contrib/llvm/tools/clang/lib/Headers/avx512dqintrin.h

   1 /*===---- avx512dqintrin.h - AVX512DQ intrinsics ---------------------------===
   2  *
   3  * Permission is hereby granted, free of charge, to any person obtaining a copy
   4  * of this software and associated documentation files (the "Software"), to deal
   5  * in the Software without restriction, including without limitation the rights
   6  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
   7  * copies of the Software, and to permit persons to whom the Software is
   8  * furnished to do so, subject to the following conditions:
   9  *
  10  * The above copyright notice and this permission notice shall be included in
  11  * all copies or substantial portions of the Software.
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  16  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  17  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  18  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  19  * THE SOFTWARE.
  20  *
  21  *===-----------------------------------------------------------------------===
  22  */
  23
  24 #ifndef __IMMINTRIN_H
  25 #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead."
  26 #endif
  27
  28 #ifndef __AVX512DQINTRIN_H
  29 #define __AVX512DQINTRIN_H
  30
  31 /* Define the default attributes for the functions in this file. */
  32 #define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), __min_vector_width__(512)))
  33 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq")))
  34
  35 static __inline __mmask8 __DEFAULT_FN_ATTRS
  36 _knot_mask8(__mmask8 __M)
  37 {
  38   return __builtin_ia32_knotqi(__M);
  39 }
  40
  41 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
  42 _kand_mask8(__mmask8 __A, __mmask8 __B)
  43 {
  44   return (__mmask8)__builtin_ia32_kandqi((__mmask8)__A, (__mmask8)__B);
  45 }
  46
  47 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
  48 _kandn_mask8(__mmask8 __A, __mmask8 __B)
  49 {
  50   return (__mmask8)__builtin_ia32_kandnqi((__mmask8)__A, (__mmask8)__B);
  51 }
  52
  53 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
  54 _kor_mask8(__mmask8 __A, __mmask8 __B)
  55 {
  56   return (__mmask8)__builtin_ia32_korqi((__mmask8)__A, (__mmask8)__B);
  57 }
  58
  59 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
  60 _kxnor_mask8(__mmask8 __A, __mmask8 __B)
  61 {
  62   return (__mmask8)__builtin_ia32_kxnorqi((__mmask8)__A, (__mmask8)__B);
  63 }
  64
  65 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
  66 _kxor_mask8(__mmask8 __A, __mmask8 __B)
  67 {
  68   return (__mmask8)__builtin_ia32_kxorqi((__mmask8)__A, (__mmask8)__B);
  69 }
  70
  71 static __inline__ unsigned char __DEFAULT_FN_ATTRS
  72 _kortestc_mask8_u8(__mmask8 __A, __mmask8 __B)
  73 {
  74   return (unsigned char)__builtin_ia32_kortestcqi(__A, __B);
  75 }
  76
  77 static __inline__ unsigned char __DEFAULT_FN_ATTRS
  78 _kortestz_mask8_u8(__mmask8 __A, __mmask8 __B)
  79 {
  80   return (unsigned char)__builtin_ia32_kortestzqi(__A, __B);
  81 }
  82
  83 static __inline__ unsigned char __DEFAULT_FN_ATTRS
  84 _kortest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) {
  85   *__C = (unsigned char)__builtin_ia32_kortestcqi(__A, __B);
  86   return (unsigned char)__builtin_ia32_kortestzqi(__A, __B);
  87 }
  88
  89 static __inline__ unsigned char __DEFAULT_FN_ATTRS
  90 _ktestc_mask8_u8(__mmask8 __A, __mmask8 __B)
  91 {
  92   return (unsigned char)__builtin_ia32_ktestcqi(__A, __B);
  93 }
  94
  95 static __inline__ unsigned char __DEFAULT_FN_ATTRS
  96 _ktestz_mask8_u8(__mmask8 __A, __mmask8 __B)
  97 {
  98   return (unsigned char)__builtin_ia32_ktestzqi(__A, __B);
  99 }
 100
 101 static __inline__ unsigned char __DEFAULT_FN_ATTRS
 102 _ktest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) {
 103   *__C = (unsigned char)__builtin_ia32_ktestcqi(__A, __B);
 104   return (unsigned char)__builtin_ia32_ktestzqi(__A, __B);
 105 }
 106
 107 static __inline__ unsigned char __DEFAULT_FN_ATTRS
 108 _ktestc_mask16_u8(__mmask16 __A, __mmask16 __B)
 109 {
 110   return (unsigned char)__builtin_ia32_ktestchi(__A, __B);
 111 }
 112
 113 static __inline__ unsigned char __DEFAULT_FN_ATTRS
 114 _ktestz_mask16_u8(__mmask16 __A, __mmask16 __B)
 115 {
 116   return (unsigned char)__builtin_ia32_ktestzhi(__A, __B);
 117 }
 118
 119 static __inline__ unsigned char __DEFAULT_FN_ATTRS
 120 _ktest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C) {
 121   *__C = (unsigned char)__builtin_ia32_ktestchi(__A, __B);
 122   return (unsigned char)__builtin_ia32_ktestzhi(__A, __B);
 123 }
 124
 125 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 126 _kadd_mask8(__mmask8 __A, __mmask8 __B)
 127 {
 128   return (__mmask8)__builtin_ia32_kaddqi((__mmask8)__A, (__mmask8)__B);
 129 }
 130
 131 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 132 _kadd_mask16(__mmask16 __A, __mmask16 __B)
 133 {
 134   return (__mmask16)__builtin_ia32_kaddhi((__mmask16)__A, (__mmask16)__B);
 135 }
 136
 137 #define _kshiftli_mask8(A, I) \
 138   (__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I))
 139
 140 #define _kshiftri_mask8(A, I) \
 141   (__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I))
 142
 143 static __inline__ unsigned int __DEFAULT_FN_ATTRS
 144 _cvtmask8_u32(__mmask8 __A) {
 145   return (unsigned int)__builtin_ia32_kmovb((__mmask8)__A);
 146 }
 147
 148 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 149 _cvtu32_mask8(unsigned int __A) {
 150   return (__mmask8)__builtin_ia32_kmovb((__mmask8)__A);
 151 }
 152
 153 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 154 _load_mask8(__mmask8 *__A) {
 155   return (__mmask8)__builtin_ia32_kmovb(*(__mmask8 *)__A);
 156 }
 157
 158 static __inline__ void __DEFAULT_FN_ATTRS
 159 _store_mask8(__mmask8 *__A, __mmask8 __B) {
 160   *(__mmask8 *)__A = __builtin_ia32_kmovb((__mmask8)__B);
 161 }
 162
 163 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 164 _mm512_mullo_epi64 (__m512i __A, __m512i __B) {
 165   return (__m512i) ((__v8du) __A * (__v8du) __B);
 166 }
 167
 168 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 169 _mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
 170   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
 171                                              (__v8di)_mm512_mullo_epi64(__A, __B),
 172                                              (__v8di)__W);
 173 }
 174
 175 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 176 _mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B) {
 177   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
 178                                              (__v8di)_mm512_mullo_epi64(__A, __B),
 179                                              (__v8di)_mm512_setzero_si512());
 180 }
 181
 182 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 183 _mm512_xor_pd(__m512d __A, __m512d __B) {
 184   return (__m512d)((__v8du)__A ^ (__v8du)__B);
 185 }
 186
 187 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 188 _mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 189   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 190                                               (__v8df)_mm512_xor_pd(__A, __B),
 191                                               (__v8df)__W);
 192 }
 193
 194 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 195 _mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 196   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 197                                               (__v8df)_mm512_xor_pd(__A, __B),
 198                                               (__v8df)_mm512_setzero_pd());
 199 }
 200
 201 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 202 _mm512_xor_ps (__m512 __A, __m512 __B) {
 203   return (__m512)((__v16su)__A ^ (__v16su)__B);
 204 }
 205
 206 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 207 _mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 208   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 209                                              (__v16sf)_mm512_xor_ps(__A, __B),
 210                                              (__v16sf)__W);
 211 }
 212
 213 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 214 _mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 215   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 216                                              (__v16sf)_mm512_xor_ps(__A, __B),
 217                                              (__v16sf)_mm512_setzero_ps());
 218 }
 219
 220 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 221 _mm512_or_pd(__m512d __A, __m512d __B) {
 222   return (__m512d)((__v8du)__A | (__v8du)__B);
 223 }
 224
 225 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 226 _mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 227   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 228                                               (__v8df)_mm512_or_pd(__A, __B),
 229                                               (__v8df)__W);
 230 }
 231
 232 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 233 _mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 234   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 235                                               (__v8df)_mm512_or_pd(__A, __B),
 236                                               (__v8df)_mm512_setzero_pd());
 237 }
 238
 239 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 240 _mm512_or_ps(__m512 __A, __m512 __B) {
 241   return (__m512)((__v16su)__A | (__v16su)__B);
 242 }
 243
 244 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 245 _mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 246   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 247                                              (__v16sf)_mm512_or_ps(__A, __B),
 248                                              (__v16sf)__W);
 249 }
 250
 251 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 252 _mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 253   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 254                                              (__v16sf)_mm512_or_ps(__A, __B),
 255                                              (__v16sf)_mm512_setzero_ps());
 256 }
 257
 258 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 259 _mm512_and_pd(__m512d __A, __m512d __B) {
 260   return (__m512d)((__v8du)__A & (__v8du)__B);
 261 }
 262
 263 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 264 _mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 265   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 266                                               (__v8df)_mm512_and_pd(__A, __B),
 267                                               (__v8df)__W);
 268 }
 269
 270 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 271 _mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 272   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 273                                               (__v8df)_mm512_and_pd(__A, __B),
 274                                               (__v8df)_mm512_setzero_pd());
 275 }
 276
 277 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 278 _mm512_and_ps(__m512 __A, __m512 __B) {
 279   return (__m512)((__v16su)__A & (__v16su)__B);
 280 }
 281
 282 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 283 _mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 284   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 285                                              (__v16sf)_mm512_and_ps(__A, __B),
 286                                              (__v16sf)__W);
 287 }
 288
 289 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 290 _mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 291   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 292                                              (__v16sf)_mm512_and_ps(__A, __B),
 293                                              (__v16sf)_mm512_setzero_ps());
 294 }
 295
 296 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 297 _mm512_andnot_pd(__m512d __A, __m512d __B) {
 298   return (__m512d)(~(__v8du)__A & (__v8du)__B);
 299 }
 300
 301 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 302 _mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 303   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 304                                               (__v8df)_mm512_andnot_pd(__A, __B),
 305                                               (__v8df)__W);
 306 }
 307
 308 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 309 _mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 310   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 311                                               (__v8df)_mm512_andnot_pd(__A, __B),
 312                                               (__v8df)_mm512_setzero_pd());
 313 }
 314
 315 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 316 _mm512_andnot_ps(__m512 __A, __m512 __B) {
 317   return (__m512)(~(__v16su)__A & (__v16su)__B);
 318 }
 319
 320 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 321 _mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 322   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 323                                              (__v16sf)_mm512_andnot_ps(__A, __B),
 324                                              (__v16sf)__W);
 325 }
 326
 327 static __inline__ __m512 __DEFAULT_FN_ATTRS512
 328 _mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 329   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 330                                              (__v16sf)_mm512_andnot_ps(__A, __B),
 331                                              (__v16sf)_mm512_setzero_ps());
 332 }
 333
 334 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 335 _mm512_cvtpd_epi64 (__m512d __A) {
 336   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 337                 (__v8di) _mm512_setzero_si512(),
 338                 (__mmask8) -1,
 339                 _MM_FROUND_CUR_DIRECTION);
 340 }
 341
 342 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 343 _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
 344   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 345                 (__v8di) __W,
 346                 (__mmask8) __U,
 347                 _MM_FROUND_CUR_DIRECTION);
 348 }
 349
 350 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 351 _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) {
 352   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 353                 (__v8di) _mm512_setzero_si512(),
 354                 (__mmask8) __U,
 355                 _MM_FROUND_CUR_DIRECTION);
 356 }
 357
 358 #define _mm512_cvt_roundpd_epi64(A, R) \
 359   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 360                                            (__v8di)_mm512_setzero_si512(), \
 361                                            (__mmask8)-1, (int)(R))
 362
 363 #define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) \
 364   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 365                                            (__v8di)(__m512i)(W), \
 366                                            (__mmask8)(U), (int)(R))
 367
 368 #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \
 369   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 370                                            (__v8di)_mm512_setzero_si512(), \
 371                                            (__mmask8)(U), (int)(R))
 372
 373 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 374 _mm512_cvtpd_epu64 (__m512d __A) {
 375   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 376                  (__v8di) _mm512_setzero_si512(),
 377                  (__mmask8) -1,
 378                  _MM_FROUND_CUR_DIRECTION);
 379 }
 380
 381 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 382 _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
 383   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 384                  (__v8di) __W,
 385                  (__mmask8) __U,
 386                  _MM_FROUND_CUR_DIRECTION);
 387 }
 388
 389 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 390 _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) {
 391   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 392                  (__v8di) _mm512_setzero_si512(),
 393                  (__mmask8) __U,
 394                  _MM_FROUND_CUR_DIRECTION);
 395 }
 396
 397 #define _mm512_cvt_roundpd_epu64(A, R) \
 398   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 399                                             (__v8di)_mm512_setzero_si512(), \
 400                                             (__mmask8)-1, (int)(R))
 401
 402 #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \
 403   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 404                                             (__v8di)(__m512i)(W), \
 405                                             (__mmask8)(U), (int)(R))
 406
 407 #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \
 408   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 409                                             (__v8di)_mm512_setzero_si512(), \
 410                                             (__mmask8)(U), (int)(R))
 411
 412 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 413 _mm512_cvtps_epi64 (__m256 __A) {
 414   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 415                 (__v8di) _mm512_setzero_si512(),
 416                 (__mmask8) -1,
 417                 _MM_FROUND_CUR_DIRECTION);
 418 }
 419
 420 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 421 _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
 422   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 423                 (__v8di) __W,
 424                 (__mmask8) __U,
 425                 _MM_FROUND_CUR_DIRECTION);
 426 }
 427
 428 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 429 _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) {
 430   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 431                 (__v8di) _mm512_setzero_si512(),
 432                 (__mmask8) __U,
 433                 _MM_FROUND_CUR_DIRECTION);
 434 }
 435
 436 #define _mm512_cvt_roundps_epi64(A, R) \
 437   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 438                                            (__v8di)_mm512_setzero_si512(), \
 439                                            (__mmask8)-1, (int)(R))
 440
 441 #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \
 442   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 443                                            (__v8di)(__m512i)(W), \
 444                                            (__mmask8)(U), (int)(R))
 445
 446 #define _mm512_maskz_cvt_roundps_epi64(U, A, R) \
 447   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 448                                            (__v8di)_mm512_setzero_si512(), \
 449                                            (__mmask8)(U), (int)(R))
 450
 451 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 452 _mm512_cvtps_epu64 (__m256 __A) {
 453   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 454                  (__v8di) _mm512_setzero_si512(),
 455                  (__mmask8) -1,
 456                  _MM_FROUND_CUR_DIRECTION);
 457 }
 458
 459 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 460 _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
 461   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 462                  (__v8di) __W,
 463                  (__mmask8) __U,
 464                  _MM_FROUND_CUR_DIRECTION);
 465 }
 466
 467 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 468 _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) {
 469   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 470                  (__v8di) _mm512_setzero_si512(),
 471                  (__mmask8) __U,
 472                  _MM_FROUND_CUR_DIRECTION);
 473 }
 474
 475 #define _mm512_cvt_roundps_epu64(A, R) \
 476   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 477                                             (__v8di)_mm512_setzero_si512(), \
 478                                             (__mmask8)-1, (int)(R))
 479
 480 #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \
 481   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 482                                             (__v8di)(__m512i)(W), \
 483                                             (__mmask8)(U), (int)(R))
 484
 485 #define _mm512_maskz_cvt_roundps_epu64(U, A, R) \
 486   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 487                                             (__v8di)_mm512_setzero_si512(), \
 488                                             (__mmask8)(U), (int)(R))
 489
 490
 491 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 492 _mm512_cvtepi64_pd (__m512i __A) {
 493   return (__m512d)__builtin_convertvector((__v8di)__A, __v8df);
 494 }
 495
 496 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 497 _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
 498   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 499                                               (__v8df)_mm512_cvtepi64_pd(__A),
 500                                               (__v8df)__W);
 501 }
 502
 503 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 504 _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) {
 505   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 506                                               (__v8df)_mm512_cvtepi64_pd(__A),
 507                                               (__v8df)_mm512_setzero_pd());
 508 }
 509
 510 #define _mm512_cvt_roundepi64_pd(A, R) \
 511   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 512                                            (__v8df)_mm512_setzero_pd(), \
 513                                            (__mmask8)-1, (int)(R))
 514
 515 #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \
 516   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 517                                            (__v8df)(__m512d)(W), \
 518                                            (__mmask8)(U), (int)(R))
 519
 520 #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \
 521   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 522                                            (__v8df)_mm512_setzero_pd(), \
 523                                            (__mmask8)(U), (int)(R))
 524
 525 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 526 _mm512_cvtepi64_ps (__m512i __A) {
 527   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 528                (__v8sf) _mm256_setzero_ps(),
 529                (__mmask8) -1,
 530                _MM_FROUND_CUR_DIRECTION);
 531 }
 532
 533 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 534 _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
 535   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 536                (__v8sf) __W,
 537                (__mmask8) __U,
 538                _MM_FROUND_CUR_DIRECTION);
 539 }
 540
 541 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 542 _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) {
 543   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 544                (__v8sf) _mm256_setzero_ps(),
 545                (__mmask8) __U,
 546                _MM_FROUND_CUR_DIRECTION);
 547 }
 548
 549 #define _mm512_cvt_roundepi64_ps(A, R) \
 550   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 551                                           (__v8sf)_mm256_setzero_ps(), \
 552                                           (__mmask8)-1, (int)(R))
 553
 554 #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \
 555   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 556                                           (__v8sf)(__m256)(W), (__mmask8)(U), \
 557                                           (int)(R))
 558
 559 #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \
 560   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 561                                           (__v8sf)_mm256_setzero_ps(), \
 562                                           (__mmask8)(U), (int)(R))
 563
 564
 565 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 566 _mm512_cvttpd_epi64 (__m512d __A) {
 567   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 568                  (__v8di) _mm512_setzero_si512(),
 569                  (__mmask8) -1,
 570                  _MM_FROUND_CUR_DIRECTION);
 571 }
 572
 573 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 574 _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
 575   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 576                  (__v8di) __W,
 577                  (__mmask8) __U,
 578                  _MM_FROUND_CUR_DIRECTION);
 579 }
 580
 581 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 582 _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) {
 583   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 584                  (__v8di) _mm512_setzero_si512(),
 585                  (__mmask8) __U,
 586                  _MM_FROUND_CUR_DIRECTION);
 587 }
 588
 589 #define _mm512_cvtt_roundpd_epi64(A, R) \
 590   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 591                                             (__v8di)_mm512_setzero_si512(), \
 592                                             (__mmask8)-1, (int)(R))
 593
 594 #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \
 595   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 596                                             (__v8di)(__m512i)(W), \
 597                                             (__mmask8)(U), (int)(R))
 598
 599 #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \
 600   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 601                                             (__v8di)_mm512_setzero_si512(), \
 602                                             (__mmask8)(U), (int)(R))
 603
 604 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 605 _mm512_cvttpd_epu64 (__m512d __A) {
 606   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 607                   (__v8di) _mm512_setzero_si512(),
 608                   (__mmask8) -1,
 609                   _MM_FROUND_CUR_DIRECTION);
 610 }
 611
 612 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 613 _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
 614   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 615                   (__v8di) __W,
 616                   (__mmask8) __U,
 617                   _MM_FROUND_CUR_DIRECTION);
 618 }
 619
 620 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 621 _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) {
 622   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 623                   (__v8di) _mm512_setzero_si512(),
 624                   (__mmask8) __U,
 625                   _MM_FROUND_CUR_DIRECTION);
 626 }
 627
 628 #define _mm512_cvtt_roundpd_epu64(A, R) \
 629   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 630                                              (__v8di)_mm512_setzero_si512(), \
 631                                              (__mmask8)-1, (int)(R))
 632
 633 #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \
 634   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 635                                              (__v8di)(__m512i)(W), \
 636                                              (__mmask8)(U), (int)(R))
 637
 638 #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \
 639   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 640                                              (__v8di)_mm512_setzero_si512(), \
 641                                              (__mmask8)(U), (int)(R))
 642
 643 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 644 _mm512_cvttps_epi64 (__m256 __A) {
 645   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 646                  (__v8di) _mm512_setzero_si512(),
 647                  (__mmask8) -1,
 648                  _MM_FROUND_CUR_DIRECTION);
 649 }
 650
 651 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 652 _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
 653   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 654                  (__v8di) __W,
 655                  (__mmask8) __U,
 656                  _MM_FROUND_CUR_DIRECTION);
 657 }
 658
 659 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 660 _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) {
 661   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 662                  (__v8di) _mm512_setzero_si512(),
 663                  (__mmask8) __U,
 664                  _MM_FROUND_CUR_DIRECTION);
 665 }
 666
 667 #define _mm512_cvtt_roundps_epi64(A, R) \
 668   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 669                                             (__v8di)_mm512_setzero_si512(), \
 670                                             (__mmask8)-1, (int)(R))
 671
 672 #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \
 673   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 674                                             (__v8di)(__m512i)(W), \
 675                                             (__mmask8)(U), (int)(R))
 676
 677 #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \
 678   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 679                                             (__v8di)_mm512_setzero_si512(), \
 680                                             (__mmask8)(U), (int)(R))
 681
 682 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 683 _mm512_cvttps_epu64 (__m256 __A) {
 684   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 685                   (__v8di) _mm512_setzero_si512(),
 686                   (__mmask8) -1,
 687                   _MM_FROUND_CUR_DIRECTION);
 688 }
 689
 690 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 691 _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
 692   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 693                   (__v8di) __W,
 694                   (__mmask8) __U,
 695                   _MM_FROUND_CUR_DIRECTION);
 696 }
 697
 698 static __inline__ __m512i __DEFAULT_FN_ATTRS512
 699 _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) {
 700   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 701                   (__v8di) _mm512_setzero_si512(),
 702                   (__mmask8) __U,
 703                   _MM_FROUND_CUR_DIRECTION);
 704 }
 705
 706 #define _mm512_cvtt_roundps_epu64(A, R) \
 707   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 708                                              (__v8di)_mm512_setzero_si512(), \
 709                                              (__mmask8)-1, (int)(R))
 710
 711 #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \
 712   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 713                                              (__v8di)(__m512i)(W), \
 714                                              (__mmask8)(U), (int)(R))
 715
 716 #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \
 717   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 718                                              (__v8di)_mm512_setzero_si512(), \
 719                                              (__mmask8)(U), (int)(R))
 720
 721 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 722 _mm512_cvtepu64_pd (__m512i __A) {
 723   return (__m512d)__builtin_convertvector((__v8du)__A, __v8df);
 724 }
 725
 726 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 727 _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
 728   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 729                                               (__v8df)_mm512_cvtepu64_pd(__A),
 730                                               (__v8df)__W);
 731 }
 732
 733 static __inline__ __m512d __DEFAULT_FN_ATTRS512
 734 _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) {
 735   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 736                                               (__v8df)_mm512_cvtepu64_pd(__A),
 737                                               (__v8df)_mm512_setzero_pd());
 738 }
 739
 740 #define _mm512_cvt_roundepu64_pd(A, R) \
 741   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 742                                             (__v8df)_mm512_setzero_pd(), \
 743                                             (__mmask8)-1, (int)(R))
 744
 745 #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \
 746   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 747                                             (__v8df)(__m512d)(W), \
 748                                             (__mmask8)(U), (int)(R))
 749
 750
 751 #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \
 752   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 753                                             (__v8df)_mm512_setzero_pd(), \
 754                                             (__mmask8)(U), (int)(R))
 755
 756
 757 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 758 _mm512_cvtepu64_ps (__m512i __A) {
 759   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 760                 (__v8sf) _mm256_setzero_ps(),
 761                 (__mmask8) -1,
 762                 _MM_FROUND_CUR_DIRECTION);
 763 }
 764
 765 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 766 _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
 767   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 768                 (__v8sf) __W,
 769                 (__mmask8) __U,
 770                 _MM_FROUND_CUR_DIRECTION);
 771 }
 772
 773 static __inline__ __m256 __DEFAULT_FN_ATTRS512
 774 _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) {
 775   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 776                 (__v8sf) _mm256_setzero_ps(),
 777                 (__mmask8) __U,
 778                 _MM_FROUND_CUR_DIRECTION);
 779 }
 780
 781 #define _mm512_cvt_roundepu64_ps(A, R) \
 782   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 783                                            (__v8sf)_mm256_setzero_ps(), \
 784                                            (__mmask8)-1, (int)(R))
 785
 786 #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \
 787   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 788                                            (__v8sf)(__m256)(W), (__mmask8)(U), \
 789                                            (int)(R))
 790
 791 #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \
 792   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 793                                            (__v8sf)_mm256_setzero_ps(), \
 794                                            (__mmask8)(U), (int)(R))
 795
 796 #define _mm512_range_pd(A, B, C) \
 797   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 798                                           (__v8df)(__m512d)(B), (int)(C), \
 799                                           (__v8df)_mm512_setzero_pd(), \
 800                                           (__mmask8)-1, \
 801                                           _MM_FROUND_CUR_DIRECTION)
 802
 803 #define _mm512_mask_range_pd(W, U, A, B, C) \
 804   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 805                                           (__v8df)(__m512d)(B), (int)(C), \
 806                                           (__v8df)(__m512d)(W), (__mmask8)(U), \
 807                                           _MM_FROUND_CUR_DIRECTION)
 808
 809 #define _mm512_maskz_range_pd(U, A, B, C) \
 810   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 811                                           (__v8df)(__m512d)(B), (int)(C), \
 812                                           (__v8df)_mm512_setzero_pd(), \
 813                                           (__mmask8)(U), \
 814                                           _MM_FROUND_CUR_DIRECTION)
 815
 816 #define _mm512_range_round_pd(A, B, C, R) \
 817   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 818                                           (__v8df)(__m512d)(B), (int)(C), \
 819                                           (__v8df)_mm512_setzero_pd(), \
 820                                           (__mmask8)-1, (int)(R))
 821
 822 #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \
 823   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 824                                           (__v8df)(__m512d)(B), (int)(C), \
 825                                           (__v8df)(__m512d)(W), (__mmask8)(U), \
 826                                           (int)(R))
 827
 828 #define _mm512_maskz_range_round_pd(U, A, B, C, R) \
 829   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 830                                           (__v8df)(__m512d)(B), (int)(C), \
 831                                           (__v8df)_mm512_setzero_pd(), \
 832                                           (__mmask8)(U), (int)(R))
 833
 834 #define _mm512_range_ps(A, B, C) \
 835   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 836                                          (__v16sf)(__m512)(B), (int)(C), \
 837                                          (__v16sf)_mm512_setzero_ps(), \
 838                                          (__mmask16)-1, \
 839                                          _MM_FROUND_CUR_DIRECTION)
 840
 841 #define _mm512_mask_range_ps(W, U, A, B, C) \
 842   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 843                                          (__v16sf)(__m512)(B), (int)(C), \
 844                                          (__v16sf)(__m512)(W), (__mmask16)(U), \
 845                                          _MM_FROUND_CUR_DIRECTION)
 846
 847 #define _mm512_maskz_range_ps(U, A, B, C) \
 848   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 849                                          (__v16sf)(__m512)(B), (int)(C), \
 850                                          (__v16sf)_mm512_setzero_ps(), \
 851                                          (__mmask16)(U), \
 852                                          _MM_FROUND_CUR_DIRECTION)
 853
 854 #define _mm512_range_round_ps(A, B, C, R) \
 855   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 856                                          (__v16sf)(__m512)(B), (int)(C), \
 857                                          (__v16sf)_mm512_setzero_ps(), \
 858                                          (__mmask16)-1, (int)(R))
 859
 860 #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \
 861   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 862                                          (__v16sf)(__m512)(B), (int)(C), \
 863                                          (__v16sf)(__m512)(W), (__mmask16)(U), \
 864                                          (int)(R))
 865
 866 #define _mm512_maskz_range_round_ps(U, A, B, C, R) \
 867   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 868                                          (__v16sf)(__m512)(B), (int)(C), \
 869                                          (__v16sf)_mm512_setzero_ps(), \
 870                                          (__mmask16)(U), (int)(R))
 871
 872 #define _mm_range_round_ss(A, B, C, R) \
 873   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 874                                                (__v4sf)(__m128)(B), \
 875                                                (__v4sf)_mm_setzero_ps(), \
 876                                                (__mmask8) -1, (int)(C),\
 877                                                (int)(R))
 878
 879 #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION)
 880
 881 #define _mm_mask_range_round_ss(W, U, A, B, C, R) \
 882   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 883                                                (__v4sf)(__m128)(B), \
 884                                                (__v4sf)(__m128)(W),\
 885                                                (__mmask8)(U), (int)(C),\
 886                                                (int)(R))
 887
 888 #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION)
 889
 890 #define _mm_maskz_range_round_ss(U, A, B, C, R) \
 891   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 892                                                (__v4sf)(__m128)(B), \
 893                                                (__v4sf)_mm_setzero_ps(), \
 894                                                (__mmask8)(U), (int)(C),\
 895                                                (int)(R))
 896
 897 #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 898
 899 #define _mm_range_round_sd(A, B, C, R) \
 900   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 901                                                 (__v2df)(__m128d)(B), \
 902                                                 (__v2df)_mm_setzero_pd(), \
 903                                                 (__mmask8) -1, (int)(C),\
 904                                                 (int)(R))
 905
 906 #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION)
 907
 908 #define _mm_mask_range_round_sd(W, U, A, B, C, R) \
 909   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 910                                                 (__v2df)(__m128d)(B), \
 911                                                 (__v2df)(__m128d)(W),\
 912                                                 (__mmask8)(U), (int)(C),\
 913                                                 (int)(R))
 914
 915 #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 916
 917 #define _mm_maskz_range_round_sd(U, A, B, C, R) \
 918   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 919                                                 (__v2df)(__m128d)(B), \
 920                                                 (__v2df)_mm_setzero_pd(), \
 921                                                 (__mmask8)(U), (int)(C),\
 922                                                 (int)(R))
 923
 924 #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 925
 926 #define _mm512_reduce_pd(A, B) \
 927   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 928                                            (__v8df)_mm512_setzero_pd(), \
 929                                            (__mmask8)-1, \
 930                                            _MM_FROUND_CUR_DIRECTION)
 931
 932 #define _mm512_mask_reduce_pd(W, U, A, B) \
 933   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 934                                            (__v8df)(__m512d)(W), \
 935                                            (__mmask8)(U), \
 936                                            _MM_FROUND_CUR_DIRECTION)
 937
 938 #define _mm512_maskz_reduce_pd(U, A, B) \
 939   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 940                                            (__v8df)_mm512_setzero_pd(), \
 941                                            (__mmask8)(U), \
 942                                            _MM_FROUND_CUR_DIRECTION)
 943
 944 #define _mm512_reduce_ps(A, B) \
 945   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 946                                           (__v16sf)_mm512_setzero_ps(), \
 947                                           (__mmask16)-1, \
 948                                           _MM_FROUND_CUR_DIRECTION)
 949
 950 #define _mm512_mask_reduce_ps(W, U, A, B) \
 951   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 952                                           (__v16sf)(__m512)(W), \
 953                                           (__mmask16)(U), \
 954                                           _MM_FROUND_CUR_DIRECTION)
 955
 956 #define _mm512_maskz_reduce_ps(U, A, B) \
 957   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 958                                           (__v16sf)_mm512_setzero_ps(), \
 959                                           (__mmask16)(U), \
 960                                           _MM_FROUND_CUR_DIRECTION)
 961
 962 #define _mm512_reduce_round_pd(A, B, R) \
 963   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 964                                            (__v8df)_mm512_setzero_pd(), \
 965                                            (__mmask8)-1, (int)(R))
 966
 967 #define _mm512_mask_reduce_round_pd(W, U, A, B, R) \
 968   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 969                                            (__v8df)(__m512d)(W), \
 970                                            (__mmask8)(U), (int)(R))
 971
 972 #define _mm512_maskz_reduce_round_pd(U, A, B, R) \
 973   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 974                                            (__v8df)_mm512_setzero_pd(), \
 975                                            (__mmask8)(U), (int)(R))
 976
 977 #define _mm512_reduce_round_ps(A, B, R) \
 978   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 979                                           (__v16sf)_mm512_setzero_ps(), \
 980                                           (__mmask16)-1, (int)(R))
 981
 982 #define _mm512_mask_reduce_round_ps(W, U, A, B, R) \
 983   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 984                                           (__v16sf)(__m512)(W), \
 985                                           (__mmask16)(U), (int)(R))
 986
 987 #define _mm512_maskz_reduce_round_ps(U, A, B, R) \
 988   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 989                                           (__v16sf)_mm512_setzero_ps(), \
 990                                           (__mmask16)(U), (int)(R))
 991
 992 #define _mm_reduce_ss(A, B, C) \
 993   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 994                                        (__v4sf)(__m128)(B), \
 995                                        (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
 996                                        (int)(C), _MM_FROUND_CUR_DIRECTION)
 997
 998 #define _mm_mask_reduce_ss(W, U, A, B, C) \
 999   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1000                                        (__v4sf)(__m128)(B), \
1001                                        (__v4sf)(__m128)(W), (__mmask8)(U), \
1002                                        (int)(C), _MM_FROUND_CUR_DIRECTION)
1003
1004 #define _mm_maskz_reduce_ss(U, A, B, C) \
1005   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1006                                        (__v4sf)(__m128)(B), \
1007                                        (__v4sf)_mm_setzero_ps(), \
1008                                        (__mmask8)(U), (int)(C), \
1009                                        _MM_FROUND_CUR_DIRECTION)
1010
1011 #define _mm_reduce_round_ss(A, B, C, R) \
1012   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1013                                        (__v4sf)(__m128)(B), \
1014                                        (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
1015                                        (int)(C), (int)(R))
1016
1017 #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \
1018   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1019                                        (__v4sf)(__m128)(B), \
1020                                        (__v4sf)(__m128)(W), (__mmask8)(U), \
1021                                        (int)(C), (int)(R))
1022
1023 #define _mm_maskz_reduce_round_ss(U, A, B, C, R) \
1024   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1025                                        (__v4sf)(__m128)(B), \
1026                                        (__v4sf)_mm_setzero_ps(), \
1027                                        (__mmask8)(U), (int)(C), (int)(R))
1028
1029 #define _mm_reduce_sd(A, B, C) \
1030   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1031                                         (__v2df)(__m128d)(B), \
1032                                         (__v2df)_mm_setzero_pd(), \
1033                                         (__mmask8)-1, (int)(C), \
1034                                         _MM_FROUND_CUR_DIRECTION)
1035
1036 #define _mm_mask_reduce_sd(W, U, A, B, C) \
1037   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1038                                         (__v2df)(__m128d)(B), \
1039                                         (__v2df)(__m128d)(W), (__mmask8)(U), \
1040                                         (int)(C), _MM_FROUND_CUR_DIRECTION)
1041
1042 #define _mm_maskz_reduce_sd(U, A, B, C) \
1043   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1044                                         (__v2df)(__m128d)(B), \
1045                                         (__v2df)_mm_setzero_pd(), \
1046                                         (__mmask8)(U), (int)(C), \
1047                                         _MM_FROUND_CUR_DIRECTION)
1048
1049 #define _mm_reduce_round_sd(A, B, C, R) \
1050   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1051                                         (__v2df)(__m128d)(B), \
1052                                         (__v2df)_mm_setzero_pd(), \
1053                                         (__mmask8)-1, (int)(C), (int)(R))
1054
1055 #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \
1056   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1057                                         (__v2df)(__m128d)(B), \
1058                                         (__v2df)(__m128d)(W), (__mmask8)(U), \
1059                                         (int)(C), (int)(R))
1060
1061 #define _mm_maskz_reduce_round_sd(U, A, B, C, R) \
1062   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1063                                         (__v2df)(__m128d)(B), \
1064                                         (__v2df)_mm_setzero_pd(), \
1065                                         (__mmask8)(U), (int)(C), (int)(R))
1066
1067 static __inline__ __mmask16 __DEFAULT_FN_ATTRS512
1068 _mm512_movepi32_mask (__m512i __A)
1069 {
1070   return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
1071 }
1072
1073 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1074 _mm512_movm_epi32 (__mmask16 __A)
1075 {
1076   return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
1077 }
1078
1079 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1080 _mm512_movm_epi64 (__mmask8 __A)
1081 {
1082   return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
1083 }
1084
1085 static __inline__ __mmask8 __DEFAULT_FN_ATTRS512
1086 _mm512_movepi64_mask (__m512i __A)
1087 {
1088   return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
1089 }
1090
1091
1092 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1093 _mm512_broadcast_f32x2 (__m128 __A)
1094 {
1095   return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
1096                                          0, 1, 0, 1, 0, 1, 0, 1,
1097                                          0, 1, 0, 1, 0, 1, 0, 1);
1098 }
1099
1100 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1101 _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A)
1102 {
1103   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1104                                              (__v16sf)_mm512_broadcast_f32x2(__A),
1105                                              (__v16sf)__O);
1106 }
1107
1108 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1109 _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A)
1110 {
1111   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1112                                              (__v16sf)_mm512_broadcast_f32x2(__A),
1113                                              (__v16sf)_mm512_setzero_ps());
1114 }
1115
1116 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1117 _mm512_broadcast_f32x8(__m256 __A)
1118 {
1119   return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A,
1120                                          0, 1, 2, 3, 4, 5, 6, 7,
1121                                          0, 1, 2, 3, 4, 5, 6, 7);
1122 }
1123
1124 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1125 _mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A)
1126 {
1127   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1128                                            (__v16sf)_mm512_broadcast_f32x8(__A),
1129                                            (__v16sf)__O);
1130 }
1131
1132 static __inline__ __m512 __DEFAULT_FN_ATTRS512
1133 _mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A)
1134 {
1135   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1136                                            (__v16sf)_mm512_broadcast_f32x8(__A),
1137                                            (__v16sf)_mm512_setzero_ps());
1138 }
1139
1140 static __inline__ __m512d __DEFAULT_FN_ATTRS512
1141 _mm512_broadcast_f64x2(__m128d __A)
1142 {
1143   return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
1144                                           0, 1, 0, 1, 0, 1, 0, 1);
1145 }
1146
1147 static __inline__ __m512d __DEFAULT_FN_ATTRS512
1148 _mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A)
1149 {
1150   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1151                                             (__v8df)_mm512_broadcast_f64x2(__A),
1152                                             (__v8df)__O);
1153 }
1154
1155 static __inline__ __m512d __DEFAULT_FN_ATTRS512
1156 _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A)
1157 {
1158   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1159                                             (__v8df)_mm512_broadcast_f64x2(__A),
1160                                             (__v8df)_mm512_setzero_pd());
1161 }
1162
1163 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1164 _mm512_broadcast_i32x2 (__m128i __A)
1165 {
1166   return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1167                                           0, 1, 0, 1, 0, 1, 0, 1,
1168                                           0, 1, 0, 1, 0, 1, 0, 1);
1169 }
1170
1171 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1172 _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A)
1173 {
1174   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1175                                              (__v16si)_mm512_broadcast_i32x2(__A),
1176                                              (__v16si)__O);
1177 }
1178
1179 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1180 _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A)
1181 {
1182   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1183                                              (__v16si)_mm512_broadcast_i32x2(__A),
1184                                              (__v16si)_mm512_setzero_si512());
1185 }
1186
1187 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1188 _mm512_broadcast_i32x8(__m256i __A)
1189 {
1190   return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A,
1191                                           0, 1, 2, 3, 4, 5, 6, 7,
1192                                           0, 1, 2, 3, 4, 5, 6, 7);
1193 }
1194
1195 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1196 _mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A)
1197 {
1198   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1199                                            (__v16si)_mm512_broadcast_i32x8(__A),
1200                                            (__v16si)__O);
1201 }
1202
1203 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1204 _mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A)
1205 {
1206   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1207                                            (__v16si)_mm512_broadcast_i32x8(__A),
1208                                            (__v16si)_mm512_setzero_si512());
1209 }
1210
1211 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1212 _mm512_broadcast_i64x2(__m128i __A)
1213 {
1214   return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1215                                           0, 1, 0, 1, 0, 1, 0, 1);
1216 }
1217
1218 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1219 _mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A)
1220 {
1221   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1222                                             (__v8di)_mm512_broadcast_i64x2(__A),
1223                                             (__v8di)__O);
1224 }
1225
1226 static __inline__ __m512i __DEFAULT_FN_ATTRS512
1227 _mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A)
1228 {
1229   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1230                                             (__v8di)_mm512_broadcast_i64x2(__A),
1231                                             (__v8di)_mm512_setzero_si512());
1232 }
1233
1234 #define _mm512_extractf32x8_ps(A, imm) \
1235   (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1236                                            (__v8sf)_mm256_undefined_ps(), \
1237                                            (__mmask8)-1)
1238
1239 #define _mm512_mask_extractf32x8_ps(W, U, A, imm) \
1240   (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1241                                            (__v8sf)(__m256)(W), \
1242                                            (__mmask8)(U))
1243
1244 #define _mm512_maskz_extractf32x8_ps(U, A, imm) \
1245   (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1246                                            (__v8sf)_mm256_setzero_ps(), \
1247                                            (__mmask8)(U))
1248
1249 #define _mm512_extractf64x2_pd(A, imm) \
1250   (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1251                                                 (int)(imm), \
1252                                                 (__v2df)_mm_undefined_pd(), \
1253                                                 (__mmask8)-1)
1254
1255 #define _mm512_mask_extractf64x2_pd(W, U, A, imm) \
1256   (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1257                                                 (int)(imm), \
1258                                                 (__v2df)(__m128d)(W), \
1259                                                 (__mmask8)(U))
1260
1261 #define _mm512_maskz_extractf64x2_pd(U, A, imm) \
1262   (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1263                                                 (int)(imm), \
1264                                                 (__v2df)_mm_setzero_pd(), \
1265                                                 (__mmask8)(U))
1266
1267 #define _mm512_extracti32x8_epi32(A, imm) \
1268   (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1269                                             (__v8si)_mm256_undefined_si256(), \
1270                                             (__mmask8)-1)
1271
1272 #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \
1273   (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1274                                             (__v8si)(__m256i)(W), \
1275                                             (__mmask8)(U))
1276
1277 #define _mm512_maskz_extracti32x8_epi32(U, A, imm) \
1278   (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1279                                             (__v8si)_mm256_setzero_si256(), \
1280                                             (__mmask8)(U))
1281
1282 #define _mm512_extracti64x2_epi64(A, imm) \
1283   (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1284                                                 (int)(imm), \
1285                                                 (__v2di)_mm_undefined_si128(), \
1286                                                 (__mmask8)-1)
1287
1288 #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \
1289   (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1290                                                 (int)(imm), \
1291                                                 (__v2di)(__m128i)(W), \
1292                                                 (__mmask8)(U))
1293
1294 #define _mm512_maskz_extracti64x2_epi64(U, A, imm) \
1295   (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1296                                                 (int)(imm), \
1297                                                 (__v2di)_mm_setzero_si128(), \
1298                                                 (__mmask8)(U))
1299
1300 #define _mm512_insertf32x8(A, B, imm) \
1301   (__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \
1302                                      (__v8sf)(__m256)(B), (int)(imm))
1303
1304 #define _mm512_mask_insertf32x8(W, U, A, B, imm) \
1305   (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1306                                  (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1307                                  (__v16sf)(__m512)(W))
1308
1309 #define _mm512_maskz_insertf32x8(U, A, B, imm) \
1310   (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1311                                  (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1312                                  (__v16sf)_mm512_setzero_ps())
1313
1314 #define _mm512_insertf64x2(A, B, imm) \
1315   (__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \
1316                                           (__v2df)(__m128d)(B), (int)(imm))
1317
1318 #define _mm512_mask_insertf64x2(W, U, A, B, imm) \
1319   (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1320                                   (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1321                                   (__v8df)(__m512d)(W))
1322
1323 #define _mm512_maskz_insertf64x2(U, A, B, imm) \
1324   (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1325                                   (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1326                                   (__v8df)_mm512_setzero_pd())
1327
1328 #define _mm512_inserti32x8(A, B, imm) \
1329   (__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \
1330                                       (__v8si)(__m256i)(B), (int)(imm))
1331
1332 #define _mm512_mask_inserti32x8(W, U, A, B, imm) \
1333   (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1334                                  (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1335                                  (__v16si)(__m512i)(W))
1336
1337 #define _mm512_maskz_inserti32x8(U, A, B, imm) \
1338   (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1339                                  (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1340                                  (__v16si)_mm512_setzero_si512())
1341
1342 #define _mm512_inserti64x2(A, B, imm) \
1343   (__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \
1344                                           (__v2di)(__m128i)(B), (int)(imm))
1345
1346 #define _mm512_mask_inserti64x2(W, U, A, B, imm) \
1347   (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1348                                   (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1349                                   (__v8di)(__m512i)(W))
1350
1351 #define _mm512_maskz_inserti64x2(U, A, B, imm) \
1352   (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1353                                   (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1354                                   (__v8di)_mm512_setzero_si512())
1355
1356 #define _mm512_mask_fpclass_ps_mask(U, A, imm) \
1357   (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1358                                               (int)(imm), (__mmask16)(U))
1359
1360 #define _mm512_fpclass_ps_mask(A, imm) \
1361   (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1362                                               (int)(imm), (__mmask16)-1)
1363
1364 #define _mm512_mask_fpclass_pd_mask(U, A, imm) \
1365   (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1366                                              (__mmask8)(U))
1367
1368 #define _mm512_fpclass_pd_mask(A, imm) \
1369   (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1370                                              (__mmask8)-1)
1371
1372 #define _mm_fpclass_sd_mask(A, imm) \
1373   (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1374                                           (__mmask8)-1)
1375
1376 #define _mm_mask_fpclass_sd_mask(U, A, imm) \
1377   (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1378                                           (__mmask8)(U))
1379
1380 #define _mm_fpclass_ss_mask(A, imm) \
1381   (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1382                                           (__mmask8)-1)
1383
1384 #define _mm_mask_fpclass_ss_mask(U, A, imm) \
1385   (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1386                                           (__mmask8)(U))
1387
1388 #undef __DEFAULT_FN_ATTRS512
1389 #undef __DEFAULT_FN_ATTRS
1390
1391 #endif