gcc/config/i386/emmintrin.h

   1 /* Copyright (C) 2003, 2004, 2005, 2006, 2007 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 2, 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    You should have received a copy of the GNU General Public License
  16    along with GCC; see the file COPYING.  If not, write to
  17    the Free Software Foundation, 51 Franklin Street, Fifth Floor,
  18    Boston, MA 02110-1301, USA.  */
  19
  20 /* As a special exception, if you include this header file into source
  21    files compiled by GCC, this header file does not by itself cause
  22    the resulting executable to be covered by the GNU General Public
  23    License.  This exception does not however invalidate any other
  24    reasons why the executable file might be covered by the GNU General
  25    Public License.  */
  26
  27 /* Implemented from the specification included in the Intel C++ Compiler
  28    User Guide and Reference, version 9.0.  */
  29
  30 #ifndef _EMMINTRIN_H_INCLUDED
  31 #define _EMMINTRIN_H_INCLUDED
  32
  33 #ifndef __SSE2__
  34 # error "SSE2 instruction set not enabled"
  35 #else
  36
  37 /* We need definitions from the SSE header files*/
  38 #include <xmmintrin.h>
  39
  40 /* SSE2 */
  41 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  42 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  43 typedef int __v4si __attribute__ ((__vector_size__ (16)));
  44 typedef short __v8hi __attribute__ ((__vector_size__ (16)));
  45 typedef char __v16qi __attribute__ ((__vector_size__ (16)));
  46
  47 /* The Intel API is flexible enough that we must allow aliasing with other
  48    vector types, and their scalar components.  */
  49 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__));
  50 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__));
  51
  52 /* Create a selector for use with the SHUFPD instruction.  */
  53 #define _MM_SHUFFLE2(fp1,fp0) \
  54  (((fp1) << 1) | (fp0))
  55
  56 /* Create a vector with element 0 as F and the rest zero.  */
  57 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  58 _mm_set_sd (double __F)
  59 {
  60   return __extension__ (__m128d){ __F, 0.0 };
  61 }
  62
  63 /* Create a vector with both elements equal to F.  */
  64 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  65 _mm_set1_pd (double __F)
  66 {
  67   return __extension__ (__m128d){ __F, __F };
  68 }
  69
  70 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  71 _mm_set_pd1 (double __F)
  72 {
  73   return _mm_set1_pd (__F);
  74 }
  75
  76 /* Create a vector with the lower value X and upper value W.  */
  77 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  78 _mm_set_pd (double __W, double __X)
  79 {
  80   return __extension__ (__m128d){ __X, __W };
  81 }
  82
  83 /* Create a vector with the lower value W and upper value X.  */
  84 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  85 _mm_setr_pd (double __W, double __X)
  86 {
  87   return __extension__ (__m128d){ __W, __X };
  88 }
  89
  90 /* Create a vector of zeros.  */
  91 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  92 _mm_setzero_pd (void)
  93 {
  94   return __extension__ (__m128d){ 0.0, 0.0 };
  95 }
  96
  97 /* Sets the low DPFP value of A from the low value of B.  */
  98 static __inline __m128d __attribute__((__always_inline__, __artificial__))
  99 _mm_move_sd (__m128d __A, __m128d __B)
 100 {
 101   return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 102 }
 103
 104 /* Load two DPFP values from P.  The address must be 16-byte aligned.  */
 105 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 106 _mm_load_pd (double const *__P)
 107 {
 108   return *(__m128d *)__P;
 109 }
 110
 111 /* Load two DPFP values from P.  The address need not be 16-byte aligned.  */
 112 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 113 _mm_loadu_pd (double const *__P)
 114 {
 115   return __builtin_ia32_loadupd (__P);
 116 }
 117
 118 /* Create a vector with all two elements equal to *P.  */
 119 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 120 _mm_load1_pd (double const *__P)
 121 {
 122   return _mm_set1_pd (*__P);
 123 }
 124
 125 /* Create a vector with element 0 as *P and the rest zero.  */
 126 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 127 _mm_load_sd (double const *__P)
 128 {
 129   return _mm_set_sd (*__P);
 130 }
 131
 132 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 133 _mm_load_pd1 (double const *__P)
 134 {
 135   return _mm_load1_pd (__P);
 136 }
 137
 138 /* Load two DPFP values in reverse order.  The address must be aligned.  */
 139 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 140 _mm_loadr_pd (double const *__P)
 141 {
 142   __m128d __tmp = _mm_load_pd (__P);
 143   return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1));
 144 }
 145
 146 /* Store two DPFP values.  The address must be 16-byte aligned.  */
 147 static __inline void __attribute__((__always_inline__, __artificial__))
 148 _mm_store_pd (double *__P, __m128d __A)
 149 {
 150   *(__m128d *)__P = __A;
 151 }
 152
 153 /* Store two DPFP values.  The address need not be 16-byte aligned.  */
 154 static __inline void __attribute__((__always_inline__, __artificial__))
 155 _mm_storeu_pd (double *__P, __m128d __A)
 156 {
 157   __builtin_ia32_storeupd (__P, __A);
 158 }
 159
 160 /* Stores the lower DPFP value.  */
 161 static __inline void __attribute__((__always_inline__, __artificial__))
 162 _mm_store_sd (double *__P, __m128d __A)
 163 {
 164   *__P = __builtin_ia32_vec_ext_v2df (__A, 0);
 165 }
 166
 167 static __inline double __attribute__((__always_inline__, __artificial__))
 168 _mm_cvtsd_f64 (__m128d __A)
 169 {
 170   return __builtin_ia32_vec_ext_v2df (__A, 0);
 171 }
 172
 173 static __inline void __attribute__((__always_inline__, __artificial__))
 174 _mm_storel_pd (double *__P, __m128d __A)
 175 {
 176   _mm_store_sd (__P, __A);
 177 }
 178
 179 /* Stores the upper DPFP value.  */
 180 static __inline void __attribute__((__always_inline__, __artificial__))
 181 _mm_storeh_pd (double *__P, __m128d __A)
 182 {
 183   *__P = __builtin_ia32_vec_ext_v2df (__A, 1);
 184 }
 185
 186 /* Store the lower DPFP value across two words.
 187    The address must be 16-byte aligned.  */
 188 static __inline void __attribute__((__always_inline__, __artificial__))
 189 _mm_store1_pd (double *__P, __m128d __A)
 190 {
 191   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0)));
 192 }
 193
 194 static __inline void __attribute__((__always_inline__, __artificial__))
 195 _mm_store_pd1 (double *__P, __m128d __A)
 196 {
 197   _mm_store1_pd (__P, __A);
 198 }
 199
 200 /* Store two DPFP values in reverse order.  The address must be aligned.  */
 201 static __inline void __attribute__((__always_inline__, __artificial__))
 202 _mm_storer_pd (double *__P, __m128d __A)
 203 {
 204   _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1)));
 205 }
 206
 207 static __inline int __attribute__((__always_inline__, __artificial__))
 208 _mm_cvtsi128_si32 (__m128i __A)
 209 {
 210   return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0);
 211 }
 212
 213 #ifdef __x86_64__
 214 /* Intel intrinsic.  */
 215 static __inline long long __attribute__((__always_inline__, __artificial__))
 216 _mm_cvtsi128_si64 (__m128i __A)
 217 {
 218   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 219 }
 220
 221 /* Microsoft intrinsic.  */
 222 static __inline long long __attribute__((__always_inline__, __artificial__))
 223 _mm_cvtsi128_si64x (__m128i __A)
 224 {
 225   return __builtin_ia32_vec_ext_v2di ((__v2di)__A, 0);
 226 }
 227 #endif
 228
 229 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 230 _mm_add_pd (__m128d __A, __m128d __B)
 231 {
 232   return (__m128d)__builtin_ia32_addpd ((__v2df)__A, (__v2df)__B);
 233 }
 234
 235 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 236 _mm_add_sd (__m128d __A, __m128d __B)
 237 {
 238   return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B);
 239 }
 240
 241 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 242 _mm_sub_pd (__m128d __A, __m128d __B)
 243 {
 244   return (__m128d)__builtin_ia32_subpd ((__v2df)__A, (__v2df)__B);
 245 }
 246
 247 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 248 _mm_sub_sd (__m128d __A, __m128d __B)
 249 {
 250   return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B);
 251 }
 252
 253 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 254 _mm_mul_pd (__m128d __A, __m128d __B)
 255 {
 256   return (__m128d)__builtin_ia32_mulpd ((__v2df)__A, (__v2df)__B);
 257 }
 258
 259 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 260 _mm_mul_sd (__m128d __A, __m128d __B)
 261 {
 262   return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B);
 263 }
 264
 265 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 266 _mm_div_pd (__m128d __A, __m128d __B)
 267 {
 268   return (__m128d)__builtin_ia32_divpd ((__v2df)__A, (__v2df)__B);
 269 }
 270
 271 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 272 _mm_div_sd (__m128d __A, __m128d __B)
 273 {
 274   return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B);
 275 }
 276
 277 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 278 _mm_sqrt_pd (__m128d __A)
 279 {
 280   return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A);
 281 }
 282
 283 /* Return pair {sqrt (A[0), B[1]}.  */
 284 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 285 _mm_sqrt_sd (__m128d __A, __m128d __B)
 286 {
 287   __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B);
 288   return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp);
 289 }
 290
 291 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 292 _mm_min_pd (__m128d __A, __m128d __B)
 293 {
 294   return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B);
 295 }
 296
 297 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 298 _mm_min_sd (__m128d __A, __m128d __B)
 299 {
 300   return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B);
 301 }
 302
 303 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 304 _mm_max_pd (__m128d __A, __m128d __B)
 305 {
 306   return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B);
 307 }
 308
 309 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 310 _mm_max_sd (__m128d __A, __m128d __B)
 311 {
 312   return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B);
 313 }
 314
 315 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 316 _mm_and_pd (__m128d __A, __m128d __B)
 317 {
 318   return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B);
 319 }
 320
 321 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 322 _mm_andnot_pd (__m128d __A, __m128d __B)
 323 {
 324   return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B);
 325 }
 326
 327 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 328 _mm_or_pd (__m128d __A, __m128d __B)
 329 {
 330   return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B);
 331 }
 332
 333 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 334 _mm_xor_pd (__m128d __A, __m128d __B)
 335 {
 336   return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B);
 337 }
 338
 339 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 340 _mm_cmpeq_pd (__m128d __A, __m128d __B)
 341 {
 342   return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B);
 343 }
 344
 345 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 346 _mm_cmplt_pd (__m128d __A, __m128d __B)
 347 {
 348   return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B);
 349 }
 350
 351 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 352 _mm_cmple_pd (__m128d __A, __m128d __B)
 353 {
 354   return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B);
 355 }
 356
 357 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 358 _mm_cmpgt_pd (__m128d __A, __m128d __B)
 359 {
 360   return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B);
 361 }
 362
 363 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 364 _mm_cmpge_pd (__m128d __A, __m128d __B)
 365 {
 366   return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B);
 367 }
 368
 369 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 370 _mm_cmpneq_pd (__m128d __A, __m128d __B)
 371 {
 372   return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B);
 373 }
 374
 375 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 376 _mm_cmpnlt_pd (__m128d __A, __m128d __B)
 377 {
 378   return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B);
 379 }
 380
 381 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 382 _mm_cmpnle_pd (__m128d __A, __m128d __B)
 383 {
 384   return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B);
 385 }
 386
 387 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 388 _mm_cmpngt_pd (__m128d __A, __m128d __B)
 389 {
 390   return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B);
 391 }
 392
 393 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 394 _mm_cmpnge_pd (__m128d __A, __m128d __B)
 395 {
 396   return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B);
 397 }
 398
 399 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 400 _mm_cmpord_pd (__m128d __A, __m128d __B)
 401 {
 402   return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B);
 403 }
 404
 405 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 406 _mm_cmpunord_pd (__m128d __A, __m128d __B)
 407 {
 408   return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B);
 409 }
 410
 411 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 412 _mm_cmpeq_sd (__m128d __A, __m128d __B)
 413 {
 414   return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B);
 415 }
 416
 417 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 418 _mm_cmplt_sd (__m128d __A, __m128d __B)
 419 {
 420   return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B);
 421 }
 422
 423 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 424 _mm_cmple_sd (__m128d __A, __m128d __B)
 425 {
 426   return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B);
 427 }
 428
 429 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 430 _mm_cmpgt_sd (__m128d __A, __m128d __B)
 431 {
 432   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 433                                          (__v2df)
 434                                          __builtin_ia32_cmpltsd ((__v2df) __B,
 435                                                                  (__v2df)
 436                                                                  __A));
 437 }
 438
 439 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 440 _mm_cmpge_sd (__m128d __A, __m128d __B)
 441 {
 442   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 443                                          (__v2df)
 444                                          __builtin_ia32_cmplesd ((__v2df) __B,
 445                                                                  (__v2df)
 446                                                                  __A));
 447 }
 448
 449 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 450 _mm_cmpneq_sd (__m128d __A, __m128d __B)
 451 {
 452   return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B);
 453 }
 454
 455 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 456 _mm_cmpnlt_sd (__m128d __A, __m128d __B)
 457 {
 458   return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B);
 459 }
 460
 461 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 462 _mm_cmpnle_sd (__m128d __A, __m128d __B)
 463 {
 464   return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B);
 465 }
 466
 467 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 468 _mm_cmpngt_sd (__m128d __A, __m128d __B)
 469 {
 470   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 471                                          (__v2df)
 472                                          __builtin_ia32_cmpnltsd ((__v2df) __B,
 473                                                                   (__v2df)
 474                                                                   __A));
 475 }
 476
 477 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 478 _mm_cmpnge_sd (__m128d __A, __m128d __B)
 479 {
 480   return (__m128d) __builtin_ia32_movsd ((__v2df) __A,
 481                                          (__v2df)
 482                                          __builtin_ia32_cmpnlesd ((__v2df) __B,
 483                                                                   (__v2df)
 484                                                                   __A));
 485 }
 486
 487 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 488 _mm_cmpord_sd (__m128d __A, __m128d __B)
 489 {
 490   return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B);
 491 }
 492
 493 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 494 _mm_cmpunord_sd (__m128d __A, __m128d __B)
 495 {
 496   return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B);
 497 }
 498
 499 static __inline int __attribute__((__always_inline__, __artificial__))
 500 _mm_comieq_sd (__m128d __A, __m128d __B)
 501 {
 502   return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B);
 503 }
 504
 505 static __inline int __attribute__((__always_inline__, __artificial__))
 506 _mm_comilt_sd (__m128d __A, __m128d __B)
 507 {
 508   return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B);
 509 }
 510
 511 static __inline int __attribute__((__always_inline__, __artificial__))
 512 _mm_comile_sd (__m128d __A, __m128d __B)
 513 {
 514   return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B);
 515 }
 516
 517 static __inline int __attribute__((__always_inline__, __artificial__))
 518 _mm_comigt_sd (__m128d __A, __m128d __B)
 519 {
 520   return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B);
 521 }
 522
 523 static __inline int __attribute__((__always_inline__, __artificial__))
 524 _mm_comige_sd (__m128d __A, __m128d __B)
 525 {
 526   return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B);
 527 }
 528
 529 static __inline int __attribute__((__always_inline__, __artificial__))
 530 _mm_comineq_sd (__m128d __A, __m128d __B)
 531 {
 532   return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B);
 533 }
 534
 535 static __inline int __attribute__((__always_inline__, __artificial__))
 536 _mm_ucomieq_sd (__m128d __A, __m128d __B)
 537 {
 538   return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B);
 539 }
 540
 541 static __inline int __attribute__((__always_inline__, __artificial__))
 542 _mm_ucomilt_sd (__m128d __A, __m128d __B)
 543 {
 544   return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B);
 545 }
 546
 547 static __inline int __attribute__((__always_inline__, __artificial__))
 548 _mm_ucomile_sd (__m128d __A, __m128d __B)
 549 {
 550   return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B);
 551 }
 552
 553 static __inline int __attribute__((__always_inline__, __artificial__))
 554 _mm_ucomigt_sd (__m128d __A, __m128d __B)
 555 {
 556   return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B);
 557 }
 558
 559 static __inline int __attribute__((__always_inline__, __artificial__))
 560 _mm_ucomige_sd (__m128d __A, __m128d __B)
 561 {
 562   return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B);
 563 }
 564
 565 static __inline int __attribute__((__always_inline__, __artificial__))
 566 _mm_ucomineq_sd (__m128d __A, __m128d __B)
 567 {
 568   return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B);
 569 }
 570
 571 /* Create a vector of Qi, where i is the element number.  */
 572
 573 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 574 _mm_set_epi64x (long long __q1, long long __q0)
 575 {
 576   return __extension__ (__m128i)(__v2di){ __q0, __q1 };
 577 }
 578
 579 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 580 _mm_set_epi64 (__m64 __q1,  __m64 __q0)
 581 {
 582   return _mm_set_epi64x ((long long)__q1, (long long)__q0);
 583 }
 584
 585 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 586 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0)
 587 {
 588   return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 };
 589 }
 590
 591 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 592 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4,
 593                short __q3, short __q2, short __q1, short __q0)
 594 {
 595   return __extension__ (__m128i)(__v8hi){
 596     __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 };
 597 }
 598
 599 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 600 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12,
 601               char __q11, char __q10, char __q09, char __q08,
 602               char __q07, char __q06, char __q05, char __q04,
 603               char __q03, char __q02, char __q01, char __q00)
 604 {
 605   return __extension__ (__m128i)(__v16qi){
 606     __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07,
 607     __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15
 608   };
 609 }
 610
 611 /* Set all of the elements of the vector to A.  */
 612
 613 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 614 _mm_set1_epi64x (long long __A)
 615 {
 616   return _mm_set_epi64x (__A, __A);
 617 }
 618
 619 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 620 _mm_set1_epi64 (__m64 __A)
 621 {
 622   return _mm_set_epi64 (__A, __A);
 623 }
 624
 625 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 626 _mm_set1_epi32 (int __A)
 627 {
 628   return _mm_set_epi32 (__A, __A, __A, __A);
 629 }
 630
 631 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 632 _mm_set1_epi16 (short __A)
 633 {
 634   return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A);
 635 }
 636
 637 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 638 _mm_set1_epi8 (char __A)
 639 {
 640   return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
 641                        __A, __A, __A, __A, __A, __A, __A, __A);
 642 }
 643
 644 /* Create a vector of Qi, where i is the element number.
 645    The parameter order is reversed from the _mm_set_epi* functions.  */
 646
 647 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 648 _mm_setr_epi64 (__m64 __q0, __m64 __q1)
 649 {
 650   return _mm_set_epi64 (__q1, __q0);
 651 }
 652
 653 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 654 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3)
 655 {
 656   return _mm_set_epi32 (__q3, __q2, __q1, __q0);
 657 }
 658
 659 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 660 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3,
 661                 short __q4, short __q5, short __q6, short __q7)
 662 {
 663   return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0);
 664 }
 665
 666 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 667 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03,
 668                char __q04, char __q05, char __q06, char __q07,
 669                char __q08, char __q09, char __q10, char __q11,
 670                char __q12, char __q13, char __q14, char __q15)
 671 {
 672   return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08,
 673                        __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00);
 674 }
 675
 676 /* Create a vector with element 0 as *P and the rest zero.  */
 677
 678 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 679 _mm_load_si128 (__m128i const *__P)
 680 {
 681   return *__P;
 682 }
 683
 684 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 685 _mm_loadu_si128 (__m128i const *__P)
 686 {
 687   return (__m128i) __builtin_ia32_loaddqu ((char const *)__P);
 688 }
 689
 690 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 691 _mm_loadl_epi64 (__m128i const *__P)
 692 {
 693   return _mm_set_epi64 ((__m64)0LL, *(__m64 *)__P);
 694 }
 695
 696 static __inline void __attribute__((__always_inline__, __artificial__))
 697 _mm_store_si128 (__m128i *__P, __m128i __B)
 698 {
 699   *__P = __B;
 700 }
 701
 702 static __inline void __attribute__((__always_inline__, __artificial__))
 703 _mm_storeu_si128 (__m128i *__P, __m128i __B)
 704 {
 705   __builtin_ia32_storedqu ((char *)__P, (__v16qi)__B);
 706 }
 707
 708 static __inline void __attribute__((__always_inline__, __artificial__))
 709 _mm_storel_epi64 (__m128i *__P, __m128i __B)
 710 {
 711   *(long long *)__P = __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 712 }
 713
 714 static __inline __m64 __attribute__((__always_inline__, __artificial__))
 715 _mm_movepi64_pi64 (__m128i __B)
 716 {
 717   return (__m64) __builtin_ia32_vec_ext_v2di ((__v2di)__B, 0);
 718 }
 719
 720 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 721 _mm_movpi64_epi64 (__m64 __A)
 722 {
 723   return _mm_set_epi64 ((__m64)0LL, __A);
 724 }
 725
 726 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 727 _mm_move_epi64 (__m128i __A)
 728 {
 729   return _mm_set_epi64 ((__m64)0LL, _mm_movepi64_pi64 (__A));
 730 }
 731
 732 /* Create a vector of zeros.  */
 733 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 734 _mm_setzero_si128 (void)
 735 {
 736   return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 };
 737 }
 738
 739 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 740 _mm_cvtepi32_pd (__m128i __A)
 741 {
 742   return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A);
 743 }
 744
 745 static __inline __m128 __attribute__((__always_inline__, __artificial__))
 746 _mm_cvtepi32_ps (__m128i __A)
 747 {
 748   return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A);
 749 }
 750
 751 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 752 _mm_cvtpd_epi32 (__m128d __A)
 753 {
 754   return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A);
 755 }
 756
 757 static __inline __m64 __attribute__((__always_inline__, __artificial__))
 758 _mm_cvtpd_pi32 (__m128d __A)
 759 {
 760   return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A);
 761 }
 762
 763 static __inline __m128 __attribute__((__always_inline__, __artificial__))
 764 _mm_cvtpd_ps (__m128d __A)
 765 {
 766   return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A);
 767 }
 768
 769 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 770 _mm_cvttpd_epi32 (__m128d __A)
 771 {
 772   return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A);
 773 }
 774
 775 static __inline __m64 __attribute__((__always_inline__, __artificial__))
 776 _mm_cvttpd_pi32 (__m128d __A)
 777 {
 778   return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A);
 779 }
 780
 781 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 782 _mm_cvtpi32_pd (__m64 __A)
 783 {
 784   return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A);
 785 }
 786
 787 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 788 _mm_cvtps_epi32 (__m128 __A)
 789 {
 790   return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A);
 791 }
 792
 793 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 794 _mm_cvttps_epi32 (__m128 __A)
 795 {
 796   return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A);
 797 }
 798
 799 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 800 _mm_cvtps_pd (__m128 __A)
 801 {
 802   return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A);
 803 }
 804
 805 static __inline int __attribute__((__always_inline__, __artificial__))
 806 _mm_cvtsd_si32 (__m128d __A)
 807 {
 808   return __builtin_ia32_cvtsd2si ((__v2df) __A);
 809 }
 810
 811 #ifdef __x86_64__
 812 /* Intel intrinsic.  */
 813 static __inline long long __attribute__((__always_inline__, __artificial__))
 814 _mm_cvtsd_si64 (__m128d __A)
 815 {
 816   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 817 }
 818
 819 /* Microsoft intrinsic.  */
 820 static __inline long long __attribute__((__always_inline__, __artificial__))
 821 _mm_cvtsd_si64x (__m128d __A)
 822 {
 823   return __builtin_ia32_cvtsd2si64 ((__v2df) __A);
 824 }
 825 #endif
 826
 827 static __inline int __attribute__((__always_inline__, __artificial__))
 828 _mm_cvttsd_si32 (__m128d __A)
 829 {
 830   return __builtin_ia32_cvttsd2si ((__v2df) __A);
 831 }
 832
 833 #ifdef __x86_64__
 834 /* Intel intrinsic.  */
 835 static __inline long long __attribute__((__always_inline__, __artificial__))
 836 _mm_cvttsd_si64 (__m128d __A)
 837 {
 838   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 839 }
 840
 841 /* Microsoft intrinsic.  */
 842 static __inline long long __attribute__((__always_inline__, __artificial__))
 843 _mm_cvttsd_si64x (__m128d __A)
 844 {
 845   return __builtin_ia32_cvttsd2si64 ((__v2df) __A);
 846 }
 847 #endif
 848
 849 static __inline __m128 __attribute__((__always_inline__, __artificial__))
 850 _mm_cvtsd_ss (__m128 __A, __m128d __B)
 851 {
 852   return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B);
 853 }
 854
 855 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 856 _mm_cvtsi32_sd (__m128d __A, int __B)
 857 {
 858   return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B);
 859 }
 860
 861 #ifdef __x86_64__
 862 /* Intel intrinsic.  */
 863 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 864 _mm_cvtsi64_sd (__m128d __A, long long __B)
 865 {
 866   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 867 }
 868
 869 /* Microsoft intrinsic.  */
 870 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 871 _mm_cvtsi64x_sd (__m128d __A, long long __B)
 872 {
 873   return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B);
 874 }
 875 #endif
 876
 877 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 878 _mm_cvtss_sd (__m128d __A, __m128 __B)
 879 {
 880   return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B);
 881 }
 882
 883 #ifdef __OPTIMIZE__
 884 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 885 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask)
 886 {
 887   return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask);
 888 }
 889 #else
 890 #define _mm_shuffle_pd(A, B, N)                                         \
 891   ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A),                \
 892                                    (__v2df)(__m128d)(B), (int)(N)))
 893 #endif
 894
 895 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 896 _mm_unpackhi_pd (__m128d __A, __m128d __B)
 897 {
 898   return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B);
 899 }
 900
 901 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 902 _mm_unpacklo_pd (__m128d __A, __m128d __B)
 903 {
 904   return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B);
 905 }
 906
 907 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 908 _mm_loadh_pd (__m128d __A, double const *__B)
 909 {
 910   return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B);
 911 }
 912
 913 static __inline __m128d __attribute__((__always_inline__, __artificial__))
 914 _mm_loadl_pd (__m128d __A, double const *__B)
 915 {
 916   return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B);
 917 }
 918
 919 static __inline int __attribute__((__always_inline__, __artificial__))
 920 _mm_movemask_pd (__m128d __A)
 921 {
 922   return __builtin_ia32_movmskpd ((__v2df)__A);
 923 }
 924
 925 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 926 _mm_packs_epi16 (__m128i __A, __m128i __B)
 927 {
 928   return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B);
 929 }
 930
 931 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 932 _mm_packs_epi32 (__m128i __A, __m128i __B)
 933 {
 934   return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B);
 935 }
 936
 937 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 938 _mm_packus_epi16 (__m128i __A, __m128i __B)
 939 {
 940   return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B);
 941 }
 942
 943 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 944 _mm_unpackhi_epi8 (__m128i __A, __m128i __B)
 945 {
 946   return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B);
 947 }
 948
 949 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 950 _mm_unpackhi_epi16 (__m128i __A, __m128i __B)
 951 {
 952   return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B);
 953 }
 954
 955 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 956 _mm_unpackhi_epi32 (__m128i __A, __m128i __B)
 957 {
 958   return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B);
 959 }
 960
 961 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 962 _mm_unpackhi_epi64 (__m128i __A, __m128i __B)
 963 {
 964   return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B);
 965 }
 966
 967 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 968 _mm_unpacklo_epi8 (__m128i __A, __m128i __B)
 969 {
 970   return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B);
 971 }
 972
 973 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 974 _mm_unpacklo_epi16 (__m128i __A, __m128i __B)
 975 {
 976   return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B);
 977 }
 978
 979 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 980 _mm_unpacklo_epi32 (__m128i __A, __m128i __B)
 981 {
 982   return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B);
 983 }
 984
 985 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 986 _mm_unpacklo_epi64 (__m128i __A, __m128i __B)
 987 {
 988   return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B);
 989 }
 990
 991 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 992 _mm_add_epi8 (__m128i __A, __m128i __B)
 993 {
 994   return (__m128i)__builtin_ia32_paddb128 ((__v16qi)__A, (__v16qi)__B);
 995 }
 996
 997 static __inline __m128i __attribute__((__always_inline__, __artificial__))
 998 _mm_add_epi16 (__m128i __A, __m128i __B)
 999 {
1000   return (__m128i)__builtin_ia32_paddw128 ((__v8hi)__A, (__v8hi)__B);
1001 }
1002
1003 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1004 _mm_add_epi32 (__m128i __A, __m128i __B)
1005 {
1006   return (__m128i)__builtin_ia32_paddd128 ((__v4si)__A, (__v4si)__B);
1007 }
1008
1009 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1010 _mm_add_epi64 (__m128i __A, __m128i __B)
1011 {
1012   return (__m128i)__builtin_ia32_paddq128 ((__v2di)__A, (__v2di)__B);
1013 }
1014
1015 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1016 _mm_adds_epi8 (__m128i __A, __m128i __B)
1017 {
1018   return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B);
1019 }
1020
1021 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1022 _mm_adds_epi16 (__m128i __A, __m128i __B)
1023 {
1024   return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B);
1025 }
1026
1027 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1028 _mm_adds_epu8 (__m128i __A, __m128i __B)
1029 {
1030   return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B);
1031 }
1032
1033 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1034 _mm_adds_epu16 (__m128i __A, __m128i __B)
1035 {
1036   return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B);
1037 }
1038
1039 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1040 _mm_sub_epi8 (__m128i __A, __m128i __B)
1041 {
1042   return (__m128i)__builtin_ia32_psubb128 ((__v16qi)__A, (__v16qi)__B);
1043 }
1044
1045 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1046 _mm_sub_epi16 (__m128i __A, __m128i __B)
1047 {
1048   return (__m128i)__builtin_ia32_psubw128 ((__v8hi)__A, (__v8hi)__B);
1049 }
1050
1051 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1052 _mm_sub_epi32 (__m128i __A, __m128i __B)
1053 {
1054   return (__m128i)__builtin_ia32_psubd128 ((__v4si)__A, (__v4si)__B);
1055 }
1056
1057 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1058 _mm_sub_epi64 (__m128i __A, __m128i __B)
1059 {
1060   return (__m128i)__builtin_ia32_psubq128 ((__v2di)__A, (__v2di)__B);
1061 }
1062
1063 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1064 _mm_subs_epi8 (__m128i __A, __m128i __B)
1065 {
1066   return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B);
1067 }
1068
1069 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1070 _mm_subs_epi16 (__m128i __A, __m128i __B)
1071 {
1072   return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B);
1073 }
1074
1075 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1076 _mm_subs_epu8 (__m128i __A, __m128i __B)
1077 {
1078   return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B);
1079 }
1080
1081 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1082 _mm_subs_epu16 (__m128i __A, __m128i __B)
1083 {
1084   return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B);
1085 }
1086
1087 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1088 _mm_madd_epi16 (__m128i __A, __m128i __B)
1089 {
1090   return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B);
1091 }
1092
1093 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1094 _mm_mulhi_epi16 (__m128i __A, __m128i __B)
1095 {
1096   return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B);
1097 }
1098
1099 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1100 _mm_mullo_epi16 (__m128i __A, __m128i __B)
1101 {
1102   return (__m128i)__builtin_ia32_pmullw128 ((__v8hi)__A, (__v8hi)__B);
1103 }
1104
1105 static __inline __m64 __attribute__((__always_inline__, __artificial__))
1106 _mm_mul_su32 (__m64 __A, __m64 __B)
1107 {
1108   return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B);
1109 }
1110
1111 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1112 _mm_mul_epu32 (__m128i __A, __m128i __B)
1113 {
1114   return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B);
1115 }
1116
1117 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1118 _mm_slli_epi16 (__m128i __A, int __B)
1119 {
1120   return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B);
1121 }
1122
1123 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1124 _mm_slli_epi32 (__m128i __A, int __B)
1125 {
1126   return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B);
1127 }
1128
1129 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1130 _mm_slli_epi64 (__m128i __A, int __B)
1131 {
1132   return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B);
1133 }
1134
1135 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1136 _mm_srai_epi16 (__m128i __A, int __B)
1137 {
1138   return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B);
1139 }
1140
1141 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1142 _mm_srai_epi32 (__m128i __A, int __B)
1143 {
1144   return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B);
1145 }
1146
1147 #ifdef __OPTIMIZE__
1148 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1149 _mm_srli_si128 (__m128i __A, const int __N)
1150 {
1151   return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8);
1152 }
1153
1154 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1155 _mm_slli_si128 (__m128i __A, const int __N)
1156 {
1157   return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8);
1158 }
1159 #else
1160 #define _mm_srli_si128(A, N) \
1161   ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8))
1162 #define _mm_slli_si128(A, N) \
1163   ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8))
1164 #endif
1165
1166 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1167 _mm_srli_epi16 (__m128i __A, int __B)
1168 {
1169   return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B);
1170 }
1171
1172 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1173 _mm_srli_epi32 (__m128i __A, int __B)
1174 {
1175   return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B);
1176 }
1177
1178 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1179 _mm_srli_epi64 (__m128i __A, int __B)
1180 {
1181   return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B);
1182 }
1183
1184 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1185 _mm_sll_epi16 (__m128i __A, __m128i __B)
1186 {
1187   return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B);
1188 }
1189
1190 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1191 _mm_sll_epi32 (__m128i __A, __m128i __B)
1192 {
1193   return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B);
1194 }
1195
1196 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1197 _mm_sll_epi64 (__m128i __A, __m128i __B)
1198 {
1199   return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B);
1200 }
1201
1202 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1203 _mm_sra_epi16 (__m128i __A, __m128i __B)
1204 {
1205   return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B);
1206 }
1207
1208 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1209 _mm_sra_epi32 (__m128i __A, __m128i __B)
1210 {
1211   return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B);
1212 }
1213
1214 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1215 _mm_srl_epi16 (__m128i __A, __m128i __B)
1216 {
1217   return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B);
1218 }
1219
1220 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1221 _mm_srl_epi32 (__m128i __A, __m128i __B)
1222 {
1223   return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B);
1224 }
1225
1226 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1227 _mm_srl_epi64 (__m128i __A, __m128i __B)
1228 {
1229   return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B);
1230 }
1231
1232 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1233 _mm_and_si128 (__m128i __A, __m128i __B)
1234 {
1235   return (__m128i)__builtin_ia32_pand128 ((__v2di)__A, (__v2di)__B);
1236 }
1237
1238 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1239 _mm_andnot_si128 (__m128i __A, __m128i __B)
1240 {
1241   return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B);
1242 }
1243
1244 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1245 _mm_or_si128 (__m128i __A, __m128i __B)
1246 {
1247   return (__m128i)__builtin_ia32_por128 ((__v2di)__A, (__v2di)__B);
1248 }
1249
1250 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1251 _mm_xor_si128 (__m128i __A, __m128i __B)
1252 {
1253   return (__m128i)__builtin_ia32_pxor128 ((__v2di)__A, (__v2di)__B);
1254 }
1255
1256 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1257 _mm_cmpeq_epi8 (__m128i __A, __m128i __B)
1258 {
1259   return (__m128i)__builtin_ia32_pcmpeqb128 ((__v16qi)__A, (__v16qi)__B);
1260 }
1261
1262 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1263 _mm_cmpeq_epi16 (__m128i __A, __m128i __B)
1264 {
1265   return (__m128i)__builtin_ia32_pcmpeqw128 ((__v8hi)__A, (__v8hi)__B);
1266 }
1267
1268 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1269 _mm_cmpeq_epi32 (__m128i __A, __m128i __B)
1270 {
1271   return (__m128i)__builtin_ia32_pcmpeqd128 ((__v4si)__A, (__v4si)__B);
1272 }
1273
1274 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1275 _mm_cmplt_epi8 (__m128i __A, __m128i __B)
1276 {
1277   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__B, (__v16qi)__A);
1278 }
1279
1280 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1281 _mm_cmplt_epi16 (__m128i __A, __m128i __B)
1282 {
1283   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__B, (__v8hi)__A);
1284 }
1285
1286 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1287 _mm_cmplt_epi32 (__m128i __A, __m128i __B)
1288 {
1289   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__B, (__v4si)__A);
1290 }
1291
1292 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1293 _mm_cmpgt_epi8 (__m128i __A, __m128i __B)
1294 {
1295   return (__m128i)__builtin_ia32_pcmpgtb128 ((__v16qi)__A, (__v16qi)__B);
1296 }
1297
1298 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1299 _mm_cmpgt_epi16 (__m128i __A, __m128i __B)
1300 {
1301   return (__m128i)__builtin_ia32_pcmpgtw128 ((__v8hi)__A, (__v8hi)__B);
1302 }
1303
1304 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1305 _mm_cmpgt_epi32 (__m128i __A, __m128i __B)
1306 {
1307   return (__m128i)__builtin_ia32_pcmpgtd128 ((__v4si)__A, (__v4si)__B);
1308 }
1309
1310 #ifdef __OPTIMIZE__
1311 static __inline int __attribute__((__always_inline__, __artificial__))
1312 _mm_extract_epi16 (__m128i const __A, int const __N)
1313 {
1314   return __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N);
1315 }
1316
1317 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1318 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N)
1319 {
1320   return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N);
1321 }
1322 #else
1323 #define _mm_extract_epi16(A, N) \
1324   ((int) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N)))
1325 #define _mm_insert_epi16(A, D, N)                               \
1326   ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \
1327                                           (int)(D), (int)(N)))
1328 #endif
1329
1330 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1331 _mm_max_epi16 (__m128i __A, __m128i __B)
1332 {
1333   return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B);
1334 }
1335
1336 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1337 _mm_max_epu8 (__m128i __A, __m128i __B)
1338 {
1339   return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B);
1340 }
1341
1342 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1343 _mm_min_epi16 (__m128i __A, __m128i __B)
1344 {
1345   return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B);
1346 }
1347
1348 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1349 _mm_min_epu8 (__m128i __A, __m128i __B)
1350 {
1351   return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B);
1352 }
1353
1354 static __inline int __attribute__((__always_inline__, __artificial__))
1355 _mm_movemask_epi8 (__m128i __A)
1356 {
1357   return __builtin_ia32_pmovmskb128 ((__v16qi)__A);
1358 }
1359
1360 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1361 _mm_mulhi_epu16 (__m128i __A, __m128i __B)
1362 {
1363   return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B);
1364 }
1365
1366 #ifdef __OPTIMIZE__
1367 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1368 _mm_shufflehi_epi16 (__m128i __A, const int __mask)
1369 {
1370   return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask);
1371 }
1372
1373 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1374 _mm_shufflelo_epi16 (__m128i __A, const int __mask)
1375 {
1376   return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask);
1377 }
1378
1379 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1380 _mm_shuffle_epi32 (__m128i __A, const int __mask)
1381 {
1382   return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask);
1383 }
1384 #else
1385 #define _mm_shufflehi_epi16(A, N) \
1386   ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N)))
1387 #define _mm_shufflelo_epi16(A, N) \
1388   ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N)))
1389 #define _mm_shuffle_epi32(A, N) \
1390   ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N)))
1391 #endif
1392
1393 static __inline void __attribute__((__always_inline__, __artificial__))
1394 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C)
1395 {
1396   __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C);
1397 }
1398
1399 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1400 _mm_avg_epu8 (__m128i __A, __m128i __B)
1401 {
1402   return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B);
1403 }
1404
1405 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1406 _mm_avg_epu16 (__m128i __A, __m128i __B)
1407 {
1408   return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B);
1409 }
1410
1411 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1412 _mm_sad_epu8 (__m128i __A, __m128i __B)
1413 {
1414   return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B);
1415 }
1416
1417 static __inline void __attribute__((__always_inline__, __artificial__))
1418 _mm_stream_si32 (int *__A, int __B)
1419 {
1420   __builtin_ia32_movnti (__A, __B);
1421 }
1422
1423 static __inline void __attribute__((__always_inline__, __artificial__))
1424 _mm_stream_si128 (__m128i *__A, __m128i __B)
1425 {
1426   __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B);
1427 }
1428
1429 static __inline void __attribute__((__always_inline__, __artificial__))
1430 _mm_stream_pd (double *__A, __m128d __B)
1431 {
1432   __builtin_ia32_movntpd (__A, (__v2df)__B);
1433 }
1434
1435 static __inline void __attribute__((__always_inline__, __artificial__))
1436 _mm_clflush (void const *__A)
1437 {
1438   __builtin_ia32_clflush (__A);
1439 }
1440
1441 static __inline void __attribute__((__always_inline__, __artificial__))
1442 _mm_lfence (void)
1443 {
1444   __builtin_ia32_lfence ();
1445 }
1446
1447 static __inline void __attribute__((__always_inline__, __artificial__))
1448 _mm_mfence (void)
1449 {
1450   __builtin_ia32_mfence ();
1451 }
1452
1453 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1454 _mm_cvtsi32_si128 (int __A)
1455 {
1456   return _mm_set_epi32 (0, 0, 0, __A);
1457 }
1458
1459 #ifdef __x86_64__
1460 /* Intel intrinsic.  */
1461 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1462 _mm_cvtsi64_si128 (long long __A)
1463 {
1464   return _mm_set_epi64x (0, __A);
1465 }
1466
1467 /* Microsoft intrinsic.  */
1468 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1469 _mm_cvtsi64x_si128 (long long __A)
1470 {
1471   return _mm_set_epi64x (0, __A);
1472 }
1473 #endif
1474
1475 /* Casts between various SP, DP, INT vector types.  Note that these do no
1476    conversion of values, they just change the type.  */
1477 static __inline __m128 __attribute__((__always_inline__, __artificial__))
1478 _mm_castpd_ps(__m128d __A)
1479 {
1480   return (__m128) __A;
1481 }
1482
1483 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1484 _mm_castpd_si128(__m128d __A)
1485 {
1486   return (__m128i) __A;
1487 }
1488
1489 static __inline __m128d __attribute__((__always_inline__, __artificial__))
1490 _mm_castps_pd(__m128 __A)
1491 {
1492   return (__m128d) __A;
1493 }
1494
1495 static __inline __m128i __attribute__((__always_inline__, __artificial__))
1496 _mm_castps_si128(__m128 __A)
1497 {
1498   return (__m128i) __A;
1499 }
1500
1501 static __inline __m128 __attribute__((__always_inline__, __artificial__))
1502 _mm_castsi128_ps(__m128i __A)
1503 {
1504   return (__m128) __A;
1505 }
1506
1507 static __inline __m128d __attribute__((__always_inline__, __artificial__))
1508 _mm_castsi128_pd(__m128i __A)
1509 {
1510   return (__m128d) __A;
1511 }
1512
1513 #endif /* __SSE2__  */
1514
1515 #endif /* _EMMINTRIN_H_INCLUDED */