gcc/config/i386/emmintrin.h

   1 /* Copyright (C) 2003-2017 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 unsigned char __v16qu __attribute__ ((__vector_size__ (16)));
  49
  50 /* The Intel API is flexible enough that we must allow aliasing with other
  51    vector types, and their scalar components.  */
  52 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  53 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  54
  55 /* Unaligned version of the same types.  */
  56 typedef long long __m128i_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
  57 typedef double __m128d_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1)));
  58
  59 /* Create a selector for use with the SHUFPD instruction.  */
  60 #define _MM_SHUFFLE2(fp1,fp0) \
  61  (((fp1) << 1) | (fp0))
  62
  63 /* Create a vector with element 0 as F and the rest zero.  */
  64 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  65 _mm_set_sd (double __F)
  66 {
  67   return __extension__ (__m128d){ __F, 0.0 };
  68 }
  69
  70 /* Create a vector with both elements equal to F.  */
  71 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  72 _mm_set1_pd (double __F)
  73 {
  74   return __extension__ (__m128d){ __F, __F };
  75 }
  76
  77 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  78 _mm_set_pd1 (double __F)
  79 {
  80   return _mm_set1_pd (__F);
  81 }
  82
  83 /* Create a vector with the lower value X and upper value W.  */
  84 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  85 _mm_set_pd (double __W, double __X)
  86 {
  87   return __extension__ (__m128d){ __X, __W };
  88 }
  89
  90 /* Create a vector with the lower value W and upper value X.  */
  91 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  92 _mm_setr_pd (double __W, double __X)
  93 {
  94   return __extension__ (__m128d){ __W, __X };
  95 }
  96
  97 /* Create an undefined vector.  */
  98 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
  99 _mm_undefined_pd (void)
 100 {
 101   __m128d __Y = __Y;
 102   return __Y;
 103 }
 104
 105 /* Create a vector of zeros.  */
 106 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 107 _mm_setzero_pd (void)
 108 {
 109   return __extension__ (__m128d){ 0.0, 0.0 };
 110 }
 111
 112 /* Sets the low DPFP value of A from the low value of B.  */
 113 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 114 _mm_move_sd (__m128d __A, __m128d __B)
 115 {
 116   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 117 }
 118
 119 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 120 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 121 _mm_load_pd (double const *__P)
 122 {
 123   return *(__m128d *)__P;
 124 }
 125
 126 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 127 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 128 _mm_loadu_pd (double const *__P)
 129 {
 130   return *(__m128d_u *)__P;
 131 }
 132
 133 /* Create a vector with all two elements equal to *P.  */
 134 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 135 _mm_load1_pd (double const *__P)
 136 {
 137   return _mm_set1_pd (*__P);
 138 }
 139
 140 /* Create a vector with element 0 as *P and the rest zero.  */
 141 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 142 _mm_load_sd (double const *__P)
 143 {
 144   return _mm_set_sd (*__P);
 145 }
 146
 147 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 148 _mm_load_pd1 (double const *__P)
 149 {
 150   return _mm_load1_pd (__P);
 151 }
 152
 153 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 154 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 155 _mm_loadr_pd (double const *__P)
 156 {
 157   __m128d __tmp = _mm_load_pd (__P);
 158   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 159 }
 160
 161 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 162 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 163 _mm_store_pd (double *__P, __m128d __A)
 164 {
 165   *(__m128d *)__P = __A;
 166 }
 167
 168 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 169 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 170 _mm_storeu_pd (double *__P, __m128d __A)
 171 {
 172   *(__m128d_u *)__P = __A;
 173 }
 174
 175 /* Stores the lower DPFP value.  */
 176 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 177 _mm_store_sd (double *__P, __m128d __A)
 178 {
 179   *__P = ((__v2df)__A)[0];
 180 }
 181
 182 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 183 _mm_cvtsd_f64 (__m128d __A)
 184 {
 185   return ((__v2df)__A)[0];
 186 }
 187
 188 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 189 _mm_storel_pd (double *__P, __m128d __A)
 190 {
 191   _mm_store_sd (__P, __A);
 192 }
 193
 194 /* Stores the upper DPFP value.  */
 195 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 196 _mm_storeh_pd (double *__P, __m128d __A)
 197 {
 198   *__P = ((__v2df)__A)[1];
 199 }
 200
 201 /* Store the lower DPFP value across two words.
 202    The address must be 16-byte aligned.  */
 203 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 204 _mm_store1_pd (double *__P, __m128d __A)
 205 {
 206   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 207 }
 208
 209 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 210 _mm_store_pd1 (double *__P, __m128d __A)
 211 {
 212   _mm_store1_pd (__P, __A);
 213 }
 214
 215 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 216 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 217 _mm_storer_pd (double *__P, __m128d __A)
 218 {
 219   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 220 }
 221
 222 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 223 _mm_cvtsi128_si32 (__m128i __A)
 224 {
 225   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 226 }
 227
 228 #ifdef __x86_64__
 229 /* Intel intrinsic.  */
 230 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 231 _mm_cvtsi128_si64 (__m128i __A)
 232 {
 233   return ((__v2di)__A)[0];
 234 }
 235
 236 /* Microsoft intrinsic.  */
 237 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 238 _mm_cvtsi128_si64x (__m128i __A)
 239 {
 240   return ((__v2di)__A)[0];
 241 }
 242 #endif
 243
 244 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 245 _mm_add_pd (__m128d __A, __m128d __B)
 246 {
 247   return (__m128d) ((__v2df)__A + (__v2df)__B);
 248 }
 249
 250 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 251 _mm_add_sd (__m128d __A, __m128d __B)
 252 {
 253   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 254 }
 255
 256 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 257 _mm_sub_pd (__m128d __A, __m128d __B)
 258 {
 259   return (__m128d) ((__v2df)__A - (__v2df)__B);
 260 }
 261
 262 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 263 _mm_sub_sd (__m128d __A, __m128d __B)
 264 {
 265   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 266 }
 267
 268 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 269 _mm_mul_pd (__m128d __A, __m128d __B)
 270 {
 271   return (__m128d) ((__v2df)__A * (__v2df)__B);
 272 }
 273
 274 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 275 _mm_mul_sd (__m128d __A, __m128d __B)
 276 {
 277   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 278 }
 279
 280 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 281 _mm_div_pd (__m128d __A, __m128d __B)
 282 {
 283   return (__m128d) ((__v2df)__A / (__v2df)__B);
 284 }
 285
 286 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 287 _mm_div_sd (__m128d __A, __m128d __B)
 288 {
 289   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 290 }
 291
 292 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 293 _mm_sqrt_pd (__m128d __A)
 294 {
 295   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 296 }
 297
 298 /* Return pair {sqrt (B[0]), A[1]}.  */
 299 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 300 _mm_sqrt_sd (__m128d __A, __m128d __B)
 301 {
 302   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 303   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 304 }
 305
 306 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 307 _mm_min_pd (__m128d __A, __m128d __B)
 308 {
 309   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 310 }
 311
 312 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 313 _mm_min_sd (__m128d __A, __m128d __B)
 314 {
 315   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 316 }
 317
 318 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 319 _mm_max_pd (__m128d __A, __m128d __B)
 320 {
 321   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 322 }
 323
 324 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 325 _mm_max_sd (__m128d __A, __m128d __B)
 326 {
 327   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 328 }
 329
 330 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 331 _mm_and_pd (__m128d __A, __m128d __B)
 332 {
 333   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 334 }
 335
 336 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 337 _mm_andnot_pd (__m128d __A, __m128d __B)
 338 {
 339   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 340 }
 341
 342 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 343 _mm_or_pd (__m128d __A, __m128d __B)
 344 {
 345   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 346 }
 347
 348 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 349 _mm_xor_pd (__m128d __A, __m128d __B)
 350 {
 351   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 352 }
 353
 354 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 355 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 356 {
 357   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 358 }
 359
 360 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 361 _mm_cmplt_pd (__m128d __A, __m128d __B)
 362 {
 363   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 364 }
 365
 366 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 367 _mm_cmple_pd (__m128d __A, __m128d __B)
 368 {
 369   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 370 }
 371
 372 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 373 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 374 {
 375   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 376 }
 377
 378 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 379 _mm_cmpge_pd (__m128d __A, __m128d __B)
 380 {
 381   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 382 }
 383
 384 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 385 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 386 {
 387   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 388 }
 389
 390 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 391 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 392 {
 393   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 394 }
 395
 396 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 397 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 398 {
 399   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 400 }
 401
 402 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 403 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 404 {
 405   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 406 }
 407
 408 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 409 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 410 {
 411   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 412 }
 413
 414 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 415 _mm_cmpord_pd (__m128d __A, __m128d __B)
 416 {
 417   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 418 }
 419
 420 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 421 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 422 {
 423   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 424 }
 425
 426 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 427 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 428 {
 429   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 430 }
 431
 432 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 433 _mm_cmplt_sd (__m128d __A, __m128d __B)
 434 {
 435   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 436 }
 437
 438 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 439 _mm_cmple_sd (__m128d __A, __m128d __B)
 440 {
 441   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 442 }
 443
 444 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 445 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 446 {
 447   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 448                                          (__v2df)
 449                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 450                                                                  (__v2df)
 451                                                                  __A));
 452 }
 453
 454 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 455 _mm_cmpge_sd (__m128d __A, __m128d __B)
 456 {
 457   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 458                                          (__v2df)
 459                                          __builtin_ia32_cmplesd ((__v2df) __B,
 460                                                                  (__v2df)
 461                                                                  __A));
 462 }
 463
 464 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 465 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 466 {
 467   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 468 }
 469
 470 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 471 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 472 {
 473   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 474 }
 475
 476 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 477 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 478 {
 479   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 480 }
 481
 482 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 483 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 484 {
 485   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 486                                          (__v2df)
 487                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 488                                                                   (__v2df)
 489                                                                   __A));
 490 }
 491
 492 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 493 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 494 {
 495   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 496                                          (__v2df)
 497                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 498                                                                   (__v2df)
 499                                                                   __A));
 500 }
 501
 502 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 503 _mm_cmpord_sd (__m128d __A, __m128d __B)
 504 {
 505   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 506 }
 507
 508 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 509 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 510 {
 511   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 512 }
 513
 514 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 515 _mm_comieq_sd (__m128d __A, __m128d __B)
 516 {
 517   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 518 }
 519
 520 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 521 _mm_comilt_sd (__m128d __A, __m128d __B)
 522 {
 523   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 524 }
 525
 526 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 527 _mm_comile_sd (__m128d __A, __m128d __B)
 528 {
 529   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 530 }
 531
 532 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 533 _mm_comigt_sd (__m128d __A, __m128d __B)
 534 {
 535   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 536 }
 537
 538 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 539 _mm_comige_sd (__m128d __A, __m128d __B)
 540 {
 541   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 542 }
 543
 544 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 545 _mm_comineq_sd (__m128d __A, __m128d __B)
 546 {
 547   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 548 }
 549
 550 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 551 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 552 {
 553   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 554 }
 555
 556 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 557 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 558 {
 559   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 560 }
 561
 562 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 563 _mm_ucomile_sd (__m128d __A, __m128d __B)
 564 {
 565   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 566 }
 567
 568 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 569 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 570 {
 571   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 572 }
 573
 574 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 575 _mm_ucomige_sd (__m128d __A, __m128d __B)
 576 {
 577   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 578 }
 579
 580 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 581 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 582 {
 583   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 584 }
 585
 586 /* Create a vector of Qi, where i is the element number.  */
 587
 588 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 589 _mm_set_epi64x (long long __q1, long long __q0)
 590 {
 591   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 592 }
 593
 594 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 595 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 596 {
 597   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 598 }
 599
 600 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 601 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 602 {
 603   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 604 }
 605
 606 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 607 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 608                short __q3, short __q2, short __q1, short __q0)
 609 {
 610   return __extension__ (__m128i)(__v8hi){
 611     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 612 }
 613
 614 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 615 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 616               char __q11, char __q10, char __q09, char __q08,
 617               char __q07, char __q06, char __q05, char __q04,
 618               char __q03, char __q02, char __q01, char __q00)
 619 {
 620   return __extension__ (__m128i)(__v16qi){
 621     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 622     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 623   };
 624 }
 625
 626 /* Set all of the elements of the vector to A.  */
 627
 628 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 629 _mm_set1_epi64x (long long __A)
 630 {
 631   return _mm_set_epi64x (__A, __A);
 632 }
 633
 634 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 635 _mm_set1_epi64 (__m64 __A)
 636 {
 637   return _mm_set_epi64 (__A, __A);
 638 }
 639
 640 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 641 _mm_set1_epi32 (int __A)
 642 {
 643   return _mm_set_epi32 (__A, __A, __A, __A);
 644 }
 645
 646 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 647 _mm_set1_epi16 (short __A)
 648 {
 649   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 650 }
 651
 652 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 653 _mm_set1_epi8 (char __A)
 654 {
 655   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 656                        __A, __A, __A, __A, __A, __A, __A, __A);
 657 }
 658
 659 /* Create a vector of Qi, where i is the element number.
 660    The parameter order is reversed from the _mm_set_epi* functions.  */
 661
 662 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 663 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 664 {
 665   return _mm_set_epi64 (__q1, __q0);
 666 }
 667
 668 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 669 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 670 {
 671   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 672 }
 673
 674 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 675 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 676                 short __q4, short __q5, short __q6, short __q7)
 677 {
 678   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 679 }
 680
 681 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 682 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 683                char __q04, char __q05, char __q06, char __q07,
 684                char __q08, char __q09, char __q10, char __q11,
 685                char __q12, char __q13, char __q14, char __q15)
 686 {
 687   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 688                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 689 }
 690
 691 /* Create a vector with element 0 as *P and the rest zero.  */
 692
 693 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 694 _mm_load_si128 (__m128i const *__P)
 695 {
 696   return *__P;
 697 }
 698
 699 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 700 _mm_loadu_si128 (__m128i_u const *__P)
 701 {
 702   return *__P;
 703 }
 704
 705 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 706 _mm_loadl_epi64 (__m128i_u const *__P)
 707 {
 708   return _mm_set_epi64 ((__m64)0LL, *(__m64_u *)__P);
 709 }
 710
 711 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 712 _mm_store_si128 (__m128i *__P, __m128i __B)
 713 {
 714   *__P = __B;
 715 }
 716
 717 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 718 _mm_storeu_si128 (__m128i_u *__P, __m128i __B)
 719 {
 720   *__P = __B;
 721 }
 722
 723 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 724 _mm_storel_epi64 (__m128i_u *__P, __m128i __B)
 725 {
 726   *(__m64_u *)__P = (__m64) ((__v2di)__B)[0];
 727 }
 728
 729 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 730 _mm_movepi64_pi64 (__m128i __B)
 731 {
 732   return (__m64) ((__v2di)__B)[0];
 733 }
 734
 735 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 736 _mm_movpi64_epi64 (__m64 __A)
 737 {
 738   return _mm_set_epi64 ((__m64)0LL, __A);
 739 }
 740
 741 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 742 _mm_move_epi64 (__m128i __A)
 743 {
 744   return (__m128i)__builtin_ia32_movq128 ((__v2di) __A);
 745 }
 746
 747 /* Create an undefined vector.  */
 748 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 749 _mm_undefined_si128 (void)
 750 {
 751   __m128i __Y = __Y;
 752   return __Y;
 753 }
 754
 755 /* Create a vector of zeros.  */
 756 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 757 _mm_setzero_si128 (void)
 758 {
 759   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 760 }
 761
 762 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 763 _mm_cvtepi32_pd (__m128i __A)
 764 {
 765   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 766 }
 767
 768 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 769 _mm_cvtepi32_ps (__m128i __A)
 770 {
 771   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 772 }
 773
 774 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 775 _mm_cvtpd_epi32 (__m128d __A)
 776 {
 777   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 778 }
 779
 780 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 781 _mm_cvtpd_pi32 (__m128d __A)
 782 {
 783   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 784 }
 785
 786 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 787 _mm_cvtpd_ps (__m128d __A)
 788 {
 789   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 790 }
 791
 792 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 793 _mm_cvttpd_epi32 (__m128d __A)
 794 {
 795   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 796 }
 797
 798 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 799 _mm_cvttpd_pi32 (__m128d __A)
 800 {
 801   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 802 }
 803
 804 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 805 _mm_cvtpi32_pd (__m64 __A)
 806 {
 807   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 808 }
 809
 810 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 811 _mm_cvtps_epi32 (__m128 __A)
 812 {
 813   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 814 }
 815
 816 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 817 _mm_cvttps_epi32 (__m128 __A)
 818 {
 819   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 820 }
 821
 822 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 823 _mm_cvtps_pd (__m128 __A)
 824 {
 825   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 826 }
 827
 828 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 829 _mm_cvtsd_si32 (__m128d __A)
 830 {
 831   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 832 }
 833
 834 #ifdef __x86_64__
 835 /* Intel intrinsic.  */
 836 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 837 _mm_cvtsd_si64 (__m128d __A)
 838 {
 839   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 840 }
 841
 842 /* Microsoft intrinsic.  */
 843 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 844 _mm_cvtsd_si64x (__m128d __A)
 845 {
 846   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 847 }
 848 #endif
 849
 850 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 851 _mm_cvttsd_si32 (__m128d __A)
 852 {
 853   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 854 }
 855
 856 #ifdef __x86_64__
 857 /* Intel intrinsic.  */
 858 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 859 _mm_cvttsd_si64 (__m128d __A)
 860 {
 861   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 862 }
 863
 864 /* Microsoft intrinsic.  */
 865 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 866 _mm_cvttsd_si64x (__m128d __A)
 867 {
 868   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 869 }
 870 #endif
 871
 872 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 873 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 874 {
 875   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 876 }
 877
 878 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 879 _mm_cvtsi32_sd (__m128d __A, int __B)
 880 {
 881   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 882 }
 883
 884 #ifdef __x86_64__
 885 /* Intel intrinsic.  */
 886 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 887 _mm_cvtsi64_sd (__m128d __A, long long __B)
 888 {
 889   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 890 }
 891
 892 /* Microsoft intrinsic.  */
 893 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 894 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 895 {
 896   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 897 }
 898 #endif
 899
 900 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 901 _mm_cvtss_sd (__m128d __A, __m128 __B)
 902 {
 903   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 904 }
 905
 906 #ifdef __OPTIMIZE__
 907 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 908 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 909 {
 910   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 911 }
 912 #else
 913 #define _mm_shuffle_pd(A, B, N)                                         \
 914   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 915                                    (__v2df)(__m128d)(B), (int)(N)))
 916 #endif
 917
 918 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 919 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 920 {
 921   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 922 }
 923
 924 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 925 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 926 {
 927   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 928 }
 929
 930 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 931 _mm_loadh_pd (__m128d __A, double const *__B)
 932 {
 933   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 934 }
 935
 936 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 937 _mm_loadl_pd (__m128d __A, double const *__B)
 938 {
 939   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 940 }
 941
 942 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 943 _mm_movemask_pd (__m128d __A)
 944 {
 945   return __builtin_ia32_movmskpd ((__v2df)__A);
 946 }
 947
 948 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 949 _mm_packs_epi16 (__m128i __A, __m128i __B)
 950 {
 951   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 952 }
 953
 954 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 955 _mm_packs_epi32 (__m128i __A, __m128i __B)
 956 {
 957   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 958 }
 959
 960 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 961 _mm_packus_epi16 (__m128i __A, __m128i __B)
 962 {
 963   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 964 }
 965
 966 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 967 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 968 {
 969   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 970 }
 971
 972 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 973 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 974 {
 975   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 976 }
 977
 978 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 979 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 980 {
 981   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 982 }
 983
 984 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 985 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 986 {
 987   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 988 }
 989
 990 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 991 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 992 {
 993   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 994 }
 995
 996 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
 997 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 998 {
 999   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
1000 }
1001
1002 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1003 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
1004 {
1005   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
1006 }
1007
1008 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1009 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
1010 {
1011   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
1012 }
1013
1014 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1015 _mm_add_epi8 (__m128i __A, __m128i __B)
1016 {
1017   return (__m128i) ((__v16qu)__A + (__v16qu)__B);
1018 }
1019
1020 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1021 _mm_add_epi16 (__m128i __A, __m128i __B)
1022 {
1023   return (__m128i) ((__v8hu)__A + (__v8hu)__B);
1024 }
1025
1026 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1027 _mm_add_epi32 (__m128i __A, __m128i __B)
1028 {
1029   return (__m128i) ((__v4su)__A + (__v4su)__B);
1030 }
1031
1032 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1033 _mm_add_epi64 (__m128i __A, __m128i __B)
1034 {
1035   return (__m128i) ((__v2du)__A + (__v2du)__B);
1036 }
1037
1038 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1039 _mm_adds_epi8 (__m128i __A, __m128i __B)
1040 {
1041   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1042 }
1043
1044 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1045 _mm_adds_epi16 (__m128i __A, __m128i __B)
1046 {
1047   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1048 }
1049
1050 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1051 _mm_adds_epu8 (__m128i __A, __m128i __B)
1052 {
1053   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1054 }
1055
1056 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1057 _mm_adds_epu16 (__m128i __A, __m128i __B)
1058 {
1059   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1060 }
1061
1062 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1063 _mm_sub_epi8 (__m128i __A, __m128i __B)
1064 {
1065   return (__m128i) ((__v16qu)__A - (__v16qu)__B);
1066 }
1067
1068 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1069 _mm_sub_epi16 (__m128i __A, __m128i __B)
1070 {
1071   return (__m128i) ((__v8hu)__A - (__v8hu)__B);
1072 }
1073
1074 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1075 _mm_sub_epi32 (__m128i __A, __m128i __B)
1076 {
1077   return (__m128i) ((__v4su)__A - (__v4su)__B);
1078 }
1079
1080 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1081 _mm_sub_epi64 (__m128i __A, __m128i __B)
1082 {
1083   return (__m128i) ((__v2du)__A - (__v2du)__B);
1084 }
1085
1086 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1087 _mm_subs_epi8 (__m128i __A, __m128i __B)
1088 {
1089   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1090 }
1091
1092 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1093 _mm_subs_epi16 (__m128i __A, __m128i __B)
1094 {
1095   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1096 }
1097
1098 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1099 _mm_subs_epu8 (__m128i __A, __m128i __B)
1100 {
1101   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1102 }
1103
1104 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1105 _mm_subs_epu16 (__m128i __A, __m128i __B)
1106 {
1107   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1108 }
1109
1110 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1111 _mm_madd_epi16 (__m128i __A, __m128i __B)
1112 {
1113   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1114 }
1115
1116 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1117 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1118 {
1119   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1120 }
1121
1122 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1123 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1124 {
1125   return (__m128i) ((__v8hu)__A * (__v8hu)__B);
1126 }
1127
1128 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1129 _mm_mul_su32 (__m64 __A, __m64 __B)
1130 {
1131   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1132 }
1133
1134 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1135 _mm_mul_epu32 (__m128i __A, __m128i __B)
1136 {
1137   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1138 }
1139
1140 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1141 _mm_slli_epi16 (__m128i __A, int __B)
1142 {
1143   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1144 }
1145
1146 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1147 _mm_slli_epi32 (__m128i __A, int __B)
1148 {
1149   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1150 }
1151
1152 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1153 _mm_slli_epi64 (__m128i __A, int __B)
1154 {
1155   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1156 }
1157
1158 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1159 _mm_srai_epi16 (__m128i __A, int __B)
1160 {
1161   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1162 }
1163
1164 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1165 _mm_srai_epi32 (__m128i __A, int __B)
1166 {
1167   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1168 }
1169
1170 #ifdef __OPTIMIZE__
1171 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1172 _mm_bsrli_si128 (__m128i __A, const int __N)
1173 {
1174   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1175 }
1176
1177 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1178 _mm_bslli_si128 (__m128i __A, const int __N)
1179 {
1180   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1181 }
1182
1183 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1184 _mm_srli_si128 (__m128i __A, const int __N)
1185 {
1186   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1187 }
1188
1189 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1190 _mm_slli_si128 (__m128i __A, const int __N)
1191 {
1192   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1193 }
1194 #else
1195 #define _mm_bsrli_si128(A, N) \
1196   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1197 #define _mm_bslli_si128(A, N) \
1198   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1199 #define _mm_srli_si128(A, N) \
1200   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1201 #define _mm_slli_si128(A, N) \
1202   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1203 #endif
1204
1205 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1206 _mm_srli_epi16 (__m128i __A, int __B)
1207 {
1208   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1209 }
1210
1211 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1212 _mm_srli_epi32 (__m128i __A, int __B)
1213 {
1214   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1215 }
1216
1217 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1218 _mm_srli_epi64 (__m128i __A, int __B)
1219 {
1220   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1221 }
1222
1223 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1224 _mm_sll_epi16 (__m128i __A, __m128i __B)
1225 {
1226   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1227 }
1228
1229 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1230 _mm_sll_epi32 (__m128i __A, __m128i __B)
1231 {
1232   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1233 }
1234
1235 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1236 _mm_sll_epi64 (__m128i __A, __m128i __B)
1237 {
1238   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1239 }
1240
1241 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1242 _mm_sra_epi16 (__m128i __A, __m128i __B)
1243 {
1244   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1245 }
1246
1247 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1248 _mm_sra_epi32 (__m128i __A, __m128i __B)
1249 {
1250   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1251 }
1252
1253 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1254 _mm_srl_epi16 (__m128i __A, __m128i __B)
1255 {
1256   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1257 }
1258
1259 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1260 _mm_srl_epi32 (__m128i __A, __m128i __B)
1261 {
1262   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1263 }
1264
1265 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1266 _mm_srl_epi64 (__m128i __A, __m128i __B)
1267 {
1268   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1269 }
1270
1271 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1272 _mm_and_si128 (__m128i __A, __m128i __B)
1273 {
1274   return (__m128i) ((__v2du)__A & (__v2du)__B);
1275 }
1276
1277 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1278 _mm_andnot_si128 (__m128i __A, __m128i __B)
1279 {
1280   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1281 }
1282
1283 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1284 _mm_or_si128 (__m128i __A, __m128i __B)
1285 {
1286   return (__m128i) ((__v2du)__A | (__v2du)__B);
1287 }
1288
1289 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1290 _mm_xor_si128 (__m128i __A, __m128i __B)
1291 {
1292   return (__m128i) ((__v2du)__A ^ (__v2du)__B);
1293 }
1294
1295 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1296 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1297 {
1298   return (__m128i) ((__v16qi)__A == (__v16qi)__B);
1299 }
1300
1301 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1302 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1303 {
1304   return (__m128i) ((__v8hi)__A == (__v8hi)__B);
1305 }
1306
1307 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1308 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1309 {
1310   return (__m128i) ((__v4si)__A == (__v4si)__B);
1311 }
1312
1313 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1314 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1315 {
1316   return (__m128i) ((__v16qi)__A < (__v16qi)__B);
1317 }
1318
1319 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1320 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1321 {
1322   return (__m128i) ((__v8hi)__A < (__v8hi)__B);
1323 }
1324
1325 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1326 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1327 {
1328   return (__m128i) ((__v4si)__A < (__v4si)__B);
1329 }
1330
1331 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1332 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1333 {
1334   return (__m128i) ((__v16qi)__A > (__v16qi)__B);
1335 }
1336
1337 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1338 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1339 {
1340   return (__m128i) ((__v8hi)__A > (__v8hi)__B);
1341 }
1342
1343 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1344 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1345 {
1346   return (__m128i) ((__v4si)__A > (__v4si)__B);
1347 }
1348
1349 #ifdef __OPTIMIZE__
1350 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1351 _mm_extract_epi16 (__m128i const __A, int const __N)
1352 {
1353   return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1354 }
1355
1356 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1357 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1358 {
1359   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1360 }
1361 #else
1362 #define _mm_extract_epi16(A, N) \
1363   ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1364 #define _mm_insert_epi16(A, D, N)                               \
1365   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1366                                           (int)(D), (int)(N)))
1367 #endif
1368
1369 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1370 _mm_max_epi16 (__m128i __A, __m128i __B)
1371 {
1372   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1373 }
1374
1375 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1376 _mm_max_epu8 (__m128i __A, __m128i __B)
1377 {
1378   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1379 }
1380
1381 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1382 _mm_min_epi16 (__m128i __A, __m128i __B)
1383 {
1384   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1385 }
1386
1387 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1388 _mm_min_epu8 (__m128i __A, __m128i __B)
1389 {
1390   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1391 }
1392
1393 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1394 _mm_movemask_epi8 (__m128i __A)
1395 {
1396   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1397 }
1398
1399 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1400 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1401 {
1402   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1403 }
1404
1405 #ifdef __OPTIMIZE__
1406 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1407 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1408 {
1409   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1410 }
1411
1412 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1413 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1414 {
1415   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1416 }
1417
1418 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1419 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1420 {
1421   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1422 }
1423 #else
1424 #define _mm_shufflehi_epi16(A, N) \
1425   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1426 #define _mm_shufflelo_epi16(A, N) \
1427   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1428 #define _mm_shuffle_epi32(A, N) \
1429   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1430 #endif
1431
1432 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1433 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1434 {
1435   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1436 }
1437
1438 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1439 _mm_avg_epu8 (__m128i __A, __m128i __B)
1440 {
1441   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1442 }
1443
1444 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1445 _mm_avg_epu16 (__m128i __A, __m128i __B)
1446 {
1447   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1448 }
1449
1450 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1451 _mm_sad_epu8 (__m128i __A, __m128i __B)
1452 {
1453   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1454 }
1455
1456 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1457 _mm_stream_si32 (int *__A, int __B)
1458 {
1459   __builtin_ia32_movnti (__A, __B);
1460 }
1461
1462 #ifdef __x86_64__
1463 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1464 _mm_stream_si64 (long long int *__A, long long int __B)
1465 {
1466   __builtin_ia32_movnti64 (__A, __B);
1467 }
1468 #endif
1469
1470 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1471 _mm_stream_si128 (__m128i *__A, __m128i __B)
1472 {
1473   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1474 }
1475
1476 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1477 _mm_stream_pd (double *__A, __m128d __B)
1478 {
1479   __builtin_ia32_movntpd (__A, (__v2df)__B);
1480 }
1481
1482 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1483 _mm_clflush (void const *__A)
1484 {
1485   __builtin_ia32_clflush (__A);
1486 }
1487
1488 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1489 _mm_lfence (void)
1490 {
1491   __builtin_ia32_lfence ();
1492 }
1493
1494 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1495 _mm_mfence (void)
1496 {
1497   __builtin_ia32_mfence ();
1498 }
1499
1500 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1501 _mm_cvtsi32_si128 (int __A)
1502 {
1503   return _mm_set_epi32 (0, 0, 0, __A);
1504 }
1505
1506 #ifdef __x86_64__
1507 /* Intel intrinsic.  */
1508 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1509 _mm_cvtsi64_si128 (long long __A)
1510 {
1511   return _mm_set_epi64x (0, __A);
1512 }
1513
1514 /* Microsoft intrinsic.  */
1515 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1516 _mm_cvtsi64x_si128 (long long __A)
1517 {
1518   return _mm_set_epi64x (0, __A);
1519 }
1520 #endif
1521
1522 /* Casts between various SP, DP, INT vector types.  Note that these do no
1523    conversion of values, they just change the type.  */
1524 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1525 _mm_castpd_ps(__m128d __A)
1526 {
1527   return (__m128) __A;
1528 }
1529
1530 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1531 _mm_castpd_si128(__m128d __A)
1532 {
1533   return (__m128i) __A;
1534 }
1535
1536 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1537 _mm_castps_pd(__m128 __A)
1538 {
1539   return (__m128d) __A;
1540 }
1541
1542 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1543 _mm_castps_si128(__m128 __A)
1544 {
1545   return (__m128i) __A;
1546 }
1547
1548 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1549 _mm_castsi128_ps(__m128i __A)
1550 {
1551   return (__m128) __A;
1552 }
1553
1554 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__))
1555 _mm_castsi128_pd(__m128i __A)
1556 {
1557   return (__m128d) __A;
1558 }
1559
1560 #ifdef __DISABLE_SSE2__
1561 #undef __DISABLE_SSE2__
1562 #pragma GCC pop_options
1563 #endif /* __DISABLE_SSE2__ */
1564
1565 #endif /* _EMMINTRIN_H_INCLUDED */