, 7052 __m128i __Y) 7053 { 7054 return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, 7055 (__v2di) __Y, 7056 (__v2di) __W, 7057 (__mmask8) __U); 7058 } 7059 7060 extern __inline __m128i 7061 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7062 _mm_maskz_srlv_epi64 (__mmask8 __U, __m128i __X, __m128i __Y) 7063 { 7064 return (__m128i) __builtin_ia32_psrlv2di_mask ((__v2di) __X, 7065 (__v2di) __Y, 7066 (__v2di) 7067 _mm_setzero_si128 (), 7068 (__mmask8) __U); 7069 } 7070 7071 extern __inline __m256i 7072 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7073 _mm256_rolv_epi32 (__m256i __A, __m256i __B) 7074 { 7075 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, 7076 (__v8si) __B, 7077 (__v8si) 7078 _mm256_setzero_si256 (), 7079 (__mmask8) -1); 7080 } 7081 7082 extern __inline __m256i 7083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7084 _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 7085 __m256i __B) 7086 { 7087 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, 7088 (__v8si) __B, 7089 (__v8si) __W, 7090 (__mmask8) __U); 7091 } 7092 7093 extern __inline __m256i 7094 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7095 _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 7096 { 7097 return (__m256i) __builtin_ia32_prolvd256_mask ((__v8si) __A, 7098 (__v8si) __B, 7099 (__v8si) 7100 _mm256_setzero_si256 (), 7101 (__mmask8) __U); 7102 } 7103 7104 extern __inline __m128i 7105 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7106 _mm_rolv_epi32 (__m128i __A, __m128i __B) 7107 { 7108 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, 7109 (__v4si) __B, 7110 (__v4si) 7111 _mm_setzero_si128 (), 7112 (__mmask8) -1); 7113 } 7114 7115 extern __inline __m128i 7116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7117 _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 7118 __m128i __B) 7119 { 7120 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, 7121 (__v4si) __B, 7122 (__v4si) __W, 7123 (__mmask8) __U); 7124 } 7125 7126 extern __inline __m128i 7127 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7128 _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 7129 { 7130 return (__m128i) __builtin_ia32_prolvd128_mask ((__v4si) __A, 7131 (__v4si) __B, 7132 (__v4si) 7133 _mm_setzero_si128 (), 7134 (__mmask8) __U); 7135 } 7136 7137 extern __inline __m256i 7138 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7139 _mm256_rorv_epi32 (__m256i __A, __m256i __B) 7140 { 7141 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, 7142 (__v8si) __B, 7143 (__v8si) 7144 _mm256_setzero_si256 (), 7145 (__mmask8) -1); 7146 } 7147 7148 extern __inline __m256i 7149 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7150 _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 7151 __m256i __B) 7152 { 7153 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, 7154 (__v8si) __B, 7155 (__v8si) __W, 7156 (__mmask8) __U); 7157 } 7158 7159 extern __inline __m256i 7160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7161 _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B) 7162 { 7163 return (__m256i) __builtin_ia32_prorvd256_mask ((__v8si) __A, 7164 (__v8si) __B, 7165 (__v8si) 7166 _mm256_setzero_si256 (), 7167 (__mmask8) __U); 7168 } 7169 7170 extern __inline __m128i 7171 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7172 _mm_rorv_epi32 (__m128i __A, __m128i __B) 7173 { 7174 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, 7175 (__v4si) __B, 7176 (__v4si) 7177 _mm_setzero_si128 (), 7178 (__mmask8) -1); 7179 } 7180 7181 extern __inline __m128i 7182 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7183 _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 7184 __m128i __B) 7185 { 7186 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, 7187 (__v4si) __B, 7188 (__v4si) __W, 7189 (__mmask8) __U); 7190 } 7191 7192 extern __inline __m128i 7193 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7194 _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 7195 { 7196 return (__m128i) __builtin_ia32_prorvd128_mask ((__v4si) __A, 7197 (__v4si) __B, 7198 (__v4si) 7199 _mm_setzero_si128 (), 7200 (__mmask8) __U); 7201 } 7202 7203 extern __inline __m256i 7204 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7205 _mm256_rolv_epi64 (__m256i __A, __m256i __B) 7206 { 7207 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, 7208 (__v4di) __B, 7209 (__v4di) 7210 _mm256_setzero_si256 (), 7211 (__mmask8) -1); 7212 } 7213 7214 extern __inline __m256i 7215 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7216 _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7217 __m256i __B) 7218 { 7219 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, 7220 (__v4di) __B, 7221 (__v4di) __W, 7222 (__mmask8) __U); 7223 } 7224 7225 extern __inline __m256i 7226 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7227 _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7228 { 7229 return (__m256i) __builtin_ia32_prolvq256_mask ((__v4di) __A, 7230 (__v4di) __B, 7231 (__v4di) 7232 _mm256_setzero_si256 (), 7233 (__mmask8) __U); 7234 } 7235 7236 extern __inline __m128i 7237 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7238 _mm_rolv_epi64 (__m128i __A, __m128i __B) 7239 { 7240 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, 7241 (__v2di) __B, 7242 (__v2di) 7243 _mm_setzero_si128 (), 7244 (__mmask8) -1); 7245 } 7246 7247 extern __inline __m128i 7248 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7249 _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 7250 __m128i __B) 7251 { 7252 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, 7253 (__v2di) __B, 7254 (__v2di) __W, 7255 (__mmask8) __U); 7256 } 7257 7258 extern __inline __m128i 7259 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7260 _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7261 { 7262 return (__m128i) __builtin_ia32_prolvq128_mask ((__v2di) __A, 7263 (__v2di) __B, 7264 (__v2di) 7265 _mm_setzero_si128 (), 7266 (__mmask8) __U); 7267 } 7268 7269 extern __inline __m256i 7270 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7271 _mm256_rorv_epi64 (__m256i __A, __m256i __B) 7272 { 7273 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, 7274 (__v4di) __B, 7275 (__v4di) 7276 _mm256_setzero_si256 (), 7277 (__mmask8) -1); 7278 } 7279 7280 extern __inline __m256i 7281 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7282 _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7283 __m256i __B) 7284 { 7285 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, 7286 (__v4di) __B, 7287 (__v4di) __W, 7288 (__mmask8) __U); 7289 } 7290 7291 extern __inline __m256i 7292 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7293 _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7294 { 7295 return (__m256i) __builtin_ia32_prorvq256_mask ((__v4di) __A, 7296 (__v4di) __B, 7297 (__v4di) 7298 _mm256_setzero_si256 (), 7299 (__mmask8) __U); 7300 } 7301 7302 extern __inline __m128i 7303 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7304 _mm_rorv_epi64 (__m128i __A, __m128i __B) 7305 { 7306 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, 7307 (__v2di) __B, 7308 (__v2di) 7309 _mm_setzero_si128 (), 7310 (__mmask8) -1); 7311 } 7312 7313 extern __inline __m128i 7314 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7315 _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 7316 __m128i __B) 7317 { 7318 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, 7319 (__v2di) __B, 7320 (__v2di) __W, 7321 (__mmask8) __U); 7322 } 7323 7324 extern __inline __m128i 7325 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7326 _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7327 { 7328 return (__m128i) __builtin_ia32_prorvq128_mask ((__v2di) __A, 7329 (__v2di) __B, 7330 (__v2di) 7331 _mm_setzero_si128 (), 7332 (__mmask8) __U); 7333 } 7334 7335 extern __inline __m256i 7336 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7337 _mm256_srav_epi64 (__m256i __X, __m256i __Y) 7338 { 7339 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, 7340 (__v4di) __Y, 7341 (__v4di) 7342 _mm256_setzero_si256 (), 7343 (__mmask8) -1); 7344 } 7345 7346 extern __inline __m256i 7347 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7348 _mm256_mask_srav_epi64 (__m256i __W, __mmask8 __U, __m256i __X, 7349 __m256i __Y) 7350 { 7351 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, 7352 (__v4di) __Y, 7353 (__v4di) __W, 7354 (__mmask8) __U); 7355 } 7356 7357 extern __inline __m256i 7358 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7359 _mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y) 7360 { 7361 return (__m256i) __builtin_ia32_psravq256_mask ((__v4di) __X, 7362 (__v4di) __Y, 7363 (__v4di) 7364 _mm256_setzero_si256 (), 7365 (__mmask8) __U); 7366 } 7367 7368 extern __inline __m256i 7369 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7370 _mm256_mask_and_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7371 __m256i __B) 7372 { 7373 return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A, 7374 (__v4di) __B, 7375 (__v4di) __W, __U); 7376 } 7377 7378 extern __inline __m256i 7379 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7380 _mm256_maskz_and_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7381 { 7382 return (__m256i) __builtin_ia32_pandq256_mask ((__v4di) __A, 7383 (__v4di) __B, 7384 (__v4di) 7385 _mm256_setzero_pd (), 7386 __U); 7387 } 7388 7389 extern __inline __m128i 7390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7391 _mm_mask_and_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 7392 __m128i __B) 7393 { 7394 return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A, 7395 (__v2di) __B, 7396 (__v2di) __W, __U); 7397 } 7398 7399 extern __inline __m128i 7400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7401 _mm_maskz_and_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7402 { 7403 return (__m128i) __builtin_ia32_pandq128_mask ((__v2di) __A, 7404 (__v2di) __B, 7405 (__v2di) 7406 _mm_setzero_pd (), 7407 __U); 7408 } 7409 7410 extern __inline __m256i 7411 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7412 _mm256_mask_andnot_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7413 __m256i __B) 7414 { 7415 return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A, 7416 (__v4di) __B, 7417 (__v4di) __W, __U); 7418 } 7419 7420 extern __inline __m256i 7421 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7422 _mm256_maskz_andnot_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7423 { 7424 return (__m256i) __builtin_ia32_pandnq256_mask ((__v4di) __A, 7425 (__v4di) __B, 7426 (__v4di) 7427 _mm256_setzero_pd (), 7428 __U); 7429 } 7430 7431 extern __inline __m128i 7432 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7433 _mm_mask_andnot_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 7434 __m128i __B) 7435 { 7436 return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A, 7437 (__v2di) __B, 7438 (__v2di) __W, __U); 7439 } 7440 7441 extern __inline __m128i 7442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7443 _mm_maskz_andnot_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7444 { 7445 return (__m128i) __builtin_ia32_pandnq128_mask ((__v2di) __A, 7446 (__v2di) __B, 7447 (__v2di) 7448 _mm_setzero_pd (), 7449 __U); 7450 } 7451 7452 extern __inline __m256i 7453 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7454 _mm256_mask_or_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7455 __m256i __B) 7456 { 7457 return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A, 7458 (__v4di) __B, 7459 (__v4di) __W, 7460 (__mmask8) __U); 7461 } 7462 7463 extern __inline __m256i 7464 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7465 _mm256_maskz_or_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7466 { 7467 return (__m256i) __builtin_ia32_porq256_mask ((__v4di) __A, 7468 (__v4di) __B, 7469 (__v4di) 7470 _mm256_setzero_si256 (), 7471 (__mmask8) __U); 7472 } 7473 7474 extern __inline __m256i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 7475 _mm256_or_epi64 (__m256i __A, __m256i __B) 7476 { 7477 return (__m256i) ((__v4du)__A | (__v4du)__B); 7478 } 7479 7480 extern __inline __m128i 7481 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7482 _mm_mask_or_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) 7483 { 7484 return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A, 7485 (__v2di) __B, 7486 (__v2di) __W, 7487 (__mmask8) __U); 7488 } 7489 7490 extern __inline __m128i 7491 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7492 _mm_maskz_or_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7493 { 7494 return (__m128i) __builtin_ia32_porq128_mask ((__v2di) __A, 7495 (__v2di) __B, 7496 (__v2di) 7497 _mm_setzero_si128 (), 7498 (__mmask8) __U); 7499 } 7500 7501 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 7502 _mm_or_epi64 (__m128i __A, __m128i __B) 7503 { 7504 return (__m128i) ((__v2du)__A | (__v2du)__B); 7505 } 7506 7507 extern __inline __m256i 7508 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7509 _mm256_mask_xor_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 7510 __m256i __B) 7511 { 7512 return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A, 7513 (__v4di) __B, 7514 (__v4di) __W, 7515 (__mmask8) __U); 7516 } 7517 7518 extern __inline __m256i 7519 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7520 _mm256_maskz_xor_epi64 (__mmask8 __U, __m256i __A, __m256i __B) 7521 { 7522 return (__m256i) __builtin_ia32_pxorq256_mask ((__v4di) __A, 7523 (__v4di) __B, 7524 (__v4di) 7525 _mm256_setzero_si256 (), 7526 (__mmask8) __U); 7527 } 7528 7529 extern __inline __m256i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 7530 _mm256_xor_epi64 (__m256i __A, __m256i __B) 7531 { 7532 return (__m256i) ((__v4du)__A ^ (__v4du)__B); 7533 } 7534 7535 extern __inline __m128i 7536 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7537 _mm_mask_xor_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 7538 __m128i __B) 7539 { 7540 return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A, 7541 (__v2di) __B, 7542 (__v2di) __W, 7543 (__mmask8) __U); 7544 } 7545 7546 extern __inline __m128i 7547 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7548 _mm_maskz_xor_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 7549 { 7550 return (__m128i) __builtin_ia32_pxorq128_mask ((__v2di) __A, 7551 (__v2di) __B, 7552 (__v2di) 7553 _mm_setzero_si128 (), 7554 (__mmask8) __U); 7555 } 7556 7557 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 7558 _mm_xor_epi64 (__m128i __A, __m128i __B) 7559 { 7560 return (__m128i) ((__v2du)__A ^ (__v2du)__B); 7561 } 7562 7563 extern __inline __m256d 7564 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7565 _mm256_mask_max_pd (__m256d __W, __mmask8 __U, __m256d __A, 7566 __m256d __B) 7567 { 7568 return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, 7569 (__v4df) __B, 7570 (__v4df) __W, 7571 (__mmask8) __U); 7572 } 7573 7574 extern __inline __m256d 7575 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7576 _mm256_maskz_max_pd (__mmask8 __U, __m256d __A, __m256d __B) 7577 { 7578 return (__m256d) __builtin_ia32_maxpd256_mask ((__v4df) __A, 7579 (__v4df) __B, 7580 (__v4df) 7581 _mm256_setzero_pd (), 7582 (__mmask8) __U); 7583 } 7584 7585 extern __inline __m256 7586 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7587 _mm256_mask_max_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 7588 { 7589 return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, 7590 (__v8sf) __B, 7591 (__v8sf) __W, 7592 (__mmask8) __U); 7593 } 7594 7595 extern __inline __m256 7596 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7597 _mm256_maskz_max_ps (__mmask8 __U, __m256 __A, __m256 __B) 7598 { 7599 return (__m256) __builtin_ia32_maxps256_mask ((__v8sf) __A, 7600 (__v8sf) __B, 7601 (__v8sf) 7602 _mm256_setzero_ps (), 7603 (__mmask8) __U); 7604 } 7605 7606 extern __inline __m128 7607 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7608 _mm_mask_div_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 7609 { 7610 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, 7611 (__v4sf) __B, 7612 (__v4sf) __W, 7613 (__mmask8) __U); 7614 } 7615 7616 extern __inline __m128 7617 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7618 _mm_maskz_div_ps (__mmask8 __U, __m128 __A, __m128 __B) 7619 { 7620 return (__m128) __builtin_ia32_divps_mask ((__v4sf) __A, 7621 (__v4sf) __B, 7622 (__v4sf) 7623 _mm_setzero_ps (), 7624 (__mmask8) __U); 7625 } 7626 7627 extern __inline __m128d 7628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7629 _mm_mask_div_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 7630 { 7631 return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, 7632 (__v2df) __B, 7633 (__v2df) __W, 7634 (__mmask8) __U); 7635 } 7636 7637 extern __inline __m128d 7638 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7639 _mm_maskz_div_pd (__mmask8 __U, __m128d __A, __m128d __B) 7640 { 7641 return (__m128d) __builtin_ia32_divpd_mask ((__v2df) __A, 7642 (__v2df) __B, 7643 (__v2df) 7644 _mm_setzero_pd (), 7645 (__mmask8) __U); 7646 } 7647 7648 extern __inline __m256d 7649 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7650 _mm256_mask_min_pd (__m256d __W, __mmask8 __U, __m256d __A, 7651 __m256d __B) 7652 { 7653 return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, 7654 (__v4df) __B, 7655 (__v4df) __W, 7656 (__mmask8) __U); 7657 } 7658 7659 extern __inline __m256d 7660 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7661 _mm256_mask_div_pd (__m256d __W, __mmask8 __U, __m256d __A, 7662 __m256d __B) 7663 { 7664 return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, 7665 (__v4df) __B, 7666 (__v4df) __W, 7667 (__mmask8) __U); 7668 } 7669 7670 extern __inline __m256d 7671 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7672 _mm256_maskz_min_pd (__mmask8 __U, __m256d __A, __m256d __B) 7673 { 7674 return (__m256d) __builtin_ia32_minpd256_mask ((__v4df) __A, 7675 (__v4df) __B, 7676 (__v4df) 7677 _mm256_setzero_pd (), 7678 (__mmask8) __U); 7679 } 7680 7681 extern __inline __m256 7682 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7683 _mm256_mask_min_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 7684 { 7685 return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, 7686 (__v8sf) __B, 7687 (__v8sf) __W, 7688 (__mmask8) __U); 7689 } 7690 7691 extern __inline __m256d 7692 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7693 _mm256_maskz_div_pd (__mmask8 __U, __m256d __A, __m256d __B) 7694 { 7695 return (__m256d) __builtin_ia32_divpd256_mask ((__v4df) __A, 7696 (__v4df) __B, 7697 (__v4df) 7698 _mm256_setzero_pd (), 7699 (__mmask8) __U); 7700 } 7701 7702 extern __inline __m256 7703 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7704 _mm256_mask_div_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 7705 { 7706 return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, 7707 (__v8sf) __B, 7708 (__v8sf) __W, 7709 (__mmask8) __U); 7710 } 7711 7712 extern __inline __m256 7713 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7714 _mm256_maskz_min_ps (__mmask8 __U, __m256 __A, __m256 __B) 7715 { 7716 return (__m256) __builtin_ia32_minps256_mask ((__v8sf) __A, 7717 (__v8sf) __B, 7718 (__v8sf) 7719 _mm256_setzero_ps (), 7720 (__mmask8) __U); 7721 } 7722 7723 extern __inline __m256 7724 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7725 _mm256_maskz_div_ps (__mmask8 __U, __m256 __A, __m256 __B) 7726 { 7727 return (__m256) __builtin_ia32_divps256_mask ((__v8sf) __A, 7728 (__v8sf) __B, 7729 (__v8sf) 7730 _mm256_setzero_ps (), 7731 (__mmask8) __U); 7732 } 7733 7734 extern __inline __m128 7735 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7736 _mm_mask_min_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 7737 { 7738 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, 7739 (__v4sf) __B, 7740 (__v4sf) __W, 7741 (__mmask8) __U); 7742 } 7743 7744 extern __inline __m128 7745 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7746 _mm_mask_mul_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 7747 { 7748 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, 7749 (__v4sf) __B, 7750 (__v4sf) __W, 7751 (__mmask8) __U); 7752 } 7753 7754 extern __inline __m128 7755 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7756 _mm_maskz_min_ps (__mmask8 __U, __m128 __A, __m128 __B) 7757 { 7758 return (__m128) __builtin_ia32_minps_mask ((__v4sf) __A, 7759 (__v4sf) __B, 7760 (__v4sf) 7761 _mm_setzero_ps (), 7762 (__mmask8) __U); 7763 } 7764 7765 extern __inline __m128 7766 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7767 _mm_maskz_mul_ps (__mmask8 __U, __m128 __A, __m128 __B) 7768 { 7769 return (__m128) __builtin_ia32_mulps_mask ((__v4sf) __A, 7770 (__v4sf) __B, 7771 (__v4sf) 7772 _mm_setzero_ps (), 7773 (__mmask8) __U); 7774 } 7775 7776 extern __inline __m128 7777 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7778 _mm_mask_max_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 7779 { 7780 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, 7781 (__v4sf) __B, 7782 (__v4sf) __W, 7783 (__mmask8) __U); 7784 } 7785 7786 extern __inline __m128 7787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7788 _mm_maskz_max_ps (__mmask8 __U, __m128 __A, __m128 __B) 7789 { 7790 return (__m128) __builtin_ia32_maxps_mask ((__v4sf) __A, 7791 (__v4sf) __B, 7792 (__v4sf) 7793 _mm_setzero_ps (), 7794 (__mmask8) __U); 7795 } 7796 7797 extern __inline __m128d 7798 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7799 _mm_mask_min_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 7800 { 7801 return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, 7802 (__v2df) __B, 7803 (__v2df) __W, 7804 (__mmask8) __U); 7805 } 7806 7807 extern __inline __m128d 7808 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7809 _mm_maskz_min_pd (__mmask8 __U, __m128d __A, __m128d __B) 7810 { 7811 return (__m128d) __builtin_ia32_minpd_mask ((__v2df) __A, 7812 (__v2df) __B, 7813 (__v2df) 7814 _mm_setzero_pd (), 7815 (__mmask8) __U); 7816 } 7817 7818 extern __inline __m128d 7819 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7820 _mm_mask_max_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 7821 { 7822 return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, 7823 (__v2df) __B, 7824 (__v2df) __W, 7825 (__mmask8) __U); 7826 } 7827 7828 extern __inline __m128d 7829 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7830 _mm_maskz_max_pd (__mmask8 __U, __m128d __A, __m128d __B) 7831 { 7832 return (__m128d) __builtin_ia32_maxpd_mask ((__v2df) __A, 7833 (__v2df) __B, 7834 (__v2df) 7835 _mm_setzero_pd (), 7836 (__mmask8) __U); 7837 } 7838 7839 extern __inline __m128d 7840 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7841 _mm_mask_mul_pd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) 7842 { 7843 return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, 7844 (__v2df) __B, 7845 (__v2df) __W, 7846 (__mmask8) __U); 7847 } 7848 7849 extern __inline __m128d 7850 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7851 _mm_maskz_mul_pd (__mmask8 __U, __m128d __A, __m128d __B) 7852 { 7853 return (__m128d) __builtin_ia32_mulpd_mask ((__v2df) __A, 7854 (__v2df) __B, 7855 (__v2df) 7856 _mm_setzero_pd (), 7857 (__mmask8) __U); 7858 } 7859 7860 extern __inline __m256 7861 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7862 _mm256_mask_mul_ps (__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) 7863 { 7864 return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, 7865 (__v8sf) __B, 7866 (__v8sf) __W, 7867 (__mmask8) __U); 7868 } 7869 7870 extern __inline __m256 7871 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7872 _mm256_maskz_mul_ps (__mmask8 __U, __m256 __A, __m256 __B) 7873 { 7874 return (__m256) __builtin_ia32_mulps256_mask ((__v8sf) __A, 7875 (__v8sf) __B, 7876 (__v8sf) 7877 _mm256_setzero_ps (), 7878 (__mmask8) __U); 7879 } 7880 7881 extern __inline __m256d 7882 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7883 _mm256_mask_mul_pd (__m256d __W, __mmask8 __U, __m256d __A, 7884 __m256d __B) 7885 { 7886 return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, 7887 (__v4df) __B, 7888 (__v4df) __W, 7889 (__mmask8) __U); 7890 } 7891 7892 extern __inline __m256d 7893 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7894 _mm256_maskz_mul_pd (__mmask8 __U, __m256d __A, __m256d __B) 7895 { 7896 return (__m256d) __builtin_ia32_mulpd256_mask ((__v4df) __A, 7897 (__v4df) __B, 7898 (__v4df) 7899 _mm256_setzero_pd (), 7900 (__mmask8) __U); 7901 } 7902 7903 extern __inline __m256i 7904 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7905 _mm256_maskz_max_epi64 (__mmask8 __M, __m256i __A, __m256i __B) 7906 { 7907 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, 7908 (__v4di) __B, 7909 (__v4di) 7910 _mm256_setzero_si256 (), 7911 __M); 7912 } 7913 7914 extern __inline __m256i 7915 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7916 _mm256_mask_max_epi64 (__m256i __W, __mmask8 __M, __m256i __A, 7917 __m256i __B) 7918 { 7919 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, 7920 (__v4di) __B, 7921 (__v4di) __W, __M); 7922 } 7923 7924 extern __inline __m256i 7925 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7926 _mm256_min_epi64 (__m256i __A, __m256i __B) 7927 { 7928 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, 7929 (__v4di) __B, 7930 (__v4di) 7931 _mm256_setzero_si256 (), 7932 (__mmask8) -1); 7933 } 7934 7935 extern __inline __m256i 7936 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7937 _mm256_mask_min_epi64 (__m256i __W, __mmask8 __M, __m256i __A, 7938 __m256i __B) 7939 { 7940 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, 7941 (__v4di) __B, 7942 (__v4di) __W, __M); 7943 } 7944 7945 extern __inline __m256i 7946 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7947 _mm256_maskz_min_epi64 (__mmask8 __M, __m256i __A, __m256i __B) 7948 { 7949 return (__m256i) __builtin_ia32_pminsq256_mask ((__v4di) __A, 7950 (__v4di) __B, 7951 (__v4di) 7952 _mm256_setzero_si256 (), 7953 __M); 7954 } 7955 7956 extern __inline __m256i 7957 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7958 _mm256_maskz_max_epu64 (__mmask8 __M, __m256i __A, __m256i __B) 7959 { 7960 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, 7961 (__v4di) __B, 7962 (__v4di) 7963 _mm256_setzero_si256 (), 7964 __M); 7965 } 7966 7967 extern __inline __m256i 7968 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7969 _mm256_max_epi64 (__m256i __A, __m256i __B) 7970 { 7971 return (__m256i) __builtin_ia32_pmaxsq256_mask ((__v4di) __A, 7972 (__v4di) __B, 7973 (__v4di) 7974 _mm256_setzero_si256 (), 7975 (__mmask8) -1); 7976 } 7977 7978 extern __inline __m256i 7979 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7980 _mm256_max_epu64 (__m256i __A, __m256i __B) 7981 { 7982 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, 7983 (__v4di) __B, 7984 (__v4di) 7985 _mm256_setzero_si256 (), 7986 (__mmask8) -1); 7987 } 7988 7989 extern __inline __m256i 7990 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 7991 _mm256_mask_max_epu64 (__m256i __W, __mmask8 __M, __m256i __A, 7992 __m256i __B) 7993 { 7994 return (__m256i) __builtin_ia32_pmaxuq256_mask ((__v4di) __A, 7995 (__v4di) __B, 7996 (__v4di) __W, __M); 7997 } 7998 7999 extern __inline __m256i 8000 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8001 _mm256_min_epu64 (__m256i __A, __m256i __B) 8002 { 8003 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, 8004 (__v4di) __B, 8005 (__v4di) 8006 _mm256_setzero_si256 (), 8007 (__mmask8) -1); 8008 } 8009 8010 extern __inline __m256i 8011 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8012 _mm256_mask_min_epu64 (__m256i __W, __mmask8 __M, __m256i __A, 8013 __m256i __B) 8014 { 8015 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, 8016 (__v4di) __B, 8017 (__v4di) __W, __M); 8018 } 8019 8020 extern __inline __m256i 8021 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8022 _mm256_maskz_min_epu64 (__mmask8 __M, __m256i __A, __m256i __B) 8023 { 8024 return (__m256i) __builtin_ia32_pminuq256_mask ((__v4di) __A, 8025 (__v4di) __B, 8026 (__v4di) 8027 _mm256_setzero_si256 (), 8028 __M); 8029 } 8030 8031 extern __inline __m256i 8032 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8033 _mm256_maskz_max_epi32 (__mmask8 __M, __m256i __A, __m256i __B) 8034 { 8035 return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, 8036 (__v8si) __B, 8037 (__v8si) 8038 _mm256_setzero_si256 (), 8039 __M); 8040 } 8041 8042 extern __inline __m256i 8043 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8044 _mm256_mask_max_epi32 (__m256i __W, __mmask8 __M, __m256i __A, 8045 __m256i __B) 8046 { 8047 return (__m256i) __builtin_ia32_pmaxsd256_mask ((__v8si) __A, 8048 (__v8si) __B, 8049 (__v8si) __W, __M); 8050 } 8051 8052 extern __inline __m256i 8053 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8054 _mm256_maskz_min_epi32 (__mmask8 __M, __m256i __A, __m256i __B) 8055 { 8056 return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, 8057 (__v8si) __B, 8058 (__v8si) 8059 _mm256_setzero_si256 (), 8060 __M); 8061 } 8062 8063 extern __inline __m256i 8064 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8065 _mm256_mask_min_epi32 (__m256i __W, __mmask8 __M, __m256i __A, 8066 __m256i __B) 8067 { 8068 return (__m256i) __builtin_ia32_pminsd256_mask ((__v8si) __A, 8069 (__v8si) __B, 8070 (__v8si) __W, __M); 8071 } 8072 8073 extern __inline __m256i 8074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8075 _mm256_maskz_max_epu32 (__mmask8 __M, __m256i __A, __m256i __B) 8076 { 8077 return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, 8078 (__v8si) __B, 8079 (__v8si) 8080 _mm256_setzero_si256 (), 8081 __M); 8082 } 8083 8084 extern __inline __m256i 8085 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8086 _mm256_mask_max_epu32 (__m256i __W, __mmask8 __M, __m256i __A, 8087 __m256i __B) 8088 { 8089 return (__m256i) __builtin_ia32_pmaxud256_mask ((__v8si) __A, 8090 (__v8si) __B, 8091 (__v8si) __W, __M); 8092 } 8093 8094 extern __inline __m256i 8095 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8096 _mm256_maskz_min_epu32 (__mmask8 __M, __m256i __A, __m256i __B) 8097 { 8098 return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, 8099 (__v8si) __B, 8100 (__v8si) 8101 _mm256_setzero_si256 (), 8102 __M); 8103 } 8104 8105 extern __inline __m256i 8106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8107 _mm256_mask_min_epu32 (__m256i __W, __mmask8 __M, __m256i __A, 8108 __m256i __B) 8109 { 8110 return (__m256i) __builtin_ia32_pminud256_mask ((__v8si) __A, 8111 (__v8si) __B, 8112 (__v8si) __W, __M); 8113 } 8114 8115 extern __inline __m128i 8116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8117 _mm_maskz_max_epi64 (__mmask8 __M, __m128i __A, __m128i __B) 8118 { 8119 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, 8120 (__v2di) __B, 8121 (__v2di) 8122 _mm_setzero_si128 (), 8123 __M); 8124 } 8125 8126 extern __inline __m128i 8127 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8128 _mm_mask_max_epi64 (__m128i __W, __mmask8 __M, __m128i __A, 8129 __m128i __B) 8130 { 8131 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, 8132 (__v2di) __B, 8133 (__v2di) __W, __M); 8134 } 8135 8136 extern __inline __m128i 8137 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8138 _mm_min_epi64 (__m128i __A, __m128i __B) 8139 { 8140 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, 8141 (__v2di) __B, 8142 (__v2di) 8143 _mm_setzero_si128 (), 8144 (__mmask8) -1); 8145 } 8146 8147 extern __inline __m128i 8148 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8149 _mm_mask_min_epi64 (__m128i __W, __mmask8 __M, __m128i __A, 8150 __m128i __B) 8151 { 8152 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, 8153 (__v2di) __B, 8154 (__v2di) __W, __M); 8155 } 8156 8157 extern __inline __m128i 8158 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8159 _mm_maskz_min_epi64 (__mmask8 __M, __m128i __A, __m128i __B) 8160 { 8161 return (__m128i) __builtin_ia32_pminsq128_mask ((__v2di) __A, 8162 (__v2di) __B, 8163 (__v2di) 8164 _mm_setzero_si128 (), 8165 __M); 8166 } 8167 8168 extern __inline __m128i 8169 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8170 _mm_maskz_max_epu64 (__mmask8 __M, __m128i __A, __m128i __B) 8171 { 8172 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, 8173 (__v2di) __B, 8174 (__v2di) 8175 _mm_setzero_si128 (), 8176 __M); 8177 } 8178 8179 extern __inline __m128i 8180 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8181 _mm_max_epi64 (__m128i __A, __m128i __B) 8182 { 8183 return (__m128i) __builtin_ia32_pmaxsq128_mask ((__v2di) __A, 8184 (__v2di) __B, 8185 (__v2di) 8186 _mm_setzero_si128 (), 8187 (__mmask8) -1); 8188 } 8189 8190 extern __inline __m128i 8191 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8192 _mm_max_epu64 (__m128i __A, __m128i __B) 8193 { 8194 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, 8195 (__v2di) __B, 8196 (__v2di) 8197 _mm_setzero_si128 (), 8198 (__mmask8) -1); 8199 } 8200 8201 extern __inline __m128i 8202 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8203 _mm_mask_max_epu64 (__m128i __W, __mmask8 __M, __m128i __A, 8204 __m128i __B) 8205 { 8206 return (__m128i) __builtin_ia32_pmaxuq128_mask ((__v2di) __A, 8207 (__v2di) __B, 8208 (__v2di) __W, __M); 8209 } 8210 8211 extern __inline __m128i 8212 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8213 _mm_min_epu64 (__m128i __A, __m128i __B) 8214 { 8215 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, 8216 (__v2di) __B, 8217 (__v2di) 8218 _mm_setzero_si128 (), 8219 (__mmask8) -1); 8220 } 8221 8222 extern __inline __m128i 8223 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8224 _mm_mask_min_epu64 (__m128i __W, __mmask8 __M, __m128i __A, 8225 __m128i __B) 8226 { 8227 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, 8228 (__v2di) __B, 8229 (__v2di) __W, __M); 8230 } 8231 8232 extern __inline __m128i 8233 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8234 _mm_maskz_min_epu64 (__mmask8 __M, __m128i __A, __m128i __B) 8235 { 8236 return (__m128i) __builtin_ia32_pminuq128_mask ((__v2di) __A, 8237 (__v2di) __B, 8238 (__v2di) 8239 _mm_setzero_si128 (), 8240 __M); 8241 } 8242 8243 extern __inline __m128i 8244 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8245 _mm_maskz_max_epi32 (__mmask8 __M, __m128i __A, __m128i __B) 8246 { 8247 return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, 8248 (__v4si) __B, 8249 (__v4si) 8250 _mm_setzero_si128 (), 8251 __M); 8252 } 8253 8254 extern __inline __m128i 8255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8256 _mm_mask_max_epi32 (__m128i __W, __mmask8 __M, __m128i __A, 8257 __m128i __B) 8258 { 8259 return (__m128i) __builtin_ia32_pmaxsd128_mask ((__v4si) __A, 8260 (__v4si) __B, 8261 (__v4si) __W, __M); 8262 } 8263 8264 extern __inline __m128i 8265 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8266 _mm_maskz_min_epi32 (__mmask8 __M, __m128i __A, __m128i __B) 8267 { 8268 return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, 8269 (__v4si) __B, 8270 (__v4si) 8271 _mm_setzero_si128 (), 8272 __M); 8273 } 8274 8275 extern __inline __m128i 8276 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8277 _mm_mask_min_epi32 (__m128i __W, __mmask8 __M, __m128i __A, 8278 __m128i __B) 8279 { 8280 return (__m128i) __builtin_ia32_pminsd128_mask ((__v4si) __A, 8281 (__v4si) __B, 8282 (__v4si) __W, __M); 8283 } 8284 8285 extern __inline __m128i 8286 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8287 _mm_maskz_max_epu32 (__mmask8 __M, __m128i __A, __m128i __B) 8288 { 8289 return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, 8290 (__v4si) __B, 8291 (__v4si) 8292 _mm_setzero_si128 (), 8293 __M); 8294 } 8295 8296 extern __inline __m128i 8297 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8298 _mm_mask_max_epu32 (__m128i __W, __mmask8 __M, __m128i __A, 8299 __m128i __B) 8300 { 8301 return (__m128i) __builtin_ia32_pmaxud128_mask ((__v4si) __A, 8302 (__v4si) __B, 8303 (__v4si) __W, __M); 8304 } 8305 8306 extern __inline __m128i 8307 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8308 _mm_maskz_min_epu32 (__mmask8 __M, __m128i __A, __m128i __B) 8309 { 8310 return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, 8311 (__v4si) __B, 8312 (__v4si) 8313 _mm_setzero_si128 (), 8314 __M); 8315 } 8316 8317 extern __inline __m128i 8318 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8319 _mm_mask_min_epu32 (__m128i __W, __mmask8 __M, __m128i __A, 8320 __m128i __B) 8321 { 8322 return (__m128i) __builtin_ia32_pminud128_mask ((__v4si) __A, 8323 (__v4si) __B, 8324 (__v4si) __W, __M); 8325 } 8326 8327 #ifndef __AVX512CD__ 8328 #pragma GCC push_options 8329 #pragma GCC target("avx512vl,avx512cd") 8330 #define __DISABLE_AVX512VLCD__ 8331 #endif 8332 8333 extern __inline __m128i 8334 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8335 _mm_broadcastmb_epi64 (__mmask8 __A) 8336 { 8337 return (__m128i) __builtin_ia32_broadcastmb128 (__A); 8338 } 8339 8340 extern __inline __m256i 8341 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8342 _mm256_broadcastmb_epi64 (__mmask8 __A) 8343 { 8344 return (__m256i) __builtin_ia32_broadcastmb256 (__A); 8345 } 8346 8347 extern __inline __m128i 8348 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8349 _mm_broadcastmw_epi32 (__mmask16 __A) 8350 { 8351 return (__m128i) __builtin_ia32_broadcastmw128 (__A); 8352 } 8353 8354 extern __inline __m256i 8355 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8356 _mm256_broadcastmw_epi32 (__mmask16 __A) 8357 { 8358 return (__m256i) __builtin_ia32_broadcastmw256 (__A); 8359 } 8360 8361 extern __inline __m256i 8362 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8363 _mm256_lzcnt_epi32 (__m256i __A) 8364 { 8365 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A, 8366 (__v8si) 8367 _mm256_setzero_si256 (), 8368 (__mmask8) -1); 8369 } 8370 8371 extern __inline __m256i 8372 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8373 _mm256_mask_lzcnt_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 8374 { 8375 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A, 8376 (__v8si) __W, 8377 (__mmask8) __U); 8378 } 8379 8380 extern __inline __m256i 8381 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8382 _mm256_maskz_lzcnt_epi32 (__mmask8 __U, __m256i __A) 8383 { 8384 return (__m256i) __builtin_ia32_vplzcntd_256_mask ((__v8si) __A, 8385 (__v8si) 8386 _mm256_setzero_si256 (), 8387 (__mmask8) __U); 8388 } 8389 8390 extern __inline __m256i 8391 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8392 _mm256_lzcnt_epi64 (__m256i __A) 8393 { 8394 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A, 8395 (__v4di) 8396 _mm256_setzero_si256 (), 8397 (__mmask8) -1); 8398 } 8399 8400 extern __inline __m256i 8401 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8402 _mm256_mask_lzcnt_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 8403 { 8404 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A, 8405 (__v4di) __W, 8406 (__mmask8) __U); 8407 } 8408 8409 extern __inline __m256i 8410 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8411 _mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A) 8412 { 8413 return (__m256i) __builtin_ia32_vplzcntq_256_mask ((__v4di) __A, 8414 (__v4di) 8415 _mm256_setzero_si256 (), 8416 (__mmask8) __U); 8417 } 8418 8419 extern __inline __m256i 8420 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8421 _mm256_conflict_epi64 (__m256i __A) 8422 { 8423 return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A, 8424 (__v4di) 8425 _mm256_setzero_si256 (), 8426 (__mmask8) -1); 8427 } 8428 8429 extern __inline __m256i 8430 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8431 _mm256_mask_conflict_epi64 (__m256i __W, __mmask8 __U, __m256i __A) 8432 { 8433 return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A, 8434 (__v4di) __W, 8435 (__mmask8) 8436 __U); 8437 } 8438 8439 extern __inline __m256i 8440 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8441 _mm256_maskz_conflict_epi64 (__mmask8 __U, __m256i __A) 8442 { 8443 return (__m256i) __builtin_ia32_vpconflictdi_256_mask ((__v4di) __A, 8444 (__v4di) 8445 _mm256_setzero_si256 (), 8446 (__mmask8) 8447 __U); 8448 } 8449 8450 extern __inline __m256i 8451 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8452 _mm256_conflict_epi32 (__m256i __A) 8453 { 8454 return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A, 8455 (__v8si) 8456 _mm256_setzero_si256 (), 8457 (__mmask8) -1); 8458 } 8459 8460 extern __inline __m256i 8461 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8462 _mm256_mask_conflict_epi32 (__m256i __W, __mmask8 __U, __m256i __A) 8463 { 8464 return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A, 8465 (__v8si) __W, 8466 (__mmask8) 8467 __U); 8468 } 8469 8470 extern __inline __m256i 8471 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8472 _mm256_maskz_conflict_epi32 (__mmask8 __U, __m256i __A) 8473 { 8474 return (__m256i) __builtin_ia32_vpconflictsi_256_mask ((__v8si) __A, 8475 (__v8si) 8476 _mm256_setzero_si256 (), 8477 (__mmask8) 8478 __U); 8479 } 8480 8481 extern __inline __m128i 8482 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8483 _mm_lzcnt_epi32 (__m128i __A) 8484 { 8485 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A, 8486 (__v4si) 8487 _mm_setzero_si128 (), 8488 (__mmask8) -1); 8489 } 8490 8491 extern __inline __m128i 8492 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8493 _mm_mask_lzcnt_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 8494 { 8495 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A, 8496 (__v4si) __W, 8497 (__mmask8) __U); 8498 } 8499 8500 extern __inline __m128i 8501 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8502 _mm_maskz_lzcnt_epi32 (__mmask8 __U, __m128i __A) 8503 { 8504 return (__m128i) __builtin_ia32_vplzcntd_128_mask ((__v4si) __A, 8505 (__v4si) 8506 _mm_setzero_si128 (), 8507 (__mmask8) __U); 8508 } 8509 8510 extern __inline __m128i 8511 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8512 _mm_lzcnt_epi64 (__m128i __A) 8513 { 8514 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A, 8515 (__v2di) 8516 _mm_setzero_si128 (), 8517 (__mmask8) -1); 8518 } 8519 8520 extern __inline __m128i 8521 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8522 _mm_mask_lzcnt_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 8523 { 8524 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A, 8525 (__v2di) __W, 8526 (__mmask8) __U); 8527 } 8528 8529 extern __inline __m128i 8530 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8531 _mm_maskz_lzcnt_epi64 (__mmask8 __U, __m128i __A) 8532 { 8533 return (__m128i) __builtin_ia32_vplzcntq_128_mask ((__v2di) __A, 8534 (__v2di) 8535 _mm_setzero_si128 (), 8536 (__mmask8) __U); 8537 } 8538 8539 extern __inline __m128i 8540 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8541 _mm_conflict_epi64 (__m128i __A) 8542 { 8543 return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A, 8544 (__v2di) 8545 _mm_setzero_si128 (), 8546 (__mmask8) -1); 8547 } 8548 8549 extern __inline __m128i 8550 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8551 _mm_mask_conflict_epi64 (__m128i __W, __mmask8 __U, __m128i __A) 8552 { 8553 return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A, 8554 (__v2di) __W, 8555 (__mmask8) 8556 __U); 8557 } 8558 8559 extern __inline __m128i 8560 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8561 _mm_maskz_conflict_epi64 (__mmask8 __U, __m128i __A) 8562 { 8563 return (__m128i) __builtin_ia32_vpconflictdi_128_mask ((__v2di) __A, 8564 (__v2di) 8565 _mm_setzero_si128 (), 8566 (__mmask8) 8567 __U); 8568 } 8569 8570 extern __inline __m128i 8571 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8572 _mm_conflict_epi32 (__m128i __A) 8573 { 8574 return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A, 8575 (__v4si) 8576 _mm_setzero_si128 (), 8577 (__mmask8) -1); 8578 } 8579 8580 extern __inline __m128i 8581 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8582 _mm_mask_conflict_epi32 (__m128i __W, __mmask8 __U, __m128i __A) 8583 { 8584 return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A, 8585 (__v4si) __W, 8586 (__mmask8) 8587 __U); 8588 } 8589 8590 extern __inline __m128i 8591 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8592 _mm_maskz_conflict_epi32 (__mmask8 __U, __m128i __A) 8593 { 8594 return (__m128i) __builtin_ia32_vpconflictsi_128_mask ((__v4si) __A, 8595 (__v4si) 8596 _mm_setzero_si128 (), 8597 (__mmask8) 8598 __U); 8599 } 8600 8601 #ifdef __DISABLE_AVX512VLCD__ 8602 #pragma GCC pop_options 8603 #endif 8604 8605 extern __inline __m256d 8606 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8607 _mm256_mask_unpacklo_pd (__m256d __W, __mmask8 __U, __m256d __A, 8608 __m256d __B) 8609 { 8610 return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A, 8611 (__v4df) __B, 8612 (__v4df) __W, 8613 (__mmask8) __U); 8614 } 8615 8616 extern __inline __m256d 8617 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8618 _mm256_maskz_unpacklo_pd (__mmask8 __U, __m256d __A, __m256d __B) 8619 { 8620 return (__m256d) __builtin_ia32_unpcklpd256_mask ((__v4df) __A, 8621 (__v4df) __B, 8622 (__v4df) 8623 _mm256_setzero_pd (), 8624 (__mmask8) __U); 8625 } 8626 8627 extern __inline __m128d 8628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8629 _mm_mask_unpacklo_pd (__m128d __W, __mmask8 __U, __m128d __A, 8630 __m128d __B) 8631 { 8632 return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A, 8633 (__v2df) __B, 8634 (__v2df) __W, 8635 (__mmask8) __U); 8636 } 8637 8638 extern __inline __m128d 8639 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8640 _mm_maskz_unpacklo_pd (__mmask8 __U, __m128d __A, __m128d __B) 8641 { 8642 return (__m128d) __builtin_ia32_unpcklpd128_mask ((__v2df) __A, 8643 (__v2df) __B, 8644 (__v2df) 8645 _mm_setzero_pd (), 8646 (__mmask8) __U); 8647 } 8648 8649 extern __inline __m256 8650 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8651 _mm256_mask_unpacklo_ps (__m256 __W, __mmask8 __U, __m256 __A, 8652 __m256 __B) 8653 { 8654 return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A, 8655 (__v8sf) __B, 8656 (__v8sf) __W, 8657 (__mmask8) __U); 8658 } 8659 8660 extern __inline __m256d 8661 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8662 _mm256_mask_unpackhi_pd (__m256d __W, __mmask8 __U, __m256d __A, 8663 __m256d __B) 8664 { 8665 return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A, 8666 (__v4df) __B, 8667 (__v4df) __W, 8668 (__mmask8) __U); 8669 } 8670 8671 extern __inline __m256d 8672 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8673 _mm256_maskz_unpackhi_pd (__mmask8 __U, __m256d __A, __m256d __B) 8674 { 8675 return (__m256d) __builtin_ia32_unpckhpd256_mask ((__v4df) __A, 8676 (__v4df) __B, 8677 (__v4df) 8678 _mm256_setzero_pd (), 8679 (__mmask8) __U); 8680 } 8681 8682 extern __inline __m128d 8683 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8684 _mm_mask_unpackhi_pd (__m128d __W, __mmask8 __U, __m128d __A, 8685 __m128d __B) 8686 { 8687 return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A, 8688 (__v2df) __B, 8689 (__v2df) __W, 8690 (__mmask8) __U); 8691 } 8692 8693 extern __inline __m128d 8694 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8695 _mm_maskz_unpackhi_pd (__mmask8 __U, __m128d __A, __m128d __B) 8696 { 8697 return (__m128d) __builtin_ia32_unpckhpd128_mask ((__v2df) __A, 8698 (__v2df) __B, 8699 (__v2df) 8700 _mm_setzero_pd (), 8701 (__mmask8) __U); 8702 } 8703 8704 extern __inline __m256 8705 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8706 _mm256_mask_unpackhi_ps (__m256 __W, __mmask8 __U, __m256 __A, 8707 __m256 __B) 8708 { 8709 return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A, 8710 (__v8sf) __B, 8711 (__v8sf) __W, 8712 (__mmask8) __U); 8713 } 8714 8715 extern __inline __m256 8716 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8717 _mm256_maskz_unpackhi_ps (__mmask8 __U, __m256 __A, __m256 __B) 8718 { 8719 return (__m256) __builtin_ia32_unpckhps256_mask ((__v8sf) __A, 8720 (__v8sf) __B, 8721 (__v8sf) 8722 _mm256_setzero_ps (), 8723 (__mmask8) __U); 8724 } 8725 8726 extern __inline __m128 8727 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8728 _mm_mask_unpackhi_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 8729 { 8730 return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A, 8731 (__v4sf) __B, 8732 (__v4sf) __W, 8733 (__mmask8) __U); 8734 } 8735 8736 extern __inline __m128 8737 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8738 _mm_maskz_unpackhi_ps (__mmask8 __U, __m128 __A, __m128 __B) 8739 { 8740 return (__m128) __builtin_ia32_unpckhps128_mask ((__v4sf) __A, 8741 (__v4sf) __B, 8742 (__v4sf) 8743 _mm_setzero_ps (), 8744 (__mmask8) __U); 8745 } 8746 8747 extern __inline __m128 8748 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8749 _mm_mask_cvtph_ps (__m128 __W, __mmask8 __U, __m128i __A) 8750 { 8751 return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, 8752 (__v4sf) __W, 8753 (__mmask8) __U); 8754 } 8755 8756 extern __inline __m128 8757 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8758 _mm_maskz_cvtph_ps (__mmask8 __U, __m128i __A) 8759 { 8760 return (__m128) __builtin_ia32_vcvtph2ps_mask ((__v8hi) __A, 8761 (__v4sf) 8762 _mm_setzero_ps (), 8763 (__mmask8) __U); 8764 } 8765 8766 extern __inline __m256 8767 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8768 _mm256_maskz_unpacklo_ps (__mmask8 __U, __m256 __A, __m256 __B) 8769 { 8770 return (__m256) __builtin_ia32_unpcklps256_mask ((__v8sf) __A, 8771 (__v8sf) __B, 8772 (__v8sf) 8773 _mm256_setzero_ps (), 8774 (__mmask8) __U); 8775 } 8776 8777 extern __inline __m256 8778 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8779 _mm256_mask_cvtph_ps (__m256 __W, __mmask8 __U, __m128i __A) 8780 { 8781 return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, 8782 (__v8sf) __W, 8783 (__mmask8) __U); 8784 } 8785 8786 extern __inline __m256 8787 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8788 _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A) 8789 { 8790 return (__m256) __builtin_ia32_vcvtph2ps256_mask ((__v8hi) __A, 8791 (__v8sf) 8792 _mm256_setzero_ps (), 8793 (__mmask8) __U); 8794 } 8795 8796 extern __inline __m128 8797 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8798 _mm_mask_unpacklo_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) 8799 { 8800 return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A, 8801 (__v4sf) __B, 8802 (__v4sf) __W, 8803 (__mmask8) __U); 8804 } 8805 8806 extern __inline __m128 8807 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8808 _mm_maskz_unpacklo_ps (__mmask8 __U, __m128 __A, __m128 __B) 8809 { 8810 return (__m128) __builtin_ia32_unpcklps128_mask ((__v4sf) __A, 8811 (__v4sf) __B, 8812 (__v4sf) 8813 _mm_setzero_ps (), 8814 (__mmask8) __U); 8815 } 8816 8817 extern __inline __m256i 8818 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8819 _mm256_mask_sra_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 8820 __m128i __B) 8821 { 8822 return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, 8823 (__v4si) __B, 8824 (__v8si) __W, 8825 (__mmask8) __U); 8826 } 8827 8828 extern __inline __m256i 8829 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8830 _mm256_maskz_sra_epi32 (__mmask8 __U, __m256i __A, __m128i __B) 8831 { 8832 return (__m256i) __builtin_ia32_psrad256_mask ((__v8si) __A, 8833 (__v4si) __B, 8834 (__v8si) 8835 _mm256_setzero_si256 (), 8836 (__mmask8) __U); 8837 } 8838 8839 extern __inline __m128i 8840 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8841 _mm_mask_sra_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 8842 __m128i __B) 8843 { 8844 return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, 8845 (__v4si) __B, 8846 (__v4si) __W, 8847 (__mmask8) __U); 8848 } 8849 8850 extern __inline __m128i 8851 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8852 _mm_maskz_sra_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 8853 { 8854 return (__m128i) __builtin_ia32_psrad128_mask ((__v4si) __A, 8855 (__v4si) __B, 8856 (__v4si) 8857 _mm_setzero_si128 (), 8858 (__mmask8) __U); 8859 } 8860 8861 extern __inline __m256i 8862 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8863 _mm256_sra_epi64 (__m256i __A, __m128i __B) 8864 { 8865 return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, 8866 (__v2di) __B, 8867 (__v4di) 8868 _mm256_setzero_si256 (), 8869 (__mmask8) -1); 8870 } 8871 8872 extern __inline __m256i 8873 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8874 _mm256_mask_sra_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 8875 __m128i __B) 8876 { 8877 return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, 8878 (__v2di) __B, 8879 (__v4di) __W, 8880 (__mmask8) __U); 8881 } 8882 8883 extern __inline __m256i 8884 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8885 _mm256_maskz_sra_epi64 (__mmask8 __U, __m256i __A, __m128i __B) 8886 { 8887 return (__m256i) __builtin_ia32_psraq256_mask ((__v4di) __A, 8888 (__v2di) __B, 8889 (__v4di) 8890 _mm256_setzero_si256 (), 8891 (__mmask8) __U); 8892 } 8893 8894 extern __inline __m128i 8895 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8896 _mm_sra_epi64 (__m128i __A, __m128i __B) 8897 { 8898 return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, 8899 (__v2di) __B, 8900 (__v2di) 8901 _mm_setzero_si128 (), 8902 (__mmask8) -1); 8903 } 8904 8905 extern __inline __m128i 8906 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8907 _mm_mask_sra_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 8908 __m128i __B) 8909 { 8910 return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, 8911 (__v2di) __B, 8912 (__v2di) __W, 8913 (__mmask8) __U); 8914 } 8915 8916 extern __inline __m128i 8917 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8918 _mm_maskz_sra_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 8919 { 8920 return (__m128i) __builtin_ia32_psraq128_mask ((__v2di) __A, 8921 (__v2di) __B, 8922 (__v2di) 8923 _mm_setzero_si128 (), 8924 (__mmask8) __U); 8925 } 8926 8927 extern __inline __m128i 8928 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8929 _mm_mask_sll_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 8930 __m128i __B) 8931 { 8932 return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, 8933 (__v4si) __B, 8934 (__v4si) __W, 8935 (__mmask8) __U); 8936 } 8937 8938 extern __inline __m128i 8939 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8940 _mm_maskz_sll_epi32 (__mmask8 __U, __m128i __A, __m128i __B) 8941 { 8942 return (__m128i) __builtin_ia32_pslld128_mask ((__v4si) __A, 8943 (__v4si) __B, 8944 (__v4si) 8945 _mm_setzero_si128 (), 8946 (__mmask8) __U); 8947 } 8948 8949 extern __inline __m128i 8950 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8951 _mm_mask_sll_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 8952 __m128i __B) 8953 { 8954 return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, 8955 (__v2di) __B, 8956 (__v2di) __W, 8957 (__mmask8) __U); 8958 } 8959 8960 extern __inline __m128i 8961 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8962 _mm_maskz_sll_epi64 (__mmask8 __U, __m128i __A, __m128i __B) 8963 { 8964 return (__m128i) __builtin_ia32_psllq128_mask ((__v2di) __A, 8965 (__v2di) __B, 8966 (__v2di) 8967 _mm_setzero_si128 (), 8968 (__mmask8) __U); 8969 } 8970 8971 extern __inline __m256i 8972 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8973 _mm256_mask_sll_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 8974 __m128i __B) 8975 { 8976 return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, 8977 (__v4si) __B, 8978 (__v8si) __W, 8979 (__mmask8) __U); 8980 } 8981 8982 extern __inline __m256i 8983 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8984 _mm256_maskz_sll_epi32 (__mmask8 __U, __m256i __A, __m128i __B) 8985 { 8986 return (__m256i) __builtin_ia32_pslld256_mask ((__v8si) __A, 8987 (__v4si) __B, 8988 (__v8si) 8989 _mm256_setzero_si256 (), 8990 (__mmask8) __U); 8991 } 8992 8993 extern __inline __m256i 8994 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 8995 _mm256_mask_sll_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 8996 __m128i __B) 8997 { 8998 return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, 8999 (__v2di) __B, 9000 (__v4di) __W, 9001 (__mmask8) __U); 9002 } 9003 9004 extern __inline __m256i 9005 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9006 _mm256_maskz_sll_epi64 (__mmask8 __U, __m256i __A, __m128i __B) 9007 { 9008 return (__m256i) __builtin_ia32_psllq256_mask ((__v4di) __A, 9009 (__v2di) __B, 9010 (__v4di) 9011 _mm256_setzero_si256 (), 9012 (__mmask8) __U); 9013 } 9014 9015 extern __inline __m256 9016 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9017 _mm256_mask_permutexvar_ps (__m256 __W, __mmask8 __U, __m256i __X, 9018 __m256 __Y) 9019 { 9020 return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, 9021 (__v8si) __X, 9022 (__v8sf) __W, 9023 (__mmask8) __U); 9024 } 9025 9026 extern __inline __m256 9027 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9028 _mm256_maskz_permutexvar_ps (__mmask8 __U, __m256i __X, __m256 __Y) 9029 { 9030 return (__m256) __builtin_ia32_permvarsf256_mask ((__v8sf) __Y, 9031 (__v8si) __X, 9032 (__v8sf) 9033 _mm256_setzero_ps (), 9034 (__mmask8) __U); 9035 } 9036 9037 extern __inline __m256d 9038 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9039 _mm256_permutexvar_pd (__m256i __X, __m256d __Y) 9040 { 9041 return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, 9042 (__v4di) __X, 9043 (__v4df) 9044 _mm256_setzero_pd (), 9045 (__mmask8) -1); 9046 } 9047 9048 extern __inline __m256d 9049 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9050 _mm256_mask_permutexvar_pd (__m256d __W, __mmask8 __U, __m256i __X, 9051 __m256d __Y) 9052 { 9053 return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, 9054 (__v4di) __X, 9055 (__v4df) __W, 9056 (__mmask8) __U); 9057 } 9058 9059 extern __inline __m256d 9060 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9061 _mm256_maskz_permutexvar_pd (__mmask8 __U, __m256i __X, __m256d __Y) 9062 { 9063 return (__m256d) __builtin_ia32_permvardf256_mask ((__v4df) __Y, 9064 (__v4di) __X, 9065 (__v4df) 9066 _mm256_setzero_pd (), 9067 (__mmask8) __U); 9068 } 9069 9070 extern __inline __m256d 9071 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9072 _mm256_mask_permutevar_pd (__m256d __W, __mmask8 __U, __m256d __A, 9073 __m256i __C) 9074 { 9075 return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, 9076 (__v4di) __C, 9077 (__v4df) __W, 9078 (__mmask8) 9079 __U); 9080 } 9081 9082 extern __inline __m256d 9083 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9084 _mm256_maskz_permutevar_pd (__mmask8 __U, __m256d __A, __m256i __C) 9085 { 9086 return (__m256d) __builtin_ia32_vpermilvarpd256_mask ((__v4df) __A, 9087 (__v4di) __C, 9088 (__v4df) 9089 _mm256_setzero_pd (), 9090 (__mmask8) 9091 __U); 9092 } 9093 9094 extern __inline __m256 9095 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9096 _mm256_mask_permutevar_ps (__m256 __W, __mmask8 __U, __m256 __A, 9097 __m256i __C) 9098 { 9099 return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, 9100 (__v8si) __C, 9101 (__v8sf) __W, 9102 (__mmask8) __U); 9103 } 9104 9105 extern __inline __m256 9106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9107 _mm256_maskz_permutevar_ps (__mmask8 __U, __m256 __A, __m256i __C) 9108 { 9109 return (__m256) __builtin_ia32_vpermilvarps256_mask ((__v8sf) __A, 9110 (__v8si) __C, 9111 (__v8sf) 9112 _mm256_setzero_ps (), 9113 (__mmask8) __U); 9114 } 9115 9116 extern __inline __m128d 9117 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9118 _mm_mask_permutevar_pd (__m128d __W, __mmask8 __U, __m128d __A, 9119 __m128i __C) 9120 { 9121 return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, 9122 (__v2di) __C, 9123 (__v2df) __W, 9124 (__mmask8) __U); 9125 } 9126 9127 extern __inline __m128d 9128 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9129 _mm_maskz_permutevar_pd (__mmask8 __U, __m128d __A, __m128i __C) 9130 { 9131 return (__m128d) __builtin_ia32_vpermilvarpd_mask ((__v2df) __A, 9132 (__v2di) __C, 9133 (__v2df) 9134 _mm_setzero_pd (), 9135 (__mmask8) __U); 9136 } 9137 9138 extern __inline __m128 9139 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9140 _mm_mask_permutevar_ps (__m128 __W, __mmask8 __U, __m128 __A, 9141 __m128i __C) 9142 { 9143 return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, 9144 (__v4si) __C, 9145 (__v4sf) __W, 9146 (__mmask8) __U); 9147 } 9148 9149 extern __inline __m128 9150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9151 _mm_maskz_permutevar_ps (__mmask8 __U, __m128 __A, __m128i __C) 9152 { 9153 return (__m128) __builtin_ia32_vpermilvarps_mask ((__v4sf) __A, 9154 (__v4si) __C, 9155 (__v4sf) 9156 _mm_setzero_ps (), 9157 (__mmask8) __U); 9158 } 9159 9160 extern __inline __m256i 9161 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9162 _mm256_maskz_mullo_epi32 (__mmask8 __M, __m256i __A, __m256i __B) 9163 { 9164 return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A, 9165 (__v8si) __B, 9166 (__v8si) 9167 _mm256_setzero_si256 (), 9168 __M); 9169 } 9170 9171 extern __inline __m256i 9172 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9173 _mm256_maskz_permutexvar_epi64 (__mmask8 __M, __m256i __X, __m256i __Y) 9174 { 9175 return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, 9176 (__v4di) __X, 9177 (__v4di) 9178 _mm256_setzero_si256 (), 9179 __M); 9180 } 9181 9182 extern __inline __m256i 9183 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9184 _mm256_mask_mullo_epi32 (__m256i __W, __mmask8 __M, __m256i __A, 9185 __m256i __B) 9186 { 9187 return (__m256i) __builtin_ia32_pmulld256_mask ((__v8si) __A, 9188 (__v8si) __B, 9189 (__v8si) __W, __M); 9190 } 9191 9192 extern __inline __m128i 9193 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9194 _mm_maskz_mullo_epi32 (__mmask8 __M, __m128i __A, __m128i __B) 9195 { 9196 return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A, 9197 (__v4si) __B, 9198 (__v4si) 9199 _mm_setzero_si128 (), 9200 __M); 9201 } 9202 9203 extern __inline __m128i 9204 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9205 _mm_mask_mullo_epi32 (__m128i __W, __mmask8 __M, __m128i __A, 9206 __m128i __B) 9207 { 9208 return (__m128i) __builtin_ia32_pmulld128_mask ((__v4si) __A, 9209 (__v4si) __B, 9210 (__v4si) __W, __M); 9211 } 9212 9213 extern __inline __m256i 9214 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9215 _mm256_mask_mul_epi32 (__m256i __W, __mmask8 __M, __m256i __X, 9216 __m256i __Y) 9217 { 9218 return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, 9219 (__v8si) __Y, 9220 (__v4di) __W, __M); 9221 } 9222 9223 extern __inline __m256i 9224 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9225 _mm256_maskz_mul_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) 9226 { 9227 return (__m256i) __builtin_ia32_pmuldq256_mask ((__v8si) __X, 9228 (__v8si) __Y, 9229 (__v4di) 9230 _mm256_setzero_si256 (), 9231 __M); 9232 } 9233 9234 extern __inline __m128i 9235 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9236 _mm_mask_mul_epi32 (__m128i __W, __mmask8 __M, __m128i __X, 9237 __m128i __Y) 9238 { 9239 return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, 9240 (__v4si) __Y, 9241 (__v2di) __W, __M); 9242 } 9243 9244 extern __inline __m128i 9245 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9246 _mm_maskz_mul_epi32 (__mmask8 __M, __m128i __X, __m128i __Y) 9247 { 9248 return (__m128i) __builtin_ia32_pmuldq128_mask ((__v4si) __X, 9249 (__v4si) __Y, 9250 (__v2di) 9251 _mm_setzero_si128 (), 9252 __M); 9253 } 9254 9255 extern __inline __m256i 9256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9257 _mm256_permutexvar_epi64 (__m256i __X, __m256i __Y) 9258 { 9259 return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, 9260 (__v4di) __X, 9261 (__v4di) 9262 _mm256_setzero_si256 (), 9263 (__mmask8) -1); 9264 } 9265 9266 extern __inline __m256i 9267 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9268 _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X, 9269 __m256i __Y) 9270 { 9271 return (__m256i) __builtin_ia32_permvardi256_mask ((__v4di) __Y, 9272 (__v4di) __X, 9273 (__v4di) __W, 9274 __M); 9275 } 9276 9277 extern __inline __m256i 9278 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9279 _mm256_mask_mul_epu32 (__m256i __W, __mmask8 __M, __m256i __X, 9280 __m256i __Y) 9281 { 9282 return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, 9283 (__v8si) __Y, 9284 (__v4di) __W, __M); 9285 } 9286 9287 extern __inline __m256i 9288 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9289 _mm256_maskz_permutexvar_epi32 (__mmask8 __M, __m256i __X, __m256i __Y) 9290 { 9291 return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, 9292 (__v8si) __X, 9293 (__v8si) 9294 _mm256_setzero_si256 (), 9295 __M); 9296 } 9297 9298 extern __inline __m256i 9299 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9300 _mm256_maskz_mul_epu32 (__mmask8 __M, __m256i __X, __m256i __Y) 9301 { 9302 return (__m256i) __builtin_ia32_pmuludq256_mask ((__v8si) __X, 9303 (__v8si) __Y, 9304 (__v4di) 9305 _mm256_setzero_si256 (), 9306 __M); 9307 } 9308 9309 extern __inline __m128i 9310 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9311 _mm_mask_mul_epu32 (__m128i __W, __mmask8 __M, __m128i __X, 9312 __m128i __Y) 9313 { 9314 return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, 9315 (__v4si) __Y, 9316 (__v2di) __W, __M); 9317 } 9318 9319 extern __inline __m128i 9320 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9321 _mm_maskz_mul_epu32 (__mmask8 __M, __m128i __X, __m128i __Y) 9322 { 9323 return (__m128i) __builtin_ia32_pmuludq128_mask ((__v4si) __X, 9324 (__v4si) __Y, 9325 (__v2di) 9326 _mm_setzero_si128 (), 9327 __M); 9328 } 9329 9330 extern __inline __m256i 9331 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9332 _mm256_permutexvar_epi32 (__m256i __X, __m256i __Y) 9333 { 9334 return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, 9335 (__v8si) __X, 9336 (__v8si) 9337 _mm256_setzero_si256 (), 9338 (__mmask8) -1); 9339 } 9340 9341 extern __inline __m256i 9342 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9343 _mm256_mask_permutexvar_epi32 (__m256i __W, __mmask8 __M, __m256i __X, 9344 __m256i __Y) 9345 { 9346 return (__m256i) __builtin_ia32_permvarsi256_mask ((__v8si) __Y, 9347 (__v8si) __X, 9348 (__v8si) __W, 9349 __M); 9350 } 9351 9352 extern __inline __mmask8 9353 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9354 _mm256_mask_cmpneq_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9355 { 9356 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9357 (__v8si) __Y, 4, 9358 (__mmask8) __M); 9359 } 9360 9361 extern __inline __mmask8 9362 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9363 _mm256_cmpneq_epu32_mask (__m256i __X, __m256i __Y) 9364 { 9365 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9366 (__v8si) __Y, 4, 9367 (__mmask8) -1); 9368 } 9369 9370 extern __inline __mmask8 9371 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9372 _mm256_mask_cmplt_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9373 { 9374 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9375 (__v8si) __Y, 1, 9376 (__mmask8) __M); 9377 } 9378 9379 extern __inline __mmask8 9380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9381 _mm256_cmplt_epu32_mask (__m256i __X, __m256i __Y) 9382 { 9383 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9384 (__v8si) __Y, 1, 9385 (__mmask8) -1); 9386 } 9387 9388 extern __inline __mmask8 9389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9390 _mm256_mask_cmpge_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9391 { 9392 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9393 (__v8si) __Y, 5, 9394 (__mmask8) __M); 9395 } 9396 9397 extern __inline __mmask8 9398 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9399 _mm256_cmpge_epu32_mask (__m256i __X, __m256i __Y) 9400 { 9401 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9402 (__v8si) __Y, 5, 9403 (__mmask8) -1); 9404 } 9405 9406 extern __inline __mmask8 9407 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9408 _mm256_mask_cmple_epu32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9409 { 9410 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9411 (__v8si) __Y, 2, 9412 (__mmask8) __M); 9413 } 9414 9415 extern __inline __mmask8 9416 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9417 _mm256_cmple_epu32_mask (__m256i __X, __m256i __Y) 9418 { 9419 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 9420 (__v8si) __Y, 2, 9421 (__mmask8) -1); 9422 } 9423 9424 extern __inline __mmask8 9425 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9426 _mm256_mask_cmpneq_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9427 { 9428 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9429 (__v4di) __Y, 4, 9430 (__mmask8) __M); 9431 } 9432 9433 extern __inline __mmask8 9434 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9435 _mm256_cmpneq_epu64_mask (__m256i __X, __m256i __Y) 9436 { 9437 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9438 (__v4di) __Y, 4, 9439 (__mmask8) -1); 9440 } 9441 9442 extern __inline __mmask8 9443 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9444 _mm256_mask_cmplt_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9445 { 9446 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9447 (__v4di) __Y, 1, 9448 (__mmask8) __M); 9449 } 9450 9451 extern __inline __mmask8 9452 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9453 _mm256_cmplt_epu64_mask (__m256i __X, __m256i __Y) 9454 { 9455 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9456 (__v4di) __Y, 1, 9457 (__mmask8) -1); 9458 } 9459 9460 extern __inline __mmask8 9461 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9462 _mm256_mask_cmpge_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9463 { 9464 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9465 (__v4di) __Y, 5, 9466 (__mmask8) __M); 9467 } 9468 9469 extern __inline __mmask8 9470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9471 _mm256_cmpge_epu64_mask (__m256i __X, __m256i __Y) 9472 { 9473 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9474 (__v4di) __Y, 5, 9475 (__mmask8) -1); 9476 } 9477 9478 extern __inline __mmask8 9479 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9480 _mm256_mask_cmple_epu64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9481 { 9482 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9483 (__v4di) __Y, 2, 9484 (__mmask8) __M); 9485 } 9486 9487 extern __inline __mmask8 9488 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9489 _mm256_cmple_epu64_mask (__m256i __X, __m256i __Y) 9490 { 9491 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 9492 (__v4di) __Y, 2, 9493 (__mmask8) -1); 9494 } 9495 9496 extern __inline __mmask8 9497 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9498 _mm256_mask_cmpneq_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9499 { 9500 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9501 (__v8si) __Y, 4, 9502 (__mmask8) __M); 9503 } 9504 9505 extern __inline __mmask8 9506 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9507 _mm256_cmpneq_epi32_mask (__m256i __X, __m256i __Y) 9508 { 9509 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9510 (__v8si) __Y, 4, 9511 (__mmask8) -1); 9512 } 9513 9514 extern __inline __mmask8 9515 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9516 _mm256_mask_cmplt_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9517 { 9518 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9519 (__v8si) __Y, 1, 9520 (__mmask8) __M); 9521 } 9522 9523 extern __inline __mmask8 9524 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9525 _mm256_cmplt_epi32_mask (__m256i __X, __m256i __Y) 9526 { 9527 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9528 (__v8si) __Y, 1, 9529 (__mmask8) -1); 9530 } 9531 9532 extern __inline __mmask8 9533 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9534 _mm256_mask_cmpge_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9535 { 9536 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9537 (__v8si) __Y, 5, 9538 (__mmask8) __M); 9539 } 9540 9541 extern __inline __mmask8 9542 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9543 _mm256_cmpge_epi32_mask (__m256i __X, __m256i __Y) 9544 { 9545 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9546 (__v8si) __Y, 5, 9547 (__mmask8) -1); 9548 } 9549 9550 extern __inline __mmask8 9551 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9552 _mm256_mask_cmple_epi32_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9553 { 9554 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9555 (__v8si) __Y, 2, 9556 (__mmask8) __M); 9557 } 9558 9559 extern __inline __mmask8 9560 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9561 _mm256_cmple_epi32_mask (__m256i __X, __m256i __Y) 9562 { 9563 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 9564 (__v8si) __Y, 2, 9565 (__mmask8) -1); 9566 } 9567 9568 extern __inline __mmask8 9569 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9570 _mm256_mask_cmpneq_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9571 { 9572 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9573 (__v4di) __Y, 4, 9574 (__mmask8) __M); 9575 } 9576 9577 extern __inline __mmask8 9578 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9579 _mm256_cmpneq_epi64_mask (__m256i __X, __m256i __Y) 9580 { 9581 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9582 (__v4di) __Y, 4, 9583 (__mmask8) -1); 9584 } 9585 9586 extern __inline __mmask8 9587 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9588 _mm256_mask_cmplt_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9589 { 9590 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9591 (__v4di) __Y, 1, 9592 (__mmask8) __M); 9593 } 9594 9595 extern __inline __mmask8 9596 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9597 _mm256_cmplt_epi64_mask (__m256i __X, __m256i __Y) 9598 { 9599 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9600 (__v4di) __Y, 1, 9601 (__mmask8) -1); 9602 } 9603 9604 extern __inline __mmask8 9605 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9606 _mm256_mask_cmpge_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9607 { 9608 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9609 (__v4di) __Y, 5, 9610 (__mmask8) __M); 9611 } 9612 9613 extern __inline __mmask8 9614 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9615 _mm256_cmpge_epi64_mask (__m256i __X, __m256i __Y) 9616 { 9617 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9618 (__v4di) __Y, 5, 9619 (__mmask8) -1); 9620 } 9621 9622 extern __inline __mmask8 9623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9624 _mm256_mask_cmple_epi64_mask (__mmask8 __M, __m256i __X, __m256i __Y) 9625 { 9626 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9627 (__v4di) __Y, 2, 9628 (__mmask8) __M); 9629 } 9630 9631 extern __inline __mmask8 9632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9633 _mm256_cmple_epi64_mask (__m256i __X, __m256i __Y) 9634 { 9635 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 9636 (__v4di) __Y, 2, 9637 (__mmask8) -1); 9638 } 9639 9640 extern __inline __mmask8 9641 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9642 _mm_mask_cmpneq_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9643 { 9644 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9645 (__v4si) __Y, 4, 9646 (__mmask8) __M); 9647 } 9648 9649 extern __inline __mmask8 9650 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9651 _mm_cmpneq_epu32_mask (__m128i __X, __m128i __Y) 9652 { 9653 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9654 (__v4si) __Y, 4, 9655 (__mmask8) -1); 9656 } 9657 9658 extern __inline __mmask8 9659 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9660 _mm_mask_cmplt_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9661 { 9662 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9663 (__v4si) __Y, 1, 9664 (__mmask8) __M); 9665 } 9666 9667 extern __inline __mmask8 9668 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9669 _mm_cmplt_epu32_mask (__m128i __X, __m128i __Y) 9670 { 9671 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9672 (__v4si) __Y, 1, 9673 (__mmask8) -1); 9674 } 9675 9676 extern __inline __mmask8 9677 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9678 _mm_mask_cmpge_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9679 { 9680 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9681 (__v4si) __Y, 5, 9682 (__mmask8) __M); 9683 } 9684 9685 extern __inline __mmask8 9686 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9687 _mm_cmpge_epu32_mask (__m128i __X, __m128i __Y) 9688 { 9689 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9690 (__v4si) __Y, 5, 9691 (__mmask8) -1); 9692 } 9693 9694 extern __inline __mmask8 9695 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9696 _mm_mask_cmple_epu32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9697 { 9698 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9699 (__v4si) __Y, 2, 9700 (__mmask8) __M); 9701 } 9702 9703 extern __inline __mmask8 9704 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9705 _mm_cmple_epu32_mask (__m128i __X, __m128i __Y) 9706 { 9707 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 9708 (__v4si) __Y, 2, 9709 (__mmask8) -1); 9710 } 9711 9712 extern __inline __mmask8 9713 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9714 _mm_mask_cmpneq_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9715 { 9716 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9717 (__v2di) __Y, 4, 9718 (__mmask8) __M); 9719 } 9720 9721 extern __inline __mmask8 9722 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9723 _mm_cmpneq_epu64_mask (__m128i __X, __m128i __Y) 9724 { 9725 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9726 (__v2di) __Y, 4, 9727 (__mmask8) -1); 9728 } 9729 9730 extern __inline __mmask8 9731 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9732 _mm_mask_cmplt_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9733 { 9734 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9735 (__v2di) __Y, 1, 9736 (__mmask8) __M); 9737 } 9738 9739 extern __inline __mmask8 9740 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9741 _mm_cmplt_epu64_mask (__m128i __X, __m128i __Y) 9742 { 9743 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9744 (__v2di) __Y, 1, 9745 (__mmask8) -1); 9746 } 9747 9748 extern __inline __mmask8 9749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9750 _mm_mask_cmpge_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9751 { 9752 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9753 (__v2di) __Y, 5, 9754 (__mmask8) __M); 9755 } 9756 9757 extern __inline __mmask8 9758 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9759 _mm_cmpge_epu64_mask (__m128i __X, __m128i __Y) 9760 { 9761 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9762 (__v2di) __Y, 5, 9763 (__mmask8) -1); 9764 } 9765 9766 extern __inline __mmask8 9767 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9768 _mm_mask_cmple_epu64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9769 { 9770 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9771 (__v2di) __Y, 2, 9772 (__mmask8) __M); 9773 } 9774 9775 extern __inline __mmask8 9776 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9777 _mm_cmple_epu64_mask (__m128i __X, __m128i __Y) 9778 { 9779 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 9780 (__v2di) __Y, 2, 9781 (__mmask8) -1); 9782 } 9783 9784 extern __inline __mmask8 9785 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9786 _mm_mask_cmpneq_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9787 { 9788 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9789 (__v4si) __Y, 4, 9790 (__mmask8) __M); 9791 } 9792 9793 extern __inline __mmask8 9794 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9795 _mm_cmpneq_epi32_mask (__m128i __X, __m128i __Y) 9796 { 9797 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9798 (__v4si) __Y, 4, 9799 (__mmask8) -1); 9800 } 9801 9802 extern __inline __mmask8 9803 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9804 _mm_mask_cmplt_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9805 { 9806 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9807 (__v4si) __Y, 1, 9808 (__mmask8) __M); 9809 } 9810 9811 extern __inline __mmask8 9812 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9813 _mm_cmplt_epi32_mask (__m128i __X, __m128i __Y) 9814 { 9815 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9816 (__v4si) __Y, 1, 9817 (__mmask8) -1); 9818 } 9819 9820 extern __inline __mmask8 9821 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9822 _mm_mask_cmpge_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9823 { 9824 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9825 (__v4si) __Y, 5, 9826 (__mmask8) __M); 9827 } 9828 9829 extern __inline __mmask8 9830 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9831 _mm_cmpge_epi32_mask (__m128i __X, __m128i __Y) 9832 { 9833 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9834 (__v4si) __Y, 5, 9835 (__mmask8) -1); 9836 } 9837 9838 extern __inline __mmask8 9839 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9840 _mm_mask_cmple_epi32_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9841 { 9842 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9843 (__v4si) __Y, 2, 9844 (__mmask8) __M); 9845 } 9846 9847 extern __inline __mmask8 9848 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9849 _mm_cmple_epi32_mask (__m128i __X, __m128i __Y) 9850 { 9851 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 9852 (__v4si) __Y, 2, 9853 (__mmask8) -1); 9854 } 9855 9856 extern __inline __mmask8 9857 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9858 _mm_mask_cmpneq_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9859 { 9860 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9861 (__v2di) __Y, 4, 9862 (__mmask8) __M); 9863 } 9864 9865 extern __inline __mmask8 9866 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9867 _mm_cmpneq_epi64_mask (__m128i __X, __m128i __Y) 9868 { 9869 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9870 (__v2di) __Y, 4, 9871 (__mmask8) -1); 9872 } 9873 9874 extern __inline __mmask8 9875 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9876 _mm_mask_cmplt_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9877 { 9878 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9879 (__v2di) __Y, 1, 9880 (__mmask8) __M); 9881 } 9882 9883 extern __inline __mmask8 9884 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9885 _mm_cmplt_epi64_mask (__m128i __X, __m128i __Y) 9886 { 9887 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9888 (__v2di) __Y, 1, 9889 (__mmask8) -1); 9890 } 9891 9892 extern __inline __mmask8 9893 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9894 _mm_mask_cmpge_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9895 { 9896 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9897 (__v2di) __Y, 5, 9898 (__mmask8) __M); 9899 } 9900 9901 extern __inline __mmask8 9902 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9903 _mm_cmpge_epi64_mask (__m128i __X, __m128i __Y) 9904 { 9905 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9906 (__v2di) __Y, 5, 9907 (__mmask8) -1); 9908 } 9909 9910 extern __inline __mmask8 9911 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9912 _mm_mask_cmple_epi64_mask (__mmask8 __M, __m128i __X, __m128i __Y) 9913 { 9914 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9915 (__v2di) __Y, 2, 9916 (__mmask8) __M); 9917 } 9918 9919 extern __inline __mmask8 9920 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9921 _mm_cmple_epi64_mask (__m128i __X, __m128i __Y) 9922 { 9923 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 9924 (__v2di) __Y, 2, 9925 (__mmask8) -1); 9926 } 9927 9928 #ifdef __OPTIMIZE__ 9929 extern __inline __m256i 9930 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9931 _mm256_permutex_epi64 (__m256i __X, const int __I) 9932 { 9933 return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X, 9934 __I, 9935 (__v4di) 9936 _mm256_setzero_si256(), 9937 (__mmask8) -1); 9938 } 9939 9940 extern __inline __m256i 9941 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9942 _mm256_mask_permutex_epi64 (__m256i __W, __mmask8 __M, 9943 __m256i __X, const int __I) 9944 { 9945 return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X, 9946 __I, 9947 (__v4di) __W, 9948 (__mmask8) __M); 9949 } 9950 9951 extern __inline __m256i 9952 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9953 _mm256_maskz_permutex_epi64 (__mmask8 __M, __m256i __X, const int __I) 9954 { 9955 return (__m256i) __builtin_ia32_permdi256_mask ((__v4di) __X, 9956 __I, 9957 (__v4di) 9958 _mm256_setzero_si256 (), 9959 (__mmask8) __M); 9960 } 9961 9962 extern __inline __m256d 9963 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9964 _mm256_mask_shuffle_pd (__m256d __W, __mmask8 __U, __m256d __A, 9965 __m256d __B, const int __imm) 9966 { 9967 return (__m256d) __builtin_ia32_shufpd256_mask ((__v4df) __A, 9968 (__v4df) __B, __imm, 9969 (__v4df) __W, 9970 (__mmask8) __U); 9971 } 9972 9973 extern __inline __m256d 9974 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9975 _mm256_maskz_shuffle_pd (__mmask8 __U, __m256d __A, __m256d __B, 9976 const int __imm) 9977 { 9978 return (__m256d) __builtin_ia32_shufpd256_mask ((__v4df) __A, 9979 (__v4df) __B, __imm, 9980 (__v4df) 9981 _mm256_setzero_pd (), 9982 (__mmask8) __U); 9983 } 9984 9985 extern __inline __m128d 9986 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9987 _mm_mask_shuffle_pd (__m128d __W, __mmask8 __U, __m128d __A, 9988 __m128d __B, const int __imm) 9989 { 9990 return (__m128d) __builtin_ia32_shufpd128_mask ((__v2df) __A, 9991 (__v2df) __B, __imm, 9992 (__v2df) __W, 9993 (__mmask8) __U); 9994 } 9995 9996 extern __inline __m128d 9997 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 9998 _mm_maskz_shuffle_pd (__mmask8 __U, __m128d __A, __m128d __B, 9999 const int __imm) 10000 { 10001 return (__m128d) __builtin_ia32_shufpd128_mask ((__v2df) __A, 10002 (__v2df) __B, __imm, 10003 (__v2df) 10004 _mm_setzero_pd (), 10005 (__mmask8) __U); 10006 } 10007 10008 extern __inline __m256 10009 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10010 _mm256_mask_shuffle_ps (__m256 __W, __mmask8 __U, __m256 __A, 10011 __m256 __B, const int __imm) 10012 { 10013 return (__m256) __builtin_ia32_shufps256_mask ((__v8sf) __A, 10014 (__v8sf) __B, __imm, 10015 (__v8sf) __W, 10016 (__mmask8) __U); 10017 } 10018 10019 extern __inline __m256 10020 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10021 _mm256_maskz_shuffle_ps (__mmask8 __U, __m256 __A, __m256 __B, 10022 const int __imm) 10023 { 10024 return (__m256) __builtin_ia32_shufps256_mask ((__v8sf) __A, 10025 (__v8sf) __B, __imm, 10026 (__v8sf) 10027 _mm256_setzero_ps (), 10028 (__mmask8) __U); 10029 } 10030 10031 extern __inline __m128 10032 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10033 _mm_mask_shuffle_ps (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B, 10034 const int __imm) 10035 { 10036 return (__m128) __builtin_ia32_shufps128_mask ((__v4sf) __A, 10037 (__v4sf) __B, __imm, 10038 (__v4sf) __W, 10039 (__mmask8) __U); 10040 } 10041 10042 extern __inline __m128 10043 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10044 _mm_maskz_shuffle_ps (__mmask8 __U, __m128 __A, __m128 __B, 10045 const int __imm) 10046 { 10047 return (__m128) __builtin_ia32_shufps128_mask ((__v4sf) __A, 10048 (__v4sf) __B, __imm, 10049 (__v4sf) 10050 _mm_setzero_ps (), 10051 (__mmask8) __U); 10052 } 10053 10054 extern __inline __m256i 10055 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10056 _mm256_inserti32x4 (__m256i __A, __m128i __B, const int __imm) 10057 { 10058 return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A, 10059 (__v4si) __B, 10060 __imm, 10061 (__v8si) 10062 _mm256_setzero_si256 (), 10063 (__mmask8) -1); 10064 } 10065 10066 extern __inline __m256i 10067 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10068 _mm256_mask_inserti32x4 (__m256i __W, __mmask8 __U, __m256i __A, 10069 __m128i __B, const int __imm) 10070 { 10071 return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A, 10072 (__v4si) __B, 10073 __imm, 10074 (__v8si) __W, 10075 (__mmask8) 10076 __U); 10077 } 10078 10079 extern __inline __m256i 10080 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10081 _mm256_maskz_inserti32x4 (__mmask8 __U, __m256i __A, __m128i __B, 10082 const int __imm) 10083 { 10084 return (__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si) __A, 10085 (__v4si) __B, 10086 __imm, 10087 (__v8si) 10088 _mm256_setzero_si256 (), 10089 (__mmask8) 10090 __U); 10091 } 10092 10093 extern __inline __m256 10094 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10095 _mm256_insertf32x4 (__m256 __A, __m128 __B, const int __imm) 10096 { 10097 return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A, 10098 (__v4sf) __B, 10099 __imm, 10100 (__v8sf) 10101 _mm256_setzero_ps (), 10102 (__mmask8) -1); 10103 } 10104 10105 extern __inline __m256 10106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10107 _mm256_mask_insertf32x4 (__m256 __W, __mmask8 __U, __m256 __A, 10108 __m128 __B, const int __imm) 10109 { 10110 return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A, 10111 (__v4sf) __B, 10112 __imm, 10113 (__v8sf) __W, 10114 (__mmask8) __U); 10115 } 10116 10117 extern __inline __m256 10118 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10119 _mm256_maskz_insertf32x4 (__mmask8 __U, __m256 __A, __m128 __B, 10120 const int __imm) 10121 { 10122 return (__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf) __A, 10123 (__v4sf) __B, 10124 __imm, 10125 (__v8sf) 10126 _mm256_setzero_ps (), 10127 (__mmask8) __U); 10128 } 10129 10130 extern __inline __m128i 10131 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10132 _mm256_extracti32x4_epi32 (__m256i __A, const int __imm) 10133 { 10134 return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A, 10135 __imm, 10136 (__v4si) 10137 _mm_setzero_si128 (), 10138 (__mmask8) -1); 10139 } 10140 10141 extern __inline __m128i 10142 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10143 _mm256_mask_extracti32x4_epi32 (__m128i __W, __mmask8 __U, __m256i __A, 10144 const int __imm) 10145 { 10146 return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A, 10147 __imm, 10148 (__v4si) __W, 10149 (__mmask8) 10150 __U); 10151 } 10152 10153 extern __inline __m128i 10154 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10155 _mm256_maskz_extracti32x4_epi32 (__mmask8 __U, __m256i __A, 10156 const int __imm) 10157 { 10158 return (__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si) __A, 10159 __imm, 10160 (__v4si) 10161 _mm_setzero_si128 (), 10162 (__mmask8) 10163 __U); 10164 } 10165 10166 extern __inline __m128 10167 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10168 _mm256_extractf32x4_ps (__m256 __A, const int __imm) 10169 { 10170 return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A, 10171 __imm, 10172 (__v4sf) 10173 _mm_setzero_ps (), 10174 (__mmask8) -1); 10175 } 10176 10177 extern __inline __m128 10178 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10179 _mm256_mask_extractf32x4_ps (__m128 __W, __mmask8 __U, __m256 __A, 10180 const int __imm) 10181 { 10182 return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A, 10183 __imm, 10184 (__v4sf) __W, 10185 (__mmask8) 10186 __U); 10187 } 10188 10189 extern __inline __m128 10190 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10191 _mm256_maskz_extractf32x4_ps (__mmask8 __U, __m256 __A, 10192 const int __imm) 10193 { 10194 return (__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf) __A, 10195 __imm, 10196 (__v4sf) 10197 _mm_setzero_ps (), 10198 (__mmask8) 10199 __U); 10200 } 10201 10202 extern __inline __m256i 10203 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10204 _mm256_shuffle_i64x2 (__m256i __A, __m256i __B, const int __imm) 10205 { 10206 return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A, 10207 (__v4di) __B, 10208 __imm, 10209 (__v4di) 10210 _mm256_setzero_si256 (), 10211 (__mmask8) -1); 10212 } 10213 10214 extern __inline __m256i 10215 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10216 _mm256_mask_shuffle_i64x2 (__m256i __W, __mmask8 __U, __m256i __A, 10217 __m256i __B, const int __imm) 10218 { 10219 return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A, 10220 (__v4di) __B, 10221 __imm, 10222 (__v4di) __W, 10223 (__mmask8) __U); 10224 } 10225 10226 extern __inline __m256i 10227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10228 _mm256_maskz_shuffle_i64x2 (__mmask8 __U, __m256i __A, __m256i __B, 10229 const int __imm) 10230 { 10231 return (__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di) __A, 10232 (__v4di) __B, 10233 __imm, 10234 (__v4di) 10235 _mm256_setzero_si256 (), 10236 (__mmask8) __U); 10237 } 10238 10239 extern __inline __m256i 10240 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10241 _mm256_shuffle_i32x4 (__m256i __A, __m256i __B, const int __imm) 10242 { 10243 return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A, 10244 (__v8si) __B, 10245 __imm, 10246 (__v8si) 10247 _mm256_setzero_si256 (), 10248 (__mmask8) -1); 10249 } 10250 10251 extern __inline __m256i 10252 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10253 _mm256_mask_shuffle_i32x4 (__m256i __W, __mmask8 __U, __m256i __A, 10254 __m256i __B, const int __imm) 10255 { 10256 return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A, 10257 (__v8si) __B, 10258 __imm, 10259 (__v8si) __W, 10260 (__mmask8) __U); 10261 } 10262 10263 extern __inline __m256i 10264 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10265 _mm256_maskz_shuffle_i32x4 (__mmask8 __U, __m256i __A, __m256i __B, 10266 const int __imm) 10267 { 10268 return (__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si) __A, 10269 (__v8si) __B, 10270 __imm, 10271 (__v8si) 10272 _mm256_setzero_si256 (), 10273 (__mmask8) __U); 10274 } 10275 10276 extern __inline __m256d 10277 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10278 _mm256_shuffle_f64x2 (__m256d __A, __m256d __B, const int __imm) 10279 { 10280 return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A, 10281 (__v4df) __B, 10282 __imm, 10283 (__v4df) 10284 _mm256_setzero_pd (), 10285 (__mmask8) -1); 10286 } 10287 10288 extern __inline __m256d 10289 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10290 _mm256_mask_shuffle_f64x2 (__m256d __W, __mmask8 __U, __m256d __A, 10291 __m256d __B, const int __imm) 10292 { 10293 return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A, 10294 (__v4df) __B, 10295 __imm, 10296 (__v4df) __W, 10297 (__mmask8) __U); 10298 } 10299 10300 extern __inline __m256d 10301 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10302 _mm256_maskz_shuffle_f64x2 (__mmask8 __U, __m256d __A, __m256d __B, 10303 const int __imm) 10304 { 10305 return (__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df) __A, 10306 (__v4df) __B, 10307 __imm, 10308 (__v4df) 10309 _mm256_setzero_pd (), 10310 (__mmask8) __U); 10311 } 10312 10313 extern __inline __m256 10314 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10315 _mm256_shuffle_f32x4 (__m256 __A, __m256 __B, const int __imm) 10316 { 10317 return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A, 10318 (__v8sf) __B, 10319 __imm, 10320 (__v8sf) 10321 _mm256_setzero_ps (), 10322 (__mmask8) -1); 10323 } 10324 10325 extern __inline __m256 10326 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10327 _mm256_mask_shuffle_f32x4 (__m256 __W, __mmask8 __U, __m256 __A, 10328 __m256 __B, const int __imm) 10329 { 10330 return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A, 10331 (__v8sf) __B, 10332 __imm, 10333 (__v8sf) __W, 10334 (__mmask8) __U); 10335 } 10336 10337 extern __inline __m256 10338 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10339 _mm256_maskz_shuffle_f32x4 (__mmask8 __U, __m256 __A, __m256 __B, 10340 const int __imm) 10341 { 10342 return (__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf) __A, 10343 (__v8sf) __B, 10344 __imm, 10345 (__v8sf) 10346 _mm256_setzero_ps (), 10347 (__mmask8) __U); 10348 } 10349 10350 extern __inline __m256d 10351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10352 _mm256_fixupimm_pd (__m256d __A, __m256d __B, __m256i __C, 10353 const int __imm) 10354 { 10355 return (__m256d) __builtin_ia32_fixupimmpd256_mask ((__v4df) __A, 10356 (__v4df) __B, 10357 (__v4di) __C, 10358 __imm, 10359 (__mmask8) -1); 10360 } 10361 10362 extern __inline __m256d 10363 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10364 _mm256_mask_fixupimm_pd (__m256d __A, __mmask8 __U, __m256d __B, 10365 __m256i __C, const int __imm) 10366 { 10367 return (__m256d) __builtin_ia32_fixupimmpd256_mask ((__v4df) __A, 10368 (__v4df) __B, 10369 (__v4di) __C, 10370 __imm, 10371 (__mmask8) __U); 10372 } 10373 10374 extern __inline __m256d 10375 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10376 _mm256_maskz_fixupimm_pd (__mmask8 __U, __m256d __A, __m256d __B, 10377 __m256i __C, const int __imm) 10378 { 10379 return (__m256d) __builtin_ia32_fixupimmpd256_maskz ((__v4df) __A, 10380 (__v4df) __B, 10381 (__v4di) __C, 10382 __imm, 10383 (__mmask8) __U); 10384 } 10385 10386 extern __inline __m256 10387 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10388 _mm256_fixupimm_ps (__m256 __A, __m256 __B, __m256i __C, 10389 const int __imm) 10390 { 10391 return (__m256) __builtin_ia32_fixupimmps256_mask ((__v8sf) __A, 10392 (__v8sf) __B, 10393 (__v8si) __C, 10394 __imm, 10395 (__mmask8) -1); 10396 } 10397 10398 extern __inline __m256 10399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10400 _mm256_mask_fixupimm_ps (__m256 __A, __mmask8 __U, __m256 __B, 10401 __m256i __C, const int __imm) 10402 { 10403 return (__m256) __builtin_ia32_fixupimmps256_mask ((__v8sf) __A, 10404 (__v8sf) __B, 10405 (__v8si) __C, 10406 __imm, 10407 (__mmask8) __U); 10408 } 10409 10410 extern __inline __m256 10411 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10412 _mm256_maskz_fixupimm_ps (__mmask8 __U, __m256 __A, __m256 __B, 10413 __m256i __C, const int __imm) 10414 { 10415 return (__m256) __builtin_ia32_fixupimmps256_maskz ((__v8sf) __A, 10416 (__v8sf) __B, 10417 (__v8si) __C, 10418 __imm, 10419 (__mmask8) __U); 10420 } 10421 10422 extern __inline __m128d 10423 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10424 _mm_fixupimm_pd (__m128d __A, __m128d __B, __m128i __C, 10425 const int __imm) 10426 { 10427 return (__m128d) __builtin_ia32_fixupimmpd128_mask ((__v2df) __A, 10428 (__v2df) __B, 10429 (__v2di) __C, 10430 __imm, 10431 (__mmask8) -1); 10432 } 10433 10434 extern __inline __m128d 10435 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10436 _mm_mask_fixupimm_pd (__m128d __A, __mmask8 __U, __m128d __B, 10437 __m128i __C, const int __imm) 10438 { 10439 return (__m128d) __builtin_ia32_fixupimmpd128_mask ((__v2df) __A, 10440 (__v2df) __B, 10441 (__v2di) __C, 10442 __imm, 10443 (__mmask8) __U); 10444 } 10445 10446 extern __inline __m128d 10447 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10448 _mm_maskz_fixupimm_pd (__mmask8 __U, __m128d __A, __m128d __B, 10449 __m128i __C, const int __imm) 10450 { 10451 return (__m128d) __builtin_ia32_fixupimmpd128_maskz ((__v2df) __A, 10452 (__v2df) __B, 10453 (__v2di) __C, 10454 __imm, 10455 (__mmask8) __U); 10456 } 10457 10458 extern __inline __m128 10459 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10460 _mm_fixupimm_ps (__m128 __A, __m128 __B, __m128i __C, const int __imm) 10461 { 10462 return (__m128) __builtin_ia32_fixupimmps128_mask ((__v4sf) __A, 10463 (__v4sf) __B, 10464 (__v4si) __C, 10465 __imm, 10466 (__mmask8) -1); 10467 } 10468 10469 extern __inline __m128 10470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10471 _mm_mask_fixupimm_ps (__m128 __A, __mmask8 __U, __m128 __B, 10472 __m128i __C, const int __imm) 10473 { 10474 return (__m128) __builtin_ia32_fixupimmps128_mask ((__v4sf) __A, 10475 (__v4sf) __B, 10476 (__v4si) __C, 10477 __imm, 10478 (__mmask8) __U); 10479 } 10480 10481 extern __inline __m128 10482 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10483 _mm_maskz_fixupimm_ps (__mmask8 __U, __m128 __A, __m128 __B, 10484 __m128i __C, const int __imm) 10485 { 10486 return (__m128) __builtin_ia32_fixupimmps128_maskz ((__v4sf) __A, 10487 (__v4sf) __B, 10488 (__v4si) __C, 10489 __imm, 10490 (__mmask8) __U); 10491 } 10492 10493 extern __inline __m256i 10494 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10495 _mm256_mask_srli_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 10496 const int __imm) 10497 { 10498 return (__m256i) __builtin_ia32_psrldi256_mask ((__v8si) __A, __imm, 10499 (__v8si) __W, 10500 (__mmask8) __U); 10501 } 10502 10503 extern __inline __m256i 10504 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10505 _mm256_maskz_srli_epi32 (__mmask8 __U, __m256i __A, const int __imm) 10506 { 10507 return (__m256i) __builtin_ia32_psrldi256_mask ((__v8si) __A, __imm, 10508 (__v8si) 10509 _mm256_setzero_si256 (), 10510 (__mmask8) __U); 10511 } 10512 10513 extern __inline __m128i 10514 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10515 _mm_mask_srli_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 10516 const int __imm) 10517 { 10518 return (__m128i) __builtin_ia32_psrldi128_mask ((__v4si) __A, __imm, 10519 (__v4si) __W, 10520 (__mmask8) __U); 10521 } 10522 10523 extern __inline __m128i 10524 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10525 _mm_maskz_srli_epi32 (__mmask8 __U, __m128i __A, const int __imm) 10526 { 10527 return (__m128i) __builtin_ia32_psrldi128_mask ((__v4si) __A, __imm, 10528 (__v4si) 10529 _mm_setzero_si128 (), 10530 (__mmask8) __U); 10531 } 10532 10533 extern __inline __m256i 10534 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10535 _mm256_mask_srli_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 10536 const int __imm) 10537 { 10538 return (__m256i) __builtin_ia32_psrlqi256_mask ((__v4di) __A, __imm, 10539 (__v4di) __W, 10540 (__mmask8) __U); 10541 } 10542 10543 extern __inline __m256i 10544 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10545 _mm256_maskz_srli_epi64 (__mmask8 __U, __m256i __A, const int __imm) 10546 { 10547 return (__m256i) __builtin_ia32_psrlqi256_mask ((__v4di) __A, __imm, 10548 (__v4di) 10549 _mm256_setzero_si256 (), 10550 (__mmask8) __U); 10551 } 10552 10553 extern __inline __m128i 10554 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10555 _mm_mask_srli_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 10556 const int __imm) 10557 { 10558 return (__m128i) __builtin_ia32_psrlqi128_mask ((__v2di) __A, __imm, 10559 (__v2di) __W, 10560 (__mmask8) __U); 10561 } 10562 10563 extern __inline __m128i 10564 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10565 _mm_maskz_srli_epi64 (__mmask8 __U, __m128i __A, const int __imm) 10566 { 10567 return (__m128i) __builtin_ia32_psrlqi128_mask ((__v2di) __A, __imm, 10568 (__v2di) 10569 _mm_setzero_si128 (), 10570 (__mmask8) __U); 10571 } 10572 10573 extern __inline __m256i 10574 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10575 _mm256_ternarylogic_epi64 (__m256i __A, __m256i __B, __m256i __C, 10576 const int __imm) 10577 { 10578 return (__m256i) 10579 __builtin_ia32_pternlogq256_mask ((__v4di) __A, 10580 (__v4di) __B, 10581 (__v4di) __C, 10582 (unsigned char) __imm, 10583 (__mmask8) -1); 10584 } 10585 10586 extern __inline __m256i 10587 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10588 _mm256_mask_ternarylogic_epi64 (__m256i __A, __mmask8 __U, 10589 __m256i __B, __m256i __C, 10590 const int __imm) 10591 { 10592 return (__m256i) 10593 __builtin_ia32_pternlogq256_mask ((__v4di) __A, 10594 (__v4di) __B, 10595 (__v4di) __C, 10596 (unsigned char) __imm, 10597 (__mmask8) __U); 10598 } 10599 10600 extern __inline __m256i 10601 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10602 _mm256_maskz_ternarylogic_epi64 (__mmask8 __U, __m256i __A, 10603 __m256i __B, __m256i __C, 10604 const int __imm) 10605 { 10606 return (__m256i) 10607 __builtin_ia32_pternlogq256_maskz ((__v4di) __A, 10608 (__v4di) __B, 10609 (__v4di) __C, 10610 (unsigned char) __imm, 10611 (__mmask8) __U); 10612 } 10613 10614 extern __inline __m256i 10615 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10616 _mm256_ternarylogic_epi32 (__m256i __A, __m256i __B, __m256i __C, 10617 const int __imm) 10618 { 10619 return (__m256i) 10620 __builtin_ia32_pternlogd256_mask ((__v8si) __A, 10621 (__v8si) __B, 10622 (__v8si) __C, 10623 (unsigned char) __imm, 10624 (__mmask8) -1); 10625 } 10626 10627 extern __inline __m256i 10628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10629 _mm256_mask_ternarylogic_epi32 (__m256i __A, __mmask8 __U, 10630 __m256i __B, __m256i __C, 10631 const int __imm) 10632 { 10633 return (__m256i) 10634 __builtin_ia32_pternlogd256_mask ((__v8si) __A, 10635 (__v8si) __B, 10636 (__v8si) __C, 10637 (unsigned char) __imm, 10638 (__mmask8) __U); 10639 } 10640 10641 extern __inline __m256i 10642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10643 _mm256_maskz_ternarylogic_epi32 (__mmask8 __U, __m256i __A, 10644 __m256i __B, __m256i __C, 10645 const int __imm) 10646 { 10647 return (__m256i) 10648 __builtin_ia32_pternlogd256_maskz ((__v8si) __A, 10649 (__v8si) __B, 10650 (__v8si) __C, 10651 (unsigned char) __imm, 10652 (__mmask8) __U); 10653 } 10654 10655 extern __inline __m128i 10656 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10657 _mm_ternarylogic_epi64 (__m128i __A, __m128i __B, __m128i __C, 10658 const int __imm) 10659 { 10660 return (__m128i) 10661 __builtin_ia32_pternlogq128_mask ((__v2di) __A, 10662 (__v2di) __B, 10663 (__v2di) __C, 10664 (unsigned char) __imm, 10665 (__mmask8) -1); 10666 } 10667 10668 extern __inline __m128i 10669 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10670 _mm_mask_ternarylogic_epi64 (__m128i __A, __mmask8 __U, 10671 __m128i __B, __m128i __C, 10672 const int __imm) 10673 { 10674 return (__m128i) 10675 __builtin_ia32_pternlogq128_mask ((__v2di) __A, 10676 (__v2di) __B, 10677 (__v2di) __C, 10678 (unsigned char) __imm, 10679 (__mmask8) __U); 10680 } 10681 10682 extern __inline __m128i 10683 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10684 _mm_maskz_ternarylogic_epi64 (__mmask8 __U, __m128i __A, 10685 __m128i __B, __m128i __C, 10686 const int __imm) 10687 { 10688 return (__m128i) 10689 __builtin_ia32_pternlogq128_maskz ((__v2di) __A, 10690 (__v2di) __B, 10691 (__v2di) __C, 10692 (unsigned char) __imm, 10693 (__mmask8) __U); 10694 } 10695 10696 extern __inline __m128i 10697 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10698 _mm_ternarylogic_epi32 (__m128i __A, __m128i __B, __m128i __C, 10699 const int __imm) 10700 { 10701 return (__m128i) 10702 __builtin_ia32_pternlogd128_mask ((__v4si) __A, 10703 (__v4si) __B, 10704 (__v4si) __C, 10705 (unsigned char) __imm, 10706 (__mmask8) -1); 10707 } 10708 10709 extern __inline __m128i 10710 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10711 _mm_mask_ternarylogic_epi32 (__m128i __A, __mmask8 __U, 10712 __m128i __B, __m128i __C, 10713 const int __imm) 10714 { 10715 return (__m128i) 10716 __builtin_ia32_pternlogd128_mask ((__v4si) __A, 10717 (__v4si) __B, 10718 (__v4si) __C, 10719 (unsigned char) __imm, 10720 (__mmask8) __U); 10721 } 10722 10723 extern __inline __m128i 10724 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10725 _mm_maskz_ternarylogic_epi32 (__mmask8 __U, __m128i __A, 10726 __m128i __B, __m128i __C, 10727 const int __imm) 10728 { 10729 return (__m128i) 10730 __builtin_ia32_pternlogd128_maskz ((__v4si) __A, 10731 (__v4si) __B, 10732 (__v4si) __C, 10733 (unsigned char) __imm, 10734 (__mmask8) __U); 10735 } 10736 10737 extern __inline __m256 10738 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10739 _mm256_roundscale_ps (__m256 __A, const int __imm) 10740 { 10741 return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, 10742 __imm, 10743 (__v8sf) 10744 _mm256_setzero_ps (), 10745 (__mmask8) -1); 10746 } 10747 10748 extern __inline __m256 10749 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10750 _mm256_mask_roundscale_ps (__m256 __W, __mmask8 __U, __m256 __A, 10751 const int __imm) 10752 { 10753 return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, 10754 __imm, 10755 (__v8sf) __W, 10756 (__mmask8) __U); 10757 } 10758 10759 extern __inline __m256 10760 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10761 _mm256_maskz_roundscale_ps (__mmask8 __U, __m256 __A, const int __imm) 10762 { 10763 return (__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf) __A, 10764 __imm, 10765 (__v8sf) 10766 _mm256_setzero_ps (), 10767 (__mmask8) __U); 10768 } 10769 10770 extern __inline __m256d 10771 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10772 _mm256_roundscale_pd (__m256d __A, const int __imm) 10773 { 10774 return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, 10775 __imm, 10776 (__v4df) 10777 _mm256_setzero_pd (), 10778 (__mmask8) -1); 10779 } 10780 10781 extern __inline __m256d 10782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10783 _mm256_mask_roundscale_pd (__m256d __W, __mmask8 __U, __m256d __A, 10784 const int __imm) 10785 { 10786 return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, 10787 __imm, 10788 (__v4df) __W, 10789 (__mmask8) __U); 10790 } 10791 10792 extern __inline __m256d 10793 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10794 _mm256_maskz_roundscale_pd (__mmask8 __U, __m256d __A, const int __imm) 10795 { 10796 return (__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df) __A, 10797 __imm, 10798 (__v4df) 10799 _mm256_setzero_pd (), 10800 (__mmask8) __U); 10801 } 10802 10803 extern __inline __m128 10804 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10805 _mm_roundscale_ps (__m128 __A, const int __imm) 10806 { 10807 return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, 10808 __imm, 10809 (__v4sf) 10810 _mm_setzero_ps (), 10811 (__mmask8) -1); 10812 } 10813 10814 extern __inline __m128 10815 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10816 _mm_mask_roundscale_ps (__m128 __W, __mmask8 __U, __m128 __A, 10817 const int __imm) 10818 { 10819 return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, 10820 __imm, 10821 (__v4sf) __W, 10822 (__mmask8) __U); 10823 } 10824 10825 extern __inline __m128 10826 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10827 _mm_maskz_roundscale_ps (__mmask8 __U, __m128 __A, const int __imm) 10828 { 10829 return (__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf) __A, 10830 __imm, 10831 (__v4sf) 10832 _mm_setzero_ps (), 10833 (__mmask8) __U); 10834 } 10835 10836 extern __inline __m128d 10837 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10838 _mm_roundscale_pd (__m128d __A, const int __imm) 10839 { 10840 return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, 10841 __imm, 10842 (__v2df) 10843 _mm_setzero_pd (), 10844 (__mmask8) -1); 10845 } 10846 10847 extern __inline __m128d 10848 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10849 _mm_mask_roundscale_pd (__m128d __W, __mmask8 __U, __m128d __A, 10850 const int __imm) 10851 { 10852 return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, 10853 __imm, 10854 (__v2df) __W, 10855 (__mmask8) __U); 10856 } 10857 10858 extern __inline __m128d 10859 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10860 _mm_maskz_roundscale_pd (__mmask8 __U, __m128d __A, const int __imm) 10861 { 10862 return (__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df) __A, 10863 __imm, 10864 (__v2df) 10865 _mm_setzero_pd (), 10866 (__mmask8) __U); 10867 } 10868 10869 extern __inline __m256 10870 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10871 _mm256_getmant_ps (__m256 __A, _MM_MANTISSA_NORM_ENUM __B, 10872 _MM_MANTISSA_SIGN_ENUM __C) 10873 { 10874 return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A, 10875 (__C << 2) | __B, 10876 (__v8sf) 10877 _mm256_setzero_ps (), 10878 (__mmask8) -1); 10879 } 10880 10881 extern __inline __m256 10882 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10883 _mm256_mask_getmant_ps (__m256 __W, __mmask8 __U, __m256 __A, 10884 _MM_MANTISSA_NORM_ENUM __B, 10885 _MM_MANTISSA_SIGN_ENUM __C) 10886 { 10887 return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A, 10888 (__C << 2) | __B, 10889 (__v8sf) __W, 10890 (__mmask8) __U); 10891 } 10892 10893 extern __inline __m256 10894 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10895 _mm256_maskz_getmant_ps (__mmask8 __U, __m256 __A, 10896 _MM_MANTISSA_NORM_ENUM __B, 10897 _MM_MANTISSA_SIGN_ENUM __C) 10898 { 10899 return (__m256) __builtin_ia32_getmantps256_mask ((__v8sf) __A, 10900 (__C << 2) | __B, 10901 (__v8sf) 10902 _mm256_setzero_ps (), 10903 (__mmask8) __U); 10904 } 10905 10906 extern __inline __m128 10907 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10908 _mm_getmant_ps (__m128 __A, _MM_MANTISSA_NORM_ENUM __B, 10909 _MM_MANTISSA_SIGN_ENUM __C) 10910 { 10911 return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A, 10912 (__C << 2) | __B, 10913 (__v4sf) 10914 _mm_setzero_ps (), 10915 (__mmask8) -1); 10916 } 10917 10918 extern __inline __m128 10919 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10920 _mm_mask_getmant_ps (__m128 __W, __mmask8 __U, __m128 __A, 10921 _MM_MANTISSA_NORM_ENUM __B, 10922 _MM_MANTISSA_SIGN_ENUM __C) 10923 { 10924 return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A, 10925 (__C << 2) | __B, 10926 (__v4sf) __W, 10927 (__mmask8) __U); 10928 } 10929 10930 extern __inline __m128 10931 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10932 _mm_maskz_getmant_ps (__mmask8 __U, __m128 __A, 10933 _MM_MANTISSA_NORM_ENUM __B, 10934 _MM_MANTISSA_SIGN_ENUM __C) 10935 { 10936 return (__m128) __builtin_ia32_getmantps128_mask ((__v4sf) __A, 10937 (__C << 2) | __B, 10938 (__v4sf) 10939 _mm_setzero_ps (), 10940 (__mmask8) __U); 10941 } 10942 10943 extern __inline __m256d 10944 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10945 _mm256_getmant_pd (__m256d __A, _MM_MANTISSA_NORM_ENUM __B, 10946 _MM_MANTISSA_SIGN_ENUM __C) 10947 { 10948 return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A, 10949 (__C << 2) | __B, 10950 (__v4df) 10951 _mm256_setzero_pd (), 10952 (__mmask8) -1); 10953 } 10954 10955 extern __inline __m256d 10956 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10957 _mm256_mask_getmant_pd (__m256d __W, __mmask8 __U, __m256d __A, 10958 _MM_MANTISSA_NORM_ENUM __B, 10959 _MM_MANTISSA_SIGN_ENUM __C) 10960 { 10961 return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A, 10962 (__C << 2) | __B, 10963 (__v4df) __W, 10964 (__mmask8) __U); 10965 } 10966 10967 extern __inline __m256d 10968 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10969 _mm256_maskz_getmant_pd (__mmask8 __U, __m256d __A, 10970 _MM_MANTISSA_NORM_ENUM __B, 10971 _MM_MANTISSA_SIGN_ENUM __C) 10972 { 10973 return (__m256d) __builtin_ia32_getmantpd256_mask ((__v4df) __A, 10974 (__C << 2) | __B, 10975 (__v4df) 10976 _mm256_setzero_pd (), 10977 (__mmask8) __U); 10978 } 10979 10980 extern __inline __m128d 10981 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10982 _mm_getmant_pd (__m128d __A, _MM_MANTISSA_NORM_ENUM __B, 10983 _MM_MANTISSA_SIGN_ENUM __C) 10984 { 10985 return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A, 10986 (__C << 2) | __B, 10987 (__v2df) 10988 _mm_setzero_pd (), 10989 (__mmask8) -1); 10990 } 10991 10992 extern __inline __m128d 10993 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 10994 _mm_mask_getmant_pd (__m128d __W, __mmask8 __U, __m128d __A, 10995 _MM_MANTISSA_NORM_ENUM __B, 10996 _MM_MANTISSA_SIGN_ENUM __C) 10997 { 10998 return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A, 10999 (__C << 2) | __B, 11000 (__v2df) __W, 11001 (__mmask8) __U); 11002 } 11003 11004 extern __inline __m128d 11005 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11006 _mm_maskz_getmant_pd (__mmask8 __U, __m128d __A, 11007 _MM_MANTISSA_NORM_ENUM __B, 11008 _MM_MANTISSA_SIGN_ENUM __C) 11009 { 11010 return (__m128d) __builtin_ia32_getmantpd128_mask ((__v2df) __A, 11011 (__C << 2) | __B, 11012 (__v2df) 11013 _mm_setzero_pd (), 11014 (__mmask8) __U); 11015 } 11016 11017 extern __inline __m256 11018 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11019 _mm256_mmask_i32gather_ps (__m256 __v1_old, __mmask8 __mask, 11020 __m256i __index, void const *__addr, 11021 int __scale) 11022 { 11023 return (__m256) __builtin_ia32_gather3siv8sf ((__v8sf) __v1_old, 11024 __addr, 11025 (__v8si) __index, 11026 __mask, __scale); 11027 } 11028 11029 extern __inline __m128 11030 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11031 _mm_mmask_i32gather_ps (__m128 __v1_old, __mmask8 __mask, 11032 __m128i __index, void const *__addr, 11033 int __scale) 11034 { 11035 return (__m128) __builtin_ia32_gather3siv4sf ((__v4sf) __v1_old, 11036 __addr, 11037 (__v4si) __index, 11038 __mask, __scale); 11039 } 11040 11041 extern __inline __m256d 11042 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11043 _mm256_mmask_i32gather_pd (__m256d __v1_old, __mmask8 __mask, 11044 __m128i __index, void const *__addr, 11045 int __scale) 11046 { 11047 return (__m256d) __builtin_ia32_gather3siv4df ((__v4df) __v1_old, 11048 __addr, 11049 (__v4si) __index, 11050 __mask, __scale); 11051 } 11052 11053 extern __inline __m128d 11054 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11055 _mm_mmask_i32gather_pd (__m128d __v1_old, __mmask8 __mask, 11056 __m128i __index, void const *__addr, 11057 int __scale) 11058 { 11059 return (__m128d) __builtin_ia32_gather3siv2df ((__v2df) __v1_old, 11060 __addr, 11061 (__v4si) __index, 11062 __mask, __scale); 11063 } 11064 11065 extern __inline __m128 11066 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11067 _mm256_mmask_i64gather_ps (__m128 __v1_old, __mmask8 __mask, 11068 __m256i __index, void const *__addr, 11069 int __scale) 11070 { 11071 return (__m128) __builtin_ia32_gather3div8sf ((__v4sf) __v1_old, 11072 __addr, 11073 (__v4di) __index, 11074 __mask, __scale); 11075 } 11076 11077 extern __inline __m128 11078 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11079 _mm_mmask_i64gather_ps (__m128 __v1_old, __mmask8 __mask, 11080 __m128i __index, void const *__addr, 11081 int __scale) 11082 { 11083 return (__m128) __builtin_ia32_gather3div4sf ((__v4sf) __v1_old, 11084 __addr, 11085 (__v2di) __index, 11086 __mask, __scale); 11087 } 11088 11089 extern __inline __m256d 11090 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11091 _mm256_mmask_i64gather_pd (__m256d __v1_old, __mmask8 __mask, 11092 __m256i __index, void const *__addr, 11093 int __scale) 11094 { 11095 return (__m256d) __builtin_ia32_gather3div4df ((__v4df) __v1_old, 11096 __addr, 11097 (__v4di) __index, 11098 __mask, __scale); 11099 } 11100 11101 extern __inline __m128d 11102 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11103 _mm_mmask_i64gather_pd (__m128d __v1_old, __mmask8 __mask, 11104 __m128i __index, void const *__addr, 11105 int __scale) 11106 { 11107 return (__m128d) __builtin_ia32_gather3div2df ((__v2df) __v1_old, 11108 __addr, 11109 (__v2di) __index, 11110 __mask, __scale); 11111 } 11112 11113 extern __inline __m256i 11114 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11115 _mm256_mmask_i32gather_epi32 (__m256i __v1_old, __mmask8 __mask, 11116 __m256i __index, void const *__addr, 11117 int __scale) 11118 { 11119 return (__m256i) __builtin_ia32_gather3siv8si ((__v8si) __v1_old, 11120 __addr, 11121 (__v8si) __index, 11122 __mask, __scale); 11123 } 11124 11125 extern __inline __m128i 11126 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11127 _mm_mmask_i32gather_epi32 (__m128i __v1_old, __mmask8 __mask, 11128 __m128i __index, void const *__addr, 11129 int __scale) 11130 { 11131 return (__m128i) __builtin_ia32_gather3siv4si ((__v4si) __v1_old, 11132 __addr, 11133 (__v4si) __index, 11134 __mask, __scale); 11135 } 11136 11137 extern __inline __m256i 11138 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11139 _mm256_mmask_i32gather_epi64 (__m256i __v1_old, __mmask8 __mask, 11140 __m128i __index, void const *__addr, 11141 int __scale) 11142 { 11143 return (__m256i) __builtin_ia32_gather3siv4di ((__v4di) __v1_old, 11144 __addr, 11145 (__v4si) __index, 11146 __mask, __scale); 11147 } 11148 11149 extern __inline __m128i 11150 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11151 _mm_mmask_i32gather_epi64 (__m128i __v1_old, __mmask8 __mask, 11152 __m128i __index, void const *__addr, 11153 int __scale) 11154 { 11155 return (__m128i) __builtin_ia32_gather3siv2di ((__v2di) __v1_old, 11156 __addr, 11157 (__v4si) __index, 11158 __mask, __scale); 11159 } 11160 11161 extern __inline __m128i 11162 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11163 _mm256_mmask_i64gather_epi32 (__m128i __v1_old, __mmask8 __mask, 11164 __m256i __index, void const *__addr, 11165 int __scale) 11166 { 11167 return (__m128i) __builtin_ia32_gather3div8si ((__v4si) __v1_old, 11168 __addr, 11169 (__v4di) __index, 11170 __mask, __scale); 11171 } 11172 11173 extern __inline __m128i 11174 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11175 _mm_mmask_i64gather_epi32 (__m128i __v1_old, __mmask8 __mask, 11176 __m128i __index, void const *__addr, 11177 int __scale) 11178 { 11179 return (__m128i) __builtin_ia32_gather3div4si ((__v4si) __v1_old, 11180 __addr, 11181 (__v2di) __index, 11182 __mask, __scale); 11183 } 11184 11185 extern __inline __m256i 11186 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11187 _mm256_mmask_i64gather_epi64 (__m256i __v1_old, __mmask8 __mask, 11188 __m256i __index, void const *__addr, 11189 int __scale) 11190 { 11191 return (__m256i) __builtin_ia32_gather3div4di ((__v4di) __v1_old, 11192 __addr, 11193 (__v4di) __index, 11194 __mask, __scale); 11195 } 11196 11197 extern __inline __m128i 11198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11199 _mm_mmask_i64gather_epi64 (__m128i __v1_old, __mmask8 __mask, 11200 __m128i __index, void const *__addr, 11201 int __scale) 11202 { 11203 return (__m128i) __builtin_ia32_gather3div2di ((__v2di) __v1_old, 11204 __addr, 11205 (__v2di) __index, 11206 __mask, __scale); 11207 } 11208 11209 extern __inline void 11210 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11211 _mm256_i32scatter_ps (void *__addr, __m256i __index, 11212 __m256 __v1, const int __scale) 11213 { 11214 __builtin_ia32_scattersiv8sf (__addr, (__mmask8) 0xFF, 11215 (__v8si) __index, (__v8sf) __v1, 11216 __scale); 11217 } 11218 11219 extern __inline void 11220 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11221 _mm256_mask_i32scatter_ps (void *__addr, __mmask8 __mask, 11222 __m256i __index, __m256 __v1, 11223 const int __scale) 11224 { 11225 __builtin_ia32_scattersiv8sf (__addr, __mask, (__v8si) __index, 11226 (__v8sf) __v1, __scale); 11227 } 11228 11229 extern __inline void 11230 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11231 _mm_i32scatter_ps (void *__addr, __m128i __index, __m128 __v1, 11232 const int __scale) 11233 { 11234 __builtin_ia32_scattersiv4sf (__addr, (__mmask8) 0xFF, 11235 (__v4si) __index, (__v4sf) __v1, 11236 __scale); 11237 } 11238 11239 extern __inline void 11240 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11241 _mm_mask_i32scatter_ps (void *__addr, __mmask8 __mask, 11242 __m128i __index, __m128 __v1, 11243 const int __scale) 11244 { 11245 __builtin_ia32_scattersiv4sf (__addr, __mask, (__v4si) __index, 11246 (__v4sf) __v1, __scale); 11247 } 11248 11249 extern __inline void 11250 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11251 _mm256_i32scatter_pd (void *__addr, __m128i __index, 11252 __m256d __v1, const int __scale) 11253 { 11254 __builtin_ia32_scattersiv4df (__addr, (__mmask8) 0xFF, 11255 (__v4si) __index, (__v4df) __v1, 11256 __scale); 11257 } 11258 11259 extern __inline void 11260 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11261 _mm256_mask_i32scatter_pd (void *__addr, __mmask8 __mask, 11262 __m128i __index, __m256d __v1, 11263 const int __scale) 11264 { 11265 __builtin_ia32_scattersiv4df (__addr, __mask, (__v4si) __index, 11266 (__v4df) __v1, __scale); 11267 } 11268 11269 extern __inline void 11270 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11271 _mm_i32scatter_pd (void *__addr, __m128i __index, 11272 __m128d __v1, const int __scale) 11273 { 11274 __builtin_ia32_scattersiv2df (__addr, (__mmask8) 0xFF, 11275 (__v4si) __index, (__v2df) __v1, 11276 __scale); 11277 } 11278 11279 extern __inline void 11280 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11281 _mm_mask_i32scatter_pd (void *__addr, __mmask8 __mask, 11282 __m128i __index, __m128d __v1, 11283 const int __scale) 11284 { 11285 __builtin_ia32_scattersiv2df (__addr, __mask, (__v4si) __index, 11286 (__v2df) __v1, __scale); 11287 } 11288 11289 extern __inline void 11290 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11291 _mm256_i64scatter_ps (void *__addr, __m256i __index, 11292 __m128 __v1, const int __scale) 11293 { 11294 __builtin_ia32_scatterdiv8sf (__addr, (__mmask8) 0xFF, 11295 (__v4di) __index, (__v4sf) __v1, 11296 __scale); 11297 } 11298 11299 extern __inline void 11300 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11301 _mm256_mask_i64scatter_ps (void *__addr, __mmask8 __mask, 11302 __m256i __index, __m128 __v1, 11303 const int __scale) 11304 { 11305 __builtin_ia32_scatterdiv8sf (__addr, __mask, (__v4di) __index, 11306 (__v4sf) __v1, __scale); 11307 } 11308 11309 extern __inline void 11310 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11311 _mm_i64scatter_ps (void *__addr, __m128i __index, __m128 __v1, 11312 const int __scale) 11313 { 11314 __builtin_ia32_scatterdiv4sf (__addr, (__mmask8) 0xFF, 11315 (__v2di) __index, (__v4sf) __v1, 11316 __scale); 11317 } 11318 11319 extern __inline void 11320 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11321 _mm_mask_i64scatter_ps (void *__addr, __mmask8 __mask, 11322 __m128i __index, __m128 __v1, 11323 const int __scale) 11324 { 11325 __builtin_ia32_scatterdiv4sf (__addr, __mask, (__v2di) __index, 11326 (__v4sf) __v1, __scale); 11327 } 11328 11329 extern __inline void 11330 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11331 _mm256_i64scatter_pd (void *__addr, __m256i __index, 11332 __m256d __v1, const int __scale) 11333 { 11334 __builtin_ia32_scatterdiv4df (__addr, (__mmask8) 0xFF, 11335 (__v4di) __index, (__v4df) __v1, 11336 __scale); 11337 } 11338 11339 extern __inline void 11340 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11341 _mm256_mask_i64scatter_pd (void *__addr, __mmask8 __mask, 11342 __m256i __index, __m256d __v1, 11343 const int __scale) 11344 { 11345 __builtin_ia32_scatterdiv4df (__addr, __mask, (__v4di) __index, 11346 (__v4df) __v1, __scale); 11347 } 11348 11349 extern __inline void 11350 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11351 _mm_i64scatter_pd (void *__addr, __m128i __index, 11352 __m128d __v1, const int __scale) 11353 { 11354 __builtin_ia32_scatterdiv2df (__addr, (__mmask8) 0xFF, 11355 (__v2di) __index, (__v2df) __v1, 11356 __scale); 11357 } 11358 11359 extern __inline void 11360 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11361 _mm_mask_i64scatter_pd (void *__addr, __mmask8 __mask, 11362 __m128i __index, __m128d __v1, 11363 const int __scale) 11364 { 11365 __builtin_ia32_scatterdiv2df (__addr, __mask, (__v2di) __index, 11366 (__v2df) __v1, __scale); 11367 } 11368 11369 extern __inline void 11370 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11371 _mm256_i32scatter_epi32 (void *__addr, __m256i __index, 11372 __m256i __v1, const int __scale) 11373 { 11374 __builtin_ia32_scattersiv8si (__addr, (__mmask8) 0xFF, 11375 (__v8si) __index, (__v8si) __v1, 11376 __scale); 11377 } 11378 11379 extern __inline void 11380 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11381 _mm256_mask_i32scatter_epi32 (void *__addr, __mmask8 __mask, 11382 __m256i __index, __m256i __v1, 11383 const int __scale) 11384 { 11385 __builtin_ia32_scattersiv8si (__addr, __mask, (__v8si) __index, 11386 (__v8si) __v1, __scale); 11387 } 11388 11389 extern __inline void 11390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11391 _mm_i32scatter_epi32 (void *__addr, __m128i __index, 11392 __m128i __v1, const int __scale) 11393 { 11394 __builtin_ia32_scattersiv4si (__addr, (__mmask8) 0xFF, 11395 (__v4si) __index, (__v4si) __v1, 11396 __scale); 11397 } 11398 11399 extern __inline void 11400 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11401 _mm_mask_i32scatter_epi32 (void *__addr, __mmask8 __mask, 11402 __m128i __index, __m128i __v1, 11403 const int __scale) 11404 { 11405 __builtin_ia32_scattersiv4si (__addr, __mask, (__v4si) __index, 11406 (__v4si) __v1, __scale); 11407 } 11408 11409 extern __inline void 11410 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11411 _mm256_i32scatter_epi64 (void *__addr, __m128i __index, 11412 __m256i __v1, const int __scale) 11413 { 11414 __builtin_ia32_scattersiv4di (__addr, (__mmask8) 0xFF, 11415 (__v4si) __index, (__v4di) __v1, 11416 __scale); 11417 } 11418 11419 extern __inline void 11420 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11421 _mm256_mask_i32scatter_epi64 (void *__addr, __mmask8 __mask, 11422 __m128i __index, __m256i __v1, 11423 const int __scale) 11424 { 11425 __builtin_ia32_scattersiv4di (__addr, __mask, (__v4si) __index, 11426 (__v4di) __v1, __scale); 11427 } 11428 11429 extern __inline void 11430 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11431 _mm_i32scatter_epi64 (void *__addr, __m128i __index, 11432 __m128i __v1, const int __scale) 11433 { 11434 __builtin_ia32_scattersiv2di (__addr, (__mmask8) 0xFF, 11435 (__v4si) __index, (__v2di) __v1, 11436 __scale); 11437 } 11438 11439 extern __inline void 11440 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11441 _mm_mask_i32scatter_epi64 (void *__addr, __mmask8 __mask, 11442 __m128i __index, __m128i __v1, 11443 const int __scale) 11444 { 11445 __builtin_ia32_scattersiv2di (__addr, __mask, (__v4si) __index, 11446 (__v2di) __v1, __scale); 11447 } 11448 11449 extern __inline void 11450 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11451 _mm256_i64scatter_epi32 (void *__addr, __m256i __index, 11452 __m128i __v1, const int __scale) 11453 { 11454 __builtin_ia32_scatterdiv8si (__addr, (__mmask8) 0xFF, 11455 (__v4di) __index, (__v4si) __v1, 11456 __scale); 11457 } 11458 11459 extern __inline void 11460 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11461 _mm256_mask_i64scatter_epi32 (void *__addr, __mmask8 __mask, 11462 __m256i __index, __m128i __v1, 11463 const int __scale) 11464 { 11465 __builtin_ia32_scatterdiv8si (__addr, __mask, (__v4di) __index, 11466 (__v4si) __v1, __scale); 11467 } 11468 11469 extern __inline void 11470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11471 _mm_i64scatter_epi32 (void *__addr, __m128i __index, 11472 __m128i __v1, const int __scale) 11473 { 11474 __builtin_ia32_scatterdiv4si (__addr, (__mmask8) 0xFF, 11475 (__v2di) __index, (__v4si) __v1, 11476 __scale); 11477 } 11478 11479 extern __inline void 11480 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11481 _mm_mask_i64scatter_epi32 (void *__addr, __mmask8 __mask, 11482 __m128i __index, __m128i __v1, 11483 const int __scale) 11484 { 11485 __builtin_ia32_scatterdiv4si (__addr, __mask, (__v2di) __index, 11486 (__v4si) __v1, __scale); 11487 } 11488 11489 extern __inline void 11490 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11491 _mm256_i64scatter_epi64 (void *__addr, __m256i __index, 11492 __m256i __v1, const int __scale) 11493 { 11494 __builtin_ia32_scatterdiv4di (__addr, (__mmask8) 0xFF, 11495 (__v4di) __index, (__v4di) __v1, 11496 __scale); 11497 } 11498 11499 extern __inline void 11500 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11501 _mm256_mask_i64scatter_epi64 (void *__addr, __mmask8 __mask, 11502 __m256i __index, __m256i __v1, 11503 const int __scale) 11504 { 11505 __builtin_ia32_scatterdiv4di (__addr, __mask, (__v4di) __index, 11506 (__v4di) __v1, __scale); 11507 } 11508 11509 extern __inline void 11510 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11511 _mm_i64scatter_epi64 (void *__addr, __m128i __index, 11512 __m128i __v1, const int __scale) 11513 { 11514 __builtin_ia32_scatterdiv2di (__addr, (__mmask8) 0xFF, 11515 (__v2di) __index, (__v2di) __v1, 11516 __scale); 11517 } 11518 11519 extern __inline void 11520 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11521 _mm_mask_i64scatter_epi64 (void *__addr, __mmask8 __mask, 11522 __m128i __index, __m128i __v1, 11523 const int __scale) 11524 { 11525 __builtin_ia32_scatterdiv2di (__addr, __mask, (__v2di) __index, 11526 (__v2di) __v1, __scale); 11527 } 11528 11529 extern __inline __m256i 11530 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11531 _mm256_mask_shuffle_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 11532 _MM_PERM_ENUM __mask) 11533 { 11534 return (__m256i) __builtin_ia32_pshufd256_mask ((__v8si) __A, __mask, 11535 (__v8si) __W, 11536 (__mmask8) __U); 11537 } 11538 11539 extern __inline __m256i 11540 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11541 _mm256_maskz_shuffle_epi32 (__mmask8 __U, __m256i __A, 11542 _MM_PERM_ENUM __mask) 11543 { 11544 return (__m256i) __builtin_ia32_pshufd256_mask ((__v8si) __A, __mask, 11545 (__v8si) 11546 _mm256_setzero_si256 (), 11547 (__mmask8) __U); 11548 } 11549 11550 extern __inline __m128i 11551 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11552 _mm_mask_shuffle_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 11553 _MM_PERM_ENUM __mask) 11554 { 11555 return (__m128i) __builtin_ia32_pshufd128_mask ((__v4si) __A, __mask, 11556 (__v4si) __W, 11557 (__mmask8) __U); 11558 } 11559 11560 extern __inline __m128i 11561 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11562 _mm_maskz_shuffle_epi32 (__mmask8 __U, __m128i __A, 11563 _MM_PERM_ENUM __mask) 11564 { 11565 return (__m128i) __builtin_ia32_pshufd128_mask ((__v4si) __A, __mask, 11566 (__v4si) 11567 _mm_setzero_si128 (), 11568 (__mmask8) __U); 11569 } 11570 11571 extern __inline __m256i 11572 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11573 _mm256_rol_epi32 (__m256i __A, const int __B) 11574 { 11575 return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B, 11576 (__v8si) 11577 _mm256_setzero_si256 (), 11578 (__mmask8) -1); 11579 } 11580 11581 extern __inline __m256i 11582 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11583 _mm256_mask_rol_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 11584 const int __B) 11585 { 11586 return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B, 11587 (__v8si) __W, 11588 (__mmask8) __U); 11589 } 11590 11591 extern __inline __m256i 11592 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11593 _mm256_maskz_rol_epi32 (__mmask8 __U, __m256i __A, const int __B) 11594 { 11595 return (__m256i) __builtin_ia32_prold256_mask ((__v8si) __A, __B, 11596 (__v8si) 11597 _mm256_setzero_si256 (), 11598 (__mmask8) __U); 11599 } 11600 11601 extern __inline __m128i 11602 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11603 _mm_rol_epi32 (__m128i __A, const int __B) 11604 { 11605 return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B, 11606 (__v4si) 11607 _mm_setzero_si128 (), 11608 (__mmask8) -1); 11609 } 11610 11611 extern __inline __m128i 11612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11613 _mm_mask_rol_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 11614 const int __B) 11615 { 11616 return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B, 11617 (__v4si) __W, 11618 (__mmask8) __U); 11619 } 11620 11621 extern __inline __m128i 11622 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11623 _mm_maskz_rol_epi32 (__mmask8 __U, __m128i __A, const int __B) 11624 { 11625 return (__m128i) __builtin_ia32_prold128_mask ((__v4si) __A, __B, 11626 (__v4si) 11627 _mm_setzero_si128 (), 11628 (__mmask8) __U); 11629 } 11630 11631 extern __inline __m256i 11632 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11633 _mm256_ror_epi32 (__m256i __A, const int __B) 11634 { 11635 return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B, 11636 (__v8si) 11637 _mm256_setzero_si256 (), 11638 (__mmask8) -1); 11639 } 11640 11641 extern __inline __m256i 11642 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11643 _mm256_mask_ror_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 11644 const int __B) 11645 { 11646 return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B, 11647 (__v8si) __W, 11648 (__mmask8) __U); 11649 } 11650 11651 extern __inline __m256i 11652 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11653 _mm256_maskz_ror_epi32 (__mmask8 __U, __m256i __A, const int __B) 11654 { 11655 return (__m256i) __builtin_ia32_prord256_mask ((__v8si) __A, __B, 11656 (__v8si) 11657 _mm256_setzero_si256 (), 11658 (__mmask8) __U); 11659 } 11660 11661 extern __inline __m128i 11662 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11663 _mm_ror_epi32 (__m128i __A, const int __B) 11664 { 11665 return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B, 11666 (__v4si) 11667 _mm_setzero_si128 (), 11668 (__mmask8) -1); 11669 } 11670 11671 extern __inline __m128i 11672 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11673 _mm_mask_ror_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 11674 const int __B) 11675 { 11676 return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B, 11677 (__v4si) __W, 11678 (__mmask8) __U); 11679 } 11680 11681 extern __inline __m128i 11682 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11683 _mm_maskz_ror_epi32 (__mmask8 __U, __m128i __A, const int __B) 11684 { 11685 return (__m128i) __builtin_ia32_prord128_mask ((__v4si) __A, __B, 11686 (__v4si) 11687 _mm_setzero_si128 (), 11688 (__mmask8) __U); 11689 } 11690 11691 extern __inline __m256i 11692 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11693 _mm256_rol_epi64 (__m256i __A, const int __B) 11694 { 11695 return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B, 11696 (__v4di) 11697 _mm256_setzero_si256 (), 11698 (__mmask8) -1); 11699 } 11700 11701 extern __inline __m256i 11702 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11703 _mm256_mask_rol_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 11704 const int __B) 11705 { 11706 return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B, 11707 (__v4di) __W, 11708 (__mmask8) __U); 11709 } 11710 11711 extern __inline __m256i 11712 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11713 _mm256_maskz_rol_epi64 (__mmask8 __U, __m256i __A, const int __B) 11714 { 11715 return (__m256i) __builtin_ia32_prolq256_mask ((__v4di) __A, __B, 11716 (__v4di) 11717 _mm256_setzero_si256 (), 11718 (__mmask8) __U); 11719 } 11720 11721 extern __inline __m128i 11722 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11723 _mm_rol_epi64 (__m128i __A, const int __B) 11724 { 11725 return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B, 11726 (__v2di) 11727 _mm_setzero_si128 (), 11728 (__mmask8) -1); 11729 } 11730 11731 extern __inline __m128i 11732 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11733 _mm_mask_rol_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 11734 const int __B) 11735 { 11736 return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B, 11737 (__v2di) __W, 11738 (__mmask8) __U); 11739 } 11740 11741 extern __inline __m128i 11742 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11743 _mm_maskz_rol_epi64 (__mmask8 __U, __m128i __A, const int __B) 11744 { 11745 return (__m128i) __builtin_ia32_prolq128_mask ((__v2di) __A, __B, 11746 (__v2di) 11747 _mm_setzero_si128 (), 11748 (__mmask8) __U); 11749 } 11750 11751 extern __inline __m256i 11752 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11753 _mm256_ror_epi64 (__m256i __A, const int __B) 11754 { 11755 return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B, 11756 (__v4di) 11757 _mm256_setzero_si256 (), 11758 (__mmask8) -1); 11759 } 11760 11761 extern __inline __m256i 11762 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11763 _mm256_mask_ror_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 11764 const int __B) 11765 { 11766 return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B, 11767 (__v4di) __W, 11768 (__mmask8) __U); 11769 } 11770 11771 extern __inline __m256i 11772 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11773 _mm256_maskz_ror_epi64 (__mmask8 __U, __m256i __A, const int __B) 11774 { 11775 return (__m256i) __builtin_ia32_prorq256_mask ((__v4di) __A, __B, 11776 (__v4di) 11777 _mm256_setzero_si256 (), 11778 (__mmask8) __U); 11779 } 11780 11781 extern __inline __m128i 11782 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11783 _mm_ror_epi64 (__m128i __A, const int __B) 11784 { 11785 return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B, 11786 (__v2di) 11787 _mm_setzero_si128 (), 11788 (__mmask8) -1); 11789 } 11790 11791 extern __inline __m128i 11792 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11793 _mm_mask_ror_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 11794 const int __B) 11795 { 11796 return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B, 11797 (__v2di) __W, 11798 (__mmask8) __U); 11799 } 11800 11801 extern __inline __m128i 11802 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11803 _mm_maskz_ror_epi64 (__mmask8 __U, __m128i __A, const int __B) 11804 { 11805 return (__m128i) __builtin_ia32_prorq128_mask ((__v2di) __A, __B, 11806 (__v2di) 11807 _mm_setzero_si128 (), 11808 (__mmask8) __U); 11809 } 11810 11811 extern __inline __m128i 11812 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11813 _mm_alignr_epi32 (__m128i __A, __m128i __B, const int __imm) 11814 { 11815 return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A, 11816 (__v4si) __B, __imm, 11817 (__v4si) 11818 _mm_setzero_si128 (), 11819 (__mmask8) -1); 11820 } 11821 11822 extern __inline __m128i 11823 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11824 _mm_mask_alignr_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 11825 __m128i __B, const int __imm) 11826 { 11827 return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A, 11828 (__v4si) __B, __imm, 11829 (__v4si) __W, 11830 (__mmask8) __U); 11831 } 11832 11833 extern __inline __m128i 11834 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11835 _mm_maskz_alignr_epi32 (__mmask8 __U, __m128i __A, __m128i __B, 11836 const int __imm) 11837 { 11838 return (__m128i) __builtin_ia32_alignd128_mask ((__v4si) __A, 11839 (__v4si) __B, __imm, 11840 (__v4si) 11841 _mm_setzero_si128 (), 11842 (__mmask8) __U); 11843 } 11844 11845 extern __inline __m128i 11846 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11847 _mm_alignr_epi64 (__m128i __A, __m128i __B, const int __imm) 11848 { 11849 return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A, 11850 (__v2di) __B, __imm, 11851 (__v2di) 11852 _mm_setzero_si128 (), 11853 (__mmask8) -1); 11854 } 11855 11856 extern __inline __m128i 11857 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11858 _mm_mask_alignr_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 11859 __m128i __B, const int __imm) 11860 { 11861 return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A, 11862 (__v2di) __B, __imm, 11863 (__v2di) __W, 11864 (__mmask8) __U); 11865 } 11866 11867 extern __inline __m128i 11868 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11869 _mm_maskz_alignr_epi64 (__mmask8 __U, __m128i __A, __m128i __B, 11870 const int __imm) 11871 { 11872 return (__m128i) __builtin_ia32_alignq128_mask ((__v2di) __A, 11873 (__v2di) __B, __imm, 11874 (__v2di) 11875 _mm_setzero_si128 (), 11876 (__mmask8) __U); 11877 } 11878 11879 extern __inline __m256i 11880 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11881 _mm256_alignr_epi32 (__m256i __A, __m256i __B, const int __imm) 11882 { 11883 return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A, 11884 (__v8si) __B, __imm, 11885 (__v8si) 11886 _mm256_setzero_si256 (), 11887 (__mmask8) -1); 11888 } 11889 11890 extern __inline __m256i 11891 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11892 _mm256_mask_alignr_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 11893 __m256i __B, const int __imm) 11894 { 11895 return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A, 11896 (__v8si) __B, __imm, 11897 (__v8si) __W, 11898 (__mmask8) __U); 11899 } 11900 11901 extern __inline __m256i 11902 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11903 _mm256_maskz_alignr_epi32 (__mmask8 __U, __m256i __A, __m256i __B, 11904 const int __imm) 11905 { 11906 return (__m256i) __builtin_ia32_alignd256_mask ((__v8si) __A, 11907 (__v8si) __B, __imm, 11908 (__v8si) 11909 _mm256_setzero_si256 (), 11910 (__mmask8) __U); 11911 } 11912 11913 extern __inline __m256i 11914 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11915 _mm256_alignr_epi64 (__m256i __A, __m256i __B, const int __imm) 11916 { 11917 return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A, 11918 (__v4di) __B, __imm, 11919 (__v4di) 11920 _mm256_setzero_si256 (), 11921 (__mmask8) -1); 11922 } 11923 11924 extern __inline __m256i 11925 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11926 _mm256_mask_alignr_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 11927 __m256i __B, const int __imm) 11928 { 11929 return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A, 11930 (__v4di) __B, __imm, 11931 (__v4di) __W, 11932 (__mmask8) __U); 11933 } 11934 11935 extern __inline __m256i 11936 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11937 _mm256_maskz_alignr_epi64 (__mmask8 __U, __m256i __A, __m256i __B, 11938 const int __imm) 11939 { 11940 return (__m256i) __builtin_ia32_alignq256_mask ((__v4di) __A, 11941 (__v4di) __B, __imm, 11942 (__v4di) 11943 _mm256_setzero_si256 (), 11944 (__mmask8) __U); 11945 } 11946 11947 extern __inline __m128i 11948 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11949 _mm_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m128 __A, 11950 const int __I) 11951 { 11952 return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, __I, 11953 (__v8hi) __W, 11954 (__mmask8) __U); 11955 } 11956 11957 extern __inline __m128i 11958 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11959 _mm_maskz_cvtps_ph (__mmask8 __U, __m128 __A, const int __I) 11960 { 11961 return (__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf) __A, __I, 11962 (__v8hi) 11963 _mm_setzero_si128 (), 11964 (__mmask8) __U); 11965 } 11966 11967 extern __inline __m128i 11968 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11969 _mm256_mask_cvtps_ph (__m128i __W, __mmask8 __U, __m256 __A, 11970 const int __I) 11971 { 11972 return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, __I, 11973 (__v8hi) __W, 11974 (__mmask8) __U); 11975 } 11976 11977 extern __inline __m128i 11978 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11979 _mm256_maskz_cvtps_ph (__mmask8 __U, __m256 __A, const int __I) 11980 { 11981 return (__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf) __A, __I, 11982 (__v8hi) 11983 _mm_setzero_si128 (), 11984 (__mmask8) __U); 11985 } 11986 11987 extern __inline __m256i 11988 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11989 _mm256_mask_srai_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 11990 const int __imm) 11991 { 11992 return (__m256i) __builtin_ia32_psradi256_mask ((__v8si) __A, __imm, 11993 (__v8si) __W, 11994 (__mmask8) __U); 11995 } 11996 11997 extern __inline __m256i 11998 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 11999 _mm256_maskz_srai_epi32 (__mmask8 __U, __m256i __A, const int __imm) 12000 { 12001 return (__m256i) __builtin_ia32_psradi256_mask ((__v8si) __A, __imm, 12002 (__v8si) 12003 _mm256_setzero_si256 (), 12004 (__mmask8) __U); 12005 } 12006 12007 extern __inline __m128i 12008 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12009 _mm_mask_srai_epi32 (__m128i __W, __mmask8 __U, __m128i __A, 12010 const int __imm) 12011 { 12012 return (__m128i) __builtin_ia32_psradi128_mask ((__v4si) __A, __imm, 12013 (__v4si) __W, 12014 (__mmask8) __U); 12015 } 12016 12017 extern __inline __m128i 12018 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12019 _mm_maskz_srai_epi32 (__mmask8 __U, __m128i __A, const int __imm) 12020 { 12021 return (__m128i) __builtin_ia32_psradi128_mask ((__v4si) __A, __imm, 12022 (__v4si) 12023 _mm_setzero_si128 (), 12024 (__mmask8) __U); 12025 } 12026 12027 extern __inline __m256i 12028 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12029 _mm256_srai_epi64 (__m256i __A, const int __imm) 12030 { 12031 return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm, 12032 (__v4di) 12033 _mm256_setzero_si256 (), 12034 (__mmask8) -1); 12035 } 12036 12037 extern __inline __m256i 12038 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12039 _mm256_mask_srai_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 12040 const int __imm) 12041 { 12042 return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm, 12043 (__v4di) __W, 12044 (__mmask8) __U); 12045 } 12046 12047 extern __inline __m256i 12048 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12049 _mm256_maskz_srai_epi64 (__mmask8 __U, __m256i __A, const int __imm) 12050 { 12051 return (__m256i) __builtin_ia32_psraqi256_mask ((__v4di) __A, __imm, 12052 (__v4di) 12053 _mm256_setzero_si256 (), 12054 (__mmask8) __U); 12055 } 12056 12057 extern __inline __m128i 12058 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12059 _mm_srai_epi64 (__m128i __A, const int __imm) 12060 { 12061 return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm, 12062 (__v2di) 12063 _mm_setzero_si128 (), 12064 (__mmask8) -1); 12065 } 12066 12067 extern __inline __m128i 12068 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12069 _mm_mask_srai_epi64 (__m128i __W, __mmask8 __U, __m128i __A, 12070 const int __imm) 12071 { 12072 return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm, 12073 (__v2di) __W, 12074 (__mmask8) __U); 12075 } 12076 12077 extern __inline __m128i 12078 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12079 _mm_maskz_srai_epi64 (__mmask8 __U, __m128i __A, const int __imm) 12080 { 12081 return (__m128i) __builtin_ia32_psraqi128_mask ((__v2di) __A, __imm, 12082 (__v2di) 12083 _mm_setzero_si128 (), 12084 (__mmask8) __U); 12085 } 12086 12087 extern __inline __m128i 12088 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12089 _mm_mask_slli_epi32 (__m128i __W, __mmask8 __U, __m128i __A, int __B) 12090 { 12091 return (__m128i) __builtin_ia32_pslldi128_mask ((__v4si) __A, __B, 12092 (__v4si) __W, 12093 (__mmask8) __U); 12094 } 12095 12096 extern __inline __m128i 12097 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12098 _mm_maskz_slli_epi32 (__mmask8 __U, __m128i __A, int __B) 12099 { 12100 return (__m128i) __builtin_ia32_pslldi128_mask ((__v4si) __A, __B, 12101 (__v4si) 12102 _mm_setzero_si128 (), 12103 (__mmask8) __U); 12104 } 12105 12106 extern __inline __m128i 12107 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12108 _mm_mask_slli_epi64 (__m128i __W, __mmask8 __U, __m128i __A, int __B) 12109 { 12110 return (__m128i) __builtin_ia32_psllqi128_mask ((__v2di) __A, __B, 12111 (__v2di) __W, 12112 (__mmask8) __U); 12113 } 12114 12115 extern __inline __m128i 12116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12117 _mm_maskz_slli_epi64 (__mmask8 __U, __m128i __A, int __B) 12118 { 12119 return (__m128i) __builtin_ia32_psllqi128_mask ((__v2di) __A, __B, 12120 (__v2di) 12121 _mm_setzero_si128 (), 12122 (__mmask8) __U); 12123 } 12124 12125 extern __inline __m256i 12126 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12127 _mm256_mask_slli_epi32 (__m256i __W, __mmask8 __U, __m256i __A, 12128 int __B) 12129 { 12130 return (__m256i) __builtin_ia32_pslldi256_mask ((__v8si) __A, __B, 12131 (__v8si) __W, 12132 (__mmask8) __U); 12133 } 12134 12135 extern __inline __m256i 12136 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12137 _mm256_maskz_slli_epi32 (__mmask8 __U, __m256i __A, int __B) 12138 { 12139 return (__m256i) __builtin_ia32_pslldi256_mask ((__v8si) __A, __B, 12140 (__v8si) 12141 _mm256_setzero_si256 (), 12142 (__mmask8) __U); 12143 } 12144 12145 extern __inline __m256i 12146 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12147 _mm256_mask_slli_epi64 (__m256i __W, __mmask8 __U, __m256i __A, 12148 int __B) 12149 { 12150 return (__m256i) __builtin_ia32_psllqi256_mask ((__v4di) __A, __B, 12151 (__v4di) __W, 12152 (__mmask8) __U); 12153 } 12154 12155 extern __inline __m256i 12156 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12157 _mm256_maskz_slli_epi64 (__mmask8 __U, __m256i __A, int __B) 12158 { 12159 return (__m256i) __builtin_ia32_psllqi256_mask ((__v4di) __A, __B, 12160 (__v4di) 12161 _mm256_setzero_si256 (), 12162 (__mmask8) __U); 12163 } 12164 12165 extern __inline __m256d 12166 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12167 _mm256_mask_permutex_pd (__m256d __W, __mmask8 __U, __m256d __X, 12168 const int __imm) 12169 { 12170 return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __imm, 12171 (__v4df) __W, 12172 (__mmask8) __U); 12173 } 12174 12175 extern __inline __m256d 12176 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12177 _mm256_maskz_permutex_pd (__mmask8 __U, __m256d __X, const int __imm) 12178 { 12179 return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __imm, 12180 (__v4df) 12181 _mm256_setzero_pd (), 12182 (__mmask8) __U); 12183 } 12184 12185 extern __inline __m256d 12186 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12187 _mm256_mask_permute_pd (__m256d __W, __mmask8 __U, __m256d __X, 12188 const int __C) 12189 { 12190 return (__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df) __X, __C, 12191 (__v4df) __W, 12192 (__mmask8) __U); 12193 } 12194 12195 extern __inline __m256d 12196 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12197 _mm256_maskz_permute_pd (__mmask8 __U, __m256d __X, const int __C) 12198 { 12199 return (__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df) __X, __C, 12200 (__v4df) 12201 _mm256_setzero_pd (), 12202 (__mmask8) __U); 12203 } 12204 12205 extern __inline __m128d 12206 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12207 _mm_mask_permute_pd (__m128d __W, __mmask8 __U, __m128d __X, 12208 const int __C) 12209 { 12210 return (__m128d) __builtin_ia32_vpermilpd_mask ((__v2df) __X, __C, 12211 (__v2df) __W, 12212 (__mmask8) __U); 12213 } 12214 12215 extern __inline __m128d 12216 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12217 _mm_maskz_permute_pd (__mmask8 __U, __m128d __X, const int __C) 12218 { 12219 return (__m128d) __builtin_ia32_vpermilpd_mask ((__v2df) __X, __C, 12220 (__v2df) 12221 _mm_setzero_pd (), 12222 (__mmask8) __U); 12223 } 12224 12225 extern __inline __m256 12226 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12227 _mm256_mask_permute_ps (__m256 __W, __mmask8 __U, __m256 __X, 12228 const int __C) 12229 { 12230 return (__m256) __builtin_ia32_vpermilps256_mask ((__v8sf) __X, __C, 12231 (__v8sf) __W, 12232 (__mmask8) __U); 12233 } 12234 12235 extern __inline __m256 12236 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12237 _mm256_maskz_permute_ps (__mmask8 __U, __m256 __X, const int __C) 12238 { 12239 return (__m256) __builtin_ia32_vpermilps256_mask ((__v8sf) __X, __C, 12240 (__v8sf) 12241 _mm256_setzero_ps (), 12242 (__mmask8) __U); 12243 } 12244 12245 extern __inline __m128 12246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12247 _mm_mask_permute_ps (__m128 __W, __mmask8 __U, __m128 __X, 12248 const int __C) 12249 { 12250 return (__m128) __builtin_ia32_vpermilps_mask ((__v4sf) __X, __C, 12251 (__v4sf) __W, 12252 (__mmask8) __U); 12253 } 12254 12255 extern __inline __m128 12256 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12257 _mm_maskz_permute_ps (__mmask8 __U, __m128 __X, const int __C) 12258 { 12259 return (__m128) __builtin_ia32_vpermilps_mask ((__v4sf) __X, __C, 12260 (__v4sf) 12261 _mm_setzero_ps (), 12262 (__mmask8) __U); 12263 } 12264 12265 extern __inline __m256d 12266 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12267 _mm256_mask_blend_pd (__mmask8 __U, __m256d __A, __m256d __W) 12268 { 12269 return (__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) __A, 12270 (__v4df) __W, 12271 (__mmask8) __U); 12272 } 12273 12274 extern __inline __m256 12275 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12276 _mm256_mask_blend_ps (__mmask8 __U, __m256 __A, __m256 __W) 12277 { 12278 return (__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) __A, 12279 (__v8sf) __W, 12280 (__mmask8) __U); 12281 } 12282 12283 extern __inline __m256i 12284 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12285 _mm256_mask_blend_epi64 (__mmask8 __U, __m256i __A, __m256i __W) 12286 { 12287 return (__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) __A, 12288 (__v4di) __W, 12289 (__mmask8) __U); 12290 } 12291 12292 extern __inline __m256i 12293 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12294 _mm256_mask_blend_epi32 (__mmask8 __U, __m256i __A, __m256i __W) 12295 { 12296 return (__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) __A, 12297 (__v8si) __W, 12298 (__mmask8) __U); 12299 } 12300 12301 extern __inline __m128d 12302 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12303 _mm_mask_blend_pd (__mmask8 __U, __m128d __A, __m128d __W) 12304 { 12305 return (__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) __A, 12306 (__v2df) __W, 12307 (__mmask8) __U); 12308 } 12309 12310 extern __inline __m128 12311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12312 _mm_mask_blend_ps (__mmask8 __U, __m128 __A, __m128 __W) 12313 { 12314 return (__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) __A, 12315 (__v4sf) __W, 12316 (__mmask8) __U); 12317 } 12318 12319 extern __inline __m128i 12320 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12321 _mm_mask_blend_epi64 (__mmask8 __U, __m128i __A, __m128i __W) 12322 { 12323 return (__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) __A, 12324 (__v2di) __W, 12325 (__mmask8) __U); 12326 } 12327 12328 extern __inline __m128i 12329 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12330 _mm_mask_blend_epi32 (__mmask8 __U, __m128i __A, __m128i __W) 12331 { 12332 return (__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) __A, 12333 (__v4si) __W, 12334 (__mmask8) __U); 12335 } 12336 12337 extern __inline __mmask8 12338 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12339 _mm256_cmp_epi64_mask (__m256i __X, __m256i __Y, const int __P) 12340 { 12341 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 12342 (__v4di) __Y, __P, 12343 (__mmask8) -1); 12344 } 12345 12346 extern __inline __mmask8 12347 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12348 _mm256_cmp_epi32_mask (__m256i __X, __m256i __Y, const int __P) 12349 { 12350 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 12351 (__v8si) __Y, __P, 12352 (__mmask8) -1); 12353 } 12354 12355 extern __inline __mmask8 12356 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12357 _mm256_cmp_epu64_mask (__m256i __X, __m256i __Y, const int __P) 12358 { 12359 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 12360 (__v4di) __Y, __P, 12361 (__mmask8) -1); 12362 } 12363 12364 extern __inline __mmask8 12365 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12366 _mm256_cmp_epu32_mask (__m256i __X, __m256i __Y, const int __P) 12367 { 12368 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 12369 (__v8si) __Y, __P, 12370 (__mmask8) -1); 12371 } 12372 12373 extern __inline __mmask8 12374 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12375 _mm256_cmp_pd_mask (__m256d __X, __m256d __Y, const int __P) 12376 { 12377 return (__mmask8) __builtin_ia32_cmppd256_mask ((__v4df) __X, 12378 (__v4df) __Y, __P, 12379 (__mmask8) -1); 12380 } 12381 12382 extern __inline __mmask8 12383 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12384 _mm256_cmp_ps_mask (__m256 __X, __m256 __Y, const int __P) 12385 { 12386 return (__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf) __X, 12387 (__v8sf) __Y, __P, 12388 (__mmask8) -1); 12389 } 12390 12391 extern __inline __mmask8 12392 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12393 _mm256_mask_cmp_epi64_mask (__mmask8 __U, __m256i __X, __m256i __Y, 12394 const int __P) 12395 { 12396 return (__mmask8) __builtin_ia32_cmpq256_mask ((__v4di) __X, 12397 (__v4di) __Y, __P, 12398 (__mmask8) __U); 12399 } 12400 12401 extern __inline __mmask8 12402 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12403 _mm256_mask_cmp_epi32_mask (__mmask8 __U, __m256i __X, __m256i __Y, 12404 const int __P) 12405 { 12406 return (__mmask8) __builtin_ia32_cmpd256_mask ((__v8si) __X, 12407 (__v8si) __Y, __P, 12408 (__mmask8) __U); 12409 } 12410 12411 extern __inline __mmask8 12412 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12413 _mm256_mask_cmp_epu64_mask (__mmask8 __U, __m256i __X, __m256i __Y, 12414 const int __P) 12415 { 12416 return (__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di) __X, 12417 (__v4di) __Y, __P, 12418 (__mmask8) __U); 12419 } 12420 12421 extern __inline __mmask8 12422 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12423 _mm256_mask_cmp_epu32_mask (__mmask8 __U, __m256i __X, __m256i __Y, 12424 const int __P) 12425 { 12426 return (__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si) __X, 12427 (__v8si) __Y, __P, 12428 (__mmask8) __U); 12429 } 12430 12431 extern __inline __mmask8 12432 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12433 _mm256_mask_cmp_pd_mask (__mmask8 __U, __m256d __X, __m256d __Y, 12434 const int __P) 12435 { 12436 return (__mmask8) __builtin_ia32_cmppd256_mask ((__v4df) __X, 12437 (__v4df) __Y, __P, 12438 (__mmask8) __U); 12439 } 12440 12441 extern __inline __mmask8 12442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12443 _mm256_mask_cmp_ps_mask (__mmask8 __U, __m256 __X, __m256 __Y, 12444 const int __P) 12445 { 12446 return (__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf) __X, 12447 (__v8sf) __Y, __P, 12448 (__mmask8) __U); 12449 } 12450 12451 extern __inline __mmask8 12452 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12453 _mm_cmp_epi64_mask (__m128i __X, __m128i __Y, const int __P) 12454 { 12455 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 12456 (__v2di) __Y, __P, 12457 (__mmask8) -1); 12458 } 12459 12460 extern __inline __mmask8 12461 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12462 _mm_cmp_epi32_mask (__m128i __X, __m128i __Y, const int __P) 12463 { 12464 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 12465 (__v4si) __Y, __P, 12466 (__mmask8) -1); 12467 } 12468 12469 extern __inline __mmask8 12470 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12471 _mm_cmp_epu64_mask (__m128i __X, __m128i __Y, const int __P) 12472 { 12473 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 12474 (__v2di) __Y, __P, 12475 (__mmask8) -1); 12476 } 12477 12478 extern __inline __mmask8 12479 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12480 _mm_cmp_epu32_mask (__m128i __X, __m128i __Y, const int __P) 12481 { 12482 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 12483 (__v4si) __Y, __P, 12484 (__mmask8) -1); 12485 } 12486 12487 extern __inline __mmask8 12488 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12489 _mm_cmp_pd_mask (__m128d __X, __m128d __Y, const int __P) 12490 { 12491 return (__mmask8) __builtin_ia32_cmppd128_mask ((__v2df) __X, 12492 (__v2df) __Y, __P, 12493 (__mmask8) -1); 12494 } 12495 12496 extern __inline __mmask8 12497 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12498 _mm_cmp_ps_mask (__m128 __X, __m128 __Y, const int __P) 12499 { 12500 return (__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf) __X, 12501 (__v4sf) __Y, __P, 12502 (__mmask8) -1); 12503 } 12504 12505 extern __inline __mmask8 12506 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12507 _mm_mask_cmp_epi64_mask (__mmask8 __U, __m128i __X, __m128i __Y, 12508 const int __P) 12509 { 12510 return (__mmask8) __builtin_ia32_cmpq128_mask ((__v2di) __X, 12511 (__v2di) __Y, __P, 12512 (__mmask8) __U); 12513 } 12514 12515 extern __inline __mmask8 12516 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12517 _mm_mask_cmp_epi32_mask (__mmask8 __U, __m128i __X, __m128i __Y, 12518 const int __P) 12519 { 12520 return (__mmask8) __builtin_ia32_cmpd128_mask ((__v4si) __X, 12521 (__v4si) __Y, __P, 12522 (__mmask8) __U); 12523 } 12524 12525 extern __inline __mmask8 12526 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12527 _mm_mask_cmp_epu64_mask (__mmask8 __U, __m128i __X, __m128i __Y, 12528 const int __P) 12529 { 12530 return (__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di) __X, 12531 (__v2di) __Y, __P, 12532 (__mmask8) __U); 12533 } 12534 12535 extern __inline __mmask8 12536 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12537 _mm_mask_cmp_epu32_mask (__mmask8 __U, __m128i __X, __m128i __Y, 12538 const int __P) 12539 { 12540 return (__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si) __X, 12541 (__v4si) __Y, __P, 12542 (__mmask8) __U); 12543 } 12544 12545 extern __inline __mmask8 12546 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12547 _mm_mask_cmp_pd_mask (__mmask8 __U, __m128d __X, __m128d __Y, 12548 const int __P) 12549 { 12550 return (__mmask8) __builtin_ia32_cmppd128_mask ((__v2df) __X, 12551 (__v2df) __Y, __P, 12552 (__mmask8) __U); 12553 } 12554 12555 extern __inline __mmask8 12556 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12557 _mm_mask_cmp_ps_mask (__mmask8 __U, __m128 __X, __m128 __Y, 12558 const int __P) 12559 { 12560 return (__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf) __X, 12561 (__v4sf) __Y, __P, 12562 (__mmask8) __U); 12563 } 12564 12565 extern __inline __m256d 12566 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 12567 _mm256_permutex_pd (__m256d __X, const int __M) 12568 { 12569 return (__m256d) __builtin_ia32_permdf256_mask ((__v4df) __X, __M, 12570 (__v4df) 12571 _mm256_undefined_pd (), 12572 (__mmask8) -1); 12573 } 12574 12575 #else 12576 #define _mm256_permutex_pd(X, M) \ 12577 ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(X), (int)(M), \ 12578 (__v4df)(__m256d) \ 12579 _mm256_undefined_pd (), \ 12580 (__mmask8)-1)) 12581 12582 #define _mm256_permutex_epi64(X, I) \ 12583 ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X), \ 12584 (int)(I), \ 12585 (__v4di)(__m256i) \ 12586 (_mm256_setzero_si256 ()),\ 12587 (__mmask8) -1)) 12588 12589 #define _mm256_maskz_permutex_epi64(M, X, I) \ 12590 ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X), \ 12591 (int)(I), \ 12592 (__v4di)(__m256i) \ 12593 (_mm256_setzero_si256 ()),\ 12594 (__mmask8)(M))) 12595 12596 #define _mm256_mask_permutex_epi64(W, M, X, I) \ 12597 ((__m256i) __builtin_ia32_permdi256_mask ((__v4di)(__m256i)(X), \ 12598 (int)(I), \ 12599 (__v4di)(__m256i)(W), \ 12600 (__mmask8)(M))) 12601 12602 #define _mm256_insertf32x4(X, Y, C) \ 12603 ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X), \ 12604 (__v4sf)(__m128) (Y), (int) (C), \ 12605 (__v8sf)(__m256)_mm256_setzero_ps (), \ 12606 (__mmask8)-1)) 12607 12608 #define _mm256_mask_insertf32x4(W, U, X, Y, C) \ 12609 ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X), \ 12610 (__v4sf)(__m128) (Y), (int) (C), \ 12611 (__v8sf)(__m256)(W), \ 12612 (__mmask8)(U))) 12613 12614 #define _mm256_maskz_insertf32x4(U, X, Y, C) \ 12615 ((__m256) __builtin_ia32_insertf32x4_256_mask ((__v8sf)(__m256) (X), \ 12616 (__v4sf)(__m128) (Y), (int) (C), \ 12617 (__v8sf)(__m256)_mm256_setzero_ps (), \ 12618 (__mmask8)(U))) 12619 12620 #define _mm256_inserti32x4(X, Y, C) \ 12621 ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\ 12622 (__v4si)(__m128i) (Y), (int) (C), \ 12623 (__v8si)(__m256i)_mm256_setzero_si256 (), \ 12624 (__mmask8)-1)) 12625 12626 #define _mm256_mask_inserti32x4(W, U, X, Y, C) \ 12627 ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\ 12628 (__v4si)(__m128i) (Y), (int) (C), \ 12629 (__v8si)(__m256i)(W), \ 12630 (__mmask8)(U))) 12631 12632 #define _mm256_maskz_inserti32x4(U, X, Y, C) \ 12633 ((__m256i) __builtin_ia32_inserti32x4_256_mask ((__v8si)(__m256i) (X),\ 12634 (__v4si)(__m128i) (Y), (int) (C), \ 12635 (__v8si)(__m256i)_mm256_setzero_si256 (), \ 12636 (__mmask8)(U))) 12637 12638 #define _mm256_extractf32x4_ps(X, C) \ 12639 ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \ 12640 (int) (C), \ 12641 (__v4sf)(__m128)_mm_setzero_ps (), \ 12642 (__mmask8)-1)) 12643 12644 #define _mm256_mask_extractf32x4_ps(W, U, X, C) \ 12645 ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \ 12646 (int) (C), \ 12647 (__v4sf)(__m128)(W), \ 12648 (__mmask8)(U))) 12649 12650 #define _mm256_maskz_extractf32x4_ps(U, X, C) \ 12651 ((__m128) __builtin_ia32_extractf32x4_256_mask ((__v8sf)(__m256) (X), \ 12652 (int) (C), \ 12653 (__v4sf)(__m128)_mm_setzero_ps (), \ 12654 (__mmask8)(U))) 12655 12656 #define _mm256_extracti32x4_epi32(X, C) \ 12657 ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\ 12658 (int) (C), (__v4si)(__m128i)_mm_setzero_si128 (), (__mmask8)-1)) 12659 12660 #define _mm256_mask_extracti32x4_epi32(W, U, X, C) \ 12661 ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\ 12662 (int) (C), (__v4si)(__m128i)(W), (__mmask8)(U))) 12663 12664 #define _mm256_maskz_extracti32x4_epi32(U, X, C) \ 12665 ((__m128i) __builtin_ia32_extracti32x4_256_mask ((__v8si)(__m256i) (X),\ 12666 (int) (C), (__v4si)(__m128i)_mm_setzero_si128 (), (__mmask8)(U))) 12667 12668 #define _mm256_shuffle_i64x2(X, Y, C) \ 12669 ((__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X), \ 12670 (__v4di)(__m256i)(Y), (int)(C), \ 12671 (__v4di)(__m256i)_mm256_setzero_si256 (), \ 12672 (__mmask8)-1)) 12673 12674 #define _mm256_mask_shuffle_i64x2(W, U, X, Y, C) \ 12675 ((__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X), \ 12676 (__v4di)(__m256i)(Y), (int)(C), \ 12677 (__v4di)(__m256i)(W),\ 12678 (__mmask8)(U))) 12679 12680 #define _mm256_maskz_shuffle_i64x2(U, X, Y, C) \ 12681 ((__m256i) __builtin_ia32_shuf_i64x2_256_mask ((__v4di)(__m256i)(X), \ 12682 (__v4di)(__m256i)(Y), (int)(C), \ 12683 (__v4di)(__m256i)_mm256_setzero_si256 (), \ 12684 (__mmask8)(U))) 12685 12686 #define _mm256_shuffle_i32x4(X, Y, C) \ 12687 ((__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X), \ 12688 (__v8si)(__m256i)(Y), (int)(C), \ 12689 (__v8si)(__m256i) \ 12690 _mm256_setzero_si256 (), \ 12691 (__mmask8)-1)) 12692 12693 #define _mm256_mask_shuffle_i32x4(W, U, X, Y, C) \ 12694 ((__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X), \ 12695 (__v8si)(__m256i)(Y), (int)(C), \ 12696 (__v8si)(__m256i)(W), \ 12697 (__mmask8)(U))) 12698 12699 #define _mm256_maskz_shuffle_i32x4(U, X, Y, C) \ 12700 ((__m256i) __builtin_ia32_shuf_i32x4_256_mask ((__v8si)(__m256i)(X), \ 12701 (__v8si)(__m256i)(Y), (int)(C), \ 12702 (__v8si)(__m256i) \ 12703 _mm256_setzero_si256 (), \ 12704 (__mmask8)(U))) 12705 12706 #define _mm256_shuffle_f64x2(X, Y, C) \ 12707 ((__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X), \ 12708 (__v4df)(__m256d)(Y), (int)(C), \ 12709 (__v4df)(__m256d)_mm256_setzero_pd (),\ 12710 (__mmask8)-1)) 12711 12712 #define _mm256_mask_shuffle_f64x2(W, U, X, Y, C) \ 12713 ((__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X), \ 12714 (__v4df)(__m256d)(Y), (int)(C), \ 12715 (__v4df)(__m256d)(W), \ 12716 (__mmask8)(U))) 12717 12718 #define _mm256_maskz_shuffle_f64x2(U, X, Y, C) \ 12719 ((__m256d) __builtin_ia32_shuf_f64x2_256_mask ((__v4df)(__m256d)(X), \ 12720 (__v4df)(__m256d)(Y), (int)(C), \ 12721 (__v4df)(__m256d)_mm256_setzero_pd( ),\ 12722 (__mmask8)(U))) 12723 12724 #define _mm256_shuffle_f32x4(X, Y, C) \ 12725 ((__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X), \ 12726 (__v8sf)(__m256)(Y), (int)(C), \ 12727 (__v8sf)(__m256)_mm256_setzero_ps (), \ 12728 (__mmask8)-1)) 12729 12730 #define _mm256_mask_shuffle_f32x4(W, U, X, Y, C) \ 12731 ((__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X), \ 12732 (__v8sf)(__m256)(Y), (int)(C), \ 12733 (__v8sf)(__m256)(W), \ 12734 (__mmask8)(U))) 12735 12736 #define _mm256_maskz_shuffle_f32x4(U, X, Y, C) \ 12737 ((__m256) __builtin_ia32_shuf_f32x4_256_mask ((__v8sf)(__m256)(X), \ 12738 (__v8sf)(__m256)(Y), (int)(C), \ 12739 (__v8sf)(__m256)_mm256_setzero_ps (), \ 12740 (__mmask8)(U))) 12741 12742 #define _mm256_mask_shuffle_pd(W, U, A, B, C) \ 12743 ((__m256d)__builtin_ia32_shufpd256_mask ((__v4df)(__m256d)(A), \ 12744 (__v4df)(__m256d)(B), (int)(C), \ 12745 (__v4df)(__m256d)(W), \ 12746 (__mmask8)(U))) 12747 12748 #define _mm256_maskz_shuffle_pd(U, A, B, C) \ 12749 ((__m256d)__builtin_ia32_shufpd256_mask ((__v4df)(__m256d)(A), \ 12750 (__v4df)(__m256d)(B), (int)(C), \ 12751 (__v4df)(__m256d) \ 12752 _mm256_setzero_pd (), \ 12753 (__mmask8)(U))) 12754 12755 #define _mm_mask_shuffle_pd(W, U, A, B, C) \ 12756 ((__m128d)__builtin_ia32_shufpd128_mask ((__v2df)(__m128d)(A), \ 12757 (__v2df)(__m128d)(B), (int)(C), \ 12758 (__v2df)(__m128d)(W), \ 12759 (__mmask8)(U))) 12760 12761 #define _mm_maskz_shuffle_pd(U, A, B, C) \ 12762 ((__m128d)__builtin_ia32_shufpd128_mask ((__v2df)(__m128d)(A), \ 12763 (__v2df)(__m128d)(B), (int)(C), \ 12764 (__v2df)(__m128d)_mm_setzero_pd (), \ 12765 (__mmask8)(U))) 12766 12767 #define _mm256_mask_shuffle_ps(W, U, A, B, C) \ 12768 ((__m256) __builtin_ia32_shufps256_mask ((__v8sf)(__m256)(A), \ 12769 (__v8sf)(__m256)(B), (int)(C), \ 12770 (__v8sf)(__m256)(W), \ 12771 (__mmask8)(U))) 12772 12773 #define _mm256_maskz_shuffle_ps(U, A, B, C) \ 12774 ((__m256) __builtin_ia32_shufps256_mask ((__v8sf)(__m256)(A), \ 12775 (__v8sf)(__m256)(B), (int)(C), \ 12776 (__v8sf)(__m256)_mm256_setzero_ps (),\ 12777 (__mmask8)(U))) 12778 12779 #define _mm_mask_shuffle_ps(W, U, A, B, C) \ 12780 ((__m128) __builtin_ia32_shufps128_mask ((__v4sf)(__m128)(A), \ 12781 (__v4sf)(__m128)(B), (int)(C), \ 12782 (__v4sf)(__m128)(W), \ 12783 (__mmask8)(U))) 12784 12785 #define _mm_maskz_shuffle_ps(U, A, B, C) \ 12786 ((__m128) __builtin_ia32_shufps128_mask ((__v4sf)(__m128)(A), \ 12787 (__v4sf)(__m128)(B), (int)(C), \ 12788 (__v4sf)(__m128)_mm_setzero_ps (), \ 12789 (__mmask8)(U))) 12790 12791 #define _mm256_fixupimm_pd(X, Y, Z, C) \ 12792 ((__m256d)__builtin_ia32_fixupimmpd256_mask ((__v4df)(__m256d)(X), \ 12793 (__v4df)(__m256d)(Y), \ 12794 (__v4di)(__m256i)(Z), (int)(C), \ 12795 (__mmask8)(-1))) 12796 12797 #define _mm256_mask_fixupimm_pd(X, U, Y, Z, C) \ 12798 ((__m256d)__builtin_ia32_fixupimmpd256_mask ((__v4df)(__m256d)(X), \ 12799 (__v4df)(__m256d)(Y), \ 12800 (__v4di)(__m256i)(Z), (int)(C), \ 12801 (__mmask8)(U))) 12802 12803 #define _mm256_maskz_fixupimm_pd(U, X, Y, Z, C) \ 12804 ((__m256d)__builtin_ia32_fixupimmpd256_maskz ((__v4df)(__m256d)(X), \ 12805 (__v4df)(__m256d)(Y), \ 12806 (__v4di)(__m256i)(Z), (int)(C),\ 12807 (__mmask8)(U))) 12808 12809 #define _mm256_fixupimm_ps(X, Y, Z, C) \ 12810 ((__m256)__builtin_ia32_fixupimmps256_mask ((__v8sf)(__m256)(X), \ 12811 (__v8sf)(__m256)(Y), \ 12812 (__v8si)(__m256i)(Z), (int)(C), \ 12813 (__mmask8)(-1))) 12814 12815 12816 #define _mm256_mask_fixupimm_ps(X, U, Y, Z, C) \ 12817 ((__m256)__builtin_ia32_fixupimmps256_mask ((__v8sf)(__m256)(X), \ 12818 (__v8sf)(__m256)(Y), \ 12819 (__v8si)(__m256i)(Z), (int)(C), \ 12820 (__mmask8)(U))) 12821 12822 #define _mm256_maskz_fixupimm_ps(U, X, Y, Z, C) \ 12823 ((__m256)__builtin_ia32_fixupimmps256_maskz ((__v8sf)(__m256)(X), \ 12824 (__v8sf)(__m256)(Y), \ 12825 (__v8si)(__m256i)(Z), (int)(C),\ 12826 (__mmask8)(U))) 12827 12828 #define _mm_fixupimm_pd(X, Y, Z, C) \ 12829 ((__m128d)__builtin_ia32_fixupimmpd128_mask ((__v2df)(__m128d)(X), \ 12830 (__v2df)(__m128d)(Y), \ 12831 (__v2di)(__m128i)(Z), (int)(C), \ 12832 (__mmask8)(-1))) 12833 12834 12835 #define _mm_mask_fixupimm_pd(X, U, Y, Z, C) \ 12836 ((__m128d)__builtin_ia32_fixupimmpd128_mask ((__v2df)(__m128d)(X), \ 12837 (__v2df)(__m128d)(Y), \ 12838 (__v2di)(__m128i)(Z), (int)(C), \ 12839 (__mmask8)(U))) 12840 12841 #define _mm_maskz_fixupimm_pd(U, X, Y, Z, C) \ 12842 ((__m128d)__builtin_ia32_fixupimmpd128_maskz ((__v2df)(__m128d)(X), \ 12843 (__v2df)(__m128d)(Y), \ 12844 (__v2di)(__m128i)(Z), (int)(C),\ 12845 (__mmask8)(U))) 12846 12847 #define _mm_fixupimm_ps(X, Y, Z, C) \ 12848 ((__m128)__builtin_ia32_fixupimmps128_mask ((__v4sf)(__m128)(X), \ 12849 (__v4sf)(__m128)(Y), \ 12850 (__v4si)(__m128i)(Z), (int)(C), \ 12851 (__mmask8)(-1))) 12852 12853 #define _mm_mask_fixupimm_ps(X, U, Y, Z, C) \ 12854 ((__m128)__builtin_ia32_fixupimmps128_mask ((__v4sf)(__m128)(X), \ 12855 (__v4sf)(__m128)(Y), \ 12856 (__v4si)(__m128i)(Z), (int)(C),\ 12857 (__mmask8)(U))) 12858 12859 #define _mm_maskz_fixupimm_ps(U, X, Y, Z, C) \ 12860 ((__m128)__builtin_ia32_fixupimmps128_maskz ((__v4sf)(__m128)(X), \ 12861 (__v4sf)(__m128)(Y), \ 12862 (__v4si)(__m128i)(Z), (int)(C),\ 12863 (__mmask8)(U))) 12864 12865 #define _mm256_mask_srli_epi32(W, U, A, B) \ 12866 ((__m256i) __builtin_ia32_psrldi256_mask ((__v8si)(__m256i)(A), \ 12867 (int)(B), (__v8si)(__m256i)(W), (__mmask8)(U))) 12868 12869 #define _mm256_maskz_srli_epi32(U, A, B) \ 12870 ((__m256i) __builtin_ia32_psrldi256_mask ((__v8si)(__m256i)(A), \ 12871 (int)(B), (__v8si)_mm256_setzero_si256 (), (__mmask8)(U))) 12872 12873 #define _mm_mask_srli_epi32(W, U, A, B) \ 12874 ((__m128i) __builtin_ia32_psrldi128_mask ((__v4si)(__m128i)(A), \ 12875 (int)(B), (__v4si)(__m128i)(W), (__mmask8)(U))) 12876 12877 #define _mm_maskz_srli_epi32(U, A, B) \ 12878 ((__m128i) __builtin_ia32_psrldi128_mask ((__v4si)(__m128i)(A), \ 12879 (int)(B), (__v4si)_mm_setzero_si128 (), (__mmask8)(U))) 12880 12881 #define _mm256_mask_srli_epi64(W, U, A, B) \ 12882 ((__m256i) __builtin_ia32_psrlqi256_mask ((__v4di)(__m256i)(A), \ 12883 (int)(B), (__v4di)(__m256i)(W), (__mmask8)(U))) 12884 12885 #define _mm256_maskz_srli_epi64(U, A, B) \ 12886 ((__m256i) __builtin_ia32_psrlqi256_mask ((__v4di)(__m256i)(A), \ 12887 (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)(U))) 12888 12889 #define _mm_mask_srli_epi64(W, U, A, B) \ 12890 ((__m128i) __builtin_ia32_psrlqi128_mask ((__v2di)(__m128i)(A), \ 12891 (int)(B), (__v2di)(__m128i)(W), (__mmask8)(U))) 12892 12893 #define _mm_maskz_srli_epi64(U, A, B) \ 12894 ((__m128i) __builtin_ia32_psrlqi128_mask ((__v2di)(__m128i)(A), \ 12895 (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)(U))) 12896 12897 #define _mm256_mask_slli_epi32(W, U, X, C) \ 12898 ((__m256i)__builtin_ia32_pslldi256_mask ((__v8si)(__m256i)(X), (int)(C),\ 12899 (__v8si)(__m256i)(W), \ 12900 (__mmask8)(U))) 12901 12902 #define _mm256_maskz_slli_epi32(U, X, C) \ 12903 ((__m256i)__builtin_ia32_pslldi256_mask ((__v8si)(__m256i)(X), (int)(C),\ 12904 (__v8si)(__m256i)_mm256_setzero_si256 (), \ 12905 (__mmask8)(U))) 12906 12907 #define _mm256_mask_slli_epi64(W, U, X, C) \ 12908 ((__m256i)__builtin_ia32_psllqi256_mask ((__v4di)(__m256i)(X), (int)(C),\ 12909 (__v4di)(__m256i)(W), \ 12910 (__mmask8)(U))) 12911 12912 #define _mm256_maskz_slli_epi64(U, X, C) \ 12913 ((__m256i)__builtin_ia32_psllqi256_mask ((__v4di)(__m256i)(X), (int)(C),\ 12914 (__v4di)(__m256i)_mm256_setzero_si256 (), \ 12915 (__mmask8)(U))) 12916 12917 #define _mm_mask_slli_epi32(W, U, X, C) \ 12918 ((__m128i)__builtin_ia32_pslldi128_mask ((__v4si)(__m128i)(X), (int)(C),\ 12919 (__v4si)(__m128i)(W),\ 12920 (__mmask8)(U))) 12921 12922 #define _mm_maskz_slli_epi32(U, X, C) \ 12923 ((__m128i)__builtin_ia32_pslldi128_mask ((__v4si)(__m128i)(X), (int)(C),\ 12924 (__v4si)(__m128i)_mm_setzero_si128 (),\ 12925 (__mmask8)(U))) 12926 12927 #define _mm_mask_slli_epi64(W, U, X, C) \ 12928 ((__m128i)__builtin_ia32_psllqi128_mask ((__v2di)(__m128i)(X), (int)(C),\ 12929 (__v2di)(__m128i)(W),\ 12930 (__mmask8)(U))) 12931 12932 #define _mm_maskz_slli_epi64(U, X, C) \ 12933 ((__m128i)__builtin_ia32_psllqi128_mask ((__v2di)(__m128i)(X), (int)(C),\ 12934 (__v2di)(__m128i)_mm_setzero_si128 (),\ 12935 (__mmask8)(U))) 12936 12937 #define _mm256_ternarylogic_epi64(A, B, C, I) \ 12938 ((__m256i) \ 12939 __builtin_ia32_pternlogq256_mask ((__v4di) (__m256i) (A), \ 12940 (__v4di) (__m256i) (B), \ 12941 (__v4di) (__m256i) (C), \ 12942 (unsigned char) (I), \ 12943 (__mmask8) -1)) 12944 12945 #define _mm256_mask_ternarylogic_epi64(A, U, B, C, I) \ 12946 ((__m256i) \ 12947 __builtin_ia32_pternlogq256_mask ((__v4di) (__m256i) (A), \ 12948 (__v4di) (__m256i) (B), \ 12949 (__v4di) (__m256i) (C), \ 12950 (unsigned char) (I), \ 12951 (__mmask8) (U))) 12952 12953 #define _mm256_maskz_ternarylogic_epi64(U, A, B, C, I) \ 12954 ((__m256i) \ 12955 __builtin_ia32_pternlogq256_maskz ((__v4di) (__m256i) (A), \ 12956 (__v4di) (__m256i) (B), \ 12957 (__v4di) (__m256i) (C), \ 12958 (unsigned char) (I), \ 12959 (__mmask8) (U))) 12960 12961 #define _mm256_ternarylogic_epi32(A, B, C, I) \ 12962 ((__m256i) \ 12963 __builtin_ia32_pternlogd256_mask ((__v8si) (__m256i) (A), \ 12964 (__v8si) (__m256i) (B), \ 12965 (__v8si) (__m256i) (C), \ 12966 (unsigned char) (I), \ 12967 (__mmask8) -1)) 12968 12969 #define _mm256_mask_ternarylogic_epi32(A, U, B, C, I) \ 12970 ((__m256i) \ 12971 __builtin_ia32_pternlogd256_mask ((__v8si) (__m256i) (A), \ 12972 (__v8si) (__m256i) (B), \ 12973 (__v8si) (__m256i) (C), \ 12974 (unsigned char) (I), \ 12975 (__mmask8) (U))) 12976 12977 #define _mm256_maskz_ternarylogic_epi32(U, A, B, C, I) \ 12978 ((__m256i) \ 12979 __builtin_ia32_pternlogd256_maskz ((__v8si) (__m256i) (A), \ 12980 (__v8si) (__m256i) (B), \ 12981 (__v8si) (__m256i) (C), \ 12982 (unsigned char) (I), \ 12983 (__mmask8) (U))) 12984 12985 #define _mm_ternarylogic_epi64(A, B, C, I) \ 12986 ((__m128i) \ 12987 __builtin_ia32_pternlogq128_mask ((__v2di) (__m128i) (A), \ 12988 (__v2di) (__m128i) (B), \ 12989 (__v2di) (__m128i) (C), \ 12990 (unsigned char) (I), \ 12991 (__mmask8) -1)) 12992 12993 #define _mm_mask_ternarylogic_epi64(A, U, B, C, I) \ 12994 ((__m128i) \ 12995 __builtin_ia32_pternlogq128_mask ((__v2di) (__m128i) (A), \ 12996 (__v2di) (__m128i) (B), \ 12997 (__v2di) (__m128i) (C), \ 12998 (unsigned char) (I), \ 12999 (__mmask8) (U))) 13000 13001 #define _mm_maskz_ternarylogic_epi64(U, A, B, C, I) \ 13002 ((__m128i) \ 13003 __builtin_ia32_pternlogq128_maskz ((__v2di) (__m128i) (A), \ 13004 (__v2di) (__m128i) (B), \ 13005 (__v2di) (__m128i) (C), \ 13006 (unsigned char) (I), \ 13007 (__mmask8) (U))) 13008 13009 #define _mm_ternarylogic_epi32(A, B, C, I) \ 13010 ((__m128i) \ 13011 __builtin_ia32_pternlogd128_mask ((__v4si) (__m128i) (A), \ 13012 (__v4si) (__m128i) (B), \ 13013 (__v4si) (__m128i) (C), \ 13014 (unsigned char) (I), \ 13015 (__mmask8) -1)) 13016 13017 #define _mm_mask_ternarylogic_epi32(A, U, B, C, I) \ 13018 ((__m128i) \ 13019 __builtin_ia32_pternlogd128_mask ((__v4si) (__m128i) (A), \ 13020 (__v4si) (__m128i) (B), \ 13021 (__v4si) (__m128i) (C), \ 13022 (unsigned char) (I), \ 13023 (__mmask8) (U))) 13024 13025 #define _mm_maskz_ternarylogic_epi32(U, A, B, C, I) \ 13026 ((__m128i) \ 13027 __builtin_ia32_pternlogd128_maskz ((__v4si) (__m128i) (A), \ 13028 (__v4si) (__m128i) (B), \ 13029 (__v4si) (__m128i) (C), \ 13030 (unsigned char) (I), \ 13031 (__mmask8) (U))) 13032 13033 #define _mm256_roundscale_ps(A, B) \ 13034 ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A), \ 13035 (int)(B), (__v8sf)(__m256)_mm256_setzero_ps (), (__mmask8)-1)) 13036 13037 #define _mm256_mask_roundscale_ps(W, U, A, B) \ 13038 ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A), \ 13039 (int)(B), (__v8sf)(__m256)(W), (__mmask8)(U))) 13040 13041 #define _mm256_maskz_roundscale_ps(U, A, B) \ 13042 ((__m256) __builtin_ia32_rndscaleps_256_mask ((__v8sf)(__m256)(A), \ 13043 (int)(B), (__v8sf)(__m256)_mm256_setzero_ps (), (__mmask8)(U))) 13044 13045 #define _mm256_roundscale_pd(A, B) \ 13046 ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A), \ 13047 (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)-1)) 13048 13049 #define _mm256_mask_roundscale_pd(W, U, A, B) \ 13050 ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A), \ 13051 (int)(B), (__v4df)(__m256d)(W), (__mmask8)(U))) 13052 13053 #define _mm256_maskz_roundscale_pd(U, A, B) \ 13054 ((__m256d) __builtin_ia32_rndscalepd_256_mask ((__v4df)(__m256d)(A), \ 13055 (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)(U))) 13056 13057 #define _mm_roundscale_ps(A, B) \ 13058 ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A), \ 13059 (int)(B), (__v4sf)(__m128)_mm_setzero_ps (), (__mmask8)-1)) 13060 13061 #define _mm_mask_roundscale_ps(W, U, A, B) \ 13062 ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A), \ 13063 (int)(B), (__v4sf)(__m128)(W), (__mmask8)(U))) 13064 13065 #define _mm_maskz_roundscale_ps(U, A, B) \ 13066 ((__m128) __builtin_ia32_rndscaleps_128_mask ((__v4sf)(__m128)(A), \ 13067 (int)(B), (__v4sf)(__m128)_mm_setzero_ps (), (__mmask8)(U))) 13068 13069 #define _mm_roundscale_pd(A, B) \ 13070 ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A), \ 13071 (int)(B), (__v2df)(__m128d)_mm_setzero_pd (), (__mmask8)-1)) 13072 13073 #define _mm_mask_roundscale_pd(W, U, A, B) \ 13074 ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A), \ 13075 (int)(B), (__v2df)(__m128d)(W), (__mmask8)(U))) 13076 13077 #define _mm_maskz_roundscale_pd(U, A, B) \ 13078 ((__m128d) __builtin_ia32_rndscalepd_128_mask ((__v2df)(__m128d)(A), \ 13079 (int)(B), (__v2df)(__m128d)_mm_setzero_pd (), (__mmask8)(U))) 13080 13081 #define _mm256_getmant_ps(X, B, C) \ 13082 ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X), \ 13083 (int)(((C)<<2) | (B)), \ 13084 (__v8sf)(__m256)_mm256_setzero_ps (), \ 13085 (__mmask8)-1)) 13086 13087 #define _mm256_mask_getmant_ps(W, U, X, B, C) \ 13088 ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X), \ 13089 (int)(((C)<<2) | (B)), \ 13090 (__v8sf)(__m256)(W), \ 13091 (__mmask8)(U))) 13092 13093 #define _mm256_maskz_getmant_ps(U, X, B, C) \ 13094 ((__m256) __builtin_ia32_getmantps256_mask ((__v8sf)(__m256) (X), \ 13095 (int)(((C)<<2) | (B)), \ 13096 (__v8sf)(__m256)_mm256_setzero_ps (), \ 13097 (__mmask8)(U))) 13098 13099 #define _mm_getmant_ps(X, B, C) \ 13100 ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X), \ 13101 (int)(((C)<<2) | (B)), \ 13102 (__v4sf)(__m128)_mm_setzero_ps (), \ 13103 (__mmask8)-1)) 13104 13105 #define _mm_mask_getmant_ps(W, U, X, B, C) \ 13106 ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X), \ 13107 (int)(((C)<<2) | (B)), \ 13108 (__v4sf)(__m128)(W), \ 13109 (__mmask8)(U))) 13110 13111 #define _mm_maskz_getmant_ps(U, X, B, C) \ 13112 ((__m128) __builtin_ia32_getmantps128_mask ((__v4sf)(__m128) (X), \ 13113 (int)(((C)<<2) | (B)), \ 13114 (__v4sf)(__m128)_mm_setzero_ps (), \ 13115 (__mmask8)(U))) 13116 13117 #define _mm256_getmant_pd(X, B, C) \ 13118 ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X), \ 13119 (int)(((C)<<2) | (B)), \ 13120 (__v4df)(__m256d)_mm256_setzero_pd (),\ 13121 (__mmask8)-1)) 13122 13123 #define _mm256_mask_getmant_pd(W, U, X, B, C) \ 13124 ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X), \ 13125 (int)(((C)<<2) | (B)), \ 13126 (__v4df)(__m256d)(W), \ 13127 (__mmask8)(U))) 13128 13129 #define _mm256_maskz_getmant_pd(U, X, B, C) \ 13130 ((__m256d) __builtin_ia32_getmantpd256_mask ((__v4df)(__m256d) (X), \ 13131 (int)(((C)<<2) | (B)), \ 13132 (__v4df)(__m256d)_mm256_setzero_pd (),\ 13133 (__mmask8)(U))) 13134 13135 #define _mm_getmant_pd(X, B, C) \ 13136 ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X), \ 13137 (int)(((C)<<2) | (B)), \ 13138 (__v2df)(__m128d)_mm_setzero_pd (), \ 13139 (__mmask8)-1)) 13140 13141 #define _mm_mask_getmant_pd(W, U, X, B, C) \ 13142 ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X), \ 13143 (int)(((C)<<2) | (B)), \ 13144 (__v2df)(__m128d)(W), \ 13145 (__mmask8)(U))) 13146 13147 #define _mm_maskz_getmant_pd(U, X, B, C) \ 13148 ((__m128d) __builtin_ia32_getmantpd128_mask ((__v2df)(__m128d) (X), \ 13149 (int)(((C)<<2) | (B)), \ 13150 (__v2df)(__m128d)_mm_setzero_pd (), \ 13151 (__mmask8)(U))) 13152 13153 #define _mm256_mmask_i32gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13154 (__m256) __builtin_ia32_gather3siv8sf ((__v8sf)(__m256) (V1OLD), \ 13155 (void const *) (ADDR), \ 13156 (__v8si)(__m256i) (INDEX), \ 13157 (__mmask8) (MASK), \ 13158 (int) (SCALE)) 13159 13160 #define _mm_mmask_i32gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13161 (__m128) __builtin_ia32_gather3siv4sf ((__v4sf)(__m128) (V1OLD), \ 13162 (void const *) (ADDR), \ 13163 (__v4si)(__m128i) (INDEX), \ 13164 (__mmask8) (MASK), \ 13165 (int) (SCALE)) 13166 13167 #define _mm256_mmask_i32gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13168 (__m256d) __builtin_ia32_gather3siv4df ((__v4df)(__m256d) (V1OLD), \ 13169 (void const *) (ADDR), \ 13170 (__v4si)(__m128i) (INDEX), \ 13171 (__mmask8) (MASK), \ 13172 (int) (SCALE)) 13173 13174 #define _mm_mmask_i32gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13175 (__m128d) __builtin_ia32_gather3siv2df ((__v2df)(__m128d) (V1OLD), \ 13176 (void const *) (ADDR), \ 13177 (__v4si)(__m128i) (INDEX), \ 13178 (__mmask8) (MASK), \ 13179 (int) (SCALE)) 13180 13181 #define _mm256_mmask_i64gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13182 (__m128) __builtin_ia32_gather3div8sf ((__v4sf)(__m128) (V1OLD), \ 13183 (void const *) (ADDR), \ 13184 (__v4di)(__m256i) (INDEX), \ 13185 (__mmask8) (MASK), \ 13186 (int) (SCALE)) 13187 13188 #define _mm_mmask_i64gather_ps(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13189 (__m128) __builtin_ia32_gather3div4sf ((__v4sf)(__m128) (V1OLD), \ 13190 (void const *) (ADDR), \ 13191 (__v2di)(__m128i) (INDEX), \ 13192 (__mmask8) (MASK), \ 13193 (int) (SCALE)) 13194 13195 #define _mm256_mmask_i64gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13196 (__m256d) __builtin_ia32_gather3div4df ((__v4df)(__m256d) (V1OLD), \ 13197 (void const *) (ADDR), \ 13198 (__v4di)(__m256i) (INDEX), \ 13199 (__mmask8) (MASK), \ 13200 (int) (SCALE)) 13201 13202 #define _mm_mmask_i64gather_pd(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13203 (__m128d) __builtin_ia32_gather3div2df ((__v2df)(__m128d) (V1OLD), \ 13204 (void const *) (ADDR), \ 13205 (__v2di)(__m128i) (INDEX), \ 13206 (__mmask8) (MASK), \ 13207 (int) (SCALE)) 13208 13209 #define _mm256_mmask_i32gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13210 (__m256i) __builtin_ia32_gather3siv8si ((__v8si)(__m256i) (V1OLD), \ 13211 (void const *) (ADDR), \ 13212 (__v8si)(__m256i) (INDEX), \ 13213 (__mmask8) (MASK), \ 13214 (int) (SCALE)) 13215 13216 #define _mm_mmask_i32gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13217 (__m128i) __builtin_ia32_gather3siv4si ((__v4si)(__m128i) (V1OLD), \ 13218 (void const *) (ADDR), \ 13219 (__v4si)(__m128i) (INDEX), \ 13220 (__mmask8) (MASK), \ 13221 (int) (SCALE)) 13222 13223 #define _mm256_mmask_i32gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13224 (__m256i) __builtin_ia32_gather3siv4di ((__v4di)(__m256i) (V1OLD), \ 13225 (void const *) (ADDR), \ 13226 (__v4si)(__m128i) (INDEX), \ 13227 (__mmask8) (MASK), \ 13228 (int) (SCALE)) 13229 13230 #define _mm_mmask_i32gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13231 (__m128i) __builtin_ia32_gather3siv2di ((__v2di)(__m128i) (V1OLD), \ 13232 (void const *) (ADDR), \ 13233 (__v4si)(__m128i) (INDEX), \ 13234 (__mmask8) (MASK), \ 13235 (int) (SCALE)) 13236 13237 #define _mm256_mmask_i64gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13238 (__m128i) __builtin_ia32_gather3div8si ((__v4si)(__m128i) (V1OLD), \ 13239 (void const *) (ADDR), \ 13240 (__v4di)(__m256i) (INDEX), \ 13241 (__mmask8) (MASK), \ 13242 (int) (SCALE)) 13243 13244 #define _mm_mmask_i64gather_epi32(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13245 (__m128i) __builtin_ia32_gather3div4si ((__v4si)(__m128i) (V1OLD), \ 13246 (void const *) (ADDR), \ 13247 (__v2di)(__m128i) (INDEX), \ 13248 (__mmask8) (MASK), \ 13249 (int) (SCALE)) 13250 13251 #define _mm256_mmask_i64gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13252 (__m256i) __builtin_ia32_gather3div4di ((__v4di)(__m256i) (V1OLD), \ 13253 (void const *) (ADDR), \ 13254 (__v4di)(__m256i) (INDEX), \ 13255 (__mmask8) (MASK), \ 13256 (int) (SCALE)) 13257 13258 #define _mm_mmask_i64gather_epi64(V1OLD, MASK, INDEX, ADDR, SCALE) \ 13259 (__m128i) __builtin_ia32_gather3div2di ((__v2di)(__m128i) (V1OLD), \ 13260 (void const *) (ADDR), \ 13261 (__v2di)(__m128i) (INDEX), \ 13262 (__mmask8) (MASK), \ 13263 (int) (SCALE)) 13264 13265 #define _mm256_i32scatter_ps(ADDR, INDEX, V1, SCALE) \ 13266 __builtin_ia32_scattersiv8sf ((void *) (ADDR), (__mmask8)0xFF, \ 13267 (__v8si)(__m256i) (INDEX), \ 13268 (__v8sf)(__m256) (V1), (int) (SCALE)) 13269 13270 #define _mm256_mask_i32scatter_ps(ADDR, MASK, INDEX, V1, SCALE) \ 13271 __builtin_ia32_scattersiv8sf ((void *) (ADDR), (__mmask8) (MASK), \ 13272 (__v8si)(__m256i) (INDEX), \ 13273 (__v8sf)(__m256) (V1), (int) (SCALE)) 13274 13275 #define _mm_i32scatter_ps(ADDR, INDEX, V1, SCALE) \ 13276 __builtin_ia32_scattersiv4sf ((void *) (ADDR), (__mmask8)0xFF, \ 13277 (__v4si)(__m128i) (INDEX), \ 13278 (__v4sf)(__m128) (V1), (int) (SCALE)) 13279 13280 #define _mm_mask_i32scatter_ps(ADDR, MASK, INDEX, V1, SCALE) \ 13281 __builtin_ia32_scattersiv4sf ((void *) (ADDR), (__mmask8) (MASK), \ 13282 (__v4si)(__m128i) (INDEX), \ 13283 (__v4sf)(__m128) (V1), (int) (SCALE)) 13284 13285 #define _mm256_i32scatter_pd(ADDR, INDEX, V1, SCALE) \ 13286 __builtin_ia32_scattersiv4df ((void *) (ADDR), (__mmask8)0xFF, \ 13287 (__v4si)(__m128i) (INDEX), \ 13288 (__v4df)(__m256d) (V1), (int) (SCALE)) 13289 13290 #define _mm256_mask_i32scatter_pd(ADDR, MASK, INDEX, V1, SCALE) \ 13291 __builtin_ia32_scattersiv4df ((void *) (ADDR), (__mmask8) (MASK), \ 13292 (__v4si)(__m128i) (INDEX), \ 13293 (__v4df)(__m256d) (V1), (int) (SCALE)) 13294 13295 #define _mm_i32scatter_pd(ADDR, INDEX, V1, SCALE) \ 13296 __builtin_ia32_scattersiv2df ((void *) (ADDR), (__mmask8)0xFF, \ 13297 (__v4si)(__m128i) (INDEX), \ 13298 (__v2df)(__m128d) (V1), (int) (SCALE)) 13299 13300 #define _mm_mask_i32scatter_pd(ADDR, MASK, INDEX, V1, SCALE) \ 13301 __builtin_ia32_scattersiv2df ((void *) (ADDR), (__mmask8) (MASK), \ 13302 (__v4si)(__m128i) (INDEX), \ 13303 (__v2df)(__m128d) (V1), (int) (SCALE)) 13304 13305 #define _mm256_i64scatter_ps(ADDR, INDEX, V1, SCALE) \ 13306 __builtin_ia32_scatterdiv8sf ((void *) (ADDR), (__mmask8)0xFF, \ 13307 (__v4di)(__m256i) (INDEX), \ 13308 (__v4sf)(__m128) (V1), (int) (SCALE)) 13309 13310 #define _mm256_mask_i64scatter_ps(ADDR, MASK, INDEX, V1, SCALE) \ 13311 __builtin_ia32_scatterdiv8sf ((void *) (ADDR), (__mmask8) (MASK), \ 13312 (__v4di)(__m256i) (INDEX), \ 13313 (__v4sf)(__m128) (V1), (int) (SCALE)) 13314 13315 #define _mm_i64scatter_ps(ADDR, INDEX, V1, SCALE) \ 13316 __builtin_ia32_scatterdiv4sf ((void *) (ADDR), (__mmask8)0xFF, \ 13317 (__v2di)(__m128i) (INDEX), \ 13318 (__v4sf)(__m128) (V1), (int) (SCALE)) 13319 13320 #define _mm_mask_i64scatter_ps(ADDR, MASK, INDEX, V1, SCALE) \ 13321 __builtin_ia32_scatterdiv4sf ((void *) (ADDR), (__mmask8) (MASK), \ 13322 (__v2di)(__m128i) (INDEX), \ 13323 (__v4sf)(__m128) (V1), (int) (SCALE)) 13324 13325 #define _mm256_i64scatter_pd(ADDR, INDEX, V1, SCALE) \ 13326 __builtin_ia32_scatterdiv4df ((void *) (ADDR), (__mmask8)0xFF, \ 13327 (__v4di)(__m256i) (INDEX), \ 13328 (__v4df)(__m256d) (V1), (int) (SCALE)) 13329 13330 #define _mm256_mask_i64scatter_pd(ADDR, MASK, INDEX, V1, SCALE) \ 13331 __builtin_ia32_scatterdiv4df ((void *) (ADDR), (__mmask8) (MASK), \ 13332 (__v4di)(__m256i) (INDEX), \ 13333 (__v4df)(__m256d) (V1), (int) (SCALE)) 13334 13335 #define _mm_i64scatter_pd(ADDR, INDEX, V1, SCALE) \ 13336 __builtin_ia32_scatterdiv2df ((void *) (ADDR), (__mmask8)0xFF, \ 13337 (__v2di)(__m128i) (INDEX), \ 13338 (__v2df)(__m128d) (V1), (int) (SCALE)) 13339 13340 #define _mm_mask_i64scatter_pd(ADDR, MASK, INDEX, V1, SCALE) \ 13341 __builtin_ia32_scatterdiv2df ((void *) (ADDR), (__mmask8) (MASK), \ 13342 (__v2di)(__m128i) (INDEX), \ 13343 (__v2df)(__m128d) (V1), (int) (SCALE)) 13344 13345 #define _mm256_i32scatter_epi32(ADDR, INDEX, V1, SCALE) \ 13346 __builtin_ia32_scattersiv8si ((void *) (ADDR), (__mmask8)0xFF, \ 13347 (__v8si)(__m256i) (INDEX), \ 13348 (__v8si)(__m256i) (V1), (int) (SCALE)) 13349 13350 #define _mm256_mask_i32scatter_epi32(ADDR, MASK, INDEX, V1, SCALE) \ 13351 __builtin_ia32_scattersiv8si ((void *) (ADDR), (__mmask8) (MASK), \ 13352 (__v8si)(__m256i) (INDEX), \ 13353 (__v8si)(__m256i) (V1), (int) (SCALE)) 13354 13355 #define _mm_i32scatter_epi32(ADDR, INDEX, V1, SCALE) \ 13356 __builtin_ia32_scattersiv4si ((void *) (ADDR), (__mmask8)0xFF, \ 13357 (__v4si)(__m128i) (INDEX), \ 13358 (__v4si)(__m128i) (V1), (int) (SCALE)) 13359 13360 #define _mm_mask_i32scatter_epi32(ADDR, MASK, INDEX, V1, SCALE) \ 13361 __builtin_ia32_scattersiv4si ((void *) (ADDR), (__mmask8) (MASK), \ 13362 (__v4si)(__m128i) (INDEX), \ 13363 (__v4si)(__m128i) (V1), (int) (SCALE)) 13364 13365 #define _mm256_i32scatter_epi64(ADDR, INDEX, V1, SCALE) \ 13366 __builtin_ia32_scattersiv4di ((void *) (ADDR), (__mmask8)0xFF, \ 13367 (__v4si)(__m128i) (INDEX), \ 13368 (__v4di)(__m256i) (V1), (int) (SCALE)) 13369 13370 #define _mm256_mask_i32scatter_epi64(ADDR, MASK, INDEX, V1, SCALE) \ 13371 __builtin_ia32_scattersiv4di ((void *) (ADDR), (__mmask8) (MASK), \ 13372 (__v4si)(__m128i) (INDEX), \ 13373 (__v4di)(__m256i) (V1), (int) (SCALE)) 13374 13375 #define _mm_i32scatter_epi64(ADDR, INDEX, V1, SCALE) \ 13376 __builtin_ia32_scattersiv2di ((void *) (ADDR), (__mmask8)0xFF, \ 13377 (__v4si)(__m128i) (INDEX), \ 13378 (__v2di)(__m128i) (V1), (int) (SCALE)) 13379 13380 #define _mm_mask_i32scatter_epi64(ADDR, MASK, INDEX, V1, SCALE) \ 13381 __builtin_ia32_scattersiv2di ((void *) (ADDR), (__mmask8) (MASK), \ 13382 (__v4si)(__m128i) (INDEX), \ 13383 (__v2di)(__m128i) (V1), (int) (SCALE)) 13384 13385 #define _mm256_i64scatter_epi32(ADDR, INDEX, V1, SCALE) \ 13386 __builtin_ia32_scatterdiv8si ((void *) (ADDR), (__mmask8)0xFF, \ 13387 (__v4di)(__m256i) (INDEX), \ 13388 (__v4si)(__m128i) (V1), (int) (SCALE)) 13389 13390 #define _mm256_mask_i64scatter_epi32(ADDR, MASK, INDEX, V1, SCALE) \ 13391 __builtin_ia32_scatterdiv8si ((void *) (ADDR), (__mmask8) (MASK), \ 13392 (__v4di)(__m256i) (INDEX), \ 13393 (__v4si)(__m128i) (V1), (int) (SCALE)) 13394 13395 #define _mm_i64scatter_epi32(ADDR, INDEX, V1, SCALE) \ 13396 __builtin_ia32_scatterdiv4si ((void *) (ADDR), (__mmask8)0xFF, \ 13397 (__v2di)(__m128i) (INDEX), \ 13398 (__v4si)(__m128i) (V1), (int) (SCALE)) 13399 13400 #define _mm_mask_i64scatter_epi32(ADDR, MASK, INDEX, V1, SCALE) \ 13401 __builtin_ia32_scatterdiv4si ((void *) (ADDR), (__mmask8) (MASK), \ 13402 (__v2di)(__m128i) (INDEX), \ 13403 (__v4si)(__m128i) (V1), (int) (SCALE)) 13404 13405 #define _mm256_i64scatter_epi64(ADDR, INDEX, V1, SCALE) \ 13406 __builtin_ia32_scatterdiv4di ((void *) (ADDR), (__mmask8)0xFF, \ 13407 (__v4di)(__m256i) (INDEX), \ 13408 (__v4di)(__m256i) (V1), (int) (SCALE)) 13409 13410 #define _mm256_mask_i64scatter_epi64(ADDR, MASK, INDEX, V1, SCALE) \ 13411 __builtin_ia32_scatterdiv4di ((void *) (ADDR), (__mmask8) (MASK), \ 13412 (__v4di)(__m256i) (INDEX), \ 13413 (__v4di)(__m256i) (V1), (int) (SCALE)) 13414 13415 #define _mm_i64scatter_epi64(ADDR, INDEX, V1, SCALE) \ 13416 __builtin_ia32_scatterdiv2di ((void *) (ADDR), (__mmask8)0xFF, \ 13417 (__v2di)(__m128i) (INDEX), \ 13418 (__v2di)(__m128i) (V1), (int) (SCALE)) 13419 13420 #define _mm_mask_i64scatter_epi64(ADDR, MASK, INDEX, V1, SCALE) \ 13421 __builtin_ia32_scatterdiv2di ((void *) (ADDR), (__mmask8) (MASK), \ 13422 (__v2di)(__m128i) (INDEX), \ 13423 (__v2di)(__m128i) (V1), (int) (SCALE)) 13424 13425 #define _mm256_mask_shuffle_epi32(W, U, X, C) \ 13426 ((__m256i) __builtin_ia32_pshufd256_mask ((__v8si)(__m256i)(X), (int)(C), \ 13427 (__v8si)(__m256i)(W), \ 13428 (__mmask8)(U))) 13429 13430 #define _mm256_maskz_shuffle_epi32(U, X, C) \ 13431 ((__m256i) __builtin_ia32_pshufd256_mask ((__v8si)(__m256i)(X), (int)(C), \ 13432 (__v8si)(__m256i) \ 13433 _mm256_setzero_si256 (), \ 13434 (__mmask8)(U))) 13435 13436 #define _mm_mask_shuffle_epi32(W, U, X, C) \ 13437 ((__m128i) __builtin_ia32_pshufd128_mask ((__v4si)(__m128i)(X), (int)(C), \ 13438 (__v4si)(__m128i)(W), \ 13439 (__mmask8)(U))) 13440 13441 #define _mm_maskz_shuffle_epi32(U, X, C) \ 13442 ((__m128i) __builtin_ia32_pshufd128_mask ((__v4si)(__m128i)(X), (int)(C), \ 13443 (__v4si)(__m128i)_mm_setzero_si128 (), \ 13444 (__mmask8)(U))) 13445 13446 #define _mm256_rol_epi64(A, B) \ 13447 ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13448 (__v4di)(__m256i)_mm256_setzero_si256 (),\ 13449 (__mmask8)-1)) 13450 13451 #define _mm256_mask_rol_epi64(W, U, A, B) \ 13452 ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13453 (__v4di)(__m256i)(W), \ 13454 (__mmask8)(U))) 13455 13456 #define _mm256_maskz_rol_epi64(U, A, B) \ 13457 ((__m256i)__builtin_ia32_prolq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13458 (__v4di)(__m256i)_mm256_setzero_si256 (),\ 13459 (__mmask8)(U))) 13460 13461 #define _mm_rol_epi64(A, B) \ 13462 ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13463 (__v2di)(__m128i)_mm_setzero_si128 (),\ 13464 (__mmask8)-1)) 13465 13466 #define _mm_mask_rol_epi64(W, U, A, B) \ 13467 ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13468 (__v2di)(__m128i)(W), \ 13469 (__mmask8)(U))) 13470 13471 #define _mm_maskz_rol_epi64(U, A, B) \ 13472 ((__m128i)__builtin_ia32_prolq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13473 (__v2di)(__m128i)_mm_setzero_si128 (),\ 13474 (__mmask8)(U))) 13475 13476 #define _mm256_ror_epi64(A, B) \ 13477 ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13478 (__v4di)(__m256i)_mm256_setzero_si256 (),\ 13479 (__mmask8)-1)) 13480 13481 #define _mm256_mask_ror_epi64(W, U, A, B) \ 13482 ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13483 (__v4di)(__m256i)(W), \ 13484 (__mmask8)(U))) 13485 13486 #define _mm256_maskz_ror_epi64(U, A, B) \ 13487 ((__m256i)__builtin_ia32_prorq256_mask ((__v4di)(__m256i)(A), (int)(B), \ 13488 (__v4di)(__m256i)_mm256_setzero_si256 (),\ 13489 (__mmask8)(U))) 13490 13491 #define _mm_ror_epi64(A, B) \ 13492 ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13493 (__v2di)(__m128i)_mm_setzero_si128 (),\ 13494 (__mmask8)-1)) 13495 13496 #define _mm_mask_ror_epi64(W, U, A, B) \ 13497 ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13498 (__v2di)(__m128i)(W), \ 13499 (__mmask8)(U))) 13500 13501 #define _mm_maskz_ror_epi64(U, A, B) \ 13502 ((__m128i)__builtin_ia32_prorq128_mask ((__v2di)(__m128i)(A), (int)(B), \ 13503 (__v2di)(__m128i)_mm_setzero_si128 (),\ 13504 (__mmask8)(U))) 13505 13506 #define _mm256_rol_epi32(A, B) \ 13507 ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13508 (__v8si)(__m256i)_mm256_setzero_si256 (),\ 13509 (__mmask8)-1)) 13510 13511 #define _mm256_mask_rol_epi32(W, U, A, B) \ 13512 ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13513 (__v8si)(__m256i)(W), \ 13514 (__mmask8)(U))) 13515 13516 #define _mm256_maskz_rol_epi32(U, A, B) \ 13517 ((__m256i)__builtin_ia32_prold256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13518 (__v8si)(__m256i)_mm256_setzero_si256 (),\ 13519 (__mmask8)(U))) 13520 13521 #define _mm_rol_epi32(A, B) \ 13522 ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13523 (__v4si)(__m128i)_mm_setzero_si128 (),\ 13524 (__mmask8)-1)) 13525 13526 #define _mm_mask_rol_epi32(W, U, A, B) \ 13527 ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13528 (__v4si)(__m128i)(W), \ 13529 (__mmask8)(U))) 13530 13531 #define _mm_maskz_rol_epi32(U, A, B) \ 13532 ((__m128i)__builtin_ia32_prold128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13533 (__v4si)(__m128i)_mm_setzero_si128 (),\ 13534 (__mmask8)(U))) 13535 13536 #define _mm256_ror_epi32(A, B) \ 13537 ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13538 (__v8si)(__m256i)_mm256_setzero_si256 (),\ 13539 (__mmask8)-1)) 13540 13541 #define _mm256_mask_ror_epi32(W, U, A, B) \ 13542 ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13543 (__v8si)(__m256i)(W), \ 13544 (__mmask8)(U))) 13545 13546 #define _mm256_maskz_ror_epi32(U, A, B) \ 13547 ((__m256i)__builtin_ia32_prord256_mask ((__v8si)(__m256i)(A), (int)(B), \ 13548 (__v8si)(__m256i) \ 13549 _mm256_setzero_si256 (), \ 13550 (__mmask8)(U))) 13551 13552 #define _mm_ror_epi32(A, B) \ 13553 ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13554 (__v4si)(__m128i)_mm_setzero_si128 (),\ 13555 (__mmask8)-1)) 13556 13557 #define _mm_mask_ror_epi32(W, U, A, B) \ 13558 ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13559 (__v4si)(__m128i)(W), \ 13560 (__mmask8)(U))) 13561 13562 #define _mm_maskz_ror_epi32(U, A, B) \ 13563 ((__m128i)__builtin_ia32_prord128_mask ((__v4si)(__m128i)(A), (int)(B), \ 13564 (__v4si)(__m128i)_mm_setzero_si128 (),\ 13565 (__mmask8)(U))) 13566 13567 #define _mm256_alignr_epi32(X, Y, C) \ 13568 ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X), \ 13569 (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)(X), (__mmask8)-1)) 13570 13571 #define _mm256_mask_alignr_epi32(W, U, X, Y, C) \ 13572 ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X), \ 13573 (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)(W), (__mmask8)(U))) 13574 13575 #define _mm256_maskz_alignr_epi32(U, X, Y, C) \ 13576 ((__m256i)__builtin_ia32_alignd256_mask ((__v8si)(__m256i)(X), \ 13577 (__v8si)(__m256i)(Y), (int)(C), (__v8si)(__m256i)_mm256_setzero_si256 (),\ 13578 (__mmask8)(U))) 13579 13580 #define _mm256_alignr_epi64(X, Y, C) \ 13581 ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X), \ 13582 (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)(X), (__mmask8)-1)) 13583 13584 #define _mm256_mask_alignr_epi64(W, U, X, Y, C) \ 13585 ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X), \ 13586 (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)(W), (__mmask8)(U))) 13587 13588 #define _mm256_maskz_alignr_epi64(U, X, Y, C) \ 13589 ((__m256i)__builtin_ia32_alignq256_mask ((__v4di)(__m256i)(X), \ 13590 (__v4di)(__m256i)(Y), (int)(C), (__v4di)(__m256i)_mm256_setzero_si256 (),\ 13591 (__mmask8)(U))) 13592 13593 #define _mm_alignr_epi32(X, Y, C) \ 13594 ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X), \ 13595 (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)(X), (__mmask8)-1)) 13596 13597 #define _mm_mask_alignr_epi32(W, U, X, Y, C) \ 13598 ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X), \ 13599 (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)(W), (__mmask8)(U))) 13600 13601 #define _mm_maskz_alignr_epi32(U, X, Y, C) \ 13602 ((__m128i)__builtin_ia32_alignd128_mask ((__v4si)(__m128i)(X), \ 13603 (__v4si)(__m128i)(Y), (int)(C), (__v4si)(__m128i)_mm_setzero_si128 (),\ 13604 (__mmask8)(U))) 13605 13606 #define _mm_alignr_epi64(X, Y, C) \ 13607 ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X), \ 13608 (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)(X), (__mmask8)-1)) 13609 13610 #define _mm_mask_alignr_epi64(W, U, X, Y, C) \ 13611 ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X), \ 13612 (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)(W), (__mmask8)(U))) 13613 13614 #define _mm_maskz_alignr_epi64(U, X, Y, C) \ 13615 ((__m128i)__builtin_ia32_alignq128_mask ((__v2di)(__m128i)(X), \ 13616 (__v2di)(__m128i)(Y), (int)(C), (__v2di)(__m128i)_mm_setzero_si128 (),\ 13617 (__mmask8)(U))) 13618 13619 #define _mm_mask_cvtps_ph(W, U, A, I) \ 13620 ((__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf)(__m128) (A), (int) (I), \ 13621 (__v8hi)(__m128i) (W), (__mmask8) (U))) 13622 13623 #define _mm_maskz_cvtps_ph(U, A, I) \ 13624 ((__m128i) __builtin_ia32_vcvtps2ph_mask ((__v4sf)(__m128) (A), (int) (I), \ 13625 (__v8hi)(__m128i) _mm_setzero_si128 (), (__mmask8) (U))) 13626 13627 #define _mm256_mask_cvtps_ph(W, U, A, I) \ 13628 ((__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf)(__m256) (A), (int) (I), \ 13629 (__v8hi)(__m128i) (W), (__mmask8) (U))) 13630 13631 #define _mm256_maskz_cvtps_ph(U, A, I) \ 13632 ((__m128i) __builtin_ia32_vcvtps2ph256_mask ((__v8sf)(__m256) (A), (int) (I), \ 13633 (__v8hi)(__m128i) _mm_setzero_si128 (), (__mmask8) (U))) 13634 13635 #define _mm256_mask_srai_epi32(W, U, A, B) \ 13636 ((__m256i) __builtin_ia32_psradi256_mask ((__v8si)(__m256i)(A), \ 13637 (int)(B), (__v8si)(__m256i)(W), (__mmask8)(U))) 13638 13639 #define _mm256_maskz_srai_epi32(U, A, B) \ 13640 ((__m256i) __builtin_ia32_psradi256_mask ((__v8si)(__m256i)(A), \ 13641 (int)(B), (__v8si)_mm256_setzero_si256 (), (__mmask8)(U))) 13642 13643 #define _mm_mask_srai_epi32(W, U, A, B) \ 13644 ((__m128i) __builtin_ia32_psradi128_mask ((__v4si)(__m128i)(A), \ 13645 (int)(B), (__v4si)(__m128i)(W), (__mmask8)(U))) 13646 13647 #define _mm_maskz_srai_epi32(U, A, B) \ 13648 ((__m128i) __builtin_ia32_psradi128_mask ((__v4si)(__m128i)(A), \ 13649 (int)(B), (__v4si)_mm_setzero_si128 (), (__mmask8)(U))) 13650 13651 #define _mm256_srai_epi64(A, B) \ 13652 ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A), \ 13653 (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)-1)) 13654 13655 #define _mm256_mask_srai_epi64(W, U, A, B) \ 13656 ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A), \ 13657 (int)(B), (__v4di)(__m256i)(W), (__mmask8)(U))) 13658 13659 #define _mm256_maskz_srai_epi64(U, A, B) \ 13660 ((__m256i) __builtin_ia32_psraqi256_mask ((__v4di)(__m256i)(A), \ 13661 (int)(B), (__v4di)_mm256_setzero_si256 (), (__mmask8)(U))) 13662 13663 #define _mm_srai_epi64(A, B) \ 13664 ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A), \ 13665 (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)-1)) 13666 13667 #define _mm_mask_srai_epi64(W, U, A, B) \ 13668 ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A), \ 13669 (int)(B), (__v2di)(__m128i)(W), (__mmask8)(U))) 13670 13671 #define _mm_maskz_srai_epi64(U, A, B) \ 13672 ((__m128i) __builtin_ia32_psraqi128_mask ((__v2di)(__m128i)(A), \ 13673 (int)(B), (__v2di)_mm_setzero_si128 (), (__mmask8)(U))) 13674 13675 #define _mm256_mask_permutex_pd(W, U, A, B) \ 13676 ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(A), \ 13677 (int)(B), (__v4df)(__m256d)(W), (__mmask8)(U))) 13678 13679 #define _mm256_maskz_permutex_pd(U, A, B) \ 13680 ((__m256d) __builtin_ia32_permdf256_mask ((__v4df)(__m256d)(A), \ 13681 (int)(B), (__v4df)(__m256d)_mm256_setzero_pd (), (__mmask8)(U))) 13682 13683 #define _mm256_mask_permute_pd(W, U, X, C) \ 13684 ((__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df)(__m256d)(X), (int)(C), \ 13685 (__v4df)(__m256d)(W), \ 13686 (__mmask8)(U))) 13687 13688 #define _mm256_maskz_permute_pd(U, X, C) \ 13689 ((__m256d) __builtin_ia32_vpermilpd256_mask ((__v4df)(__m256d)(X), (int)(C), \ 13690 (__v4df)(__m256d)_mm256_setzero_pd (),\ 13691 (__mmask8)(U))) 13692 13693 #define _mm256_mask_permute_ps(W, U, X, C) \ 13694 ((__m256) __builtin_ia32_vpermilps256_mask ((__v8sf)(__m256)(X), (int)(C), \ 13695 (__v8sf)(__m256)(W), (__mmask8)(U))) 13696 13697 #define _mm256_maskz_permute_ps(U, X, C) \ 13698 ((__m256) __builtin_ia32_vpermilps256_mask ((__v8sf)(__m256)(X), (int)(C), \ 13699 (__v8sf)(__m256)_mm256_setzero_ps (), \ 13700 (__mmask8)(U))) 13701 13702 #define _mm_mask_permute_pd(W, U, X, C) \ 13703 ((__m128d) __builtin_ia32_vpermilpd_mask ((__v2df)(__m128d)(X), (int)(C), \ 13704 (__v2df)(__m128d)(W), (__mmask8)(U))) 13705 13706 #define _mm_maskz_permute_pd(U, X, C) \ 13707 ((__m128d) __builtin_ia32_vpermilpd_mask ((__v2df)(__m128d)(X), (int)(C), \ 13708 (__v2df)(__m128d)_mm_setzero_pd (), \ 13709 (__mmask8)(U))) 13710 13711 #define _mm_mask_permute_ps(W, U, X, C) \ 13712 ((__m128) __builtin_ia32_vpermilps_mask ((__v4sf)(__m128)(X), (int)(C), \ 13713 (__v4sf)(__m128)(W), (__mmask8)(U))) 13714 13715 #define _mm_maskz_permute_ps(U, X, C) \ 13716 ((__m128) __builtin_ia32_vpermilps_mask ((__v4sf)(__m128)(X), (int)(C), \ 13717 (__v4sf)(__m128)_mm_setzero_ps (), \ 13718 (__mmask8)(U))) 13719 13720 #define _mm256_mask_blend_pd(__U, __A, __W) \ 13721 ((__m256d) __builtin_ia32_blendmpd_256_mask ((__v4df) (__A), \ 13722 (__v4df) (__W), \ 13723 (__mmask8) (__U))) 13724 13725 #define _mm256_mask_blend_ps(__U, __A, __W) \ 13726 ((__m256) __builtin_ia32_blendmps_256_mask ((__v8sf) (__A), \ 13727 (__v8sf) (__W), \ 13728 (__mmask8) (__U))) 13729 13730 #define _mm256_mask_blend_epi64(__U, __A, __W) \ 13731 ((__m256i) __builtin_ia32_blendmq_256_mask ((__v4di) (__A), \ 13732 (__v4di) (__W), \ 13733 (__mmask8) (__U))) 13734 13735 #define _mm256_mask_blend_epi32(__U, __A, __W) \ 13736 ((__m256i) __builtin_ia32_blendmd_256_mask ((__v8si) (__A), \ 13737 (__v8si) (__W), \ 13738 (__mmask8) (__U))) 13739 13740 #define _mm_mask_blend_pd(__U, __A, __W) \ 13741 ((__m128d) __builtin_ia32_blendmpd_128_mask ((__v2df) (__A), \ 13742 (__v2df) (__W), \ 13743 (__mmask8) (__U))) 13744 13745 #define _mm_mask_blend_ps(__U, __A, __W) \ 13746 ((__m128) __builtin_ia32_blendmps_128_mask ((__v4sf) (__A), \ 13747 (__v4sf) (__W), \ 13748 (__mmask8) (__U))) 13749 13750 #define _mm_mask_blend_epi64(__U, __A, __W) \ 13751 ((__m128i) __builtin_ia32_blendmq_128_mask ((__v2di) (__A), \ 13752 (__v2di) (__W), \ 13753 (__mmask8) (__U))) 13754 13755 #define _mm_mask_blend_epi32(__U, __A, __W) \ 13756 ((__m128i) __builtin_ia32_blendmd_128_mask ((__v4si) (__A), \ 13757 (__v4si) (__W), \ 13758 (__mmask8) (__U))) 13759 13760 #define _mm256_cmp_epu32_mask(X, Y, P) \ 13761 ((__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si)(__m256i)(X), \ 13762 (__v8si)(__m256i)(Y), (int)(P),\ 13763 (__mmask8)-1)) 13764 13765 #define _mm256_cmp_epi64_mask(X, Y, P) \ 13766 ((__mmask8) __builtin_ia32_cmpq256_mask ((__v4di)(__m256i)(X), \ 13767 (__v4di)(__m256i)(Y), (int)(P),\ 13768 (__mmask8)-1)) 13769 13770 #define _mm256_cmp_epi32_mask(X, Y, P) \ 13771 ((__mmask8) __builtin_ia32_cmpd256_mask ((__v8si)(__m256i)(X), \ 13772 (__v8si)(__m256i)(Y), (int)(P),\ 13773 (__mmask8)-1)) 13774 13775 #define _mm256_cmp_epu64_mask(X, Y, P) \ 13776 ((__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di)(__m256i)(X), \ 13777 (__v4di)(__m256i)(Y), (int)(P),\ 13778 (__mmask8)-1)) 13779 13780 #define _mm256_cmp_pd_mask(X, Y, P) \ 13781 ((__mmask8) __builtin_ia32_cmppd256_mask ((__v4df)(__m256d)(X), \ 13782 (__v4df)(__m256d)(Y), (int)(P),\ 13783 (__mmask8)-1)) 13784 13785 #define _mm256_cmp_ps_mask(X, Y, P) \ 13786 ((__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf)(__m256)(X), \ 13787 (__v8sf)(__m256)(Y), (int)(P),\ 13788 (__mmask8)-1)) 13789 13790 #define _mm256_mask_cmp_epi64_mask(M, X, Y, P) \ 13791 ((__mmask8) __builtin_ia32_cmpq256_mask ((__v4di)(__m256i)(X), \ 13792 (__v4di)(__m256i)(Y), (int)(P),\ 13793 (__mmask8)(M))) 13794 13795 #define _mm256_mask_cmp_epi32_mask(M, X, Y, P) \ 13796 ((__mmask8) __builtin_ia32_cmpd256_mask ((__v8si)(__m256i)(X), \ 13797 (__v8si)(__m256i)(Y), (int)(P),\ 13798 (__mmask8)(M))) 13799 13800 #define _mm256_mask_cmp_epu64_mask(M, X, Y, P) \ 13801 ((__mmask8) __builtin_ia32_ucmpq256_mask ((__v4di)(__m256i)(X), \ 13802 (__v4di)(__m256i)(Y), (int)(P),\ 13803 (__mmask8)(M))) 13804 13805 #define _mm256_mask_cmp_epu32_mask(M, X, Y, P) \ 13806 ((__mmask8) __builtin_ia32_ucmpd256_mask ((__v8si)(__m256i)(X), \ 13807 (__v8si)(__m256i)(Y), (int)(P),\ 13808 (__mmask8)(M))) 13809 13810 #define _mm256_mask_cmp_pd_mask(M, X, Y, P) \ 13811 ((__mmask8) __builtin_ia32_cmppd256_mask ((__v4df)(__m256d)(X), \ 13812 (__v4df)(__m256d)(Y), (int)(P),\ 13813 (__mmask8)(M))) 13814 13815 #define _mm256_mask_cmp_ps_mask(M, X, Y, P) \ 13816 ((__mmask8) __builtin_ia32_cmpps256_mask ((__v8sf)(__m256)(X), \ 13817 (__v8sf)(__m256)(Y), (int)(P),\ 13818 (__mmask8)(M))) 13819 13820 #define _mm_cmp_epi64_mask(X, Y, P) \ 13821 ((__mmask8) __builtin_ia32_cmpq128_mask ((__v2di)(__m128i)(X), \ 13822 (__v2di)(__m128i)(Y), (int)(P),\ 13823 (__mmask8)-1)) 13824 13825 #define _mm_cmp_epi32_mask(X, Y, P) \ 13826 ((__mmask8) __builtin_ia32_cmpd128_mask ((__v4si)(__m128i)(X), \ 13827 (__v4si)(__m128i)(Y), (int)(P),\ 13828 (__mmask8)-1)) 13829 13830 #define _mm_cmp_epu64_mask(X, Y, P) \ 13831 ((__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di)(__m128i)(X), \ 13832 (__v2di)(__m128i)(Y), (int)(P),\ 13833 (__mmask8)-1)) 13834 13835 #define _mm_cmp_epu32_mask(X, Y, P) \ 13836 ((__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si)(__m128i)(X), \ 13837 (__v4si)(__m128i)(Y), (int)(P),\ 13838 (__mmask8)-1)) 13839 13840 #define _mm_cmp_pd_mask(X, Y, P) \ 13841 ((__mmask8) __builtin_ia32_cmppd128_mask ((__v2df)(__m128d)(X), \ 13842 (__v2df)(__m128d)(Y), (int)(P),\ 13843 (__mmask8)-1)) 13844 13845 #define _mm_cmp_ps_mask(X, Y, P) \ 13846 ((__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf)(__m128)(X), \ 13847 (__v4sf)(__m128)(Y), (int)(P),\ 13848 (__mmask8)-1)) 13849 13850 #define _mm_mask_cmp_epi64_mask(M, X, Y, P) \ 13851 ((__mmask8) __builtin_ia32_cmpq128_mask ((__v2di)(__m128i)(X), \ 13852 (__v2di)(__m128i)(Y), (int)(P),\ 13853 (__mmask8)(M))) 13854 13855 #define _mm_mask_cmp_epi32_mask(M, X, Y, P) \ 13856 ((__mmask8) __builtin_ia32_cmpd128_mask ((__v4si)(__m128i)(X), \ 13857 (__v4si)(__m128i)(Y), (int)(P),\ 13858 (__mmask8)(M))) 13859 13860 #define _mm_mask_cmp_epu64_mask(M, X, Y, P) \ 13861 ((__mmask8) __builtin_ia32_ucmpq128_mask ((__v2di)(__m128i)(X), \ 13862 (__v2di)(__m128i)(Y), (int)(P),\ 13863 (__mmask8)(M))) 13864 13865 #define _mm_mask_cmp_epu32_mask(M, X, Y, P) \ 13866 ((__mmask8) __builtin_ia32_ucmpd128_mask ((__v4si)(__m128i)(X), \ 13867 (__v4si)(__m128i)(Y), (int)(P),\ 13868 (__mmask8)(M))) 13869 13870 #define _mm_mask_cmp_pd_mask(M, X, Y, P) \ 13871 ((__mmask8) __builtin_ia32_cmppd128_mask ((__v2df)(__m128d)(X), \ 13872 (__v2df)(__m128d)(Y), (int)(P),\ 13873 (__mmask8)(M))) 13874 13875 #define _mm_mask_cmp_ps_mask(M, X, Y, P) \ 13876 ((__mmask8) __builtin_ia32_cmpps128_mask ((__v4sf)(__m128)(X), \ 13877 (__v4sf)(__m128)(Y), (int)(P),\ 13878 (__mmask8)(M))) 13879 13880 #endif 13881 13882 #define _mm256_permutexvar_ps(A, B) _mm256_permutevar8x32_ps ((B), (A)) 13883 #define _mm256_mask_cvt_roundps_ph(A, B, C, D) \ 13884 _mm256_mask_cvtps_ph ((A), (B), (C), (D)) 13885 #define _mm256_maskz_cvt_roundps_ph(A, B, C) \ 13886 _mm256_maskz_cvtps_ph ((A), (B), (C)) 13887 #define _mm_mask_cvt_roundps_ph(A, B, C, D) \ 13888 _mm_mask_cvtps_ph ((A), (B), (C), (D)) 13889 #define _mm_maskz_cvt_roundps_ph(A, B, C) _mm_maskz_cvtps_ph ((A), (B), (C)) 13890 13891 #ifdef __DISABLE_AVX512VL__ 13892 #undef __DISABLE_AVX512VL__ 13893 #pragma GCC pop_options 13894 #endif /* __DISABLE_AVX512VL__ */ 13895 13896 #endif /* _AVX512VLINTRIN_H_INCLUDED */ 13897
Indexes created Fri May 08 00:25:09 UTC 2026