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_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq")))
  33
  34 static __inline__ __m512i __DEFAULT_FN_ATTRS
  35 _mm512_mullo_epi64 (__m512i __A, __m512i __B) {
  36   return (__m512i) ((__v8du) __A * (__v8du) __B);
  37 }
  38
  39 static __inline__ __m512i __DEFAULT_FN_ATTRS
  40 _mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
  41   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
  42                                              (__v8di)_mm512_mullo_epi64(__A, __B),
  43                                              (__v8di)__W);
  44 }
  45
  46 static __inline__ __m512i __DEFAULT_FN_ATTRS
  47 _mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B) {
  48   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
  49                                              (__v8di)_mm512_mullo_epi64(__A, __B),
  50                                              (__v8di)_mm512_setzero_si512());
  51 }
  52
  53 static __inline__ __m512d __DEFAULT_FN_ATTRS
  54 _mm512_xor_pd(__m512d __A, __m512d __B) {
  55   return (__m512d)((__v8du)__A ^ (__v8du)__B);
  56 }
  57
  58 static __inline__ __m512d __DEFAULT_FN_ATTRS
  59 _mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
  60   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
  61                                               (__v8df)_mm512_xor_pd(__A, __B),
  62                                               (__v8df)__W);
  63 }
  64
  65 static __inline__ __m512d __DEFAULT_FN_ATTRS
  66 _mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B) {
  67   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
  68                                               (__v8df)_mm512_xor_pd(__A, __B),
  69                                               (__v8df)_mm512_setzero_pd());
  70 }
  71
  72 static __inline__ __m512 __DEFAULT_FN_ATTRS
  73 _mm512_xor_ps (__m512 __A, __m512 __B) {
  74   return (__m512)((__v16su)__A ^ (__v16su)__B);
  75 }
  76
  77 static __inline__ __m512 __DEFAULT_FN_ATTRS
  78 _mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
  79   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
  80                                              (__v16sf)_mm512_xor_ps(__A, __B),
  81                                              (__v16sf)__W);
  82 }
  83
  84 static __inline__ __m512 __DEFAULT_FN_ATTRS
  85 _mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B) {
  86   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
  87                                              (__v16sf)_mm512_xor_ps(__A, __B),
  88                                              (__v16sf)_mm512_setzero_ps());
  89 }
  90
  91 static __inline__ __m512d __DEFAULT_FN_ATTRS
  92 _mm512_or_pd(__m512d __A, __m512d __B) {
  93   return (__m512d)((__v8du)__A | (__v8du)__B);
  94 }
  95
  96 static __inline__ __m512d __DEFAULT_FN_ATTRS
  97 _mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
  98   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
  99                                               (__v8df)_mm512_or_pd(__A, __B),
 100                                               (__v8df)__W);
 101 }
 102
 103 static __inline__ __m512d __DEFAULT_FN_ATTRS
 104 _mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 105   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 106                                               (__v8df)_mm512_or_pd(__A, __B),
 107                                               (__v8df)_mm512_setzero_pd());
 108 }
 109
 110 static __inline__ __m512 __DEFAULT_FN_ATTRS
 111 _mm512_or_ps(__m512 __A, __m512 __B) {
 112   return (__m512)((__v16su)__A | (__v16su)__B);
 113 }
 114
 115 static __inline__ __m512 __DEFAULT_FN_ATTRS
 116 _mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 117   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 118                                              (__v16sf)_mm512_or_ps(__A, __B),
 119                                              (__v16sf)__W);
 120 }
 121
 122 static __inline__ __m512 __DEFAULT_FN_ATTRS
 123 _mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 124   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 125                                              (__v16sf)_mm512_or_ps(__A, __B),
 126                                              (__v16sf)_mm512_setzero_ps());
 127 }
 128
 129 static __inline__ __m512d __DEFAULT_FN_ATTRS
 130 _mm512_and_pd(__m512d __A, __m512d __B) {
 131   return (__m512d)((__v8du)__A & (__v8du)__B);
 132 }
 133
 134 static __inline__ __m512d __DEFAULT_FN_ATTRS
 135 _mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 136   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 137                                               (__v8df)_mm512_and_pd(__A, __B),
 138                                               (__v8df)__W);
 139 }
 140
 141 static __inline__ __m512d __DEFAULT_FN_ATTRS
 142 _mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 143   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 144                                               (__v8df)_mm512_and_pd(__A, __B),
 145                                               (__v8df)_mm512_setzero_pd());
 146 }
 147
 148 static __inline__ __m512 __DEFAULT_FN_ATTRS
 149 _mm512_and_ps(__m512 __A, __m512 __B) {
 150   return (__m512)((__v16su)__A & (__v16su)__B);
 151 }
 152
 153 static __inline__ __m512 __DEFAULT_FN_ATTRS
 154 _mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 155   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 156                                              (__v16sf)_mm512_and_ps(__A, __B),
 157                                              (__v16sf)__W);
 158 }
 159
 160 static __inline__ __m512 __DEFAULT_FN_ATTRS
 161 _mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 162   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 163                                              (__v16sf)_mm512_and_ps(__A, __B),
 164                                              (__v16sf)_mm512_setzero_ps());
 165 }
 166
 167 static __inline__ __m512d __DEFAULT_FN_ATTRS
 168 _mm512_andnot_pd(__m512d __A, __m512d __B) {
 169   return (__m512d)(~(__v8du)__A & (__v8du)__B);
 170 }
 171
 172 static __inline__ __m512d __DEFAULT_FN_ATTRS
 173 _mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
 174   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 175                                               (__v8df)_mm512_andnot_pd(__A, __B),
 176                                               (__v8df)__W);
 177 }
 178
 179 static __inline__ __m512d __DEFAULT_FN_ATTRS
 180 _mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B) {
 181   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
 182                                               (__v8df)_mm512_andnot_pd(__A, __B),
 183                                               (__v8df)_mm512_setzero_pd());
 184 }
 185
 186 static __inline__ __m512 __DEFAULT_FN_ATTRS
 187 _mm512_andnot_ps(__m512 __A, __m512 __B) {
 188   return (__m512)(~(__v16su)__A & (__v16su)__B);
 189 }
 190
 191 static __inline__ __m512 __DEFAULT_FN_ATTRS
 192 _mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
 193   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 194                                              (__v16sf)_mm512_andnot_ps(__A, __B),
 195                                              (__v16sf)__W);
 196 }
 197
 198 static __inline__ __m512 __DEFAULT_FN_ATTRS
 199 _mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B) {
 200   return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
 201                                              (__v16sf)_mm512_andnot_ps(__A, __B),
 202                                              (__v16sf)_mm512_setzero_ps());
 203 }
 204
 205 static __inline__ __m512i __DEFAULT_FN_ATTRS
 206 _mm512_cvtpd_epi64 (__m512d __A) {
 207   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 208                 (__v8di) _mm512_setzero_si512(),
 209                 (__mmask8) -1,
 210                 _MM_FROUND_CUR_DIRECTION);
 211 }
 212
 213 static __inline__ __m512i __DEFAULT_FN_ATTRS
 214 _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
 215   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 216                 (__v8di) __W,
 217                 (__mmask8) __U,
 218                 _MM_FROUND_CUR_DIRECTION);
 219 }
 220
 221 static __inline__ __m512i __DEFAULT_FN_ATTRS
 222 _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) {
 223   return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
 224                 (__v8di) _mm512_setzero_si512(),
 225                 (__mmask8) __U,
 226                 _MM_FROUND_CUR_DIRECTION);
 227 }
 228
 229 #define _mm512_cvt_roundpd_epi64(A, R) __extension__ ({              \
 230   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 231                                            (__v8di)_mm512_setzero_si512(), \
 232                                            (__mmask8)-1, (int)(R)); })
 233
 234 #define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) __extension__ ({ \
 235   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 236                                            (__v8di)(__m512i)(W), \
 237                                            (__mmask8)(U), (int)(R)); })
 238
 239 #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) __extension__ ({   \
 240   (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
 241                                            (__v8di)_mm512_setzero_si512(), \
 242                                            (__mmask8)(U), (int)(R)); })
 243
 244 static __inline__ __m512i __DEFAULT_FN_ATTRS
 245 _mm512_cvtpd_epu64 (__m512d __A) {
 246   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 247                  (__v8di) _mm512_setzero_si512(),
 248                  (__mmask8) -1,
 249                  _MM_FROUND_CUR_DIRECTION);
 250 }
 251
 252 static __inline__ __m512i __DEFAULT_FN_ATTRS
 253 _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
 254   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 255                  (__v8di) __W,
 256                  (__mmask8) __U,
 257                  _MM_FROUND_CUR_DIRECTION);
 258 }
 259
 260 static __inline__ __m512i __DEFAULT_FN_ATTRS
 261 _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) {
 262   return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
 263                  (__v8di) _mm512_setzero_si512(),
 264                  (__mmask8) __U,
 265                  _MM_FROUND_CUR_DIRECTION);
 266 }
 267
 268 #define _mm512_cvt_roundpd_epu64(A, R) __extension__ ({               \
 269   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 270                                             (__v8di)_mm512_setzero_si512(), \
 271                                             (__mmask8)-1, (int)(R)); })
 272
 273 #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) __extension__ ({ \
 274   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 275                                             (__v8di)(__m512i)(W), \
 276                                             (__mmask8)(U), (int)(R)); })
 277
 278 #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) __extension__ ({     \
 279   (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
 280                                             (__v8di)_mm512_setzero_si512(), \
 281                                             (__mmask8)(U), (int)(R)); })
 282
 283 static __inline__ __m512i __DEFAULT_FN_ATTRS
 284 _mm512_cvtps_epi64 (__m256 __A) {
 285   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 286                 (__v8di) _mm512_setzero_si512(),
 287                 (__mmask8) -1,
 288                 _MM_FROUND_CUR_DIRECTION);
 289 }
 290
 291 static __inline__ __m512i __DEFAULT_FN_ATTRS
 292 _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
 293   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 294                 (__v8di) __W,
 295                 (__mmask8) __U,
 296                 _MM_FROUND_CUR_DIRECTION);
 297 }
 298
 299 static __inline__ __m512i __DEFAULT_FN_ATTRS
 300 _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) {
 301   return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
 302                 (__v8di) _mm512_setzero_si512(),
 303                 (__mmask8) __U,
 304                 _MM_FROUND_CUR_DIRECTION);
 305 }
 306
 307 #define _mm512_cvt_roundps_epi64(A, R) __extension__ ({             \
 308   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 309                                            (__v8di)_mm512_setzero_si512(), \
 310                                            (__mmask8)-1, (int)(R)); })
 311
 312 #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) __extension__ ({ \
 313   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 314                                            (__v8di)(__m512i)(W), \
 315                                            (__mmask8)(U), (int)(R)); })
 316
 317 #define _mm512_maskz_cvt_roundps_epi64(U, A, R) __extension__ ({   \
 318   (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
 319                                            (__v8di)_mm512_setzero_si512(), \
 320                                            (__mmask8)(U), (int)(R)); })
 321
 322 static __inline__ __m512i __DEFAULT_FN_ATTRS
 323 _mm512_cvtps_epu64 (__m256 __A) {
 324   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 325                  (__v8di) _mm512_setzero_si512(),
 326                  (__mmask8) -1,
 327                  _MM_FROUND_CUR_DIRECTION);
 328 }
 329
 330 static __inline__ __m512i __DEFAULT_FN_ATTRS
 331 _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
 332   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 333                  (__v8di) __W,
 334                  (__mmask8) __U,
 335                  _MM_FROUND_CUR_DIRECTION);
 336 }
 337
 338 static __inline__ __m512i __DEFAULT_FN_ATTRS
 339 _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) {
 340   return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
 341                  (__v8di) _mm512_setzero_si512(),
 342                  (__mmask8) __U,
 343                  _MM_FROUND_CUR_DIRECTION);
 344 }
 345
 346 #define _mm512_cvt_roundps_epu64(A, R) __extension__ ({              \
 347   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 348                                             (__v8di)_mm512_setzero_si512(), \
 349                                             (__mmask8)-1, (int)(R)); })
 350
 351 #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) __extension__ ({ \
 352   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 353                                             (__v8di)(__m512i)(W), \
 354                                             (__mmask8)(U), (int)(R)); })
 355
 356 #define _mm512_maskz_cvt_roundps_epu64(U, A, R) __extension__ ({   \
 357   (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
 358                                             (__v8di)_mm512_setzero_si512(), \
 359                                             (__mmask8)(U), (int)(R)); })
 360
 361
 362 static __inline__ __m512d __DEFAULT_FN_ATTRS
 363 _mm512_cvtepi64_pd (__m512i __A) {
 364   return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A,
 365                 (__v8df) _mm512_setzero_pd(),
 366                 (__mmask8) -1,
 367                 _MM_FROUND_CUR_DIRECTION);
 368 }
 369
 370 static __inline__ __m512d __DEFAULT_FN_ATTRS
 371 _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
 372   return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A,
 373                 (__v8df) __W,
 374                 (__mmask8) __U,
 375                 _MM_FROUND_CUR_DIRECTION);
 376 }
 377
 378 static __inline__ __m512d __DEFAULT_FN_ATTRS
 379 _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) {
 380   return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A,
 381                 (__v8df) _mm512_setzero_pd(),
 382                 (__mmask8) __U,
 383                 _MM_FROUND_CUR_DIRECTION);
 384 }
 385
 386 #define _mm512_cvt_roundepi64_pd(A, R) __extension__ ({          \
 387   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 388                                            (__v8df)_mm512_setzero_pd(), \
 389                                            (__mmask8)-1, (int)(R)); })
 390
 391 #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) __extension__ ({ \
 392   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 393                                            (__v8df)(__m512d)(W), \
 394                                            (__mmask8)(U), (int)(R)); })
 395
 396 #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) __extension__ ({ \
 397   (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
 398                                            (__v8df)_mm512_setzero_pd(), \
 399                                            (__mmask8)(U), (int)(R)); })
 400
 401 static __inline__ __m256 __DEFAULT_FN_ATTRS
 402 _mm512_cvtepi64_ps (__m512i __A) {
 403   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 404                (__v8sf) _mm256_setzero_ps(),
 405                (__mmask8) -1,
 406                _MM_FROUND_CUR_DIRECTION);
 407 }
 408
 409 static __inline__ __m256 __DEFAULT_FN_ATTRS
 410 _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
 411   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 412                (__v8sf) __W,
 413                (__mmask8) __U,
 414                _MM_FROUND_CUR_DIRECTION);
 415 }
 416
 417 static __inline__ __m256 __DEFAULT_FN_ATTRS
 418 _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) {
 419   return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
 420                (__v8sf) _mm256_setzero_ps(),
 421                (__mmask8) __U,
 422                _MM_FROUND_CUR_DIRECTION);
 423 }
 424
 425 #define _mm512_cvt_roundepi64_ps(A, R) __extension__ ({        \
 426   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 427                                           (__v8sf)_mm256_setzero_ps(), \
 428                                           (__mmask8)-1, (int)(R)); })
 429
 430 #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) __extension__ ({ \
 431   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 432                                           (__v8sf)(__m256)(W), (__mmask8)(U), \
 433                                           (int)(R)); })
 434
 435 #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) __extension__ ({ \
 436   (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
 437                                           (__v8sf)_mm256_setzero_ps(), \
 438                                           (__mmask8)(U), (int)(R)); })
 439
 440
 441 static __inline__ __m512i __DEFAULT_FN_ATTRS
 442 _mm512_cvttpd_epi64 (__m512d __A) {
 443   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 444                  (__v8di) _mm512_setzero_si512(),
 445                  (__mmask8) -1,
 446                  _MM_FROUND_CUR_DIRECTION);
 447 }
 448
 449 static __inline__ __m512i __DEFAULT_FN_ATTRS
 450 _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
 451   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 452                  (__v8di) __W,
 453                  (__mmask8) __U,
 454                  _MM_FROUND_CUR_DIRECTION);
 455 }
 456
 457 static __inline__ __m512i __DEFAULT_FN_ATTRS
 458 _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) {
 459   return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
 460                  (__v8di) _mm512_setzero_si512(),
 461                  (__mmask8) __U,
 462                  _MM_FROUND_CUR_DIRECTION);
 463 }
 464
 465 #define _mm512_cvtt_roundpd_epi64(A, R) __extension__ ({             \
 466   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 467                                             (__v8di)_mm512_setzero_si512(), \
 468                                             (__mmask8)-1, (int)(R)); })
 469
 470 #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) __extension__ ({ \
 471   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 472                                             (__v8di)(__m512i)(W), \
 473                                             (__mmask8)(U), (int)(R)); })
 474
 475 #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) __extension__ ({ \
 476   (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
 477                                             (__v8di)_mm512_setzero_si512(), \
 478                                             (__mmask8)(U), (int)(R)); })
 479
 480 static __inline__ __m512i __DEFAULT_FN_ATTRS
 481 _mm512_cvttpd_epu64 (__m512d __A) {
 482   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 483                   (__v8di) _mm512_setzero_si512(),
 484                   (__mmask8) -1,
 485                   _MM_FROUND_CUR_DIRECTION);
 486 }
 487
 488 static __inline__ __m512i __DEFAULT_FN_ATTRS
 489 _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
 490   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 491                   (__v8di) __W,
 492                   (__mmask8) __U,
 493                   _MM_FROUND_CUR_DIRECTION);
 494 }
 495
 496 static __inline__ __m512i __DEFAULT_FN_ATTRS
 497 _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) {
 498   return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
 499                   (__v8di) _mm512_setzero_si512(),
 500                   (__mmask8) __U,
 501                   _MM_FROUND_CUR_DIRECTION);
 502 }
 503
 504 #define _mm512_cvtt_roundpd_epu64(A, R) __extension__ ({              \
 505   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 506                                              (__v8di)_mm512_setzero_si512(), \
 507                                              (__mmask8)-1, (int)(R)); })
 508
 509 #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) __extension__ ({ \
 510   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 511                                              (__v8di)(__m512i)(W), \
 512                                              (__mmask8)(U), (int)(R)); })
 513
 514 #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) __extension__ ({   \
 515   (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
 516                                              (__v8di)_mm512_setzero_si512(), \
 517                                              (__mmask8)(U), (int)(R)); })
 518
 519 static __inline__ __m512i __DEFAULT_FN_ATTRS
 520 _mm512_cvttps_epi64 (__m256 __A) {
 521   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 522                  (__v8di) _mm512_setzero_si512(),
 523                  (__mmask8) -1,
 524                  _MM_FROUND_CUR_DIRECTION);
 525 }
 526
 527 static __inline__ __m512i __DEFAULT_FN_ATTRS
 528 _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
 529   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 530                  (__v8di) __W,
 531                  (__mmask8) __U,
 532                  _MM_FROUND_CUR_DIRECTION);
 533 }
 534
 535 static __inline__ __m512i __DEFAULT_FN_ATTRS
 536 _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) {
 537   return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
 538                  (__v8di) _mm512_setzero_si512(),
 539                  (__mmask8) __U,
 540                  _MM_FROUND_CUR_DIRECTION);
 541 }
 542
 543 #define _mm512_cvtt_roundps_epi64(A, R) __extension__ ({            \
 544   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 545                                             (__v8di)_mm512_setzero_si512(), \
 546                                             (__mmask8)-1, (int)(R)); })
 547
 548 #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) __extension__ ({ \
 549   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 550                                             (__v8di)(__m512i)(W), \
 551                                             (__mmask8)(U), (int)(R)); })
 552
 553 #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) __extension__ ({  \
 554   (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
 555                                             (__v8di)_mm512_setzero_si512(), \
 556                                             (__mmask8)(U), (int)(R)); })
 557
 558 static __inline__ __m512i __DEFAULT_FN_ATTRS
 559 _mm512_cvttps_epu64 (__m256 __A) {
 560   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 561                   (__v8di) _mm512_setzero_si512(),
 562                   (__mmask8) -1,
 563                   _MM_FROUND_CUR_DIRECTION);
 564 }
 565
 566 static __inline__ __m512i __DEFAULT_FN_ATTRS
 567 _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
 568   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 569                   (__v8di) __W,
 570                   (__mmask8) __U,
 571                   _MM_FROUND_CUR_DIRECTION);
 572 }
 573
 574 static __inline__ __m512i __DEFAULT_FN_ATTRS
 575 _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) {
 576   return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
 577                   (__v8di) _mm512_setzero_si512(),
 578                   (__mmask8) __U,
 579                   _MM_FROUND_CUR_DIRECTION);
 580 }
 581
 582 #define _mm512_cvtt_roundps_epu64(A, R) __extension__ ({            \
 583   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 584                                              (__v8di)_mm512_setzero_si512(), \
 585                                              (__mmask8)-1, (int)(R)); })
 586
 587 #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) __extension__ ({ \
 588   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 589                                              (__v8di)(__m512i)(W), \
 590                                              (__mmask8)(U), (int)(R)); })
 591
 592 #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) __extension__ ({  \
 593   (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
 594                                              (__v8di)_mm512_setzero_si512(), \
 595                                              (__mmask8)(U), (int)(R)); })
 596
 597 static __inline__ __m512d __DEFAULT_FN_ATTRS
 598 _mm512_cvtepu64_pd (__m512i __A) {
 599   return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A,
 600                  (__v8df) _mm512_setzero_pd(),
 601                  (__mmask8) -1,
 602                  _MM_FROUND_CUR_DIRECTION);
 603 }
 604
 605 static __inline__ __m512d __DEFAULT_FN_ATTRS
 606 _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
 607   return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A,
 608                  (__v8df) __W,
 609                  (__mmask8) __U,
 610                  _MM_FROUND_CUR_DIRECTION);
 611 }
 612
 613 static __inline__ __m512d __DEFAULT_FN_ATTRS
 614 _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) {
 615   return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A,
 616                  (__v8df) _mm512_setzero_pd(),
 617                  (__mmask8) __U,
 618                  _MM_FROUND_CUR_DIRECTION);
 619 }
 620
 621 #define _mm512_cvt_roundepu64_pd(A, R) __extension__ ({          \
 622   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 623                                             (__v8df)_mm512_setzero_pd(), \
 624                                             (__mmask8)-1, (int)(R)); })
 625
 626 #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) __extension__ ({ \
 627   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 628                                             (__v8df)(__m512d)(W), \
 629                                             (__mmask8)(U), (int)(R)); })
 630
 631
 632 #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) __extension__ ({ \
 633   (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
 634                                             (__v8df)_mm512_setzero_pd(), \
 635                                             (__mmask8)(U), (int)(R)); })
 636
 637
 638 static __inline__ __m256 __DEFAULT_FN_ATTRS
 639 _mm512_cvtepu64_ps (__m512i __A) {
 640   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 641                 (__v8sf) _mm256_setzero_ps(),
 642                 (__mmask8) -1,
 643                 _MM_FROUND_CUR_DIRECTION);
 644 }
 645
 646 static __inline__ __m256 __DEFAULT_FN_ATTRS
 647 _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
 648   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 649                 (__v8sf) __W,
 650                 (__mmask8) __U,
 651                 _MM_FROUND_CUR_DIRECTION);
 652 }
 653
 654 static __inline__ __m256 __DEFAULT_FN_ATTRS
 655 _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) {
 656   return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
 657                 (__v8sf) _mm256_setzero_ps(),
 658                 (__mmask8) __U,
 659                 _MM_FROUND_CUR_DIRECTION);
 660 }
 661
 662 #define _mm512_cvt_roundepu64_ps(A, R) __extension__ ({         \
 663   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 664                                            (__v8sf)_mm256_setzero_ps(), \
 665                                            (__mmask8)-1, (int)(R)); })
 666
 667 #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) __extension__ ({ \
 668   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 669                                            (__v8sf)(__m256)(W), (__mmask8)(U), \
 670                                            (int)(R)); })
 671
 672 #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) __extension__ ({ \
 673   (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
 674                                            (__v8sf)_mm256_setzero_ps(), \
 675                                            (__mmask8)(U), (int)(R)); })
 676
 677 #define _mm512_range_pd(A, B, C) __extension__ ({                     \
 678   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 679                                           (__v8df)(__m512d)(B), (int)(C), \
 680                                           (__v8df)_mm512_setzero_pd(), \
 681                                           (__mmask8)-1, \
 682                                           _MM_FROUND_CUR_DIRECTION); })
 683
 684 #define _mm512_mask_range_pd(W, U, A, B, C) __extension__ ({      \
 685   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 686                                           (__v8df)(__m512d)(B), (int)(C), \
 687                                           (__v8df)(__m512d)(W), (__mmask8)(U), \
 688                                           _MM_FROUND_CUR_DIRECTION); })
 689
 690 #define _mm512_maskz_range_pd(U, A, B, C) __extension__ ({           \
 691   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 692                                           (__v8df)(__m512d)(B), (int)(C), \
 693                                           (__v8df)_mm512_setzero_pd(), \
 694                                           (__mmask8)(U), \
 695                                           _MM_FROUND_CUR_DIRECTION); })
 696
 697 #define _mm512_range_round_pd(A, B, C, R) __extension__ ({           \
 698   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 699                                           (__v8df)(__m512d)(B), (int)(C), \
 700                                           (__v8df)_mm512_setzero_pd(), \
 701                                           (__mmask8)-1, (int)(R)); })
 702
 703 #define _mm512_mask_range_round_pd(W, U, A, B, C, R) __extension__ ({ \
 704   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 705                                           (__v8df)(__m512d)(B), (int)(C), \
 706                                           (__v8df)(__m512d)(W), (__mmask8)(U), \
 707                                           (int)(R)); })
 708
 709 #define _mm512_maskz_range_round_pd(U, A, B, C, R) __extension__ ({ \
 710   (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
 711                                           (__v8df)(__m512d)(B), (int)(C), \
 712                                           (__v8df)_mm512_setzero_pd(), \
 713                                           (__mmask8)(U), (int)(R)); })
 714
 715 #define _mm512_range_ps(A, B, C) __extension__ ({                       \
 716   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 717                                          (__v16sf)(__m512)(B), (int)(C), \
 718                                          (__v16sf)_mm512_setzero_ps(), \
 719                                          (__mmask16)-1, \
 720                                          _MM_FROUND_CUR_DIRECTION); })
 721
 722 #define _mm512_mask_range_ps(W, U, A, B, C) __extension__ ({         \
 723   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 724                                          (__v16sf)(__m512)(B), (int)(C), \
 725                                          (__v16sf)(__m512)(W), (__mmask16)(U), \
 726                                          _MM_FROUND_CUR_DIRECTION); })
 727
 728 #define _mm512_maskz_range_ps(U, A, B, C) __extension__ ({      \
 729   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 730                                          (__v16sf)(__m512)(B), (int)(C), \
 731                                          (__v16sf)_mm512_setzero_ps(), \
 732                                          (__mmask16)(U), \
 733                                          _MM_FROUND_CUR_DIRECTION); })
 734
 735 #define _mm512_range_round_ps(A, B, C, R) __extension__ ({         \
 736   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 737                                          (__v16sf)(__m512)(B), (int)(C), \
 738                                          (__v16sf)_mm512_setzero_ps(), \
 739                                          (__mmask16)-1, (int)(R)); })
 740
 741 #define _mm512_mask_range_round_ps(W, U, A, B, C, R) __extension__ ({ \
 742   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 743                                          (__v16sf)(__m512)(B), (int)(C), \
 744                                          (__v16sf)(__m512)(W), (__mmask16)(U), \
 745                                          (int)(R)); })
 746
 747 #define _mm512_maskz_range_round_ps(U, A, B, C, R) __extension__ ({ \
 748   (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
 749                                          (__v16sf)(__m512)(B), (int)(C), \
 750                                          (__v16sf)_mm512_setzero_ps(), \
 751                                          (__mmask16)(U), (int)(R)); })
 752
 753 #define _mm_range_round_ss(A, B, C, R) __extension__ ({           \
 754   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 755                                                (__v4sf)(__m128)(B), \
 756                                                (__v4sf)_mm_setzero_ps(), \
 757                                                (__mmask8) -1, (int)(C),\
 758                                                (int)(R)); })
 759
 760 #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION)
 761
 762 #define _mm_mask_range_round_ss(W, U, A, B, C, R) __extension__ ({ \
 763   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 764                                                (__v4sf)(__m128)(B), \
 765                                                (__v4sf)(__m128)(W),\
 766                                                (__mmask8)(U), (int)(C),\
 767                                                (int)(R)); })
 768
 769 #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION)
 770
 771 #define _mm_maskz_range_round_ss(U, A, B, C, R) __extension__ ({ \
 772   (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
 773                                                (__v4sf)(__m128)(B), \
 774                                                (__v4sf)_mm_setzero_ps(), \
 775                                                (__mmask8)(U), (int)(C),\
 776                                                (int)(R)); })
 777
 778 #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 779
 780 #define _mm_range_round_sd(A, B, C, R) __extension__ ({           \
 781   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 782                                                 (__v2df)(__m128d)(B), \
 783                                                 (__v2df)_mm_setzero_pd(), \
 784                                                 (__mmask8) -1, (int)(C),\
 785                                                 (int)(R)); })
 786
 787 #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION)
 788
 789 #define _mm_mask_range_round_sd(W, U, A, B, C, R) __extension__ ({ \
 790   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 791                                                 (__v2df)(__m128d)(B), \
 792                                                 (__v2df)(__m128d)(W),\
 793                                                 (__mmask8)(U), (int)(C),\
 794                                                 (int)(R)); })
 795
 796 #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 797
 798 #define _mm_maskz_range_round_sd(U, A, B, C, R) __extension__ ({ \
 799   (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
 800                                                 (__v2df)(__m128d)(B), \
 801                                                 (__v2df)_mm_setzero_pd(), \
 802                                                 (__mmask8)(U), (int)(C),\
 803                                                 (int)(R)); })
 804
 805 #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
 806
 807 #define _mm512_reduce_pd(A, B) __extension__ ({             \
 808   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 809                                            (__v8df)_mm512_setzero_pd(), \
 810                                            (__mmask8)-1, \
 811                                            _MM_FROUND_CUR_DIRECTION); })
 812
 813 #define _mm512_mask_reduce_pd(W, U, A, B) __extension__ ({ \
 814   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 815                                            (__v8df)(__m512d)(W), \
 816                                            (__mmask8)(U), \
 817                                            _MM_FROUND_CUR_DIRECTION); })
 818
 819 #define _mm512_maskz_reduce_pd(U, A, B) __extension__ ({  \
 820   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 821                                            (__v8df)_mm512_setzero_pd(), \
 822                                            (__mmask8)(U), \
 823                                            _MM_FROUND_CUR_DIRECTION); })
 824
 825 #define _mm512_reduce_ps(A, B) __extension__ ({              \
 826   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 827                                           (__v16sf)_mm512_setzero_ps(), \
 828                                           (__mmask16)-1, \
 829                                           _MM_FROUND_CUR_DIRECTION); })
 830
 831 #define _mm512_mask_reduce_ps(W, U, A, B) __extension__ ({   \
 832   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 833                                           (__v16sf)(__m512)(W), \
 834                                           (__mmask16)(U), \
 835                                           _MM_FROUND_CUR_DIRECTION); })
 836
 837 #define _mm512_maskz_reduce_ps(U, A, B) __extension__ ({       \
 838   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 839                                           (__v16sf)_mm512_setzero_ps(), \
 840                                           (__mmask16)(U), \
 841                                           _MM_FROUND_CUR_DIRECTION); })
 842
 843 #define _mm512_reduce_round_pd(A, B, R) __extension__ ({\
 844   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 845                                            (__v8df)_mm512_setzero_pd(), \
 846                                            (__mmask8)-1, (int)(R)); })
 847
 848 #define _mm512_mask_reduce_round_pd(W, U, A, B, R) __extension__ ({\
 849   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 850                                            (__v8df)(__m512d)(W), \
 851                                            (__mmask8)(U), (int)(R)); })
 852
 853 #define _mm512_maskz_reduce_round_pd(U, A, B, R) __extension__ ({\
 854   (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
 855                                            (__v8df)_mm512_setzero_pd(), \
 856                                            (__mmask8)(U), (int)(R)); })
 857
 858 #define _mm512_reduce_round_ps(A, B, R) __extension__ ({\
 859   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 860                                           (__v16sf)_mm512_setzero_ps(), \
 861                                           (__mmask16)-1, (int)(R)); })
 862
 863 #define _mm512_mask_reduce_round_ps(W, U, A, B, R) __extension__ ({\
 864   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 865                                           (__v16sf)(__m512)(W), \
 866                                           (__mmask16)(U), (int)(R)); })
 867
 868 #define _mm512_maskz_reduce_round_ps(U, A, B, R) __extension__ ({\
 869   (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
 870                                           (__v16sf)_mm512_setzero_ps(), \
 871                                           (__mmask16)(U), (int)(R)); })
 872
 873 #define _mm_reduce_ss(A, B, C) __extension__ ({              \
 874   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 875                                        (__v4sf)(__m128)(B), \
 876                                        (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
 877                                        (int)(C), _MM_FROUND_CUR_DIRECTION); })
 878
 879 #define _mm_mask_reduce_ss(W, U, A, B, C) __extension__ ({   \
 880   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 881                                        (__v4sf)(__m128)(B), \
 882                                        (__v4sf)(__m128)(W), (__mmask8)(U), \
 883                                        (int)(C), _MM_FROUND_CUR_DIRECTION); })
 884
 885 #define _mm_maskz_reduce_ss(U, A, B, C) __extension__ ({       \
 886   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 887                                        (__v4sf)(__m128)(B), \
 888                                        (__v4sf)_mm_setzero_ps(), \
 889                                        (__mmask8)(U), (int)(C), \
 890                                        _MM_FROUND_CUR_DIRECTION); })
 891
 892 #define _mm_reduce_round_ss(A, B, C, R) __extension__ ({              \
 893   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 894                                        (__v4sf)(__m128)(B), \
 895                                        (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
 896                                        (int)(C), (int)(R)); })
 897
 898 #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) __extension__ ({   \
 899   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 900                                        (__v4sf)(__m128)(B), \
 901                                        (__v4sf)(__m128)(W), (__mmask8)(U), \
 902                                        (int)(C), (int)(R)); })
 903
 904 #define _mm_maskz_reduce_round_ss(U, A, B, C, R) __extension__ ({       \
 905   (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
 906                                        (__v4sf)(__m128)(B), \
 907                                        (__v4sf)_mm_setzero_ps(), \
 908                                        (__mmask8)(U), (int)(C), (int)(R)); })
 909
 910 #define _mm_reduce_sd(A, B, C) __extension__ ({              \
 911   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 912                                         (__v2df)(__m128d)(B), \
 913                                         (__v2df)_mm_setzero_pd(), \
 914                                         (__mmask8)-1, (int)(C), \
 915                                         _MM_FROUND_CUR_DIRECTION); })
 916
 917 #define _mm_mask_reduce_sd(W, U, A, B, C) __extension__ ({   \
 918   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 919                                         (__v2df)(__m128d)(B), \
 920                                         (__v2df)(__m128d)(W), (__mmask8)(U), \
 921                                         (int)(C), _MM_FROUND_CUR_DIRECTION); })
 922
 923 #define _mm_maskz_reduce_sd(U, A, B, C) __extension__ ({       \
 924   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 925                                         (__v2df)(__m128d)(B), \
 926                                         (__v2df)_mm_setzero_pd(), \
 927                                         (__mmask8)(U), (int)(C), \
 928                                         _MM_FROUND_CUR_DIRECTION); })
 929
 930 #define _mm_reduce_round_sd(A, B, C, R) __extension__ ({              \
 931   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 932                                         (__v2df)(__m128d)(B), \
 933                                         (__v2df)_mm_setzero_pd(), \
 934                                         (__mmask8)-1, (int)(C), (int)(R)); })
 935
 936 #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) __extension__ ({   \
 937   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 938                                         (__v2df)(__m128d)(B), \
 939                                         (__v2df)(__m128d)(W), (__mmask8)(U), \
 940                                         (int)(C), (int)(R)); })
 941
 942 #define _mm_maskz_reduce_round_sd(U, A, B, C, R) __extension__ ({       \
 943   (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
 944                                         (__v2df)(__m128d)(B), \
 945                                         (__v2df)_mm_setzero_pd(), \
 946                                         (__mmask8)(U), (int)(C), (int)(R)); })
 947
 948 static __inline__ __mmask16 __DEFAULT_FN_ATTRS
 949 _mm512_movepi32_mask (__m512i __A)
 950 {
 951   return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
 952 }
 953
 954 static __inline__ __m512i __DEFAULT_FN_ATTRS
 955 _mm512_movm_epi32 (__mmask16 __A)
 956 {
 957   return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
 958 }
 959
 960 static __inline__ __m512i __DEFAULT_FN_ATTRS
 961 _mm512_movm_epi64 (__mmask8 __A)
 962 {
 963   return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
 964 }
 965
 966 static __inline__ __mmask8 __DEFAULT_FN_ATTRS
 967 _mm512_movepi64_mask (__m512i __A)
 968 {
 969   return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
 970 }
 971
 972
 973 static __inline__ __m512 __DEFAULT_FN_ATTRS
 974 _mm512_broadcast_f32x2 (__m128 __A)
 975 {
 976   return (__m512)__builtin_shufflevector((__v4sf)__A,
 977                                          (__v4sf)_mm_undefined_ps(),
 978                                          0, 1, 0, 1, 0, 1, 0, 1,
 979                                          0, 1, 0, 1, 0, 1, 0, 1);
 980 }
 981
 982 static __inline__ __m512 __DEFAULT_FN_ATTRS
 983 _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A)
 984 {
 985   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
 986                                              (__v16sf)_mm512_broadcast_f32x2(__A),
 987                                              (__v16sf)__O);
 988 }
 989
 990 static __inline__ __m512 __DEFAULT_FN_ATTRS
 991 _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A)
 992 {
 993   return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
 994                                              (__v16sf)_mm512_broadcast_f32x2(__A),
 995                                              (__v16sf)_mm512_setzero_ps());
 996 }
 997
 998 static __inline__ __m512 __DEFAULT_FN_ATTRS
 999 _mm512_broadcast_f32x8(__m256 __A)
1000 {
1001   return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A,
1002                                          0, 1, 2, 3, 4, 5, 6, 7,
1003                                          0, 1, 2, 3, 4, 5, 6, 7);
1004 }
1005
1006 static __inline__ __m512 __DEFAULT_FN_ATTRS
1007 _mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A)
1008 {
1009   return (__m512)__builtin_ia32_selectps_512((__mmask8)__M,
1010                                            (__v16sf)_mm512_broadcast_f32x8(__A),
1011                                            (__v16sf)__O);
1012 }
1013
1014 static __inline__ __m512 __DEFAULT_FN_ATTRS
1015 _mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A)
1016 {
1017   return (__m512)__builtin_ia32_selectps_512((__mmask8)__M,
1018                                            (__v16sf)_mm512_broadcast_f32x8(__A),
1019                                            (__v16sf)_mm512_setzero_ps());
1020 }
1021
1022 static __inline__ __m512d __DEFAULT_FN_ATTRS
1023 _mm512_broadcast_f64x2(__m128d __A)
1024 {
1025   return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
1026                                           0, 1, 0, 1, 0, 1, 0, 1);
1027 }
1028
1029 static __inline__ __m512d __DEFAULT_FN_ATTRS
1030 _mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A)
1031 {
1032   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1033                                             (__v8df)_mm512_broadcast_f64x2(__A),
1034                                             (__v8df)__O);
1035 }
1036
1037 static __inline__ __m512d __DEFAULT_FN_ATTRS
1038 _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A)
1039 {
1040   return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1041                                             (__v8df)_mm512_broadcast_f64x2(__A),
1042                                             (__v8df)_mm512_setzero_pd());
1043 }
1044
1045 static __inline__ __m512i __DEFAULT_FN_ATTRS
1046 _mm512_broadcast_i32x2 (__m128i __A)
1047 {
1048   return (__m512i)__builtin_shufflevector((__v4si)__A,
1049                                           (__v4si)_mm_undefined_si128(),
1050                                           0, 1, 0, 1, 0, 1, 0, 1,
1051                                           0, 1, 0, 1, 0, 1, 0, 1);
1052 }
1053
1054 static __inline__ __m512i __DEFAULT_FN_ATTRS
1055 _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A)
1056 {
1057   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1058                                              (__v16si)_mm512_broadcast_i32x2(__A),
1059                                              (__v16si)__O);
1060 }
1061
1062 static __inline__ __m512i __DEFAULT_FN_ATTRS
1063 _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A)
1064 {
1065   return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1066                                              (__v16si)_mm512_broadcast_i32x2(__A),
1067                                              (__v16si)_mm512_setzero_si512());
1068 }
1069
1070 static __inline__ __m512i __DEFAULT_FN_ATTRS
1071 _mm512_broadcast_i32x8(__m256i __A)
1072 {
1073   return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A,
1074                                           0, 1, 2, 3, 4, 5, 6, 7,
1075                                           0, 1, 2, 3, 4, 5, 6, 7);
1076 }
1077
1078 static __inline__ __m512i __DEFAULT_FN_ATTRS
1079 _mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A)
1080 {
1081   return (__m512i)__builtin_ia32_selectd_512((__mmask8)__M,
1082                                            (__v16si)_mm512_broadcast_i32x8(__A),
1083                                            (__v16si)__O);
1084 }
1085
1086 static __inline__ __m512i __DEFAULT_FN_ATTRS
1087 _mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A)
1088 {
1089   return (__m512i)__builtin_ia32_selectd_512((__mmask8)__M,
1090                                            (__v16si)_mm512_broadcast_i32x8(__A),
1091                                            (__v16si)_mm512_setzero_si512());
1092 }
1093
1094 static __inline__ __m512i __DEFAULT_FN_ATTRS
1095 _mm512_broadcast_i64x2(__m128i __A)
1096 {
1097   return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1098                                           0, 1, 0, 1, 0, 1, 0, 1);
1099 }
1100
1101 static __inline__ __m512i __DEFAULT_FN_ATTRS
1102 _mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A)
1103 {
1104   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1105                                             (__v8di)_mm512_broadcast_i64x2(__A),
1106                                             (__v8di)__O);
1107 }
1108
1109 static __inline__ __m512i __DEFAULT_FN_ATTRS
1110 _mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A)
1111 {
1112   return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1113                                             (__v8di)_mm512_broadcast_i64x2(__A),
1114                                             (__v8di)_mm512_setzero_si512());
1115 }
1116
1117 #define _mm512_extractf32x8_ps(A, imm) __extension__ ({ \
1118   (__m256)__builtin_shufflevector((__v16sf)(__m512)(A),           \
1119                                   (__v16sf)_mm512_undefined_ps(), \
1120                                   ((imm) & 1) ?  8 : 0,           \
1121                                   ((imm) & 1) ?  9 : 1,           \
1122                                   ((imm) & 1) ? 10 : 2,           \
1123                                   ((imm) & 1) ? 11 : 3,           \
1124                                   ((imm) & 1) ? 12 : 4,           \
1125                                   ((imm) & 1) ? 13 : 5,           \
1126                                   ((imm) & 1) ? 14 : 6,           \
1127                                   ((imm) & 1) ? 15 : 7); })
1128
1129 #define _mm512_mask_extractf32x8_ps(W, U, A, imm) __extension__ ({ \
1130   (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
1131                                    (__v8sf)_mm512_extractf32x8_ps((A), (imm)), \
1132                                    (__v8sf)(W)); })
1133
1134 #define _mm512_maskz_extractf32x8_ps(U, A, imm) __extension__ ({ \
1135   (__m256)__builtin_ia32_selectps_256((__mmask8)(U), \
1136                                    (__v8sf)_mm512_extractf32x8_ps((A), (imm)), \
1137                                    (__v8sf)_mm256_setzero_ps()); })
1138
1139 #define _mm512_extractf64x2_pd(A, imm) __extension__ ({ \
1140   (__m128d)__builtin_shufflevector((__v8df)(__m512d)(A),          \
1141                                    (__v8df)_mm512_undefined_pd(), \
1142                                    0 + ((imm) & 0x3) * 2,         \
1143                                    1 + ((imm) & 0x3) * 2); })
1144
1145 #define _mm512_mask_extractf64x2_pd(W, U, A, imm) __extension__ ({ \
1146   (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
1147                                    (__v2df)_mm512_extractf64x2_pd((A), (imm)), \
1148                                    (__v2df)(W)); })
1149
1150 #define _mm512_maskz_extractf64x2_pd(U, A, imm) __extension__ ({ \
1151   (__m128d)__builtin_ia32_selectpd_128((__mmask8)(U), \
1152                                    (__v2df)_mm512_extractf64x2_pd((A), (imm)), \
1153                                    (__v2df)_mm_setzero_pd()); })
1154
1155 #define _mm512_extracti32x8_epi32(A, imm) __extension__ ({ \
1156   (__m256i)__builtin_shufflevector((__v16si)(__m512i)(A),             \
1157                                    (__v16si)_mm512_undefined_epi32(), \
1158                                    ((imm) & 1) ?  8 : 0,              \
1159                                    ((imm) & 1) ?  9 : 1,              \
1160                                    ((imm) & 1) ? 10 : 2,              \
1161                                    ((imm) & 1) ? 11 : 3,              \
1162                                    ((imm) & 1) ? 12 : 4,              \
1163                                    ((imm) & 1) ? 13 : 5,              \
1164                                    ((imm) & 1) ? 14 : 6,              \
1165                                    ((imm) & 1) ? 15 : 7); })
1166
1167 #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) __extension__ ({ \
1168   (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \
1169                                 (__v8si)_mm512_extracti32x8_epi32((A), (imm)), \
1170                                 (__v8si)(W)); })
1171
1172 #define _mm512_maskz_extracti32x8_epi32(U, A, imm) __extension__ ({ \
1173   (__m256i)__builtin_ia32_selectd_256((__mmask8)(U), \
1174                                 (__v8si)_mm512_extracti32x8_epi32((A), (imm)), \
1175                                 (__v8si)_mm256_setzero_si256()); })
1176
1177 #define _mm512_extracti64x2_epi64(A, imm) __extension__ ({ \
1178   (__m128i)__builtin_shufflevector((__v8di)(__m512i)(A),          \
1179                                    (__v8di)_mm512_undefined_epi32(), \
1180                                    0 + ((imm) & 0x3) * 2,           \
1181                                    1 + ((imm) & 0x3) * 2); })
1182
1183 #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) __extension__ ({ \
1184   (__m128d)__builtin_ia32_selectq_128((__mmask8)(U), \
1185                                 (__v2di)_mm512_extracti64x2_epi64((A), (imm)), \
1186                                 (__v2di)(W)); })
1187
1188 #define _mm512_maskz_extracti64x2_epi64(U, A, imm) __extension__ ({ \
1189   (__m128d)__builtin_ia32_selectq_128((__mmask8)(U), \
1190                                 (__v2di)_mm512_extracti64x2_epi64((A), (imm)), \
1191                                 (__v2di)_mm_setzero_di()); })
1192
1193 #define _mm512_insertf32x8(A, B, imm) __extension__ ({ \
1194   (__m512)__builtin_shufflevector((__v16sf)(__m512)(A), \
1195                                   (__v16sf)_mm512_castps256_ps512((__m256)(B)),\
1196                                   ((imm) & 0x1) ?  0 : 16, \
1197                                   ((imm) & 0x1) ?  1 : 17, \
1198                                   ((imm) & 0x1) ?  2 : 18, \
1199                                   ((imm) & 0x1) ?  3 : 19, \
1200                                   ((imm) & 0x1) ?  4 : 20, \
1201                                   ((imm) & 0x1) ?  5 : 21, \
1202                                   ((imm) & 0x1) ?  6 : 22, \
1203                                   ((imm) & 0x1) ?  7 : 23, \
1204                                   ((imm) & 0x1) ? 16 :  8, \
1205                                   ((imm) & 0x1) ? 17 :  9, \
1206                                   ((imm) & 0x1) ? 18 : 10, \
1207                                   ((imm) & 0x1) ? 19 : 11, \
1208                                   ((imm) & 0x1) ? 20 : 12, \
1209                                   ((imm) & 0x1) ? 21 : 13, \
1210                                   ((imm) & 0x1) ? 22 : 14, \
1211                                   ((imm) & 0x1) ? 23 : 15); })
1212
1213 #define _mm512_mask_insertf32x8(W, U, A, B, imm) __extension__ ({ \
1214   (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1215                                  (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1216                                  (__v16sf)(W)); })
1217
1218 #define _mm512_maskz_insertf32x8(U, A, B, imm) __extension__ ({ \
1219   (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1220                                  (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1221                                  (__v16sf)_mm512_setzero_ps()); })
1222
1223 #define _mm512_insertf64x2(A, B, imm) __extension__ ({ \
1224   (__m512d)__builtin_shufflevector((__v8df)(__m512d)(A), \
1225                                   (__v8df)_mm512_castpd128_pd512((__m128d)(B)),\
1226                                   (((imm) & 0x3) == 0) ? 8 : 0, \
1227                                   (((imm) & 0x3) == 0) ? 9 : 1, \
1228                                   (((imm) & 0x3) == 1) ? 8 : 2, \
1229                                   (((imm) & 0x3) == 1) ? 9 : 3, \
1230                                   (((imm) & 0x3) == 2) ? 8 : 4, \
1231                                   (((imm) & 0x3) == 2) ? 9 : 5, \
1232                                   (((imm) & 0x3) == 3) ? 8 : 6, \
1233                                   (((imm) & 0x3) == 3) ? 9 : 7); })
1234
1235 #define _mm512_mask_insertf64x2(W, U, A, B, imm) __extension__ ({ \
1236   (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1237                                   (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1238                                   (__v8df)(W)); })
1239
1240 #define _mm512_maskz_insertf64x2(U, A, B, imm) __extension__ ({ \
1241   (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1242                                   (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1243                                   (__v8df)_mm512_setzero_pd()); })
1244
1245 #define _mm512_inserti32x8(A, B, imm) __extension__ ({ \
1246   (__m512i)__builtin_shufflevector((__v16si)(__m512i)(A), \
1247                                  (__v16si)_mm512_castsi256_si512((__m256i)(B)),\
1248                                  ((imm) & 0x1) ?  0 : 16, \
1249                                  ((imm) & 0x1) ?  1 : 17, \
1250                                  ((imm) & 0x1) ?  2 : 18, \
1251                                  ((imm) & 0x1) ?  3 : 19, \
1252                                  ((imm) & 0x1) ?  4 : 20, \
1253                                  ((imm) & 0x1) ?  5 : 21, \
1254                                  ((imm) & 0x1) ?  6 : 22, \
1255                                  ((imm) & 0x1) ?  7 : 23, \
1256                                  ((imm) & 0x1) ? 16 :  8, \
1257                                  ((imm) & 0x1) ? 17 :  9, \
1258                                  ((imm) & 0x1) ? 18 : 10, \
1259                                  ((imm) & 0x1) ? 19 : 11, \
1260                                  ((imm) & 0x1) ? 20 : 12, \
1261                                  ((imm) & 0x1) ? 21 : 13, \
1262                                  ((imm) & 0x1) ? 22 : 14, \
1263                                  ((imm) & 0x1) ? 23 : 15); })
1264
1265 #define _mm512_mask_inserti32x8(W, U, A, B, imm) __extension__ ({ \
1266   (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1267                                  (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1268                                  (__v16si)(W)); })
1269
1270 #define _mm512_maskz_inserti32x8(U, A, B, imm) __extension__ ({ \
1271   (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1272                                  (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1273                                  (__v16si)_mm512_setzero_si512()); })
1274
1275 #define _mm512_inserti64x2(A, B, imm) __extension__ ({ \
1276   (__m512i)__builtin_shufflevector((__v8di)(__m512i)(A), \
1277                                   (__v8di)_mm512_castsi128_si512((__m128i)(B)),\
1278                                   (((imm) & 0x3) == 0) ? 8 : 0, \
1279                                   (((imm) & 0x3) == 0) ? 9 : 1, \
1280                                   (((imm) & 0x3) == 1) ? 8 : 2, \
1281                                   (((imm) & 0x3) == 1) ? 9 : 3, \
1282                                   (((imm) & 0x3) == 2) ? 8 : 4, \
1283                                   (((imm) & 0x3) == 2) ? 9 : 5, \
1284                                   (((imm) & 0x3) == 3) ? 8 : 6, \
1285                                   (((imm) & 0x3) == 3) ? 9 : 7); })
1286
1287 #define _mm512_mask_inserti64x2(W, U, A, B, imm) __extension__ ({ \
1288   (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1289                                   (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1290                                   (__v8di)(W)); })
1291
1292 #define _mm512_maskz_inserti64x2(U, A, B, imm) __extension__ ({ \
1293   (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1294                                   (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1295                                   (__v8di)_mm512_setzero_si512()); })
1296
1297 #define _mm512_mask_fpclass_ps_mask(U, A, imm) __extension__ ({ \
1298   (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1299                                               (int)(imm), (__mmask16)(U)); })
1300
1301 #define _mm512_fpclass_ps_mask(A, imm) __extension__ ({ \
1302   (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1303                                               (int)(imm), (__mmask16)-1); })
1304
1305 #define _mm512_mask_fpclass_pd_mask(U, A, imm) __extension__ ({ \
1306   (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1307                                              (__mmask8)(U)); })
1308
1309 #define _mm512_fpclass_pd_mask(A, imm) __extension__ ({ \
1310   (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1311                                              (__mmask8)-1); })
1312
1313 #define _mm_fpclass_sd_mask(A, imm) __extension__ ({ \
1314   (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1315                                           (__mmask8)-1); })
1316
1317 #define _mm_mask_fpclass_sd_mask(U, A, imm) __extension__ ({ \
1318   (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1319                                           (__mmask8)(U)); })
1320
1321 #define _mm_fpclass_ss_mask(A, imm) __extension__ ({ \
1322   (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1323                                           (__mmask8)-1); })
1324
1325 #define _mm_mask_fpclass_ss_mask(U, A, imm) __extension__ ({ \
1326   (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1327                                           (__mmask8)(U)); })
1328
1329 #undef __DEFAULT_FN_ATTRS
1330
1331 #endif