clang/lib/Headers/avx512vldqintrin.h

   1 /*===---- avx512vldqintrin.h - AVX512VL and AVX512DQ intrinsics ------------===
   2  *
   3  * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4  * See https://llvm.org/LICENSE.txt for license information.
   5  * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6  *
   7  *===-----------------------------------------------------------------------===
   8  */
   9
  10 #ifndef __IMMINTRIN_H
  11 #error "Never use <avx512vldqintrin.h> directly; include <immintrin.h> instead."
  12 #endif
  13
  14 #ifndef __AVX512VLDQINTRIN_H
  15 #define __AVX512VLDQINTRIN_H
  16
  17 /* Define the default attributes for the functions in this file. */
  18 #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512dq"), __min_vector_width__(128)))
  19 #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512dq"), __min_vector_width__(256)))
  20
  21 static __inline__ __m256i __DEFAULT_FN_ATTRS256
  22 _mm256_mullo_epi64 (__m256i __A, __m256i __B) {
  23   return (__m256i) ((__v4du) __A * (__v4du) __B);
  24 }
  25
  26 static __inline__ __m256i __DEFAULT_FN_ATTRS256
  27 _mm256_mask_mullo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
  28   return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
  29                                              (__v4di)_mm256_mullo_epi64(__A, __B),
  30                                              (__v4di)__W);
  31 }
  32
  33 static __inline__ __m256i __DEFAULT_FN_ATTRS256
  34 _mm256_maskz_mullo_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
  35   return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
  36                                              (__v4di)_mm256_mullo_epi64(__A, __B),
  37                                              (__v4di)_mm256_setzero_si256());
  38 }
  39
  40 static __inline__ __m128i __DEFAULT_FN_ATTRS128
  41 _mm_mullo_epi64 (__m128i __A, __m128i __B) {
  42   return (__m128i) ((__v2du) __A * (__v2du) __B);
  43 }
  44
  45 static __inline__ __m128i __DEFAULT_FN_ATTRS128
  46 _mm_mask_mullo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
  47   return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
  48                                              (__v2di)_mm_mullo_epi64(__A, __B),
  49                                              (__v2di)__W);
  50 }
  51
  52 static __inline__ __m128i __DEFAULT_FN_ATTRS128
  53 _mm_maskz_mullo_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
  54   return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
  55                                              (__v2di)_mm_mullo_epi64(__A, __B),
  56                                              (__v2di)_mm_setzero_si128());
  57 }
  58
  59 static __inline__ __m256d __DEFAULT_FN_ATTRS256
  60 _mm256_mask_andnot_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
  61   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
  62                                               (__v4df)_mm256_andnot_pd(__A, __B),
  63                                               (__v4df)__W);
  64 }
  65
  66 static __inline__ __m256d __DEFAULT_FN_ATTRS256
  67 _mm256_maskz_andnot_pd(__mmask8 __U, __m256d __A, __m256d __B) {
  68   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
  69                                               (__v4df)_mm256_andnot_pd(__A, __B),
  70                                               (__v4df)_mm256_setzero_pd());
  71 }
  72
  73 static __inline__ __m128d __DEFAULT_FN_ATTRS128
  74 _mm_mask_andnot_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
  75   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
  76                                               (__v2df)_mm_andnot_pd(__A, __B),
  77                                               (__v2df)__W);
  78 }
  79
  80 static __inline__ __m128d __DEFAULT_FN_ATTRS128
  81 _mm_maskz_andnot_pd(__mmask8 __U, __m128d __A, __m128d __B) {
  82   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
  83                                               (__v2df)_mm_andnot_pd(__A, __B),
  84                                               (__v2df)_mm_setzero_pd());
  85 }
  86
  87 static __inline__ __m256 __DEFAULT_FN_ATTRS256
  88 _mm256_mask_andnot_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
  89   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
  90                                              (__v8sf)_mm256_andnot_ps(__A, __B),
  91                                              (__v8sf)__W);
  92 }
  93
  94 static __inline__ __m256 __DEFAULT_FN_ATTRS256
  95 _mm256_maskz_andnot_ps(__mmask8 __U, __m256 __A, __m256 __B) {
  96   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
  97                                              (__v8sf)_mm256_andnot_ps(__A, __B),
  98                                              (__v8sf)_mm256_setzero_ps());
  99 }
 100
 101 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 102 _mm_mask_andnot_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
 103   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 104                                              (__v4sf)_mm_andnot_ps(__A, __B),
 105                                              (__v4sf)__W);
 106 }
 107
 108 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 109 _mm_maskz_andnot_ps(__mmask8 __U, __m128 __A, __m128 __B) {
 110   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 111                                              (__v4sf)_mm_andnot_ps(__A, __B),
 112                                              (__v4sf)_mm_setzero_ps());
 113 }
 114
 115 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 116 _mm256_mask_and_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
 117   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 118                                               (__v4df)_mm256_and_pd(__A, __B),
 119                                               (__v4df)__W);
 120 }
 121
 122 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 123 _mm256_maskz_and_pd(__mmask8 __U, __m256d __A, __m256d __B) {
 124   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 125                                               (__v4df)_mm256_and_pd(__A, __B),
 126                                               (__v4df)_mm256_setzero_pd());
 127 }
 128
 129 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 130 _mm_mask_and_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
 131   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 132                                               (__v2df)_mm_and_pd(__A, __B),
 133                                               (__v2df)__W);
 134 }
 135
 136 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 137 _mm_maskz_and_pd(__mmask8 __U, __m128d __A, __m128d __B) {
 138   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 139                                               (__v2df)_mm_and_pd(__A, __B),
 140                                               (__v2df)_mm_setzero_pd());
 141 }
 142
 143 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 144 _mm256_mask_and_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
 145   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 146                                              (__v8sf)_mm256_and_ps(__A, __B),
 147                                              (__v8sf)__W);
 148 }
 149
 150 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 151 _mm256_maskz_and_ps(__mmask8 __U, __m256 __A, __m256 __B) {
 152   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 153                                              (__v8sf)_mm256_and_ps(__A, __B),
 154                                              (__v8sf)_mm256_setzero_ps());
 155 }
 156
 157 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 158 _mm_mask_and_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
 159   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 160                                              (__v4sf)_mm_and_ps(__A, __B),
 161                                              (__v4sf)__W);
 162 }
 163
 164 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 165 _mm_maskz_and_ps(__mmask8 __U, __m128 __A, __m128 __B) {
 166   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 167                                              (__v4sf)_mm_and_ps(__A, __B),
 168                                              (__v4sf)_mm_setzero_ps());
 169 }
 170
 171 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 172 _mm256_mask_xor_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
 173   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 174                                               (__v4df)_mm256_xor_pd(__A, __B),
 175                                               (__v4df)__W);
 176 }
 177
 178 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 179 _mm256_maskz_xor_pd(__mmask8 __U, __m256d __A, __m256d __B) {
 180   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 181                                               (__v4df)_mm256_xor_pd(__A, __B),
 182                                               (__v4df)_mm256_setzero_pd());
 183 }
 184
 185 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 186 _mm_mask_xor_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
 187   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 188                                               (__v2df)_mm_xor_pd(__A, __B),
 189                                               (__v2df)__W);
 190 }
 191
 192 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 193 _mm_maskz_xor_pd (__mmask8 __U, __m128d __A, __m128d __B) {
 194   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 195                                               (__v2df)_mm_xor_pd(__A, __B),
 196                                               (__v2df)_mm_setzero_pd());
 197 }
 198
 199 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 200 _mm256_mask_xor_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
 201   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 202                                              (__v8sf)_mm256_xor_ps(__A, __B),
 203                                              (__v8sf)__W);
 204 }
 205
 206 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 207 _mm256_maskz_xor_ps(__mmask8 __U, __m256 __A, __m256 __B) {
 208   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 209                                              (__v8sf)_mm256_xor_ps(__A, __B),
 210                                              (__v8sf)_mm256_setzero_ps());
 211 }
 212
 213 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 214 _mm_mask_xor_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
 215   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 216                                              (__v4sf)_mm_xor_ps(__A, __B),
 217                                              (__v4sf)__W);
 218 }
 219
 220 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 221 _mm_maskz_xor_ps(__mmask8 __U, __m128 __A, __m128 __B) {
 222   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 223                                              (__v4sf)_mm_xor_ps(__A, __B),
 224                                              (__v4sf)_mm_setzero_ps());
 225 }
 226
 227 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 228 _mm256_mask_or_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
 229   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 230                                               (__v4df)_mm256_or_pd(__A, __B),
 231                                               (__v4df)__W);
 232 }
 233
 234 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 235 _mm256_maskz_or_pd(__mmask8 __U, __m256d __A, __m256d __B) {
 236   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 237                                               (__v4df)_mm256_or_pd(__A, __B),
 238                                               (__v4df)_mm256_setzero_pd());
 239 }
 240
 241 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 242 _mm_mask_or_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
 243   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 244                                               (__v2df)_mm_or_pd(__A, __B),
 245                                               (__v2df)__W);
 246 }
 247
 248 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 249 _mm_maskz_or_pd(__mmask8 __U, __m128d __A, __m128d __B) {
 250   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 251                                               (__v2df)_mm_or_pd(__A, __B),
 252                                               (__v2df)_mm_setzero_pd());
 253 }
 254
 255 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 256 _mm256_mask_or_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
 257   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 258                                              (__v8sf)_mm256_or_ps(__A, __B),
 259                                              (__v8sf)__W);
 260 }
 261
 262 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 263 _mm256_maskz_or_ps(__mmask8 __U, __m256 __A, __m256 __B) {
 264   return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
 265                                              (__v8sf)_mm256_or_ps(__A, __B),
 266                                              (__v8sf)_mm256_setzero_ps());
 267 }
 268
 269 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 270 _mm_mask_or_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
 271   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 272                                              (__v4sf)_mm_or_ps(__A, __B),
 273                                              (__v4sf)__W);
 274 }
 275
 276 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 277 _mm_maskz_or_ps(__mmask8 __U, __m128 __A, __m128 __B) {
 278   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 279                                              (__v4sf)_mm_or_ps(__A, __B),
 280                                              (__v4sf)_mm_setzero_ps());
 281 }
 282
 283 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 284 _mm_cvtpd_epi64 (__m128d __A) {
 285   return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A,
 286                 (__v2di) _mm_setzero_si128(),
 287                 (__mmask8) -1);
 288 }
 289
 290 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 291 _mm_mask_cvtpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) {
 292   return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A,
 293                 (__v2di) __W,
 294                 (__mmask8) __U);
 295 }
 296
 297 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 298 _mm_maskz_cvtpd_epi64 (__mmask8 __U, __m128d __A) {
 299   return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A,
 300                 (__v2di) _mm_setzero_si128(),
 301                 (__mmask8) __U);
 302 }
 303
 304 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 305 _mm256_cvtpd_epi64 (__m256d __A) {
 306   return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A,
 307                 (__v4di) _mm256_setzero_si256(),
 308                 (__mmask8) -1);
 309 }
 310
 311 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 312 _mm256_mask_cvtpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) {
 313   return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A,
 314                 (__v4di) __W,
 315                 (__mmask8) __U);
 316 }
 317
 318 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 319 _mm256_maskz_cvtpd_epi64 (__mmask8 __U, __m256d __A) {
 320   return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A,
 321                 (__v4di) _mm256_setzero_si256(),
 322                 (__mmask8) __U);
 323 }
 324
 325 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 326 _mm_cvtpd_epu64 (__m128d __A) {
 327   return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A,
 328                 (__v2di) _mm_setzero_si128(),
 329                 (__mmask8) -1);
 330 }
 331
 332 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 333 _mm_mask_cvtpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) {
 334   return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A,
 335                 (__v2di) __W,
 336                 (__mmask8) __U);
 337 }
 338
 339 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 340 _mm_maskz_cvtpd_epu64 (__mmask8 __U, __m128d __A) {
 341   return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A,
 342                 (__v2di) _mm_setzero_si128(),
 343                 (__mmask8) __U);
 344 }
 345
 346 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 347 _mm256_cvtpd_epu64 (__m256d __A) {
 348   return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A,
 349                 (__v4di) _mm256_setzero_si256(),
 350                 (__mmask8) -1);
 351 }
 352
 353 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 354 _mm256_mask_cvtpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) {
 355   return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A,
 356                 (__v4di) __W,
 357                 (__mmask8) __U);
 358 }
 359
 360 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 361 _mm256_maskz_cvtpd_epu64 (__mmask8 __U, __m256d __A) {
 362   return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A,
 363                 (__v4di) _mm256_setzero_si256(),
 364                 (__mmask8) __U);
 365 }
 366
 367 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 368 _mm_cvtps_epi64 (__m128 __A) {
 369   return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A,
 370                 (__v2di) _mm_setzero_si128(),
 371                 (__mmask8) -1);
 372 }
 373
 374 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 375 _mm_mask_cvtps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) {
 376   return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A,
 377                 (__v2di) __W,
 378                 (__mmask8) __U);
 379 }
 380
 381 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 382 _mm_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) {
 383   return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A,
 384                 (__v2di) _mm_setzero_si128(),
 385                 (__mmask8) __U);
 386 }
 387
 388 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 389 _mm256_cvtps_epi64 (__m128 __A) {
 390   return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A,
 391                 (__v4di) _mm256_setzero_si256(),
 392                 (__mmask8) -1);
 393 }
 394
 395 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 396 _mm256_mask_cvtps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) {
 397   return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A,
 398                 (__v4di) __W,
 399                 (__mmask8) __U);
 400 }
 401
 402 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 403 _mm256_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) {
 404   return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A,
 405                 (__v4di) _mm256_setzero_si256(),
 406                 (__mmask8) __U);
 407 }
 408
 409 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 410 _mm_cvtps_epu64 (__m128 __A) {
 411   return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A,
 412                 (__v2di) _mm_setzero_si128(),
 413                 (__mmask8) -1);
 414 }
 415
 416 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 417 _mm_mask_cvtps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) {
 418   return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A,
 419                 (__v2di) __W,
 420                 (__mmask8) __U);
 421 }
 422
 423 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 424 _mm_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) {
 425   return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A,
 426                 (__v2di) _mm_setzero_si128(),
 427                 (__mmask8) __U);
 428 }
 429
 430 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 431 _mm256_cvtps_epu64 (__m128 __A) {
 432   return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A,
 433                 (__v4di) _mm256_setzero_si256(),
 434                 (__mmask8) -1);
 435 }
 436
 437 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 438 _mm256_mask_cvtps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) {
 439   return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A,
 440                 (__v4di) __W,
 441                 (__mmask8) __U);
 442 }
 443
 444 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 445 _mm256_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) {
 446   return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A,
 447                 (__v4di) _mm256_setzero_si256(),
 448                 (__mmask8) __U);
 449 }
 450
 451 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 452 _mm_cvtepi64_pd (__m128i __A) {
 453   return (__m128d)__builtin_convertvector((__v2di)__A, __v2df);
 454 }
 455
 456 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 457 _mm_mask_cvtepi64_pd (__m128d __W, __mmask8 __U, __m128i __A) {
 458   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 459                                               (__v2df)_mm_cvtepi64_pd(__A),
 460                                               (__v2df)__W);
 461 }
 462
 463 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 464 _mm_maskz_cvtepi64_pd (__mmask8 __U, __m128i __A) {
 465   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 466                                               (__v2df)_mm_cvtepi64_pd(__A),
 467                                               (__v2df)_mm_setzero_pd());
 468 }
 469
 470 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 471 _mm256_cvtepi64_pd (__m256i __A) {
 472   return (__m256d)__builtin_convertvector((__v4di)__A, __v4df);
 473 }
 474
 475 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 476 _mm256_mask_cvtepi64_pd (__m256d __W, __mmask8 __U, __m256i __A) {
 477   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 478                                               (__v4df)_mm256_cvtepi64_pd(__A),
 479                                               (__v4df)__W);
 480 }
 481
 482 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 483 _mm256_maskz_cvtepi64_pd (__mmask8 __U, __m256i __A) {
 484   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 485                                               (__v4df)_mm256_cvtepi64_pd(__A),
 486                                               (__v4df)_mm256_setzero_pd());
 487 }
 488
 489 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 490 _mm_cvtepi64_ps (__m128i __A) {
 491   return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A,
 492                 (__v4sf) _mm_setzero_ps(),
 493                 (__mmask8) -1);
 494 }
 495
 496 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 497 _mm_mask_cvtepi64_ps (__m128 __W, __mmask8 __U, __m128i __A) {
 498   return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A,
 499                 (__v4sf) __W,
 500                 (__mmask8) __U);
 501 }
 502
 503 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 504 _mm_maskz_cvtepi64_ps (__mmask8 __U, __m128i __A) {
 505   return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A,
 506                 (__v4sf) _mm_setzero_ps(),
 507                 (__mmask8) __U);
 508 }
 509
 510 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 511 _mm256_cvtepi64_ps (__m256i __A) {
 512   return (__m128)__builtin_convertvector((__v4di)__A, __v4sf);
 513 }
 514
 515 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 516 _mm256_mask_cvtepi64_ps (__m128 __W, __mmask8 __U, __m256i __A) {
 517   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 518                                              (__v4sf)_mm256_cvtepi64_ps(__A),
 519                                              (__v4sf)__W);
 520 }
 521
 522 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 523 _mm256_maskz_cvtepi64_ps (__mmask8 __U, __m256i __A) {
 524   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 525                                              (__v4sf)_mm256_cvtepi64_ps(__A),
 526                                              (__v4sf)_mm_setzero_ps());
 527 }
 528
 529 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 530 _mm_cvttpd_epi64 (__m128d __A) {
 531   return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A,
 532                 (__v2di) _mm_setzero_si128(),
 533                 (__mmask8) -1);
 534 }
 535
 536 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 537 _mm_mask_cvttpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) {
 538   return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A,
 539                 (__v2di) __W,
 540                 (__mmask8) __U);
 541 }
 542
 543 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 544 _mm_maskz_cvttpd_epi64 (__mmask8 __U, __m128d __A) {
 545   return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A,
 546                 (__v2di) _mm_setzero_si128(),
 547                 (__mmask8) __U);
 548 }
 549
 550 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 551 _mm256_cvttpd_epi64 (__m256d __A) {
 552   return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A,
 553                 (__v4di) _mm256_setzero_si256(),
 554                 (__mmask8) -1);
 555 }
 556
 557 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 558 _mm256_mask_cvttpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) {
 559   return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A,
 560                 (__v4di) __W,
 561                 (__mmask8) __U);
 562 }
 563
 564 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 565 _mm256_maskz_cvttpd_epi64 (__mmask8 __U, __m256d __A) {
 566   return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A,
 567                 (__v4di) _mm256_setzero_si256(),
 568                 (__mmask8) __U);
 569 }
 570
 571 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 572 _mm_cvttpd_epu64 (__m128d __A) {
 573   return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A,
 574                 (__v2di) _mm_setzero_si128(),
 575                 (__mmask8) -1);
 576 }
 577
 578 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 579 _mm_mask_cvttpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) {
 580   return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A,
 581                 (__v2di) __W,
 582                 (__mmask8) __U);
 583 }
 584
 585 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 586 _mm_maskz_cvttpd_epu64 (__mmask8 __U, __m128d __A) {
 587   return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A,
 588                 (__v2di) _mm_setzero_si128(),
 589                 (__mmask8) __U);
 590 }
 591
 592 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 593 _mm256_cvttpd_epu64 (__m256d __A) {
 594   return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A,
 595                 (__v4di) _mm256_setzero_si256(),
 596                 (__mmask8) -1);
 597 }
 598
 599 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 600 _mm256_mask_cvttpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) {
 601   return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A,
 602                 (__v4di) __W,
 603                 (__mmask8) __U);
 604 }
 605
 606 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 607 _mm256_maskz_cvttpd_epu64 (__mmask8 __U, __m256d __A) {
 608   return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A,
 609                 (__v4di) _mm256_setzero_si256(),
 610                 (__mmask8) __U);
 611 }
 612
 613 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 614 _mm_cvttps_epi64 (__m128 __A) {
 615   return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A,
 616                 (__v2di) _mm_setzero_si128(),
 617                 (__mmask8) -1);
 618 }
 619
 620 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 621 _mm_mask_cvttps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) {
 622   return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A,
 623                 (__v2di) __W,
 624                 (__mmask8) __U);
 625 }
 626
 627 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 628 _mm_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) {
 629   return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A,
 630                 (__v2di) _mm_setzero_si128(),
 631                 (__mmask8) __U);
 632 }
 633
 634 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 635 _mm256_cvttps_epi64 (__m128 __A) {
 636   return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A,
 637                 (__v4di) _mm256_setzero_si256(),
 638                 (__mmask8) -1);
 639 }
 640
 641 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 642 _mm256_mask_cvttps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) {
 643   return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A,
 644                 (__v4di) __W,
 645                 (__mmask8) __U);
 646 }
 647
 648 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 649 _mm256_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) {
 650   return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A,
 651                 (__v4di) _mm256_setzero_si256(),
 652                 (__mmask8) __U);
 653 }
 654
 655 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 656 _mm_cvttps_epu64 (__m128 __A) {
 657   return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A,
 658                 (__v2di) _mm_setzero_si128(),
 659                 (__mmask8) -1);
 660 }
 661
 662 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 663 _mm_mask_cvttps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) {
 664   return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A,
 665                 (__v2di) __W,
 666                 (__mmask8) __U);
 667 }
 668
 669 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 670 _mm_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) {
 671   return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A,
 672                 (__v2di) _mm_setzero_si128(),
 673                 (__mmask8) __U);
 674 }
 675
 676 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 677 _mm256_cvttps_epu64 (__m128 __A) {
 678   return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A,
 679                 (__v4di) _mm256_setzero_si256(),
 680                 (__mmask8) -1);
 681 }
 682
 683 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 684 _mm256_mask_cvttps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) {
 685   return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A,
 686                 (__v4di) __W,
 687                 (__mmask8) __U);
 688 }
 689
 690 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 691 _mm256_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) {
 692   return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A,
 693                 (__v4di) _mm256_setzero_si256(),
 694                 (__mmask8) __U);
 695 }
 696
 697 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 698 _mm_cvtepu64_pd (__m128i __A) {
 699   return (__m128d)__builtin_convertvector((__v2du)__A, __v2df);
 700 }
 701
 702 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 703 _mm_mask_cvtepu64_pd (__m128d __W, __mmask8 __U, __m128i __A) {
 704   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 705                                               (__v2df)_mm_cvtepu64_pd(__A),
 706                                               (__v2df)__W);
 707 }
 708
 709 static __inline__ __m128d __DEFAULT_FN_ATTRS128
 710 _mm_maskz_cvtepu64_pd (__mmask8 __U, __m128i __A) {
 711   return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
 712                                               (__v2df)_mm_cvtepu64_pd(__A),
 713                                               (__v2df)_mm_setzero_pd());
 714 }
 715
 716 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 717 _mm256_cvtepu64_pd (__m256i __A) {
 718   return (__m256d)__builtin_convertvector((__v4du)__A, __v4df);
 719 }
 720
 721 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 722 _mm256_mask_cvtepu64_pd (__m256d __W, __mmask8 __U, __m256i __A) {
 723   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 724                                               (__v4df)_mm256_cvtepu64_pd(__A),
 725                                               (__v4df)__W);
 726 }
 727
 728 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 729 _mm256_maskz_cvtepu64_pd (__mmask8 __U, __m256i __A) {
 730   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
 731                                               (__v4df)_mm256_cvtepu64_pd(__A),
 732                                               (__v4df)_mm256_setzero_pd());
 733 }
 734
 735 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 736 _mm_cvtepu64_ps (__m128i __A) {
 737   return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A,
 738                 (__v4sf) _mm_setzero_ps(),
 739                 (__mmask8) -1);
 740 }
 741
 742 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 743 _mm_mask_cvtepu64_ps (__m128 __W, __mmask8 __U, __m128i __A) {
 744   return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A,
 745                 (__v4sf) __W,
 746                 (__mmask8) __U);
 747 }
 748
 749 static __inline__ __m128 __DEFAULT_FN_ATTRS128
 750 _mm_maskz_cvtepu64_ps (__mmask8 __U, __m128i __A) {
 751   return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A,
 752                 (__v4sf) _mm_setzero_ps(),
 753                 (__mmask8) __U);
 754 }
 755
 756 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 757 _mm256_cvtepu64_ps (__m256i __A) {
 758   return (__m128)__builtin_convertvector((__v4du)__A, __v4sf);
 759 }
 760
 761 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 762 _mm256_mask_cvtepu64_ps (__m128 __W, __mmask8 __U, __m256i __A) {
 763   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 764                                              (__v4sf)_mm256_cvtepu64_ps(__A),
 765                                              (__v4sf)__W);
 766 }
 767
 768 static __inline__ __m128 __DEFAULT_FN_ATTRS256
 769 _mm256_maskz_cvtepu64_ps (__mmask8 __U, __m256i __A) {
 770   return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
 771                                              (__v4sf)_mm256_cvtepu64_ps(__A),
 772                                              (__v4sf)_mm_setzero_ps());
 773 }
 774
 775 #define _mm_range_pd(A, B, C) \
 776   ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \
 777                                            (__v2df)(__m128d)(B), (int)(C), \
 778                                            (__v2df)_mm_setzero_pd(), \
 779                                            (__mmask8)-1))
 780
 781 #define _mm_mask_range_pd(W, U, A, B, C) \
 782   ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \
 783                                            (__v2df)(__m128d)(B), (int)(C), \
 784                                            (__v2df)(__m128d)(W), \
 785                                            (__mmask8)(U)))
 786
 787 #define _mm_maskz_range_pd(U, A, B, C) \
 788   ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \
 789                                            (__v2df)(__m128d)(B), (int)(C), \
 790                                            (__v2df)_mm_setzero_pd(), \
 791                                            (__mmask8)(U)))
 792
 793 #define _mm256_range_pd(A, B, C) \
 794   ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \
 795                                            (__v4df)(__m256d)(B), (int)(C), \
 796                                            (__v4df)_mm256_setzero_pd(), \
 797                                            (__mmask8)-1))
 798
 799 #define _mm256_mask_range_pd(W, U, A, B, C) \
 800   ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \
 801                                            (__v4df)(__m256d)(B), (int)(C), \
 802                                            (__v4df)(__m256d)(W), \
 803                                            (__mmask8)(U)))
 804
 805 #define _mm256_maskz_range_pd(U, A, B, C) \
 806   ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \
 807                                            (__v4df)(__m256d)(B), (int)(C), \
 808                                            (__v4df)_mm256_setzero_pd(), \
 809                                            (__mmask8)(U)))
 810
 811 #define _mm_range_ps(A, B, C) \
 812   ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \
 813                                           (__v4sf)(__m128)(B), (int)(C), \
 814                                           (__v4sf)_mm_setzero_ps(), \
 815                                           (__mmask8)-1))
 816
 817 #define _mm_mask_range_ps(W, U, A, B, C) \
 818   ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \
 819                                           (__v4sf)(__m128)(B), (int)(C), \
 820                                           (__v4sf)(__m128)(W), (__mmask8)(U)))
 821
 822 #define _mm_maskz_range_ps(U, A, B, C) \
 823   ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \
 824                                           (__v4sf)(__m128)(B), (int)(C), \
 825                                           (__v4sf)_mm_setzero_ps(), \
 826                                           (__mmask8)(U)))
 827
 828 #define _mm256_range_ps(A, B, C) \
 829   ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \
 830                                           (__v8sf)(__m256)(B), (int)(C), \
 831                                           (__v8sf)_mm256_setzero_ps(), \
 832                                           (__mmask8)-1))
 833
 834 #define _mm256_mask_range_ps(W, U, A, B, C) \
 835   ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \
 836                                           (__v8sf)(__m256)(B), (int)(C), \
 837                                           (__v8sf)(__m256)(W), (__mmask8)(U)))
 838
 839 #define _mm256_maskz_range_ps(U, A, B, C) \
 840   ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \
 841                                           (__v8sf)(__m256)(B), (int)(C), \
 842                                           (__v8sf)_mm256_setzero_ps(), \
 843                                           (__mmask8)(U)))
 844
 845 #define _mm_reduce_pd(A, B) \
 846   ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \
 847                                             (__v2df)_mm_setzero_pd(), \
 848                                             (__mmask8)-1))
 849
 850 #define _mm_mask_reduce_pd(W, U, A, B) \
 851   ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \
 852                                             (__v2df)(__m128d)(W), \
 853                                             (__mmask8)(U)))
 854
 855 #define _mm_maskz_reduce_pd(U, A, B) \
 856   ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \
 857                                             (__v2df)_mm_setzero_pd(), \
 858                                             (__mmask8)(U)))
 859
 860 #define _mm256_reduce_pd(A, B) \
 861   ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \
 862                                             (__v4df)_mm256_setzero_pd(), \
 863                                             (__mmask8)-1))
 864
 865 #define _mm256_mask_reduce_pd(W, U, A, B) \
 866   ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \
 867                                             (__v4df)(__m256d)(W), \
 868                                             (__mmask8)(U)))
 869
 870 #define _mm256_maskz_reduce_pd(U, A, B) \
 871   ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \
 872                                             (__v4df)_mm256_setzero_pd(), \
 873                                             (__mmask8)(U)))
 874
 875 #define _mm_reduce_ps(A, B) \
 876   ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \
 877                                            (__v4sf)_mm_setzero_ps(), \
 878                                            (__mmask8)-1))
 879
 880 #define _mm_mask_reduce_ps(W, U, A, B) \
 881   ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \
 882                                            (__v4sf)(__m128)(W), \
 883                                            (__mmask8)(U)))
 884
 885 #define _mm_maskz_reduce_ps(U, A, B) \
 886   ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \
 887                                            (__v4sf)_mm_setzero_ps(), \
 888                                            (__mmask8)(U)))
 889
 890 #define _mm256_reduce_ps(A, B) \
 891   ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \
 892                                            (__v8sf)_mm256_setzero_ps(), \
 893                                            (__mmask8)-1))
 894
 895 #define _mm256_mask_reduce_ps(W, U, A, B) \
 896   ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \
 897                                            (__v8sf)(__m256)(W), \
 898                                            (__mmask8)(U)))
 899
 900 #define _mm256_maskz_reduce_ps(U, A, B) \
 901   ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \
 902                                            (__v8sf)_mm256_setzero_ps(), \
 903                                            (__mmask8)(U)))
 904
 905 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
 906 _mm_movepi32_mask (__m128i __A)
 907 {
 908   return (__mmask8) __builtin_ia32_cvtd2mask128 ((__v4si) __A);
 909 }
 910
 911 static __inline__ __mmask8 __DEFAULT_FN_ATTRS256
 912 _mm256_movepi32_mask (__m256i __A)
 913 {
 914   return (__mmask8) __builtin_ia32_cvtd2mask256 ((__v8si) __A);
 915 }
 916
 917 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 918 _mm_movm_epi32 (__mmask8 __A)
 919 {
 920   return (__m128i) __builtin_ia32_cvtmask2d128 (__A);
 921 }
 922
 923 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 924 _mm256_movm_epi32 (__mmask8 __A)
 925 {
 926   return (__m256i) __builtin_ia32_cvtmask2d256 (__A);
 927 }
 928
 929 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 930 _mm_movm_epi64 (__mmask8 __A)
 931 {
 932   return (__m128i) __builtin_ia32_cvtmask2q128 (__A);
 933 }
 934
 935 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 936 _mm256_movm_epi64 (__mmask8 __A)
 937 {
 938   return (__m256i) __builtin_ia32_cvtmask2q256 (__A);
 939 }
 940
 941 static __inline__ __mmask8 __DEFAULT_FN_ATTRS128
 942 _mm_movepi64_mask (__m128i __A)
 943 {
 944   return (__mmask8) __builtin_ia32_cvtq2mask128 ((__v2di) __A);
 945 }
 946
 947 static __inline__ __mmask8 __DEFAULT_FN_ATTRS256
 948 _mm256_movepi64_mask (__m256i __A)
 949 {
 950   return (__mmask8) __builtin_ia32_cvtq2mask256 ((__v4di) __A);
 951 }
 952
 953 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 954 _mm256_broadcast_f32x2 (__m128 __A)
 955 {
 956   return (__m256)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
 957                                          0, 1, 0, 1, 0, 1, 0, 1);
 958 }
 959
 960 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 961 _mm256_mask_broadcast_f32x2 (__m256 __O, __mmask8 __M, __m128 __A)
 962 {
 963   return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
 964                                              (__v8sf)_mm256_broadcast_f32x2(__A),
 965                                              (__v8sf)__O);
 966 }
 967
 968 static __inline__ __m256 __DEFAULT_FN_ATTRS256
 969 _mm256_maskz_broadcast_f32x2 (__mmask8 __M, __m128 __A)
 970 {
 971   return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
 972                                              (__v8sf)_mm256_broadcast_f32x2(__A),
 973                                              (__v8sf)_mm256_setzero_ps());
 974 }
 975
 976 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 977 _mm256_broadcast_f64x2(__m128d __A)
 978 {
 979   return (__m256d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
 980                                           0, 1, 0, 1);
 981 }
 982
 983 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 984 _mm256_mask_broadcast_f64x2(__m256d __O, __mmask8 __M, __m128d __A)
 985 {
 986   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M,
 987                                             (__v4df)_mm256_broadcast_f64x2(__A),
 988                                             (__v4df)__O);
 989 }
 990
 991 static __inline__ __m256d __DEFAULT_FN_ATTRS256
 992 _mm256_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A)
 993 {
 994   return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M,
 995                                             (__v4df)_mm256_broadcast_f64x2(__A),
 996                                             (__v4df)_mm256_setzero_pd());
 997 }
 998
 999 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1000 _mm_broadcast_i32x2 (__m128i __A)
1001 {
1002   return (__m128i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1003                                           0, 1, 0, 1);
1004 }
1005
1006 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1007 _mm_mask_broadcast_i32x2 (__m128i __O, __mmask8 __M, __m128i __A)
1008 {
1009   return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
1010                                              (__v4si)_mm_broadcast_i32x2(__A),
1011                                              (__v4si)__O);
1012 }
1013
1014 static __inline__ __m128i __DEFAULT_FN_ATTRS128
1015 _mm_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A)
1016 {
1017   return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
1018                                              (__v4si)_mm_broadcast_i32x2(__A),
1019                                              (__v4si)_mm_setzero_si128());
1020 }
1021
1022 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1023 _mm256_broadcast_i32x2 (__m128i __A)
1024 {
1025   return (__m256i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1026                                           0, 1, 0, 1, 0, 1, 0, 1);
1027 }
1028
1029 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1030 _mm256_mask_broadcast_i32x2 (__m256i __O, __mmask8 __M, __m128i __A)
1031 {
1032   return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
1033                                              (__v8si)_mm256_broadcast_i32x2(__A),
1034                                              (__v8si)__O);
1035 }
1036
1037 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1038 _mm256_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A)
1039 {
1040   return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
1041                                              (__v8si)_mm256_broadcast_i32x2(__A),
1042                                              (__v8si)_mm256_setzero_si256());
1043 }
1044
1045 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1046 _mm256_broadcast_i64x2(__m128i __A)
1047 {
1048   return (__m256i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1049                                           0, 1, 0, 1);
1050 }
1051
1052 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1053 _mm256_mask_broadcast_i64x2(__m256i __O, __mmask8 __M, __m128i __A)
1054 {
1055   return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
1056                                             (__v4di)_mm256_broadcast_i64x2(__A),
1057                                             (__v4di)__O);
1058 }
1059
1060 static __inline__ __m256i __DEFAULT_FN_ATTRS256
1061 _mm256_maskz_broadcast_i64x2 (__mmask8 __M, __m128i __A)
1062 {
1063   return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
1064                                             (__v4di)_mm256_broadcast_i64x2(__A),
1065                                             (__v4di)_mm256_setzero_si256());
1066 }
1067
1068 #define _mm256_extractf64x2_pd(A, imm) \
1069   ((__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \
1070                                                  (int)(imm), \
1071                                                  (__v2df)_mm_undefined_pd(), \
1072                                                  (__mmask8)-1))
1073
1074 #define _mm256_mask_extractf64x2_pd(W, U, A, imm) \
1075   ((__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \
1076                                                  (int)(imm), \
1077                                                  (__v2df)(__m128d)(W), \
1078                                                  (__mmask8)(U)))
1079
1080 #define _mm256_maskz_extractf64x2_pd(U, A, imm) \
1081   ((__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \
1082                                                  (int)(imm), \
1083                                                  (__v2df)_mm_setzero_pd(), \
1084                                                  (__mmask8)(U)))
1085
1086 #define _mm256_extracti64x2_epi64(A, imm) \
1087   ((__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \
1088                                                 (int)(imm), \
1089                                                 (__v2di)_mm_undefined_si128(), \
1090                                                 (__mmask8)-1))
1091
1092 #define _mm256_mask_extracti64x2_epi64(W, U, A, imm) \
1093   ((__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \
1094                                                  (int)(imm), \
1095                                                  (__v2di)(__m128i)(W), \
1096                                                  (__mmask8)(U)))
1097
1098 #define _mm256_maskz_extracti64x2_epi64(U, A, imm) \
1099   ((__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \
1100                                                  (int)(imm), \
1101                                                  (__v2di)_mm_setzero_si128(), \
1102                                                  (__mmask8)(U)))
1103
1104 #define _mm256_insertf64x2(A, B, imm) \
1105   ((__m256d)__builtin_ia32_insertf64x2_256((__v4df)(__m256d)(A), \
1106                                            (__v2df)(__m128d)(B), (int)(imm)))
1107
1108 #define _mm256_mask_insertf64x2(W, U, A, B, imm) \
1109   ((__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
1110                                   (__v4df)_mm256_insertf64x2((A), (B), (imm)), \
1111                                   (__v4df)(__m256d)(W)))
1112
1113 #define _mm256_maskz_insertf64x2(U, A, B, imm) \
1114   ((__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \
1115                                   (__v4df)_mm256_insertf64x2((A), (B), (imm)), \
1116                                   (__v4df)_mm256_setzero_pd()))
1117
1118 #define _mm256_inserti64x2(A, B, imm) \
1119   ((__m256i)__builtin_ia32_inserti64x2_256((__v4di)(__m256i)(A), \
1120                                            (__v2di)(__m128i)(B), (int)(imm)))
1121
1122 #define _mm256_mask_inserti64x2(W, U, A, B, imm) \
1123   ((__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
1124                                    (__v4di)_mm256_inserti64x2((A), (B), (imm)), \
1125                                    (__v4di)(__m256i)(W)))
1126
1127 #define _mm256_maskz_inserti64x2(U, A, B, imm) \
1128   ((__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \
1129                                    (__v4di)_mm256_inserti64x2((A), (B), (imm)), \
1130                                    (__v4di)_mm256_setzero_si256()))
1131
1132 #define _mm_mask_fpclass_pd_mask(U, A, imm) \
1133   ((__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \
1134                                               (__mmask8)(U)))
1135
1136 #define _mm_fpclass_pd_mask(A, imm) \
1137   ((__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \
1138                                               (__mmask8)-1))
1139
1140 #define _mm256_mask_fpclass_pd_mask(U, A, imm) \
1141   ((__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \
1142                                               (__mmask8)(U)))
1143
1144 #define _mm256_fpclass_pd_mask(A, imm) \
1145   ((__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \
1146                                               (__mmask8)-1))
1147
1148 #define _mm_mask_fpclass_ps_mask(U, A, imm) \
1149   ((__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \
1150                                               (__mmask8)(U)))
1151
1152 #define _mm_fpclass_ps_mask(A, imm) \
1153   ((__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \
1154                                               (__mmask8)-1))
1155
1156 #define _mm256_mask_fpclass_ps_mask(U, A, imm) \
1157   ((__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \
1158                                               (__mmask8)(U)))
1159
1160 #define _mm256_fpclass_ps_mask(A, imm) \
1161   ((__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \
1162                                               (__mmask8)-1))
1163
1164 #undef __DEFAULT_FN_ATTRS128
1165 #undef __DEFAULT_FN_ATTRS256
1166
1167 #endif