external/bsd/llvm/dist/clang/lib/Headers/emmintrin.h

   1 /*===---- emmintrin.h - SSE2 intrinsics ------------------------------------===
   2  *
   3  * Permission is hereby granted, free of charge, to any person obtaining a copy
   4  * of this software and associated documentation files (the "Software"), to deal
   5  * in the Software without restriction, including without limitation the rights
   6  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
   7  * copies of the Software, and to permit persons to whom the Software is
   8  * furnished to do so, subject to the following conditions:
   9  *
  10  * The above copyright notice and this permission notice shall be included in
  11  * all copies or substantial portions of the Software.
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  16  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  17  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  18  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  19  * THE SOFTWARE.
  20  *
  21  *===-----------------------------------------------------------------------===
  22  */
  23
  24 #ifndef __EMMINTRIN_H
  25 #define __EMMINTRIN_H
  26
  27 #ifndef __SSE2__
  28 #error "SSE2 instruction set not enabled"
  29 #else
  30
  31 #include <xmmintrin.h>
  32
  33 typedef double __m128d __attribute__((__vector_size__(16)));
  34 typedef long long __m128i __attribute__((__vector_size__(16)));
  35
  36 /* Type defines.  */
  37 typedef double __v2df __attribute__ ((__vector_size__ (16)));
  38 typedef long long __v2di __attribute__ ((__vector_size__ (16)));
  39 typedef short __v8hi __attribute__((__vector_size__(16)));
  40 typedef char __v16qi __attribute__((__vector_size__(16)));
  41
  42 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  43 _mm_add_sd(__m128d __a, __m128d __b)
  44 {
  45   __a[0] += __b[0];
  46   return __a;
  47 }
  48
  49 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  50 _mm_add_pd(__m128d __a, __m128d __b)
  51 {
  52   return __a + __b;
  53 }
  54
  55 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  56 _mm_sub_sd(__m128d __a, __m128d __b)
  57 {
  58   __a[0] -= __b[0];
  59   return __a;
  60 }
  61
  62 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  63 _mm_sub_pd(__m128d __a, __m128d __b)
  64 {
  65   return __a - __b;
  66 }
  67
  68 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  69 _mm_mul_sd(__m128d __a, __m128d __b)
  70 {
  71   __a[0] *= __b[0];
  72   return __a;
  73 }
  74
  75 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  76 _mm_mul_pd(__m128d __a, __m128d __b)
  77 {
  78   return __a * __b;
  79 }
  80
  81 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  82 _mm_div_sd(__m128d __a, __m128d __b)
  83 {
  84   __a[0] /= __b[0];
  85   return __a;
  86 }
  87
  88 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  89 _mm_div_pd(__m128d __a, __m128d __b)
  90 {
  91   return __a / __b;
  92 }
  93
  94 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
  95 _mm_sqrt_sd(__m128d __a, __m128d __b)
  96 {
  97   __m128d __c = __builtin_ia32_sqrtsd(__b);
  98   return (__m128d) { __c[0], __a[1] };
  99 }
 100
 101 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 102 _mm_sqrt_pd(__m128d __a)
 103 {
 104   return __builtin_ia32_sqrtpd(__a);
 105 }
 106
 107 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 108 _mm_min_sd(__m128d __a, __m128d __b)
 109 {
 110   return __builtin_ia32_minsd(__a, __b);
 111 }
 112
 113 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 114 _mm_min_pd(__m128d __a, __m128d __b)
 115 {
 116   return __builtin_ia32_minpd(__a, __b);
 117 }
 118
 119 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 120 _mm_max_sd(__m128d __a, __m128d __b)
 121 {
 122   return __builtin_ia32_maxsd(__a, __b);
 123 }
 124
 125 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 126 _mm_max_pd(__m128d __a, __m128d __b)
 127 {
 128   return __builtin_ia32_maxpd(__a, __b);
 129 }
 130
 131 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 132 _mm_and_pd(__m128d __a, __m128d __b)
 133 {
 134   return (__m128d)((__v4si)__a & (__v4si)__b);
 135 }
 136
 137 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 138 _mm_andnot_pd(__m128d __a, __m128d __b)
 139 {
 140   return (__m128d)(~(__v4si)__a & (__v4si)__b);
 141 }
 142
 143 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 144 _mm_or_pd(__m128d __a, __m128d __b)
 145 {
 146   return (__m128d)((__v4si)__a | (__v4si)__b);
 147 }
 148
 149 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 150 _mm_xor_pd(__m128d __a, __m128d __b)
 151 {
 152   return (__m128d)((__v4si)__a ^ (__v4si)__b);
 153 }
 154
 155 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 156 _mm_cmpeq_pd(__m128d __a, __m128d __b)
 157 {
 158   return (__m128d)__builtin_ia32_cmpeqpd(__a, __b);
 159 }
 160
 161 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 162 _mm_cmplt_pd(__m128d __a, __m128d __b)
 163 {
 164   return (__m128d)__builtin_ia32_cmpltpd(__a, __b);
 165 }
 166
 167 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 168 _mm_cmple_pd(__m128d __a, __m128d __b)
 169 {
 170   return (__m128d)__builtin_ia32_cmplepd(__a, __b);
 171 }
 172
 173 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 174 _mm_cmpgt_pd(__m128d __a, __m128d __b)
 175 {
 176   return (__m128d)__builtin_ia32_cmpltpd(__b, __a);
 177 }
 178
 179 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 180 _mm_cmpge_pd(__m128d __a, __m128d __b)
 181 {
 182   return (__m128d)__builtin_ia32_cmplepd(__b, __a);
 183 }
 184
 185 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 186 _mm_cmpord_pd(__m128d __a, __m128d __b)
 187 {
 188   return (__m128d)__builtin_ia32_cmpordpd(__a, __b);
 189 }
 190
 191 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 192 _mm_cmpunord_pd(__m128d __a, __m128d __b)
 193 {
 194   return (__m128d)__builtin_ia32_cmpunordpd(__a, __b);
 195 }
 196
 197 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 198 _mm_cmpneq_pd(__m128d __a, __m128d __b)
 199 {
 200   return (__m128d)__builtin_ia32_cmpneqpd(__a, __b);
 201 }
 202
 203 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 204 _mm_cmpnlt_pd(__m128d __a, __m128d __b)
 205 {
 206   return (__m128d)__builtin_ia32_cmpnltpd(__a, __b);
 207 }
 208
 209 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 210 _mm_cmpnle_pd(__m128d __a, __m128d __b)
 211 {
 212   return (__m128d)__builtin_ia32_cmpnlepd(__a, __b);
 213 }
 214
 215 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 216 _mm_cmpngt_pd(__m128d __a, __m128d __b)
 217 {
 218   return (__m128d)__builtin_ia32_cmpnltpd(__b, __a);
 219 }
 220
 221 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 222 _mm_cmpnge_pd(__m128d __a, __m128d __b)
 223 {
 224   return (__m128d)__builtin_ia32_cmpnlepd(__b, __a);
 225 }
 226
 227 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 228 _mm_cmpeq_sd(__m128d __a, __m128d __b)
 229 {
 230   return (__m128d)__builtin_ia32_cmpeqsd(__a, __b);
 231 }
 232
 233 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 234 _mm_cmplt_sd(__m128d __a, __m128d __b)
 235 {
 236   return (__m128d)__builtin_ia32_cmpltsd(__a, __b);
 237 }
 238
 239 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 240 _mm_cmple_sd(__m128d __a, __m128d __b)
 241 {
 242   return (__m128d)__builtin_ia32_cmplesd(__a, __b);
 243 }
 244
 245 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 246 _mm_cmpgt_sd(__m128d __a, __m128d __b)
 247 {
 248   __m128d __c = __builtin_ia32_cmpltsd(__b, __a);
 249   return (__m128d) { __c[0], __a[1] };
 250 }
 251
 252 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 253 _mm_cmpge_sd(__m128d __a, __m128d __b)
 254 {
 255   __m128d __c = __builtin_ia32_cmplesd(__b, __a);
 256   return (__m128d) { __c[0], __a[1] };
 257 }
 258
 259 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 260 _mm_cmpord_sd(__m128d __a, __m128d __b)
 261 {
 262   return (__m128d)__builtin_ia32_cmpordsd(__a, __b);
 263 }
 264
 265 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 266 _mm_cmpunord_sd(__m128d __a, __m128d __b)
 267 {
 268   return (__m128d)__builtin_ia32_cmpunordsd(__a, __b);
 269 }
 270
 271 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 272 _mm_cmpneq_sd(__m128d __a, __m128d __b)
 273 {
 274   return (__m128d)__builtin_ia32_cmpneqsd(__a, __b);
 275 }
 276
 277 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 278 _mm_cmpnlt_sd(__m128d __a, __m128d __b)
 279 {
 280   return (__m128d)__builtin_ia32_cmpnltsd(__a, __b);
 281 }
 282
 283 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 284 _mm_cmpnle_sd(__m128d __a, __m128d __b)
 285 {
 286   return (__m128d)__builtin_ia32_cmpnlesd(__a, __b);
 287 }
 288
 289 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 290 _mm_cmpngt_sd(__m128d __a, __m128d __b)
 291 {
 292   __m128d __c = __builtin_ia32_cmpnltsd(__b, __a);
 293   return (__m128d) { __c[0], __a[1] };
 294 }
 295
 296 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 297 _mm_cmpnge_sd(__m128d __a, __m128d __b)
 298 {
 299   __m128d __c = __builtin_ia32_cmpnlesd(__b, __a);
 300   return (__m128d) { __c[0], __a[1] };
 301 }
 302
 303 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 304 _mm_comieq_sd(__m128d __a, __m128d __b)
 305 {
 306   return __builtin_ia32_comisdeq(__a, __b);
 307 }
 308
 309 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 310 _mm_comilt_sd(__m128d __a, __m128d __b)
 311 {
 312   return __builtin_ia32_comisdlt(__a, __b);
 313 }
 314
 315 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 316 _mm_comile_sd(__m128d __a, __m128d __b)
 317 {
 318   return __builtin_ia32_comisdle(__a, __b);
 319 }
 320
 321 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 322 _mm_comigt_sd(__m128d __a, __m128d __b)
 323 {
 324   return __builtin_ia32_comisdgt(__a, __b);
 325 }
 326
 327 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 328 _mm_comige_sd(__m128d __a, __m128d __b)
 329 {
 330   return __builtin_ia32_comisdge(__a, __b);
 331 }
 332
 333 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 334 _mm_comineq_sd(__m128d __a, __m128d __b)
 335 {
 336   return __builtin_ia32_comisdneq(__a, __b);
 337 }
 338
 339 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 340 _mm_ucomieq_sd(__m128d __a, __m128d __b)
 341 {
 342   return __builtin_ia32_ucomisdeq(__a, __b);
 343 }
 344
 345 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 346 _mm_ucomilt_sd(__m128d __a, __m128d __b)
 347 {
 348   return __builtin_ia32_ucomisdlt(__a, __b);
 349 }
 350
 351 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 352 _mm_ucomile_sd(__m128d __a, __m128d __b)
 353 {
 354   return __builtin_ia32_ucomisdle(__a, __b);
 355 }
 356
 357 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 358 _mm_ucomigt_sd(__m128d __a, __m128d __b)
 359 {
 360   return __builtin_ia32_ucomisdgt(__a, __b);
 361 }
 362
 363 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 364 _mm_ucomige_sd(__m128d __a, __m128d __b)
 365 {
 366   return __builtin_ia32_ucomisdge(__a, __b);
 367 }
 368
 369 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 370 _mm_ucomineq_sd(__m128d __a, __m128d __b)
 371 {
 372   return __builtin_ia32_ucomisdneq(__a, __b);
 373 }
 374
 375 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
 376 _mm_cvtpd_ps(__m128d __a)
 377 {
 378   return __builtin_ia32_cvtpd2ps(__a);
 379 }
 380
 381 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 382 _mm_cvtps_pd(__m128 __a)
 383 {
 384   return __builtin_ia32_cvtps2pd(__a);
 385 }
 386
 387 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 388 _mm_cvtepi32_pd(__m128i __a)
 389 {
 390   return __builtin_ia32_cvtdq2pd((__v4si)__a);
 391 }
 392
 393 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 394 _mm_cvtpd_epi32(__m128d __a)
 395 {
 396   return __builtin_ia32_cvtpd2dq(__a);
 397 }
 398
 399 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 400 _mm_cvtsd_si32(__m128d __a)
 401 {
 402   return __builtin_ia32_cvtsd2si(__a);
 403 }
 404
 405 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
 406 _mm_cvtsd_ss(__m128 __a, __m128d __b)
 407 {
 408   __a[0] = __b[0];
 409   return __a;
 410 }
 411
 412 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 413 _mm_cvtsi32_sd(__m128d __a, int __b)
 414 {
 415   __a[0] = __b;
 416   return __a;
 417 }
 418
 419 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 420 _mm_cvtss_sd(__m128d __a, __m128 __b)
 421 {
 422   __a[0] = __b[0];
 423   return __a;
 424 }
 425
 426 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 427 _mm_cvttpd_epi32(__m128d __a)
 428 {
 429   return (__m128i)__builtin_ia32_cvttpd2dq(__a);
 430 }
 431
 432 static __inline__ int __attribute__((__always_inline__, __nodebug__))
 433 _mm_cvttsd_si32(__m128d __a)
 434 {
 435   return __a[0];
 436 }
 437
 438 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 439 _mm_cvtpd_pi32(__m128d __a)
 440 {
 441   return (__m64)__builtin_ia32_cvtpd2pi(__a);
 442 }
 443
 444 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 445 _mm_cvttpd_pi32(__m128d __a)
 446 {
 447   return (__m64)__builtin_ia32_cvttpd2pi(__a);
 448 }
 449
 450 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 451 _mm_cvtpi32_pd(__m64 __a)
 452 {
 453   return __builtin_ia32_cvtpi2pd((__v2si)__a);
 454 }
 455
 456 static __inline__ double __attribute__((__always_inline__, __nodebug__))
 457 _mm_cvtsd_f64(__m128d __a)
 458 {
 459   return __a[0];
 460 }
 461
 462 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 463 _mm_load_pd(double const *__dp)
 464 {
 465   return *(__m128d*)__dp;
 466 }
 467
 468 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 469 _mm_load1_pd(double const *__dp)
 470 {
 471   struct __mm_load1_pd_struct {
 472     double __u;
 473   } __attribute__((__packed__, __may_alias__));
 474   double __u = ((struct __mm_load1_pd_struct*)__dp)->__u;
 475   return (__m128d){ __u, __u };
 476 }
 477
 478 #define        _mm_load_pd1(dp)        _mm_load1_pd(dp)
 479
 480 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 481 _mm_loadr_pd(double const *__dp)
 482 {
 483   __m128d __u = *(__m128d*)__dp;
 484   return __builtin_shufflevector(__u, __u, 1, 0);
 485 }
 486
 487 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 488 _mm_loadu_pd(double const *__dp)
 489 {
 490   struct __loadu_pd {
 491     __m128d __v;
 492   } __attribute__((packed, may_alias));
 493   return ((struct __loadu_pd*)__dp)->__v;
 494 }
 495
 496 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 497 _mm_load_sd(double const *__dp)
 498 {
 499   struct __mm_load_sd_struct {
 500     double __u;
 501   } __attribute__((__packed__, __may_alias__));
 502   double __u = ((struct __mm_load_sd_struct*)__dp)->__u;
 503   return (__m128d){ __u, 0 };
 504 }
 505
 506 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 507 _mm_loadh_pd(__m128d __a, double const *__dp)
 508 {
 509   struct __mm_loadh_pd_struct {
 510     double __u;
 511   } __attribute__((__packed__, __may_alias__));
 512   double __u = ((struct __mm_loadh_pd_struct*)__dp)->__u;
 513   return (__m128d){ __a[0], __u };
 514 }
 515
 516 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 517 _mm_loadl_pd(__m128d __a, double const *__dp)
 518 {
 519   struct __mm_loadl_pd_struct {
 520     double __u;
 521   } __attribute__((__packed__, __may_alias__));
 522   double __u = ((struct __mm_loadl_pd_struct*)__dp)->__u;
 523   return (__m128d){ __u, __a[1] };
 524 }
 525
 526 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 527 _mm_set_sd(double __w)
 528 {
 529   return (__m128d){ __w, 0 };
 530 }
 531
 532 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 533 _mm_set1_pd(double __w)
 534 {
 535   return (__m128d){ __w, __w };
 536 }
 537
 538 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 539 _mm_set_pd(double __w, double __x)
 540 {
 541   return (__m128d){ __x, __w };
 542 }
 543
 544 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 545 _mm_setr_pd(double __w, double __x)
 546 {
 547   return (__m128d){ __w, __x };
 548 }
 549
 550 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 551 _mm_setzero_pd(void)
 552 {
 553   return (__m128d){ 0, 0 };
 554 }
 555
 556 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 557 _mm_move_sd(__m128d __a, __m128d __b)
 558 {
 559   return (__m128d){ __b[0], __a[1] };
 560 }
 561
 562 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 563 _mm_store_sd(double *__dp, __m128d __a)
 564 {
 565   struct __mm_store_sd_struct {
 566     double __u;
 567   } __attribute__((__packed__, __may_alias__));
 568   ((struct __mm_store_sd_struct*)__dp)->__u = __a[0];
 569 }
 570
 571 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 572 _mm_store1_pd(double *__dp, __m128d __a)
 573 {
 574   struct __mm_store1_pd_struct {
 575     double __u[2];
 576   } __attribute__((__packed__, __may_alias__));
 577   ((struct __mm_store1_pd_struct*)__dp)->__u[0] = __a[0];
 578   ((struct __mm_store1_pd_struct*)__dp)->__u[1] = __a[0];
 579 }
 580
 581 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 582 _mm_store_pd(double *__dp, __m128d __a)
 583 {
 584   *(__m128d *)__dp = __a;
 585 }
 586
 587 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 588 _mm_storeu_pd(double *__dp, __m128d __a)
 589 {
 590   __builtin_ia32_storeupd(__dp, __a);
 591 }
 592
 593 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 594 _mm_storer_pd(double *__dp, __m128d __a)
 595 {
 596   __a = __builtin_shufflevector(__a, __a, 1, 0);
 597   *(__m128d *)__dp = __a;
 598 }
 599
 600 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 601 _mm_storeh_pd(double *__dp, __m128d __a)
 602 {
 603   struct __mm_storeh_pd_struct {
 604     double __u;
 605   } __attribute__((__packed__, __may_alias__));
 606   ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[1];
 607 }
 608
 609 static __inline__ void __attribute__((__always_inline__, __nodebug__))
 610 _mm_storel_pd(double *__dp, __m128d __a)
 611 {
 612   struct __mm_storeh_pd_struct {
 613     double __u;
 614   } __attribute__((__packed__, __may_alias__));
 615   ((struct __mm_storeh_pd_struct*)__dp)->__u = __a[0];
 616 }
 617
 618 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 619 _mm_add_epi8(__m128i __a, __m128i __b)
 620 {
 621   return (__m128i)((__v16qi)__a + (__v16qi)__b);
 622 }
 623
 624 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 625 _mm_add_epi16(__m128i __a, __m128i __b)
 626 {
 627   return (__m128i)((__v8hi)__a + (__v8hi)__b);
 628 }
 629
 630 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 631 _mm_add_epi32(__m128i __a, __m128i __b)
 632 {
 633   return (__m128i)((__v4si)__a + (__v4si)__b);
 634 }
 635
 636 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 637 _mm_add_si64(__m64 __a, __m64 __b)
 638 {
 639   return __a + __b;
 640 }
 641
 642 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 643 _mm_add_epi64(__m128i __a, __m128i __b)
 644 {
 645   return __a + __b;
 646 }
 647
 648 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 649 _mm_adds_epi8(__m128i __a, __m128i __b)
 650 {
 651   return (__m128i)__builtin_ia32_paddsb128((__v16qi)__a, (__v16qi)__b);
 652 }
 653
 654 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 655 _mm_adds_epi16(__m128i __a, __m128i __b)
 656 {
 657   return (__m128i)__builtin_ia32_paddsw128((__v8hi)__a, (__v8hi)__b);
 658 }
 659
 660 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 661 _mm_adds_epu8(__m128i __a, __m128i __b)
 662 {
 663   return (__m128i)__builtin_ia32_paddusb128((__v16qi)__a, (__v16qi)__b);
 664 }
 665
 666 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 667 _mm_adds_epu16(__m128i __a, __m128i __b)
 668 {
 669   return (__m128i)__builtin_ia32_paddusw128((__v8hi)__a, (__v8hi)__b);
 670 }
 671
 672 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 673 _mm_avg_epu8(__m128i __a, __m128i __b)
 674 {
 675   return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
 676 }
 677
 678 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 679 _mm_avg_epu16(__m128i __a, __m128i __b)
 680 {
 681   return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
 682 }
 683
 684 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 685 _mm_madd_epi16(__m128i __a, __m128i __b)
 686 {
 687   return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)__a, (__v8hi)__b);
 688 }
 689
 690 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 691 _mm_max_epi16(__m128i __a, __m128i __b)
 692 {
 693   return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)__a, (__v8hi)__b);
 694 }
 695
 696 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 697 _mm_max_epu8(__m128i __a, __m128i __b)
 698 {
 699   return (__m128i)__builtin_ia32_pmaxub128((__v16qi)__a, (__v16qi)__b);
 700 }
 701
 702 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 703 _mm_min_epi16(__m128i __a, __m128i __b)
 704 {
 705   return (__m128i)__builtin_ia32_pminsw128((__v8hi)__a, (__v8hi)__b);
 706 }
 707
 708 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 709 _mm_min_epu8(__m128i __a, __m128i __b)
 710 {
 711   return (__m128i)__builtin_ia32_pminub128((__v16qi)__a, (__v16qi)__b);
 712 }
 713
 714 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 715 _mm_mulhi_epi16(__m128i __a, __m128i __b)
 716 {
 717   return (__m128i)__builtin_ia32_pmulhw128((__v8hi)__a, (__v8hi)__b);
 718 }
 719
 720 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 721 _mm_mulhi_epu16(__m128i __a, __m128i __b)
 722 {
 723   return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)__a, (__v8hi)__b);
 724 }
 725
 726 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 727 _mm_mullo_epi16(__m128i __a, __m128i __b)
 728 {
 729   return (__m128i)((__v8hi)__a * (__v8hi)__b);
 730 }
 731
 732 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 733 _mm_mul_su32(__m64 __a, __m64 __b)
 734 {
 735   return __builtin_ia32_pmuludq((__v2si)__a, (__v2si)__b);
 736 }
 737
 738 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 739 _mm_mul_epu32(__m128i __a, __m128i __b)
 740 {
 741   return __builtin_ia32_pmuludq128((__v4si)__a, (__v4si)__b);
 742 }
 743
 744 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 745 _mm_sad_epu8(__m128i __a, __m128i __b)
 746 {
 747   return __builtin_ia32_psadbw128((__v16qi)__a, (__v16qi)__b);
 748 }
 749
 750 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 751 _mm_sub_epi8(__m128i __a, __m128i __b)
 752 {
 753   return (__m128i)((__v16qi)__a - (__v16qi)__b);
 754 }
 755
 756 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 757 _mm_sub_epi16(__m128i __a, __m128i __b)
 758 {
 759   return (__m128i)((__v8hi)__a - (__v8hi)__b);
 760 }
 761
 762 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 763 _mm_sub_epi32(__m128i __a, __m128i __b)
 764 {
 765   return (__m128i)((__v4si)__a - (__v4si)__b);
 766 }
 767
 768 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
 769 _mm_sub_si64(__m64 __a, __m64 __b)
 770 {
 771   return __a - __b;
 772 }
 773
 774 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 775 _mm_sub_epi64(__m128i __a, __m128i __b)
 776 {
 777   return __a - __b;
 778 }
 779
 780 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 781 _mm_subs_epi8(__m128i __a, __m128i __b)
 782 {
 783   return (__m128i)__builtin_ia32_psubsb128((__v16qi)__a, (__v16qi)__b);
 784 }
 785
 786 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 787 _mm_subs_epi16(__m128i __a, __m128i __b)
 788 {
 789   return (__m128i)__builtin_ia32_psubsw128((__v8hi)__a, (__v8hi)__b);
 790 }
 791
 792 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 793 _mm_subs_epu8(__m128i __a, __m128i __b)
 794 {
 795   return (__m128i)__builtin_ia32_psubusb128((__v16qi)__a, (__v16qi)__b);
 796 }
 797
 798 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 799 _mm_subs_epu16(__m128i __a, __m128i __b)
 800 {
 801   return (__m128i)__builtin_ia32_psubusw128((__v8hi)__a, (__v8hi)__b);
 802 }
 803
 804 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 805 _mm_and_si128(__m128i __a, __m128i __b)
 806 {
 807   return __a & __b;
 808 }
 809
 810 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 811 _mm_andnot_si128(__m128i __a, __m128i __b)
 812 {
 813   return ~__a & __b;
 814 }
 815
 816 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 817 _mm_or_si128(__m128i __a, __m128i __b)
 818 {
 819   return __a | __b;
 820 }
 821
 822 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 823 _mm_xor_si128(__m128i __a, __m128i __b)
 824 {
 825   return __a ^ __b;
 826 }
 827
 828 #define _mm_slli_si128(a, count) __extension__ ({ \
 829   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
 830   __m128i __a = (a); \
 831    _Pragma("clang diagnostic pop"); \
 832   (__m128i)__builtin_ia32_pslldqi128(__a, (count)*8); })
 833
 834 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 835 _mm_slli_epi16(__m128i __a, int __count)
 836 {
 837   return (__m128i)__builtin_ia32_psllwi128((__v8hi)__a, __count);
 838 }
 839
 840 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 841 _mm_sll_epi16(__m128i __a, __m128i __count)
 842 {
 843   return (__m128i)__builtin_ia32_psllw128((__v8hi)__a, (__v8hi)__count);
 844 }
 845
 846 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 847 _mm_slli_epi32(__m128i __a, int __count)
 848 {
 849   return (__m128i)__builtin_ia32_pslldi128((__v4si)__a, __count);
 850 }
 851
 852 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 853 _mm_sll_epi32(__m128i __a, __m128i __count)
 854 {
 855   return (__m128i)__builtin_ia32_pslld128((__v4si)__a, (__v4si)__count);
 856 }
 857
 858 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 859 _mm_slli_epi64(__m128i __a, int __count)
 860 {
 861   return __builtin_ia32_psllqi128(__a, __count);
 862 }
 863
 864 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 865 _mm_sll_epi64(__m128i __a, __m128i __count)
 866 {
 867   return __builtin_ia32_psllq128(__a, __count);
 868 }
 869
 870 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 871 _mm_srai_epi16(__m128i __a, int __count)
 872 {
 873   return (__m128i)__builtin_ia32_psrawi128((__v8hi)__a, __count);
 874 }
 875
 876 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 877 _mm_sra_epi16(__m128i __a, __m128i __count)
 878 {
 879   return (__m128i)__builtin_ia32_psraw128((__v8hi)__a, (__v8hi)__count);
 880 }
 881
 882 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 883 _mm_srai_epi32(__m128i __a, int __count)
 884 {
 885   return (__m128i)__builtin_ia32_psradi128((__v4si)__a, __count);
 886 }
 887
 888 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 889 _mm_sra_epi32(__m128i __a, __m128i __count)
 890 {
 891   return (__m128i)__builtin_ia32_psrad128((__v4si)__a, (__v4si)__count);
 892 }
 893
 894
 895 #define _mm_srli_si128(a, count) __extension__ ({ \
 896   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
 897   __m128i __a = (a); \
 898   _Pragma("clang diagnostic pop"); \
 899   (__m128i)__builtin_ia32_psrldqi128(__a, (count)*8); })
 900
 901 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 902 _mm_srli_epi16(__m128i __a, int __count)
 903 {
 904   return (__m128i)__builtin_ia32_psrlwi128((__v8hi)__a, __count);
 905 }
 906
 907 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 908 _mm_srl_epi16(__m128i __a, __m128i __count)
 909 {
 910   return (__m128i)__builtin_ia32_psrlw128((__v8hi)__a, (__v8hi)__count);
 911 }
 912
 913 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 914 _mm_srli_epi32(__m128i __a, int __count)
 915 {
 916   return (__m128i)__builtin_ia32_psrldi128((__v4si)__a, __count);
 917 }
 918
 919 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 920 _mm_srl_epi32(__m128i __a, __m128i __count)
 921 {
 922   return (__m128i)__builtin_ia32_psrld128((__v4si)__a, (__v4si)__count);
 923 }
 924
 925 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 926 _mm_srli_epi64(__m128i __a, int __count)
 927 {
 928   return __builtin_ia32_psrlqi128(__a, __count);
 929 }
 930
 931 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 932 _mm_srl_epi64(__m128i __a, __m128i __count)
 933 {
 934   return __builtin_ia32_psrlq128(__a, __count);
 935 }
 936
 937 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 938 _mm_cmpeq_epi8(__m128i __a, __m128i __b)
 939 {
 940   return (__m128i)((__v16qi)__a == (__v16qi)__b);
 941 }
 942
 943 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 944 _mm_cmpeq_epi16(__m128i __a, __m128i __b)
 945 {
 946   return (__m128i)((__v8hi)__a == (__v8hi)__b);
 947 }
 948
 949 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 950 _mm_cmpeq_epi32(__m128i __a, __m128i __b)
 951 {
 952   return (__m128i)((__v4si)__a == (__v4si)__b);
 953 }
 954
 955 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 956 _mm_cmpgt_epi8(__m128i __a, __m128i __b)
 957 {
 958   /* This function always performs a signed comparison, but __v16qi is a char
 959      which may be signed or unsigned. */
 960   typedef signed char __v16qs __attribute__((__vector_size__(16)));
 961   return (__m128i)((__v16qs)__a > (__v16qs)__b);
 962 }
 963
 964 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 965 _mm_cmpgt_epi16(__m128i __a, __m128i __b)
 966 {
 967   return (__m128i)((__v8hi)__a > (__v8hi)__b);
 968 }
 969
 970 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 971 _mm_cmpgt_epi32(__m128i __a, __m128i __b)
 972 {
 973   return (__m128i)((__v4si)__a > (__v4si)__b);
 974 }
 975
 976 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 977 _mm_cmplt_epi8(__m128i __a, __m128i __b)
 978 {
 979   return _mm_cmpgt_epi8(__b, __a);
 980 }
 981
 982 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 983 _mm_cmplt_epi16(__m128i __a, __m128i __b)
 984 {
 985   return _mm_cmpgt_epi16(__b, __a);
 986 }
 987
 988 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
 989 _mm_cmplt_epi32(__m128i __a, __m128i __b)
 990 {
 991   return _mm_cmpgt_epi32(__b, __a);
 992 }
 993
 994 #ifdef __x86_64__
 995 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
 996 _mm_cvtsi64_sd(__m128d __a, long long __b)
 997 {
 998   __a[0] = __b;
 999   return __a;
1000 }
1001
1002 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
1003 _mm_cvtsd_si64(__m128d __a)
1004 {
1005   return __builtin_ia32_cvtsd2si64(__a);
1006 }
1007
1008 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
1009 _mm_cvttsd_si64(__m128d __a)
1010 {
1011   return __a[0];
1012 }
1013 #endif
1014
1015 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1016 _mm_cvtepi32_ps(__m128i __a)
1017 {
1018   return __builtin_ia32_cvtdq2ps((__v4si)__a);
1019 }
1020
1021 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1022 _mm_cvtps_epi32(__m128 __a)
1023 {
1024   return (__m128i)__builtin_ia32_cvtps2dq(__a);
1025 }
1026
1027 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1028 _mm_cvttps_epi32(__m128 __a)
1029 {
1030   return (__m128i)__builtin_ia32_cvttps2dq(__a);
1031 }
1032
1033 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1034 _mm_cvtsi32_si128(int __a)
1035 {
1036   return (__m128i)(__v4si){ __a, 0, 0, 0 };
1037 }
1038
1039 #ifdef __x86_64__
1040 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1041 _mm_cvtsi64_si128(long long __a)
1042 {
1043   return (__m128i){ __a, 0 };
1044 }
1045 #endif
1046
1047 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1048 _mm_cvtsi128_si32(__m128i __a)
1049 {
1050   __v4si __b = (__v4si)__a;
1051   return __b[0];
1052 }
1053
1054 #ifdef __x86_64__
1055 static __inline__ long long __attribute__((__always_inline__, __nodebug__))
1056 _mm_cvtsi128_si64(__m128i __a)
1057 {
1058   return __a[0];
1059 }
1060 #endif
1061
1062 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1063 _mm_load_si128(__m128i const *__p)
1064 {
1065   return *__p;
1066 }
1067
1068 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1069 _mm_loadu_si128(__m128i const *__p)
1070 {
1071   struct __loadu_si128 {
1072     __m128i __v;
1073   } __attribute__((packed, may_alias));
1074   return ((struct __loadu_si128*)__p)->__v;
1075 }
1076
1077 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1078 _mm_loadl_epi64(__m128i const *__p)
1079 {
1080   struct __mm_loadl_epi64_struct {
1081     long long __u;
1082   } __attribute__((__packed__, __may_alias__));
1083   return (__m128i) { ((struct __mm_loadl_epi64_struct*)__p)->__u, 0};
1084 }
1085
1086 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1087 _mm_set_epi64x(long long q1, long long q0)
1088 {
1089   return (__m128i){ q0, q1 };
1090 }
1091
1092 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1093 _mm_set_epi64(__m64 q1, __m64 q0)
1094 {
1095   return (__m128i){ (long long)q0, (long long)q1 };
1096 }
1097
1098 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1099 _mm_set_epi32(int i3, int i2, int i1, int i0)
1100 {
1101   return (__m128i)(__v4si){ i0, i1, i2, i3};
1102 }
1103
1104 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1105 _mm_set_epi16(short w7, short w6, short w5, short w4, short w3, short w2, short w1, short w0)
1106 {
1107   return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1108 }
1109
1110 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1111 _mm_set_epi8(char b15, char b14, char b13, char b12, char b11, char b10, char b9, char b8, char b7, char b6, char b5, char b4, char b3, char b2, char b1, char b0)
1112 {
1113   return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1114 }
1115
1116 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1117 _mm_set1_epi64x(long long __q)
1118 {
1119   return (__m128i){ __q, __q };
1120 }
1121
1122 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1123 _mm_set1_epi64(__m64 __q)
1124 {
1125   return (__m128i){ (long long)__q, (long long)__q };
1126 }
1127
1128 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1129 _mm_set1_epi32(int __i)
1130 {
1131   return (__m128i)(__v4si){ __i, __i, __i, __i };
1132 }
1133
1134 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1135 _mm_set1_epi16(short __w)
1136 {
1137   return (__m128i)(__v8hi){ __w, __w, __w, __w, __w, __w, __w, __w };
1138 }
1139
1140 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1141 _mm_set1_epi8(char __b)
1142 {
1143   return (__m128i)(__v16qi){ __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b };
1144 }
1145
1146 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1147 _mm_setr_epi64(__m64 q0, __m64 q1)
1148 {
1149   return (__m128i){ (long long)q0, (long long)q1 };
1150 }
1151
1152 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1153 _mm_setr_epi32(int i0, int i1, int i2, int i3)
1154 {
1155   return (__m128i)(__v4si){ i0, i1, i2, i3};
1156 }
1157
1158 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1159 _mm_setr_epi16(short w0, short w1, short w2, short w3, short w4, short w5, short w6, short w7)
1160 {
1161   return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 };
1162 }
1163
1164 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1165 _mm_setr_epi8(char b0, char b1, char b2, char b3, char b4, char b5, char b6, char b7, char b8, char b9, char b10, char b11, char b12, char b13, char b14, char b15)
1166 {
1167   return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15 };
1168 }
1169
1170 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1171 _mm_setzero_si128(void)
1172 {
1173   return (__m128i){ 0LL, 0LL };
1174 }
1175
1176 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1177 _mm_store_si128(__m128i *__p, __m128i __b)
1178 {
1179   *__p = __b;
1180 }
1181
1182 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1183 _mm_storeu_si128(__m128i *__p, __m128i __b)
1184 {
1185   __builtin_ia32_storedqu((char *)__p, (__v16qi)__b);
1186 }
1187
1188 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1189 _mm_maskmoveu_si128(__m128i __d, __m128i __n, char *__p)
1190 {
1191   __builtin_ia32_maskmovdqu((__v16qi)__d, (__v16qi)__n, __p);
1192 }
1193
1194 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1195 _mm_storel_epi64(__m128i *__p, __m128i __a)
1196 {
1197   struct __mm_storel_epi64_struct {
1198     long long __u;
1199   } __attribute__((__packed__, __may_alias__));
1200   ((struct __mm_storel_epi64_struct*)__p)->__u = __a[0];
1201 }
1202
1203 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1204 _mm_stream_pd(double *__p, __m128d __a)
1205 {
1206   __builtin_ia32_movntpd(__p, __a);
1207 }
1208
1209 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1210 _mm_stream_si128(__m128i *__p, __m128i __a)
1211 {
1212   __builtin_ia32_movntdq(__p, __a);
1213 }
1214
1215 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1216 _mm_stream_si32(int *__p, int __a)
1217 {
1218   __builtin_ia32_movnti(__p, __a);
1219 }
1220
1221 #ifdef __x86_64__
1222 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1223 _mm_stream_si64(long long *__p, long long __a)
1224 {
1225   __builtin_ia32_movnti64(__p, __a);
1226 }
1227 #endif
1228
1229 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1230 _mm_clflush(void const *__p)
1231 {
1232   __builtin_ia32_clflush(__p);
1233 }
1234
1235 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1236 _mm_lfence(void)
1237 {
1238   __builtin_ia32_lfence();
1239 }
1240
1241 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1242 _mm_mfence(void)
1243 {
1244   __builtin_ia32_mfence();
1245 }
1246
1247 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1248 _mm_packs_epi16(__m128i __a, __m128i __b)
1249 {
1250   return (__m128i)__builtin_ia32_packsswb128((__v8hi)__a, (__v8hi)__b);
1251 }
1252
1253 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1254 _mm_packs_epi32(__m128i __a, __m128i __b)
1255 {
1256   return (__m128i)__builtin_ia32_packssdw128((__v4si)__a, (__v4si)__b);
1257 }
1258
1259 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1260 _mm_packus_epi16(__m128i __a, __m128i __b)
1261 {
1262   return (__m128i)__builtin_ia32_packuswb128((__v8hi)__a, (__v8hi)__b);
1263 }
1264
1265 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1266 _mm_extract_epi16(__m128i __a, int __imm)
1267 {
1268   __v8hi __b = (__v8hi)__a;
1269   return (unsigned short)__b[__imm & 7];
1270 }
1271
1272 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1273 _mm_insert_epi16(__m128i __a, int __b, int __imm)
1274 {
1275   __v8hi __c = (__v8hi)__a;
1276   __c[__imm & 7] = __b;
1277   return (__m128i)__c;
1278 }
1279
1280 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1281 _mm_movemask_epi8(__m128i __a)
1282 {
1283   return __builtin_ia32_pmovmskb128((__v16qi)__a);
1284 }
1285
1286 #define _mm_shuffle_epi32(a, imm) __extension__ ({ \
1287   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
1288   __m128i __a = (a); \
1289   _Pragma("clang diagnostic pop"); \
1290   (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si) _mm_set1_epi32(0), \
1291                                    (imm) & 0x3, ((imm) & 0xc) >> 2, \
1292                                    ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6); })
1293
1294 #define _mm_shufflelo_epi16(a, imm) __extension__ ({ \
1295   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
1296   __m128i __a = (a); \
1297   _Pragma("clang diagnostic pop"); \
1298   (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1299                                    (imm) & 0x3, ((imm) & 0xc) >> 2, \
1300                                    ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \
1301                                    4, 5, 6, 7); })
1302
1303 #define _mm_shufflehi_epi16(a, imm) __extension__ ({ \
1304   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
1305   __m128i __a = (a); \
1306   _Pragma("clang diagnostic pop"); \
1307   (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi) _mm_set1_epi16(0), \
1308                                    0, 1, 2, 3, \
1309                                    4 + (((imm) & 0x03) >> 0), \
1310                                    4 + (((imm) & 0x0c) >> 2), \
1311                                    4 + (((imm) & 0x30) >> 4), \
1312                                    4 + (((imm) & 0xc0) >> 6)); })
1313
1314 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1315 _mm_unpackhi_epi8(__m128i __a, __m128i __b)
1316 {
1317   return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 8, 16+8, 9, 16+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15);
1318 }
1319
1320 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1321 _mm_unpackhi_epi16(__m128i __a, __m128i __b)
1322 {
1323   return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 4, 8+4, 5, 8+5, 6, 8+6, 7, 8+7);
1324 }
1325
1326 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1327 _mm_unpackhi_epi32(__m128i __a, __m128i __b)
1328 {
1329   return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 2, 4+2, 3, 4+3);
1330 }
1331
1332 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1333 _mm_unpackhi_epi64(__m128i __a, __m128i __b)
1334 {
1335   return (__m128i)__builtin_shufflevector(__a, __b, 1, 2+1);
1336 }
1337
1338 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1339 _mm_unpacklo_epi8(__m128i __a, __m128i __b)
1340 {
1341   return (__m128i)__builtin_shufflevector((__v16qi)__a, (__v16qi)__b, 0, 16+0, 1, 16+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7);
1342 }
1343
1344 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1345 _mm_unpacklo_epi16(__m128i __a, __m128i __b)
1346 {
1347   return (__m128i)__builtin_shufflevector((__v8hi)__a, (__v8hi)__b, 0, 8+0, 1, 8+1, 2, 8+2, 3, 8+3);
1348 }
1349
1350 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1351 _mm_unpacklo_epi32(__m128i __a, __m128i __b)
1352 {
1353   return (__m128i)__builtin_shufflevector((__v4si)__a, (__v4si)__b, 0, 4+0, 1, 4+1);
1354 }
1355
1356 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1357 _mm_unpacklo_epi64(__m128i __a, __m128i __b)
1358 {
1359   return (__m128i)__builtin_shufflevector(__a, __b, 0, 2+0);
1360 }
1361
1362 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__))
1363 _mm_movepi64_pi64(__m128i __a)
1364 {
1365   return (__m64)__a[0];
1366 }
1367
1368 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1369 _mm_movpi64_epi64(__m64 __a)
1370 {
1371   return (__m128i){ (long long)__a, 0 };
1372 }
1373
1374 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1375 _mm_move_epi64(__m128i __a)
1376 {
1377   return __builtin_shufflevector(__a, (__m128i){ 0 }, 0, 2);
1378 }
1379
1380 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1381 _mm_unpackhi_pd(__m128d __a, __m128d __b)
1382 {
1383   return __builtin_shufflevector(__a, __b, 1, 2+1);
1384 }
1385
1386 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1387 _mm_unpacklo_pd(__m128d __a, __m128d __b)
1388 {
1389   return __builtin_shufflevector(__a, __b, 0, 2+0);
1390 }
1391
1392 static __inline__ int __attribute__((__always_inline__, __nodebug__))
1393 _mm_movemask_pd(__m128d __a)
1394 {
1395   return __builtin_ia32_movmskpd(__a);
1396 }
1397
1398 #define _mm_shuffle_pd(a, b, i) __extension__ ({ \
1399   _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wshadow\""); \
1400   __m128d __a = (a); \
1401   __m128d __b = (b); \
1402   _Pragma("clang diagnostic pop"); \
1403   __builtin_shufflevector(__a, __b, (i) & 1, (((i) & 2) >> 1) + 2); })
1404
1405 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1406 _mm_castpd_ps(__m128d __a)
1407 {
1408   return (__m128)__a;
1409 }
1410
1411 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1412 _mm_castpd_si128(__m128d __a)
1413 {
1414   return (__m128i)__a;
1415 }
1416
1417 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1418 _mm_castps_pd(__m128 __a)
1419 {
1420   return (__m128d)__a;
1421 }
1422
1423 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__))
1424 _mm_castps_si128(__m128 __a)
1425 {
1426   return (__m128i)__a;
1427 }
1428
1429 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
1430 _mm_castsi128_ps(__m128i __a)
1431 {
1432   return (__m128)__a;
1433 }
1434
1435 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
1436 _mm_castsi128_pd(__m128i __a)
1437 {
1438   return (__m128d)__a;
1439 }
1440
1441 static __inline__ void __attribute__((__always_inline__, __nodebug__))
1442 _mm_pause(void)
1443 {
1444   __asm__ volatile ("pause");
1445 }
1446
1447 #define _MM_SHUFFLE2(x, y) (((x) << 1) | (y))
1448
1449 #endif /* __SSE2__ */
1450
1451 #endif /* __EMMINTRIN_H */