gcc/config/i386/emmintrin.h

   1 /* Copyright (C) 2003-2024 Free Software Foundation, Inc.
   2
   3    This file is part of GCC.
   4
   5    GCC is free software; you can redistribute it and/or modify
   6    it under the terms of the GNU General Public License as published by
   7    the Free Software Foundation; either version 3, or (at your option)
   8    any later version.
   9
  10    GCC is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13    GNU General Public License for more details.
  14
  15    Under Section 7 of GPL version 3, you are granted additional
  16    permissions described in the GCC Runtime Library Exception, version
  17    3.1, as published by the Free Software Foundation.
  18
  19    You should have received a copy of the GNU General Public License and
  20    a copy of the GCC Runtime Library Exception along with this program;
  21    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  22    <http://www.gnu.org/licenses/>.  */
  23
  24 /* Implemented from the specification included in the Intel C++ Compiler
  25    User Guide and Reference, version 9.0.  */
  26
  27 #ifndef _EMMINTRIN_H_INCLUDED
  28 #define _EMMINTRIN_H_INCLUDED
  29
  30 /* We need definitions from the SSE header files*/
  31 #include <xmmintrin.h>
  32
  33 #ifndef __SSE2__
  34 #pragma GCC push_options
  35 #pragma GCC target("sse2")
  36 #define __DISABLE_SSE2__
  37 #endif /* __SSE2__ */
  38
  39 /* SSE2 */
  40 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  41 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  42 typedef unsigned long long __v2du __attribute__ ((__vector_size__ (16)));
  43 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  44 typedef unsigned int __v4su __attribute__ ((__vector_size__ (16)));
  45 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  46 typedef unsigned short __v8hu __attribute__ ((__vector_size__ (16)));
  47 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  48 typedef signed char __v16qs __attribute__ ((__vector_size__ (16)));
  49 typedef unsigned char __v16qu __attribute__ ((__vector_size__ (16)));
  50
  51 /* The Intel API is flexible enough that we must allow aliasing with other
  52    vector types, and their scalar components.  */
  53 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  54 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  55
  56 /* Unaligned version of the same types.  */
  57 typedef long long __m128i_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
  58 typedef double __m128d_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
  59 typedef double __x86_double_u __attribute__ ((__may_alias__, __aligned__ (1)));
  60
  61 /* Create a selector for use with the SHUFPD instruction.  */
  62 #define _MM_SHUFFLE2(fp1,fp0) \
  63  (((fp1) << 1) | (fp0))
  64
  65 /* Create a vector with element 0 as F and the rest zero.  */
  66 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  67 _mm_set_sd (double __F)
  68 {
  69   return __extension__ (__m128d){ __F, 0.0 };
  70 }
  71
  72 /* Create a vector with both elements equal to F.  */
  73 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  74 _mm_set1_pd (double __F)
  75 {
  76   return __extension__ (__m128d){ __F, __F };
  77 }
  78
  79 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  80 _mm_set_pd1 (double __F)
  81 {
  82   return _mm_set1_pd (__F);
  83 }
  84
  85 /* Create a vector with the lower value X and upper value W.  */
  86 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  87 _mm_set_pd (double __W, double __X)
  88 {
  89   return __extension__ (__m128d){ __X, __W };
  90 }
  91
  92 /* Create a vector with the lower value W and upper value X.  */
  93 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  94 _mm_setr_pd (double __W, double __X)
  95 {
  96   return __extension__ (__m128d){ __W, __X };
  97 }
  98
  99 /* Create an undefined vector.  */
 100 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 101 _mm_undefined_pd (void)
 102 {
 103 #pragma GCC diagnostic push
 104 #pragma GCC diagnostic ignored "-Winit-self"
 105   __m128d __Y = __Y;
 106 #pragma GCC diagnostic pop
 107   return __Y;
 108 }
 109
 110 /* Create a vector of zeros.  */
 111 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 112 _mm_setzero_pd (void)
 113 {
 114   return __extension__ (__m128d){ 0.0, 0.0 };
 115 }
 116
 117 /* Sets the low DPFP value of A from the low value of B.  */
 118 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 119 _mm_move_sd (__m128d __A, __m128d __B)
 120 {
 121   return __extension__ (__m128d) __builtin_shuffle ((__v2df)__A, (__v2df)__B, (__v2di){2, 1});
 122 }
 123
 124 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 125 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 126 _mm_load_pd (double const *__P)
 127 {
 128   return *(__m128d *)__P;
 129 }
 130
 131 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 132 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 133 _mm_loadu_pd (double const *__P)
 134 {
 135   return *(__m128d_u *)__P;
 136 }
 137
 138 /* Create a vector with all two elements equal to *P.  */
 139 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 140 _mm_load1_pd (double const *__P)
 141 {
 142   return _mm_set1_pd (*__P);
 143 }
 144
 145 /* Create a vector with element 0 as *P and the rest zero.  */
 146 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 147 _mm_load_sd (double const *__P)
 148 {
 149   return __extension__ (__m128d) { *(__x86_double_u *)__P, 0.0 };
 150 }
 151
 152 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 153 _mm_load_pd1 (double const *__P)
 154 {
 155   return _mm_load1_pd (__P);
 156 }
 157
 158 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 159 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 160 _mm_loadr_pd (double const *__P)
 161 {
 162   __m128d __tmp = _mm_load_pd (__P);
 163   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 164 }
 165
 166 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 167 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 168 _mm_store_pd (double *__P, __m128d __A)
 169 {
 170   *(__m128d *)__P = __A;
 171 }
 172
 173 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 174 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 175 _mm_storeu_pd (double *__P, __m128d __A)
 176 {
 177   *(__m128d_u *)__P = __A;
 178 }
 179
 180 /* Stores the lower DPFP value.  */
 181 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 182 _mm_store_sd (double *__P, __m128d __A)
 183 {
 184   *(__x86_double_u *)__P = ((__v2df)__A)[0] ;
 185 }
 186
 187 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 188 _mm_cvtsd_f64 (__m128d __A)
 189 {
 190   return ((__v2df)__A)[0];
 191 }
 192
 193 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 194 _mm_storel_pd (double *__P, __m128d __A)
 195 {
 196   *__P = ((__v2df)__A)[0];
 197 }
 198
 199 /* Stores the upper DPFP value.  */
 200 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 201 _mm_storeh_pd (double *__P, __m128d __A)
 202 {
 203   *__P = ((__v2df)__A)[1];
 204 }
 205
 206 /* Store the lower DPFP value across two words.
 207    The address must be 16-byte aligned.  */
 208 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 209 _mm_store1_pd (double *__P, __m128d __A)
 210 {
 211   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 212 }
 213
 214 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 215 _mm_store_pd1 (double *__P, __m128d __A)
 216 {
 217   _mm_store1_pd (__P, __A);
 218 }
 219
 220 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 221 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 222 _mm_storer_pd (double *__P, __m128d __A)
 223 {
 224   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 225 }
 226
 227 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 228 _mm_cvtsi128_si32 (__m128i __A)
 229 {
 230   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 231 }
 232
 233 #ifdef __x86_64__
 234 /* Intel intrinsic.  */
 235 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 236 _mm_cvtsi128_si64 (__m128i __A)
 237 {
 238   return ((__v2di)__A)[0];
 239 }
 240
 241 /* Microsoft intrinsic.  */
 242 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 243 _mm_cvtsi128_si64x (__m128i __A)
 244 {
 245   return ((__v2di)__A)[0];
 246 }
 247 #endif
 248
 249 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 250 _mm_add_pd (__m128d __A, __m128d __B)
 251 {
 252   return (__m128d) ((__v2df)__A + (__v2df)__B);
 253 }
 254
 255 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 256 _mm_add_sd (__m128d __A, __m128d __B)
 257 {
 258   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 259 }
 260
 261 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 262 _mm_sub_pd (__m128d __A, __m128d __B)
 263 {
 264   return (__m128d) ((__v2df)__A - (__v2df)__B);
 265 }
 266
 267 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 268 _mm_sub_sd (__m128d __A, __m128d __B)
 269 {
 270   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 271 }
 272
 273 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 274 _mm_mul_pd (__m128d __A, __m128d __B)
 275 {
 276   return (__m128d) ((__v2df)__A * (__v2df)__B);
 277 }
 278
 279 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 280 _mm_mul_sd (__m128d __A, __m128d __B)
 281 {
 282   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 283 }
 284
 285 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 286 _mm_div_pd (__m128d __A, __m128d __B)
 287 {
 288   return (__m128d) ((__v2df)__A / (__v2df)__B);
 289 }
 290
 291 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 292 _mm_div_sd (__m128d __A, __m128d __B)
 293 {
 294   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 295 }
 296
 297 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 298 _mm_sqrt_pd (__m128d __A)
 299 {
 300   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 301 }
 302
 303 /* Return pair {sqrt (B[0]), A[1]}.  */
 304 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 305 _mm_sqrt_sd (__m128d __A, __m128d __B)
 306 {
 307   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 308   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 309 }
 310
 311 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 312 _mm_min_pd (__m128d __A, __m128d __B)
 313 {
 314   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 315 }
 316
 317 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 318 _mm_min_sd (__m128d __A, __m128d __B)
 319 {
 320   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 321 }
 322
 323 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 324 _mm_max_pd (__m128d __A, __m128d __B)
 325 {
 326   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 327 }
 328
 329 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 330 _mm_max_sd (__m128d __A, __m128d __B)
 331 {
 332   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 333 }
 334
 335 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 336 _mm_and_pd (__m128d __A, __m128d __B)
 337 {
 338   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 339 }
 340
 341 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 342 _mm_andnot_pd (__m128d __A, __m128d __B)
 343 {
 344   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 345 }
 346
 347 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 348 _mm_or_pd (__m128d __A, __m128d __B)
 349 {
 350   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 351 }
 352
 353 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 354 _mm_xor_pd (__m128d __A, __m128d __B)
 355 {
 356   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 357 }
 358
 359 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 360 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 361 {
 362   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 363 }
 364
 365 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 366 _mm_cmplt_pd (__m128d __A, __m128d __B)
 367 {
 368   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 369 }
 370
 371 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 372 _mm_cmple_pd (__m128d __A, __m128d __B)
 373 {
 374   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 375 }
 376
 377 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 378 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 379 {
 380   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 381 }
 382
 383 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 384 _mm_cmpge_pd (__m128d __A, __m128d __B)
 385 {
 386   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 387 }
 388
 389 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 390 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 391 {
 392   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 393 }
 394
 395 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 396 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 397 {
 398   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 399 }
 400
 401 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 402 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 403 {
 404   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 405 }
 406
 407 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 408 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 409 {
 410   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 411 }
 412
 413 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 414 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 415 {
 416   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 417 }
 418
 419 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 420 _mm_cmpord_pd (__m128d __A, __m128d __B)
 421 {
 422   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 423 }
 424
 425 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 426 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 427 {
 428   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 429 }
 430
 431 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 432 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 433 {
 434   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 435 }
 436
 437 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 438 _mm_cmplt_sd (__m128d __A, __m128d __B)
 439 {
 440   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 441 }
 442
 443 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 444 _mm_cmple_sd (__m128d __A, __m128d __B)
 445 {
 446   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 447 }
 448
 449 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 450 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 451 {
 452   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 453                                          (__v2df)
 454                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 455                                                                  (__v2df)
 456                                                                  __A));
 457 }
 458
 459 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 460 _mm_cmpge_sd (__m128d __A, __m128d __B)
 461 {
 462   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 463                                          (__v2df)
 464                                          __builtin_ia32_cmplesd ((__v2df) __B,
 465                                                                  (__v2df)
 466                                                                  __A));
 467 }
 468
 469 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 470 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 471 {
 472   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 473 }
 474
 475 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 476 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 477 {
 478   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 479 }
 480
 481 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 482 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 483 {
 484   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 485 }
 486
 487 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 488 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 489 {
 490   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 491                                          (__v2df)
 492                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 493                                                                   (__v2df)
 494                                                                   __A));
 495 }
 496
 497 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 498 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 499 {
 500   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 501                                          (__v2df)
 502                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 503                                                                   (__v2df)
 504                                                                   __A));
 505 }
 506
 507 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 508 _mm_cmpord_sd (__m128d __A, __m128d __B)
 509 {
 510   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 511 }
 512
 513 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 514 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 515 {
 516   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 517 }
 518
 519 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 520 _mm_comieq_sd (__m128d __A, __m128d __B)
 521 {
 522   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 523 }
 524
 525 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 526 _mm_comilt_sd (__m128d __A, __m128d __B)
 527 {
 528   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 529 }
 530
 531 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 532 _mm_comile_sd (__m128d __A, __m128d __B)
 533 {
 534   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 535 }
 536
 537 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 538 _mm_comigt_sd (__m128d __A, __m128d __B)
 539 {
 540   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 541 }
 542
 543 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 544 _mm_comige_sd (__m128d __A, __m128d __B)
 545 {
 546   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 547 }
 548
 549 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 550 _mm_comineq_sd (__m128d __A, __m128d __B)
 551 {
 552   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 553 }
 554
 555 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 556 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 557 {
 558   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 559 }
 560
 561 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 562 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 563 {
 564   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 565 }
 566
 567 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 568 _mm_ucomile_sd (__m128d __A, __m128d __B)
 569 {
 570   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 571 }
 572
 573 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 574 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 575 {
 576   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 577 }
 578
 579 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 580 _mm_ucomige_sd (__m128d __A, __m128d __B)
 581 {
 582   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 583 }
 584
 585 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 586 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 587 {
 588   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 589 }
 590
 591 /* Create a vector of Qi, where i is the element number.  */
 592
 593 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 594 _mm_set_epi64x (long long __q1, long long __q0)
 595 {
 596   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 597 }
 598
 599 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 600 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 601 {
 602   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 603 }
 604
 605 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 606 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 607 {
 608   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 609 }
 610
 611 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 612 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 613                short __q3, short __q2, short __q1, short __q0)
 614 {
 615   return __extension__ (__m128i)(__v8hi){
 616     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 617 }
 618
 619 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 620 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 621               char __q11, char __q10, char __q09, char __q08,
 622               char __q07, char __q06, char __q05, char __q04,
 623               char __q03, char __q02, char __q01, char __q00)
 624 {
 625   return __extension__ (__m128i)(__v16qi){
 626     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 627     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 628   };
 629 }
 630
 631 /* Set all of the elements of the vector to A.  */
 632
 633 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 634 _mm_set1_epi64x (long long __A)
 635 {
 636   return _mm_set_epi64x (__A, __A);
 637 }
 638
 639 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 640 _mm_set1_epi64 (__m64 __A)
 641 {
 642   return _mm_set_epi64 (__A, __A);
 643 }
 644
 645 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 646 _mm_set1_epi32 (int __A)
 647 {
 648   return _mm_set_epi32 (__A, __A, __A, __A);
 649 }
 650
 651 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 652 _mm_set1_epi16 (short __A)
 653 {
 654   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 655 }
 656
 657 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 658 _mm_set1_epi8 (char __A)
 659 {
 660   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 661                        __A, __A, __A, __A, __A, __A, __A, __A);
 662 }
 663
 664 /* Create a vector of Qi, where i is the element number.
 665    The parameter order is reversed from the _mm_set_epi* functions.  */
 666
 667 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 668 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 669 {
 670   return _mm_set_epi64 (__q1, __q0);
 671 }
 672
 673 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 674 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 675 {
 676   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 677 }
 678
 679 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 680 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 681                 short __q4, short __q5, short __q6, short __q7)
 682 {
 683   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 684 }
 685
 686 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 687 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 688                char __q04, char __q05, char __q06, char __q07,
 689                char __q08, char __q09, char __q10, char __q11,
 690                char __q12, char __q13, char __q14, char __q15)
 691 {
 692   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 693                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 694 }
 695
 696 /* Create a vector with element 0 as *P and the rest zero.  */
 697
 698 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 699 _mm_load_si128 (__m128i const *__P)
 700 {
 701   return *__P;
 702 }
 703
 704 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 705 _mm_loadu_si128 (__m128i_u const *__P)
 706 {
 707   return *__P;
 708 }
 709
 710 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 711 _mm_loadl_epi64 (__m128i_u const *__P)
 712 {
 713   return _mm_set_epi64 ((__m64)0LL, *(__m64_u *)__P);
 714 }
 715
 716 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 717 _mm_loadu_si64 (void const *__P)
 718 {
 719   return _mm_loadl_epi64 ((__m128i_u *)__P);
 720 }
 721
 722 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 723 _mm_loadu_si32 (void const *__P)
 724 {
 725   return _mm_set_epi32 (0, 0, 0, (*(__m32_u *)__P)[0]);
 726 }
 727
 728 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 729 _mm_loadu_si16 (void const *__P)
 730 {
 731   return _mm_set_epi16 (0, 0, 0, 0, 0, 0, 0, (*(__m16_u *)__P)[0]);
 732 }
 733
 734 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 735 _mm_store_si128 (__m128i *__P, __m128i __B)
 736 {
 737   *__P = __B;
 738 }
 739
 740 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 741 _mm_storeu_si128 (__m128i_u *__P, __m128i __B)
 742 {
 743   *__P = __B;
 744 }
 745
 746 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 747 _mm_storel_epi64 (__m128i_u *__P, __m128i __B)
 748 {
 749   *(__m64_u *)__P = (__m64) ((__v2di)__B)[0];
 750 }
 751
 752 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 753 _mm_storeu_si64 (void *__P, __m128i __B)
 754 {
 755   _mm_storel_epi64 ((__m128i_u *)__P, __B);
 756 }
 757
 758 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 759 _mm_storeu_si32 (void *__P, __m128i __B)
 760 {
 761   *(__m32_u *)__P = (__m32) ((__v4si)__B)[0];
 762 }
 763
 764 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 765 _mm_storeu_si16 (void *__P, __m128i __B)
 766 {
 767   *(__m16_u *)__P = (__m16) ((__v8hi)__B)[0];
 768 }
 769
 770 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 771 _mm_movepi64_pi64 (__m128i __B)
 772 {
 773   return (__m64) ((__v2di)__B)[0];
 774 }
 775
 776 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 777 _mm_movpi64_epi64 (__m64 __A)
 778 {
 779   return _mm_set_epi64 ((__m64)0LL, __A);
 780 }
 781
 782 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 783 _mm_move_epi64 (__m128i __A)
 784 {
 785   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 786 }
 787
 788 /* Create an undefined vector.  */
 789 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 790 _mm_undefined_si128 (void)
 791 {
 792 #pragma GCC diagnostic push
 793 #pragma GCC diagnostic ignored "-Winit-self"
 794   __m128i __Y = __Y;
 795 #pragma GCC diagnostic pop
 796   return __Y;
 797 }
 798
 799 /* Create a vector of zeros.  */
 800 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 801 _mm_setzero_si128 (void)
 802 {
 803   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 804 }
 805
 806 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 807 _mm_cvtepi32_pd (__m128i __A)
 808 {
 809   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 810 }
 811
 812 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 813 _mm_cvtepi32_ps (__m128i __A)
 814 {
 815   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 816 }
 817
 818 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 819 _mm_cvtpd_epi32 (__m128d __A)
 820 {
 821   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 822 }
 823
 824 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 825 _mm_cvtpd_pi32 (__m128d __A)
 826 {
 827   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 828 }
 829
 830 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 831 _mm_cvtpd_ps (__m128d __A)
 832 {
 833   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 834 }
 835
 836 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 837 _mm_cvttpd_epi32 (__m128d __A)
 838 {
 839   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 840 }
 841
 842 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 843 _mm_cvttpd_pi32 (__m128d __A)
 844 {
 845   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 846 }
 847
 848 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 849 _mm_cvtpi32_pd (__m64 __A)
 850 {
 851   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 852 }
 853
 854 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 855 _mm_cvtps_epi32 (__m128 __A)
 856 {
 857   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 858 }
 859
 860 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 861 _mm_cvttps_epi32 (__m128 __A)
 862 {
 863   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 864 }
 865
 866 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 867 _mm_cvtps_pd (__m128 __A)
 868 {
 869   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 870 }
 871
 872 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 873 _mm_cvtsd_si32 (__m128d __A)
 874 {
 875   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 876 }
 877
 878 #ifdef __x86_64__
 879 /* Intel intrinsic.  */
 880 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 881 _mm_cvtsd_si64 (__m128d __A)
 882 {
 883   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 884 }
 885
 886 /* Microsoft intrinsic.  */
 887 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 888 _mm_cvtsd_si64x (__m128d __A)
 889 {
 890   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 891 }
 892 #endif
 893
 894 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 895 _mm_cvttsd_si32 (__m128d __A)
 896 {
 897   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 898 }
 899
 900 #ifdef __x86_64__
 901 /* Intel intrinsic.  */
 902 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 903 _mm_cvttsd_si64 (__m128d __A)
 904 {
 905   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 906 }
 907
 908 /* Microsoft intrinsic.  */
 909 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 910 _mm_cvttsd_si64x (__m128d __A)
 911 {
 912   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 913 }
 914 #endif
 915
 916 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 917 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 918 {
 919   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 920 }
 921
 922 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 923 _mm_cvtsi32_sd (__m128d __A, int __B)
 924 {
 925   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 926 }
 927
 928 #ifdef __x86_64__
 929 /* Intel intrinsic.  */
 930 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 931 _mm_cvtsi64_sd (__m128d __A, long long __B)
 932 {
 933   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 934 }
 935
 936 /* Microsoft intrinsic.  */
 937 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 938 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 939 {
 940   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 941 }
 942 #endif
 943
 944 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 945 _mm_cvtss_sd (__m128d __A, __m128 __B)
 946 {
 947   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 948 }
 949
 950 #ifdef __OPTIMIZE__
 951 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 952 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 953 {
 954   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 955 }
 956 #else
 957 #define _mm_shuffle_pd(A, B, N)                                         \
 958   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 959                                    (__v2df)(__m128d)(B), (int)(N)))
 960 #endif
 961
 962 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 963 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 964 {
 965   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 966 }
 967
 968 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 969 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 970 {
 971   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 972 }
 973
 974 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 975 _mm_loadh_pd (__m128d __A, double const *__B)
 976 {
 977   return __extension__ (__m128d) { ((__v2df)__A)[0], *(__x86_double_u*)__B };
 978 }
 979
 980 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 981 _mm_loadl_pd (__m128d __A, double const *__B)
 982 {
 983   return __extension__ (__m128d) { *(__x86_double_u*)__B, ((__v2df)__A)[1] };
 984 }
 985
 986 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 987 _mm_movemask_pd (__m128d __A)
 988 {
 989   return __builtin_ia32_movmskpd ((__v2df)__A);
 990 }
 991
 992 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 993 _mm_packs_epi16 (__m128i __A, __m128i __B)
 994 {
 995   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 996 }
 997
 998 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 999 _mm_packs_epi32 (__m128i __A, __m128i __B)
1000 {
1001   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
1002 }
1003
1004 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1005 _mm_packus_epi16 (__m128i __A, __m128i __B)
1006 {
1007   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
1008 }
1009
1010 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1011 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
1012 {
1013   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
1014 }
1015
1016 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1017 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
1018 {
1019   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
1020 }
1021
1022 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1023 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
1024 {
1025   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
1026 }
1027
1028 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1029 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
1030 {
1031   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
1032 }
1033
1034 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1035 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
1036 {
1037   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
1038 }
1039
1040 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1041 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
1042 {
1043   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
1044 }
1045
1046 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1047 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
1048 {
1049   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
1050 }
1051
1052 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1053 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
1054 {
1055   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
1056 }
1057
1058 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1059 _mm_add_epi8 (__m128i __A, __m128i __B)
1060 {
1061   return (__m128i) ((__v16qu)__A + (__v16qu)__B);
1062 }
1063
1064 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1065 _mm_add_epi16 (__m128i __A, __m128i __B)
1066 {
1067   return (__m128i) ((__v8hu)__A + (__v8hu)__B);
1068 }
1069
1070 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1071 _mm_add_epi32 (__m128i __A, __m128i __B)
1072 {
1073   return (__m128i) ((__v4su)__A + (__v4su)__B);
1074 }
1075
1076 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1077 _mm_add_epi64 (__m128i __A, __m128i __B)
1078 {
1079   return (__m128i) ((__v2du)__A + (__v2du)__B);
1080 }
1081
1082 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1083 _mm_adds_epi8 (__m128i __A, __m128i __B)
1084 {
1085   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1086 }
1087
1088 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1089 _mm_adds_epi16 (__m128i __A, __m128i __B)
1090 {
1091   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1092 }
1093
1094 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1095 _mm_adds_epu8 (__m128i __A, __m128i __B)
1096 {
1097   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1098 }
1099
1100 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1101 _mm_adds_epu16 (__m128i __A, __m128i __B)
1102 {
1103   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1104 }
1105
1106 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1107 _mm_sub_epi8 (__m128i __A, __m128i __B)
1108 {
1109   return (__m128i) ((__v16qu)__A - (__v16qu)__B);
1110 }
1111
1112 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1113 _mm_sub_epi16 (__m128i __A, __m128i __B)
1114 {
1115   return (__m128i) ((__v8hu)__A - (__v8hu)__B);
1116 }
1117
1118 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1119 _mm_sub_epi32 (__m128i __A, __m128i __B)
1120 {
1121   return (__m128i) ((__v4su)__A - (__v4su)__B);
1122 }
1123
1124 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1125 _mm_sub_epi64 (__m128i __A, __m128i __B)
1126 {
1127   return (__m128i) ((__v2du)__A - (__v2du)__B);
1128 }
1129
1130 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1131 _mm_subs_epi8 (__m128i __A, __m128i __B)
1132 {
1133   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1134 }
1135
1136 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1137 _mm_subs_epi16 (__m128i __A, __m128i __B)
1138 {
1139   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1140 }
1141
1142 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1143 _mm_subs_epu8 (__m128i __A, __m128i __B)
1144 {
1145   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1146 }
1147
1148 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1149 _mm_subs_epu16 (__m128i __A, __m128i __B)
1150 {
1151   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1152 }
1153
1154 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1155 _mm_madd_epi16 (__m128i __A, __m128i __B)
1156 {
1157   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1158 }
1159
1160 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1161 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1162 {
1163   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1164 }
1165
1166 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1167 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1168 {
1169   return (__m128i) ((__v8hu)__A * (__v8hu)__B);
1170 }
1171
1172 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1173 _mm_mul_su32 (__m64 __A, __m64 __B)
1174 {
1175   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1176 }
1177
1178 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1179 _mm_mul_epu32 (__m128i __A, __m128i __B)
1180 {
1181   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1182 }
1183
1184 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1185 _mm_slli_epi16 (__m128i __A, int __B)
1186 {
1187   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1188 }
1189
1190 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1191 _mm_slli_epi32 (__m128i __A, int __B)
1192 {
1193   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1194 }
1195
1196 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1197 _mm_slli_epi64 (__m128i __A, int __B)
1198 {
1199   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1200 }
1201
1202 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1203 _mm_srai_epi16 (__m128i __A, int __B)
1204 {
1205   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1206 }
1207
1208 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1209 _mm_srai_epi32 (__m128i __A, int __B)
1210 {
1211   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1212 }
1213
1214 #ifdef __OPTIMIZE__
1215 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1216 _mm_bsrli_si128 (__m128i __A, const int __N)
1217 {
1218   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1219 }
1220
1221 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1222 _mm_bslli_si128 (__m128i __A, const int __N)
1223 {
1224   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1225 }
1226
1227 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1228 _mm_srli_si128 (__m128i __A, const int __N)
1229 {
1230   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1231 }
1232
1233 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1234 _mm_slli_si128 (__m128i __A, const int __N)
1235 {
1236   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1237 }
1238 #else
1239 #define _mm_bsrli_si128(A, N) \
1240   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1241 #define _mm_bslli_si128(A, N) \
1242   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1243 #define _mm_srli_si128(A, N) \
1244   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1245 #define _mm_slli_si128(A, N) \
1246   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1247 #endif
1248
1249 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1250 _mm_srli_epi16 (__m128i __A, int __B)
1251 {
1252   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1253 }
1254
1255 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1256 _mm_srli_epi32 (__m128i __A, int __B)
1257 {
1258   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1259 }
1260
1261 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1262 _mm_srli_epi64 (__m128i __A, int __B)
1263 {
1264   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1265 }
1266
1267 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1268 _mm_sll_epi16 (__m128i __A, __m128i __B)
1269 {
1270   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1271 }
1272
1273 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1274 _mm_sll_epi32 (__m128i __A, __m128i __B)
1275 {
1276   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1277 }
1278
1279 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1280 _mm_sll_epi64 (__m128i __A, __m128i __B)
1281 {
1282   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1283 }
1284
1285 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1286 _mm_sra_epi16 (__m128i __A, __m128i __B)
1287 {
1288   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1289 }
1290
1291 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1292 _mm_sra_epi32 (__m128i __A, __m128i __B)
1293 {
1294   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1295 }
1296
1297 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1298 _mm_srl_epi16 (__m128i __A, __m128i __B)
1299 {
1300   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1301 }
1302
1303 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1304 _mm_srl_epi32 (__m128i __A, __m128i __B)
1305 {
1306   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1307 }
1308
1309 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1310 _mm_srl_epi64 (__m128i __A, __m128i __B)
1311 {
1312   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1313 }
1314
1315 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1316 _mm_and_si128 (__m128i __A, __m128i __B)
1317 {
1318   return (__m128i) ((__v2du)__A & (__v2du)__B);
1319 }
1320
1321 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1322 _mm_andnot_si128 (__m128i __A, __m128i __B)
1323 {
1324   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1325 }
1326
1327 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1328 _mm_or_si128 (__m128i __A, __m128i __B)
1329 {
1330   return (__m128i) ((__v2du)__A | (__v2du)__B);
1331 }
1332
1333 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1334 _mm_xor_si128 (__m128i __A, __m128i __B)
1335 {
1336   return (__m128i) ((__v2du)__A ^ (__v2du)__B);
1337 }
1338
1339 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1340 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1341 {
1342   return (__m128i) ((__v16qi)__A == (__v16qi)__B);
1343 }
1344
1345 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1346 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1347 {
1348   return (__m128i) ((__v8hi)__A == (__v8hi)__B);
1349 }
1350
1351 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1352 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1353 {
1354   return (__m128i) ((__v4si)__A == (__v4si)__B);
1355 }
1356
1357 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1358 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1359 {
1360   return (__m128i) ((__v16qs)__A < (__v16qs)__B);
1361 }
1362
1363 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1364 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1365 {
1366   return (__m128i) ((__v8hi)__A < (__v8hi)__B);
1367 }
1368
1369 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1370 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1371 {
1372   return (__m128i) ((__v4si)__A < (__v4si)__B);
1373 }
1374
1375 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1376 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1377 {
1378   return (__m128i) ((__v16qs)__A > (__v16qs)__B);
1379 }
1380
1381 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1382 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1383 {
1384   return (__m128i) ((__v8hi)__A > (__v8hi)__B);
1385 }
1386
1387 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1388 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1389 {
1390   return (__m128i) ((__v4si)__A > (__v4si)__B);
1391 }
1392
1393 #ifdef __OPTIMIZE__
1394 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1395 _mm_cmp_pd (__m128d __X, __m128d __Y, const int __P)
1396 {
1397   return (__m128d) __builtin_ia32_cmppd ((__v2df)__X, (__v2df)__Y, __P);
1398 }
1399
1400 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1401 _mm_cmp_sd (__m128d __X, __m128d __Y, const int __P)
1402 {
1403   return (__m128d) __builtin_ia32_cmpsd ((__v2df)__X, (__v2df)__Y, __P);
1404 }
1405 #else
1406 #define _mm_cmp_pd(X, Y, P)                                             \
1407   ((__m128d) __builtin_ia32_cmppd ((__v2df)(__m128d)(X),                \
1408                                    (__v2df)(__m128d)(Y), (int)(P)))
1409
1410 #define _mm_cmp_sd(X, Y, P)                                             \
1411   ((__m128d) __builtin_ia32_cmpsd ((__v2df)(__m128d)(X),                \
1412                                    (__v2df)(__m128d)(Y), (int)(P)))
1413 #endif
1414
1415 #ifdef __OPTIMIZE__
1416 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1417 _mm_extract_epi16 (__m128i const __A, int const __N)
1418 {
1419   return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1420 }
1421
1422 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1423 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1424 {
1425   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1426 }
1427 #else
1428 #define _mm_extract_epi16(A, N) \
1429   ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1430 #define _mm_insert_epi16(A, D, N)                               \
1431   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1432                                           (int)(D), (int)(N)))
1433 #endif
1434
1435 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1436 _mm_max_epi16 (__m128i __A, __m128i __B)
1437 {
1438   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1439 }
1440
1441 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1442 _mm_max_epu8 (__m128i __A, __m128i __B)
1443 {
1444   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1445 }
1446
1447 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1448 _mm_min_epi16 (__m128i __A, __m128i __B)
1449 {
1450   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1451 }
1452
1453 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1454 _mm_min_epu8 (__m128i __A, __m128i __B)
1455 {
1456   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1457 }
1458
1459 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1460 _mm_movemask_epi8 (__m128i __A)
1461 {
1462   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1463 }
1464
1465 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1466 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1467 {
1468   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1469 }
1470
1471 #ifdef __OPTIMIZE__
1472 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1473 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1474 {
1475   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1476 }
1477
1478 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1479 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1480 {
1481   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1482 }
1483
1484 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1485 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1486 {
1487   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1488 }
1489 #else
1490 #define _mm_shufflehi_epi16(A, N) \
1491   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1492 #define _mm_shufflelo_epi16(A, N) \
1493   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1494 #define _mm_shuffle_epi32(A, N) \
1495   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1496 #endif
1497
1498 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1499 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1500 {
1501   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1502 }
1503
1504 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1505 _mm_avg_epu8 (__m128i __A, __m128i __B)
1506 {
1507   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1508 }
1509
1510 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1511 _mm_avg_epu16 (__m128i __A, __m128i __B)
1512 {
1513   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1514 }
1515
1516 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1517 _mm_sad_epu8 (__m128i __A, __m128i __B)
1518 {
1519   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1520 }
1521
1522 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1523 _mm_stream_si32 (int *__A, int __B)
1524 {
1525   __builtin_ia32_movnti (__A, __B);
1526 }
1527
1528 #ifdef __x86_64__
1529 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1530 _mm_stream_si64 (long long int *__A, long long int __B)
1531 {
1532   __builtin_ia32_movnti64 (__A, __B);
1533 }
1534 #endif
1535
1536 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1537 _mm_stream_si128 (__m128i *__A, __m128i __B)
1538 {
1539   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1540 }
1541
1542 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1543 _mm_stream_pd (double *__A, __m128d __B)
1544 {
1545   __builtin_ia32_movntpd (__A, (__v2df)__B);
1546 }
1547
1548 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1549 _mm_clflush (void const *__A)
1550 {
1551   __builtin_ia32_clflush (__A);
1552 }
1553
1554 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1555 _mm_lfence (void)
1556 {
1557   __builtin_ia32_lfence ();
1558 }
1559
1560 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1561 _mm_mfence (void)
1562 {
1563   __builtin_ia32_mfence ();
1564 }
1565
1566 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1567 _mm_cvtsi32_si128 (int __A)
1568 {
1569   return _mm_set_epi32 (0, 0, 0, __A);
1570 }
1571
1572 #ifdef __x86_64__
1573 /* Intel intrinsic.  */
1574 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1575 _mm_cvtsi64_si128 (long long __A)
1576 {
1577   return _mm_set_epi64x (0, __A);
1578 }
1579
1580 /* Microsoft intrinsic.  */
1581 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1582 _mm_cvtsi64x_si128 (long long __A)
1583 {
1584   return _mm_set_epi64x (0, __A);
1585 }
1586 #endif
1587
1588 /* Casts between various SP, DP, INT vector types.  Note that these do no
1589    conversion of values, they just change the type.  */
1590 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1591 _mm_castpd_ps(__m128d __A)
1592 {
1593   return (__m128) __A;
1594 }
1595
1596 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1597 _mm_castpd_si128(__m128d __A)
1598 {
1599   return (__m128i) __A;
1600 }
1601
1602 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1603 _mm_castps_pd(__m128 __A)
1604 {
1605   return (__m128d) __A;
1606 }
1607
1608 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1609 _mm_castps_si128(__m128 __A)
1610 {
1611   return (__m128i) __A;
1612 }
1613
1614 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1615 _mm_castsi128_ps(__m128i __A)
1616 {
1617   return (__m128) __A;
1618 }
1619
1620 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1621 _mm_castsi128_pd(__m128i __A)
1622 {
1623   return (__m128d) __A;
1624 }
1625
1626 #ifdef __DISABLE_SSE2__
1627 #undef __DISABLE_SSE2__
1628 #pragma GCC pop_options
1629 #endif /* __DISABLE_SSE2__ */
1630
1631 #endif /* _EMMINTRIN_H_INCLUDED */