1 /* Copyright (C) 2003-2018 Free Software Foundation, Inc. 2 3 This file is part of GCC. 4 5 GCC is free software; you can redistribute it and/or modify 6 it under the terms of the GNU General Public License as published by 7 the Free Software Foundation; either version 3, or (at your option) 8 any later version. 9 10 GCC is distributed in the hope that it will be useful, 11 but WITHOUT ANY WARRANTY; without even the implied warranty of 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 GNU General Public License for more details. 14 15 Under Section 7 of GPL version 3, you are granted additional 16 permissions described in the GCC Runtime Library Exception, version 17 3.1, as published by the Free Software Foundation. 18 19 You should have received a copy of the GNU General Public License and 20 a copy of the GCC Runtime Library Exception along with this program; 21 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 22 <http://www.gnu.org/licenses/>. */ 23 24 /* Implemented from the specification included in the Intel C++ Compiler 25 User Guide and Reference, version 9.0. */ 26 27 #ifndef _EMMINTRIN_H_INCLUDED 28 #define _EMMINTRIN_H_INCLUDED 29 30 /* We need definitions from the SSE header files*/ 31 #include <xmmintrin.h> 32 33 #ifndef __SSE2__ 34 #pragma GCC push_options 35 #pragma GCC target("sse2") 36 #define __DISABLE_SSE2__ 37 #endif /* __SSE2__ */ 38 39 /* SSE2 */ 40 typedef double __v2df __attribute__ ((__vector_size__ (16))); 41 typedef long long __v2di __attribute__ ((__vector_size__ (16))); 42 typedef unsigned long long __v2du __attribute__ ((__vector_size__ (16))); 43 typedef int __v4si __attribute__ ((__vector_size__ (16))); 44 typedef unsigned int __v4su __attribute__ ((__vector_size__ (16))); 45 typedef short __v8hi __attribute__ ((__vector_size__ (16))); 46 typedef unsigned short __v8hu __attribute__ ((__vector_size__ (16))); 47 typedef char __v16qi __attribute__ ((__vector_size__ (16))); 48 typedef signed char __v16qs __attribute__ ((__vector_size__ (16))); 49 typedef unsigned char __v16qu __attribute__ ((__vector_size__ (16))); 50 51 /* The Intel API is flexible enough that we must allow aliasing with other 52 vector types, and their scalar components. */ 53 typedef long long __m128i __attribute__ ((__vector_size__ (16), __may_alias__)); 54 typedef double __m128d __attribute__ ((__vector_size__ (16), __may_alias__)); 55 56 /* Unaligned version of the same types. */ 57 typedef long long __m128i_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1))); 58 typedef double __m128d_u __attribute__ ((__vector_size__ (16), __may_alias__, __aligned__ (1))); 59 60 /* Create a selector for use with the SHUFPD instruction. */ 61 #define _MM_SHUFFLE2(fp1,fp0) \ 62 (((fp1) << 1) | (fp0)) 63 64 /* Create a vector with element 0 as F and the rest zero. */ 65 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 66 _mm_set_sd (double __F) 67 { 68 return __extension__ (__m128d){ __F, 0.0 }; 69 } 70 71 /* Create a vector with both elements equal to F. */ 72 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 73 _mm_set1_pd (double __F) 74 { 75 return __extension__ (__m128d){ __F, __F }; 76 } 77 78 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 79 _mm_set_pd1 (double __F) 80 { 81 return _mm_set1_pd (__F); 82 } 83 84 /* Create a vector with the lower value X and upper value W. */ 85 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 86 _mm_set_pd (double __W, double __X) 87 { 88 return __extension__ (__m128d){ __X, __W }; 89 } 90 91 /* Create a vector with the lower value W and upper value X. */ 92 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 93 _mm_setr_pd (double __W, double __X) 94 { 95 return __extension__ (__m128d){ __W, __X }; 96 } 97 98 /* Create an undefined vector. */ 99 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 100 _mm_undefined_pd (void) 101 { 102 __m128d __Y = __Y; 103 return __Y; 104 } 105 106 /* Create a vector of zeros. */ 107 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 108 _mm_setzero_pd (void) 109 { 110 return __extension__ (__m128d){ 0.0, 0.0 }; 111 } 112 113 /* Sets the low DPFP value of A from the low value of B. */ 114 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 115 _mm_move_sd (__m128d __A, __m128d __B) 116 { 117 return (__m128d) __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B); 118 } 119 120 /* Load two DPFP values from P. The address must be 16-byte aligned. */ 121 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 122 _mm_load_pd (double const *__P) 123 { 124 return *(__m128d *)__P; 125 } 126 127 /* Load two DPFP values from P. The address need not be 16-byte aligned. */ 128 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 129 _mm_loadu_pd (double const *__P) 130 { 131 return *(__m128d_u *)__P; 132 } 133 134 /* Create a vector with all two elements equal to *P. */ 135 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 136 _mm_load1_pd (double const *__P) 137 { 138 return _mm_set1_pd (*__P); 139 } 140 141 /* Create a vector with element 0 as *P and the rest zero. */ 142 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 143 _mm_load_sd (double const *__P) 144 { 145 return _mm_set_sd (*__P); 146 } 147 148 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 149 _mm_load_pd1 (double const *__P) 150 { 151 return _mm_load1_pd (__P); 152 } 153 154 /* Load two DPFP values in reverse order. The address must be aligned. */ 155 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 156 _mm_loadr_pd (double const *__P) 157 { 158 __m128d __tmp = _mm_load_pd (__P); 159 return __builtin_ia32_shufpd (__tmp, __tmp, _MM_SHUFFLE2 (0,1)); 160 } 161 162 /* Store two DPFP values. The address must be 16-byte aligned. */ 163 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 164 _mm_store_pd (double *__P, __m128d __A) 165 { 166 *(__m128d *)__P = __A; 167 } 168 169 /* Store two DPFP values. The address need not be 16-byte aligned. */ 170 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 171 _mm_storeu_pd (double *__P, __m128d __A) 172 { 173 *(__m128d_u *)__P = __A; 174 } 175 176 /* Stores the lower DPFP value. */ 177 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 178 _mm_store_sd (double *__P, __m128d __A) 179 { 180 *__P = ((__v2df)__A)[0]; 181 } 182 183 extern __inline double __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 184 _mm_cvtsd_f64 (__m128d __A) 185 { 186 return ((__v2df)__A)[0]; 187 } 188 189 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 190 _mm_storel_pd (double *__P, __m128d __A) 191 { 192 _mm_store_sd (__P, __A); 193 } 194 195 /* Stores the upper DPFP value. */ 196 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 197 _mm_storeh_pd (double *__P, __m128d __A) 198 { 199 *__P = ((__v2df)__A)[1]; 200 } 201 202 /* Store the lower DPFP value across two words. 203 The address must be 16-byte aligned. */ 204 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 205 _mm_store1_pd (double *__P, __m128d __A) 206 { 207 _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,0))); 208 } 209 210 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 211 _mm_store_pd1 (double *__P, __m128d __A) 212 { 213 _mm_store1_pd (__P, __A); 214 } 215 216 /* Store two DPFP values in reverse order. The address must be aligned. */ 217 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 218 _mm_storer_pd (double *__P, __m128d __A) 219 { 220 _mm_store_pd (__P, __builtin_ia32_shufpd (__A, __A, _MM_SHUFFLE2 (0,1))); 221 } 222 223 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 224 _mm_cvtsi128_si32 (__m128i __A) 225 { 226 return __builtin_ia32_vec_ext_v4si ((__v4si)__A, 0); 227 } 228 229 #ifdef __x86_64__ 230 /* Intel intrinsic. */ 231 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 232 _mm_cvtsi128_si64 (__m128i __A) 233 { 234 return ((__v2di)__A)[0]; 235 } 236 237 /* Microsoft intrinsic. */ 238 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 239 _mm_cvtsi128_si64x (__m128i __A) 240 { 241 return ((__v2di)__A)[0]; 242 } 243 #endif 244 245 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 246 _mm_add_pd (__m128d __A, __m128d __B) 247 { 248 return (__m128d) ((__v2df)__A + (__v2df)__B); 249 } 250 251 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 252 _mm_add_sd (__m128d __A, __m128d __B) 253 { 254 return (__m128d)__builtin_ia32_addsd ((__v2df)__A, (__v2df)__B); 255 } 256 257 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 258 _mm_sub_pd (__m128d __A, __m128d __B) 259 { 260 return (__m128d) ((__v2df)__A - (__v2df)__B); 261 } 262 263 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 264 _mm_sub_sd (__m128d __A, __m128d __B) 265 { 266 return (__m128d)__builtin_ia32_subsd ((__v2df)__A, (__v2df)__B); 267 } 268 269 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 270 _mm_mul_pd (__m128d __A, __m128d __B) 271 { 272 return (__m128d) ((__v2df)__A * (__v2df)__B); 273 } 274 275 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 276 _mm_mul_sd (__m128d __A, __m128d __B) 277 { 278 return (__m128d)__builtin_ia32_mulsd ((__v2df)__A, (__v2df)__B); 279 } 280 281 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 282 _mm_div_pd (__m128d __A, __m128d __B) 283 { 284 return (__m128d) ((__v2df)__A / (__v2df)__B); 285 } 286 287 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 288 _mm_div_sd (__m128d __A, __m128d __B) 289 { 290 return (__m128d)__builtin_ia32_divsd ((__v2df)__A, (__v2df)__B); 291 } 292 293 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 294 _mm_sqrt_pd (__m128d __A) 295 { 296 return (__m128d)__builtin_ia32_sqrtpd ((__v2df)__A); 297 } 298 299 /* Return pair {sqrt (B[0]), A[1]}. */ 300 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 301 _mm_sqrt_sd (__m128d __A, __m128d __B) 302 { 303 __v2df __tmp = __builtin_ia32_movsd ((__v2df)__A, (__v2df)__B); 304 return (__m128d)__builtin_ia32_sqrtsd ((__v2df)__tmp); 305 } 306 307 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 308 _mm_min_pd (__m128d __A, __m128d __B) 309 { 310 return (__m128d)__builtin_ia32_minpd ((__v2df)__A, (__v2df)__B); 311 } 312 313 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 314 _mm_min_sd (__m128d __A, __m128d __B) 315 { 316 return (__m128d)__builtin_ia32_minsd ((__v2df)__A, (__v2df)__B); 317 } 318 319 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 320 _mm_max_pd (__m128d __A, __m128d __B) 321 { 322 return (__m128d)__builtin_ia32_maxpd ((__v2df)__A, (__v2df)__B); 323 } 324 325 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 326 _mm_max_sd (__m128d __A, __m128d __B) 327 { 328 return (__m128d)__builtin_ia32_maxsd ((__v2df)__A, (__v2df)__B); 329 } 330 331 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 332 _mm_and_pd (__m128d __A, __m128d __B) 333 { 334 return (__m128d)__builtin_ia32_andpd ((__v2df)__A, (__v2df)__B); 335 } 336 337 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 338 _mm_andnot_pd (__m128d __A, __m128d __B) 339 { 340 return (__m128d)__builtin_ia32_andnpd ((__v2df)__A, (__v2df)__B); 341 } 342 343 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 344 _mm_or_pd (__m128d __A, __m128d __B) 345 { 346 return (__m128d)__builtin_ia32_orpd ((__v2df)__A, (__v2df)__B); 347 } 348 349 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 350 _mm_xor_pd (__m128d __A, __m128d __B) 351 { 352 return (__m128d)__builtin_ia32_xorpd ((__v2df)__A, (__v2df)__B); 353 } 354 355 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 356 _mm_cmpeq_pd (__m128d __A, __m128d __B) 357 { 358 return (__m128d)__builtin_ia32_cmpeqpd ((__v2df)__A, (__v2df)__B); 359 } 360 361 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 362 _mm_cmplt_pd (__m128d __A, __m128d __B) 363 { 364 return (__m128d)__builtin_ia32_cmpltpd ((__v2df)__A, (__v2df)__B); 365 } 366 367 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 368 _mm_cmple_pd (__m128d __A, __m128d __B) 369 { 370 return (__m128d)__builtin_ia32_cmplepd ((__v2df)__A, (__v2df)__B); 371 } 372 373 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 374 _mm_cmpgt_pd (__m128d __A, __m128d __B) 375 { 376 return (__m128d)__builtin_ia32_cmpgtpd ((__v2df)__A, (__v2df)__B); 377 } 378 379 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 380 _mm_cmpge_pd (__m128d __A, __m128d __B) 381 { 382 return (__m128d)__builtin_ia32_cmpgepd ((__v2df)__A, (__v2df)__B); 383 } 384 385 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 386 _mm_cmpneq_pd (__m128d __A, __m128d __B) 387 { 388 return (__m128d)__builtin_ia32_cmpneqpd ((__v2df)__A, (__v2df)__B); 389 } 390 391 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 392 _mm_cmpnlt_pd (__m128d __A, __m128d __B) 393 { 394 return (__m128d)__builtin_ia32_cmpnltpd ((__v2df)__A, (__v2df)__B); 395 } 396 397 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 398 _mm_cmpnle_pd (__m128d __A, __m128d __B) 399 { 400 return (__m128d)__builtin_ia32_cmpnlepd ((__v2df)__A, (__v2df)__B); 401 } 402 403 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 404 _mm_cmpngt_pd (__m128d __A, __m128d __B) 405 { 406 return (__m128d)__builtin_ia32_cmpngtpd ((__v2df)__A, (__v2df)__B); 407 } 408 409 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 410 _mm_cmpnge_pd (__m128d __A, __m128d __B) 411 { 412 return (__m128d)__builtin_ia32_cmpngepd ((__v2df)__A, (__v2df)__B); 413 } 414 415 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 416 _mm_cmpord_pd (__m128d __A, __m128d __B) 417 { 418 return (__m128d)__builtin_ia32_cmpordpd ((__v2df)__A, (__v2df)__B); 419 } 420 421 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 422 _mm_cmpunord_pd (__m128d __A, __m128d __B) 423 { 424 return (__m128d)__builtin_ia32_cmpunordpd ((__v2df)__A, (__v2df)__B); 425 } 426 427 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 428 _mm_cmpeq_sd (__m128d __A, __m128d __B) 429 { 430 return (__m128d)__builtin_ia32_cmpeqsd ((__v2df)__A, (__v2df)__B); 431 } 432 433 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 434 _mm_cmplt_sd (__m128d __A, __m128d __B) 435 { 436 return (__m128d)__builtin_ia32_cmpltsd ((__v2df)__A, (__v2df)__B); 437 } 438 439 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 440 _mm_cmple_sd (__m128d __A, __m128d __B) 441 { 442 return (__m128d)__builtin_ia32_cmplesd ((__v2df)__A, (__v2df)__B); 443 } 444 445 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 446 _mm_cmpgt_sd (__m128d __A, __m128d __B) 447 { 448 return (__m128d) __builtin_ia32_movsd ((__v2df) __A, 449 (__v2df) 450 __builtin_ia32_cmpltsd ((__v2df) __B, 451 (__v2df) 452 __A)); 453 } 454 455 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 456 _mm_cmpge_sd (__m128d __A, __m128d __B) 457 { 458 return (__m128d) __builtin_ia32_movsd ((__v2df) __A, 459 (__v2df) 460 __builtin_ia32_cmplesd ((__v2df) __B, 461 (__v2df) 462 __A)); 463 } 464 465 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 466 _mm_cmpneq_sd (__m128d __A, __m128d __B) 467 { 468 return (__m128d)__builtin_ia32_cmpneqsd ((__v2df)__A, (__v2df)__B); 469 } 470 471 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 472 _mm_cmpnlt_sd (__m128d __A, __m128d __B) 473 { 474 return (__m128d)__builtin_ia32_cmpnltsd ((__v2df)__A, (__v2df)__B); 475 } 476 477 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 478 _mm_cmpnle_sd (__m128d __A, __m128d __B) 479 { 480 return (__m128d)__builtin_ia32_cmpnlesd ((__v2df)__A, (__v2df)__B); 481 } 482 483 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 484 _mm_cmpngt_sd (__m128d __A, __m128d __B) 485 { 486 return (__m128d) __builtin_ia32_movsd ((__v2df) __A, 487 (__v2df) 488 __builtin_ia32_cmpnltsd ((__v2df) __B, 489 (__v2df) 490 __A)); 491 } 492 493 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 494 _mm_cmpnge_sd (__m128d __A, __m128d __B) 495 { 496 return (__m128d) __builtin_ia32_movsd ((__v2df) __A, 497 (__v2df) 498 __builtin_ia32_cmpnlesd ((__v2df) __B, 499 (__v2df) 500 __A)); 501 } 502 503 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 504 _mm_cmpord_sd (__m128d __A, __m128d __B) 505 { 506 return (__m128d)__builtin_ia32_cmpordsd ((__v2df)__A, (__v2df)__B); 507 } 508 509 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 510 _mm_cmpunord_sd (__m128d __A, __m128d __B) 511 { 512 return (__m128d)__builtin_ia32_cmpunordsd ((__v2df)__A, (__v2df)__B); 513 } 514 515 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 516 _mm_comieq_sd (__m128d __A, __m128d __B) 517 { 518 return __builtin_ia32_comisdeq ((__v2df)__A, (__v2df)__B); 519 } 520 521 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 522 _mm_comilt_sd (__m128d __A, __m128d __B) 523 { 524 return __builtin_ia32_comisdlt ((__v2df)__A, (__v2df)__B); 525 } 526 527 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 528 _mm_comile_sd (__m128d __A, __m128d __B) 529 { 530 return __builtin_ia32_comisdle ((__v2df)__A, (__v2df)__B); 531 } 532 533 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 534 _mm_comigt_sd (__m128d __A, __m128d __B) 535 { 536 return __builtin_ia32_comisdgt ((__v2df)__A, (__v2df)__B); 537 } 538 539 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 540 _mm_comige_sd (__m128d __A, __m128d __B) 541 { 542 return __builtin_ia32_comisdge ((__v2df)__A, (__v2df)__B); 543 } 544 545 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 546 _mm_comineq_sd (__m128d __A, __m128d __B) 547 { 548 return __builtin_ia32_comisdneq ((__v2df)__A, (__v2df)__B); 549 } 550 551 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 552 _mm_ucomieq_sd (__m128d __A, __m128d __B) 553 { 554 return __builtin_ia32_ucomisdeq ((__v2df)__A, (__v2df)__B); 555 } 556 557 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 558 _mm_ucomilt_sd (__m128d __A, __m128d __B) 559 { 560 return __builtin_ia32_ucomisdlt ((__v2df)__A, (__v2df)__B); 561 } 562 563 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 564 _mm_ucomile_sd (__m128d __A, __m128d __B) 565 { 566 return __builtin_ia32_ucomisdle ((__v2df)__A, (__v2df)__B); 567 } 568 569 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 570 _mm_ucomigt_sd (__m128d __A, __m128d __B) 571 { 572 return __builtin_ia32_ucomisdgt ((__v2df)__A, (__v2df)__B); 573 } 574 575 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 576 _mm_ucomige_sd (__m128d __A, __m128d __B) 577 { 578 return __builtin_ia32_ucomisdge ((__v2df)__A, (__v2df)__B); 579 } 580 581 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 582 _mm_ucomineq_sd (__m128d __A, __m128d __B) 583 { 584 return __builtin_ia32_ucomisdneq ((__v2df)__A, (__v2df)__B); 585 } 586 587 /* Create a vector of Qi, where i is the element number. */ 588 589 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 590 _mm_set_epi64x (long long __q1, long long __q0) 591 { 592 return __extension__ (__m128i)(__v2di){ __q0, __q1 }; 593 } 594 595 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 596 _mm_set_epi64 (__m64 __q1, __m64 __q0) 597 { 598 return _mm_set_epi64x ((long long)__q1, (long long)__q0); 599 } 600 601 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 602 _mm_set_epi32 (int __q3, int __q2, int __q1, int __q0) 603 { 604 return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 }; 605 } 606 607 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 608 _mm_set_epi16 (short __q7, short __q6, short __q5, short __q4, 609 short __q3, short __q2, short __q1, short __q0) 610 { 611 return __extension__ (__m128i)(__v8hi){ 612 __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 }; 613 } 614 615 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 616 _mm_set_epi8 (char __q15, char __q14, char __q13, char __q12, 617 char __q11, char __q10, char __q09, char __q08, 618 char __q07, char __q06, char __q05, char __q04, 619 char __q03, char __q02, char __q01, char __q00) 620 { 621 return __extension__ (__m128i)(__v16qi){ 622 __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07, 623 __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15 624 }; 625 } 626 627 /* Set all of the elements of the vector to A. */ 628 629 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 630 _mm_set1_epi64x (long long __A) 631 { 632 return _mm_set_epi64x (__A, __A); 633 } 634 635 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 636 _mm_set1_epi64 (__m64 __A) 637 { 638 return _mm_set_epi64 (__A, __A); 639 } 640 641 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 642 _mm_set1_epi32 (int __A) 643 { 644 return _mm_set_epi32 (__A, __A, __A, __A); 645 } 646 647 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 648 _mm_set1_epi16 (short __A) 649 { 650 return _mm_set_epi16 (__A, __A, __A, __A, __A, __A, __A, __A); 651 } 652 653 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 654 _mm_set1_epi8 (char __A) 655 { 656 return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A, 657 __A, __A, __A, __A, __A, __A, __A, __A); 658 } 659 660 /* Create a vector of Qi, where i is the element number. 661 The parameter order is reversed from the _mm_set_epi* functions. */ 662 663 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 664 _mm_setr_epi64 (__m64 __q0, __m64 __q1) 665 { 666 return _mm_set_epi64 (__q1, __q0); 667 } 668 669 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 670 _mm_setr_epi32 (int __q0, int __q1, int __q2, int __q3) 671 { 672 return _mm_set_epi32 (__q3, __q2, __q1, __q0); 673 } 674 675 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 676 _mm_setr_epi16 (short __q0, short __q1, short __q2, short __q3, 677 short __q4, short __q5, short __q6, short __q7) 678 { 679 return _mm_set_epi16 (__q7, __q6, __q5, __q4, __q3, __q2, __q1, __q0); 680 } 681 682 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 683 _mm_setr_epi8 (char __q00, char __q01, char __q02, char __q03, 684 char __q04, char __q05, char __q06, char __q07, 685 char __q08, char __q09, char __q10, char __q11, 686 char __q12, char __q13, char __q14, char __q15) 687 { 688 return _mm_set_epi8 (__q15, __q14, __q13, __q12, __q11, __q10, __q09, __q08, 689 __q07, __q06, __q05, __q04, __q03, __q02, __q01, __q00); 690 } 691 692 /* Create a vector with element 0 as *P and the rest zero. */ 693 694 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 695 _mm_load_si128 (__m128i const *__P) 696 { 697 return *__P; 698 } 699 700 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 701 _mm_loadu_si128 (__m128i_u const *__P) 702 { 703 return *__P; 704 } 705 706 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 707 _mm_loadl_epi64 (__m128i_u const *__P) 708 { 709 return _mm_set_epi64 ((__m64)0LL, *(__m64_u *)__P); 710 } 711 712 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 713 _mm_store_si128 (__m128i *__P, __m128i __B) 714 { 715 *__P = __B; 716 } 717 718 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 719 _mm_storeu_si128 (__m128i_u *__P, __m128i __B) 720 { 721 *__P = __B; 722 } 723 724 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 725 _mm_storel_epi64 (__m128i_u *__P, __m128i __B) 726 { 727 *(__m64_u *)__P = (__m64) ((__v2di)__B)[0]; 728 } 729 730 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 731 _mm_movepi64_pi64 (__m128i __B) 732 { 733 return (__m64) ((__v2di)__B)[0]; 734 } 735 736 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 737 _mm_movpi64_epi64 (__m64 __A) 738 { 739 return _mm_set_epi64 ((__m64)0LL, __A); 740 } 741 742 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 743 _mm_move_epi64 (__m128i __A) 744 { 745 return (__m128i)__builtin_ia32_movq128 ((__v2di) __A); 746 } 747 748 /* Create an undefined vector. */ 749 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 750 _mm_undefined_si128 (void) 751 { 752 __m128i __Y = __Y; 753 return __Y; 754 } 755 756 /* Create a vector of zeros. */ 757 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 758 _mm_setzero_si128 (void) 759 { 760 return __extension__ (__m128i)(__v4si){ 0, 0, 0, 0 }; 761 } 762 763 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 764 _mm_cvtepi32_pd (__m128i __A) 765 { 766 return (__m128d)__builtin_ia32_cvtdq2pd ((__v4si) __A); 767 } 768 769 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 770 _mm_cvtepi32_ps (__m128i __A) 771 { 772 return (__m128)__builtin_ia32_cvtdq2ps ((__v4si) __A); 773 } 774 775 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 776 _mm_cvtpd_epi32 (__m128d __A) 777 { 778 return (__m128i)__builtin_ia32_cvtpd2dq ((__v2df) __A); 779 } 780 781 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 782 _mm_cvtpd_pi32 (__m128d __A) 783 { 784 return (__m64)__builtin_ia32_cvtpd2pi ((__v2df) __A); 785 } 786 787 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 788 _mm_cvtpd_ps (__m128d __A) 789 { 790 return (__m128)__builtin_ia32_cvtpd2ps ((__v2df) __A); 791 } 792 793 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 794 _mm_cvttpd_epi32 (__m128d __A) 795 { 796 return (__m128i)__builtin_ia32_cvttpd2dq ((__v2df) __A); 797 } 798 799 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 800 _mm_cvttpd_pi32 (__m128d __A) 801 { 802 return (__m64)__builtin_ia32_cvttpd2pi ((__v2df) __A); 803 } 804 805 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 806 _mm_cvtpi32_pd (__m64 __A) 807 { 808 return (__m128d)__builtin_ia32_cvtpi2pd ((__v2si) __A); 809 } 810 811 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 812 _mm_cvtps_epi32 (__m128 __A) 813 { 814 return (__m128i)__builtin_ia32_cvtps2dq ((__v4sf) __A); 815 } 816 817 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 818 _mm_cvttps_epi32 (__m128 __A) 819 { 820 return (__m128i)__builtin_ia32_cvttps2dq ((__v4sf) __A); 821 } 822 823 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 824 _mm_cvtps_pd (__m128 __A) 825 { 826 return (__m128d)__builtin_ia32_cvtps2pd ((__v4sf) __A); 827 } 828 829 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 830 _mm_cvtsd_si32 (__m128d __A) 831 { 832 return __builtin_ia32_cvtsd2si ((__v2df) __A); 833 } 834 835 #ifdef __x86_64__ 836 /* Intel intrinsic. */ 837 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 838 _mm_cvtsd_si64 (__m128d __A) 839 { 840 return __builtin_ia32_cvtsd2si64 ((__v2df) __A); 841 } 842 843 /* Microsoft intrinsic. */ 844 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 845 _mm_cvtsd_si64x (__m128d __A) 846 { 847 return __builtin_ia32_cvtsd2si64 ((__v2df) __A); 848 } 849 #endif 850 851 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 852 _mm_cvttsd_si32 (__m128d __A) 853 { 854 return __builtin_ia32_cvttsd2si ((__v2df) __A); 855 } 856 857 #ifdef __x86_64__ 858 /* Intel intrinsic. */ 859 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 860 _mm_cvttsd_si64 (__m128d __A) 861 { 862 return __builtin_ia32_cvttsd2si64 ((__v2df) __A); 863 } 864 865 /* Microsoft intrinsic. */ 866 extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 867 _mm_cvttsd_si64x (__m128d __A) 868 { 869 return __builtin_ia32_cvttsd2si64 ((__v2df) __A); 870 } 871 #endif 872 873 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 874 _mm_cvtsd_ss (__m128 __A, __m128d __B) 875 { 876 return (__m128)__builtin_ia32_cvtsd2ss ((__v4sf) __A, (__v2df) __B); 877 } 878 879 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 880 _mm_cvtsi32_sd (__m128d __A, int __B) 881 { 882 return (__m128d)__builtin_ia32_cvtsi2sd ((__v2df) __A, __B); 883 } 884 885 #ifdef __x86_64__ 886 /* Intel intrinsic. */ 887 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 888 _mm_cvtsi64_sd (__m128d __A, long long __B) 889 { 890 return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B); 891 } 892 893 /* Microsoft intrinsic. */ 894 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 895 _mm_cvtsi64x_sd (__m128d __A, long long __B) 896 { 897 return (__m128d)__builtin_ia32_cvtsi642sd ((__v2df) __A, __B); 898 } 899 #endif 900 901 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 902 _mm_cvtss_sd (__m128d __A, __m128 __B) 903 { 904 return (__m128d)__builtin_ia32_cvtss2sd ((__v2df) __A, (__v4sf)__B); 905 } 906 907 #ifdef __OPTIMIZE__ 908 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 909 _mm_shuffle_pd(__m128d __A, __m128d __B, const int __mask) 910 { 911 return (__m128d)__builtin_ia32_shufpd ((__v2df)__A, (__v2df)__B, __mask); 912 } 913 #else 914 #define _mm_shuffle_pd(A, B, N) \ 915 ((__m128d)__builtin_ia32_shufpd ((__v2df)(__m128d)(A), \ 916 (__v2df)(__m128d)(B), (int)(N))) 917 #endif 918 919 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 920 _mm_unpackhi_pd (__m128d __A, __m128d __B) 921 { 922 return (__m128d)__builtin_ia32_unpckhpd ((__v2df)__A, (__v2df)__B); 923 } 924 925 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 926 _mm_unpacklo_pd (__m128d __A, __m128d __B) 927 { 928 return (__m128d)__builtin_ia32_unpcklpd ((__v2df)__A, (__v2df)__B); 929 } 930 931 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 932 _mm_loadh_pd (__m128d __A, double const *__B) 933 { 934 return (__m128d)__builtin_ia32_loadhpd ((__v2df)__A, __B); 935 } 936 937 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 938 _mm_loadl_pd (__m128d __A, double const *__B) 939 { 940 return (__m128d)__builtin_ia32_loadlpd ((__v2df)__A, __B); 941 } 942 943 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 944 _mm_movemask_pd (__m128d __A) 945 { 946 return __builtin_ia32_movmskpd ((__v2df)__A); 947 } 948 949 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 950 _mm_packs_epi16 (__m128i __A, __m128i __B) 951 { 952 return (__m128i)__builtin_ia32_packsswb128 ((__v8hi)__A, (__v8hi)__B); 953 } 954 955 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 956 _mm_packs_epi32 (__m128i __A, __m128i __B) 957 { 958 return (__m128i)__builtin_ia32_packssdw128 ((__v4si)__A, (__v4si)__B); 959 } 960 961 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 962 _mm_packus_epi16 (__m128i __A, __m128i __B) 963 { 964 return (__m128i)__builtin_ia32_packuswb128 ((__v8hi)__A, (__v8hi)__B); 965 } 966 967 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 968 _mm_unpackhi_epi8 (__m128i __A, __m128i __B) 969 { 970 return (__m128i)__builtin_ia32_punpckhbw128 ((__v16qi)__A, (__v16qi)__B); 971 } 972 973 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 974 _mm_unpackhi_epi16 (__m128i __A, __m128i __B) 975 { 976 return (__m128i)__builtin_ia32_punpckhwd128 ((__v8hi)__A, (__v8hi)__B); 977 } 978 979 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 980 _mm_unpackhi_epi32 (__m128i __A, __m128i __B) 981 { 982 return (__m128i)__builtin_ia32_punpckhdq128 ((__v4si)__A, (__v4si)__B); 983 } 984 985 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 986 _mm_unpackhi_epi64 (__m128i __A, __m128i __B) 987 { 988 return (__m128i)__builtin_ia32_punpckhqdq128 ((__v2di)__A, (__v2di)__B); 989 } 990 991 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 992 _mm_unpacklo_epi8 (__m128i __A, __m128i __B) 993 { 994 return (__m128i)__builtin_ia32_punpcklbw128 ((__v16qi)__A, (__v16qi)__B); 995 } 996 997 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 998 _mm_unpacklo_epi16 (__m128i __A, __m128i __B) 999 { 1000 return (__m128i)__builtin_ia32_punpcklwd128 ((__v8hi)__A, (__v8hi)__B); 1001 } 1002 1003 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1004 _mm_unpacklo_epi32 (__m128i __A, __m128i __B) 1005 { 1006 return (__m128i)__builtin_ia32_punpckldq128 ((__v4si)__A, (__v4si)__B); 1007 } 1008 1009 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1010 _mm_unpacklo_epi64 (__m128i __A, __m128i __B) 1011 { 1012 return (__m128i)__builtin_ia32_punpcklqdq128 ((__v2di)__A, (__v2di)__B); 1013 } 1014 1015 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1016 _mm_add_epi8 (__m128i __A, __m128i __B) 1017 { 1018 return (__m128i) ((__v16qu)__A + (__v16qu)__B); 1019 } 1020 1021 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1022 _mm_add_epi16 (__m128i __A, __m128i __B) 1023 { 1024 return (__m128i) ((__v8hu)__A + (__v8hu)__B); 1025 } 1026 1027 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1028 _mm_add_epi32 (__m128i __A, __m128i __B) 1029 { 1030 return (__m128i) ((__v4su)__A + (__v4su)__B); 1031 } 1032 1033 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1034 _mm_add_epi64 (__m128i __A, __m128i __B) 1035 { 1036 return (__m128i) ((__v2du)__A + (__v2du)__B); 1037 } 1038 1039 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1040 _mm_adds_epi8 (__m128i __A, __m128i __B) 1041 { 1042 return (__m128i)__builtin_ia32_paddsb128 ((__v16qi)__A, (__v16qi)__B); 1043 } 1044 1045 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1046 _mm_adds_epi16 (__m128i __A, __m128i __B) 1047 { 1048 return (__m128i)__builtin_ia32_paddsw128 ((__v8hi)__A, (__v8hi)__B); 1049 } 1050 1051 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1052 _mm_adds_epu8 (__m128i __A, __m128i __B) 1053 { 1054 return (__m128i)__builtin_ia32_paddusb128 ((__v16qi)__A, (__v16qi)__B); 1055 } 1056 1057 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1058 _mm_adds_epu16 (__m128i __A, __m128i __B) 1059 { 1060 return (__m128i)__builtin_ia32_paddusw128 ((__v8hi)__A, (__v8hi)__B); 1061 } 1062 1063 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1064 _mm_sub_epi8 (__m128i __A, __m128i __B) 1065 { 1066 return (__m128i) ((__v16qu)__A - (__v16qu)__B); 1067 } 1068 1069 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1070 _mm_sub_epi16 (__m128i __A, __m128i __B) 1071 { 1072 return (__m128i) ((__v8hu)__A - (__v8hu)__B); 1073 } 1074 1075 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1076 _mm_sub_epi32 (__m128i __A, __m128i __B) 1077 { 1078 return (__m128i) ((__v4su)__A - (__v4su)__B); 1079 } 1080 1081 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1082 _mm_sub_epi64 (__m128i __A, __m128i __B) 1083 { 1084 return (__m128i) ((__v2du)__A - (__v2du)__B); 1085 } 1086 1087 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1088 _mm_subs_epi8 (__m128i __A, __m128i __B) 1089 { 1090 return (__m128i)__builtin_ia32_psubsb128 ((__v16qi)__A, (__v16qi)__B); 1091 } 1092 1093 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1094 _mm_subs_epi16 (__m128i __A, __m128i __B) 1095 { 1096 return (__m128i)__builtin_ia32_psubsw128 ((__v8hi)__A, (__v8hi)__B); 1097 } 1098 1099 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1100 _mm_subs_epu8 (__m128i __A, __m128i __B) 1101 { 1102 return (__m128i)__builtin_ia32_psubusb128 ((__v16qi)__A, (__v16qi)__B); 1103 } 1104 1105 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1106 _mm_subs_epu16 (__m128i __A, __m128i __B) 1107 { 1108 return (__m128i)__builtin_ia32_psubusw128 ((__v8hi)__A, (__v8hi)__B); 1109 } 1110 1111 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1112 _mm_madd_epi16 (__m128i __A, __m128i __B) 1113 { 1114 return (__m128i)__builtin_ia32_pmaddwd128 ((__v8hi)__A, (__v8hi)__B); 1115 } 1116 1117 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1118 _mm_mulhi_epi16 (__m128i __A, __m128i __B) 1119 { 1120 return (__m128i)__builtin_ia32_pmulhw128 ((__v8hi)__A, (__v8hi)__B); 1121 } 1122 1123 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1124 _mm_mullo_epi16 (__m128i __A, __m128i __B) 1125 { 1126 return (__m128i) ((__v8hu)__A * (__v8hu)__B); 1127 } 1128 1129 extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1130 _mm_mul_su32 (__m64 __A, __m64 __B) 1131 { 1132 return (__m64)__builtin_ia32_pmuludq ((__v2si)__A, (__v2si)__B); 1133 } 1134 1135 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1136 _mm_mul_epu32 (__m128i __A, __m128i __B) 1137 { 1138 return (__m128i)__builtin_ia32_pmuludq128 ((__v4si)__A, (__v4si)__B); 1139 } 1140 1141 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1142 _mm_slli_epi16 (__m128i __A, int __B) 1143 { 1144 return (__m128i)__builtin_ia32_psllwi128 ((__v8hi)__A, __B); 1145 } 1146 1147 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1148 _mm_slli_epi32 (__m128i __A, int __B) 1149 { 1150 return (__m128i)__builtin_ia32_pslldi128 ((__v4si)__A, __B); 1151 } 1152 1153 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1154 _mm_slli_epi64 (__m128i __A, int __B) 1155 { 1156 return (__m128i)__builtin_ia32_psllqi128 ((__v2di)__A, __B); 1157 } 1158 1159 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1160 _mm_srai_epi16 (__m128i __A, int __B) 1161 { 1162 return (__m128i)__builtin_ia32_psrawi128 ((__v8hi)__A, __B); 1163 } 1164 1165 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1166 _mm_srai_epi32 (__m128i __A, int __B) 1167 { 1168 return (__m128i)__builtin_ia32_psradi128 ((__v4si)__A, __B); 1169 } 1170 1171 #ifdef __OPTIMIZE__ 1172 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1173 _mm_bsrli_si128 (__m128i __A, const int __N) 1174 { 1175 return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8); 1176 } 1177 1178 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1179 _mm_bslli_si128 (__m128i __A, const int __N) 1180 { 1181 return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8); 1182 } 1183 1184 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1185 _mm_srli_si128 (__m128i __A, const int __N) 1186 { 1187 return (__m128i)__builtin_ia32_psrldqi128 (__A, __N * 8); 1188 } 1189 1190 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1191 _mm_slli_si128 (__m128i __A, const int __N) 1192 { 1193 return (__m128i)__builtin_ia32_pslldqi128 (__A, __N * 8); 1194 } 1195 #else 1196 #define _mm_bsrli_si128(A, N) \ 1197 ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8)) 1198 #define _mm_bslli_si128(A, N) \ 1199 ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8)) 1200 #define _mm_srli_si128(A, N) \ 1201 ((__m128i)__builtin_ia32_psrldqi128 ((__m128i)(A), (int)(N) * 8)) 1202 #define _mm_slli_si128(A, N) \ 1203 ((__m128i)__builtin_ia32_pslldqi128 ((__m128i)(A), (int)(N) * 8)) 1204 #endif 1205 1206 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1207 _mm_srli_epi16 (__m128i __A, int __B) 1208 { 1209 return (__m128i)__builtin_ia32_psrlwi128 ((__v8hi)__A, __B); 1210 } 1211 1212 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1213 _mm_srli_epi32 (__m128i __A, int __B) 1214 { 1215 return (__m128i)__builtin_ia32_psrldi128 ((__v4si)__A, __B); 1216 } 1217 1218 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1219 _mm_srli_epi64 (__m128i __A, int __B) 1220 { 1221 return (__m128i)__builtin_ia32_psrlqi128 ((__v2di)__A, __B); 1222 } 1223 1224 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1225 _mm_sll_epi16 (__m128i __A, __m128i __B) 1226 { 1227 return (__m128i)__builtin_ia32_psllw128((__v8hi)__A, (__v8hi)__B); 1228 } 1229 1230 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1231 _mm_sll_epi32 (__m128i __A, __m128i __B) 1232 { 1233 return (__m128i)__builtin_ia32_pslld128((__v4si)__A, (__v4si)__B); 1234 } 1235 1236 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1237 _mm_sll_epi64 (__m128i __A, __m128i __B) 1238 { 1239 return (__m128i)__builtin_ia32_psllq128((__v2di)__A, (__v2di)__B); 1240 } 1241 1242 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1243 _mm_sra_epi16 (__m128i __A, __m128i __B) 1244 { 1245 return (__m128i)__builtin_ia32_psraw128 ((__v8hi)__A, (__v8hi)__B); 1246 } 1247 1248 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1249 _mm_sra_epi32 (__m128i __A, __m128i __B) 1250 { 1251 return (__m128i)__builtin_ia32_psrad128 ((__v4si)__A, (__v4si)__B); 1252 } 1253 1254 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1255 _mm_srl_epi16 (__m128i __A, __m128i __B) 1256 { 1257 return (__m128i)__builtin_ia32_psrlw128 ((__v8hi)__A, (__v8hi)__B); 1258 } 1259 1260 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1261 _mm_srl_epi32 (__m128i __A, __m128i __B) 1262 { 1263 return (__m128i)__builtin_ia32_psrld128 ((__v4si)__A, (__v4si)__B); 1264 } 1265 1266 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1267 _mm_srl_epi64 (__m128i __A, __m128i __B) 1268 { 1269 return (__m128i)__builtin_ia32_psrlq128 ((__v2di)__A, (__v2di)__B); 1270 } 1271 1272 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1273 _mm_and_si128 (__m128i __A, __m128i __B) 1274 { 1275 return (__m128i) ((__v2du)__A & (__v2du)__B); 1276 } 1277 1278 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1279 _mm_andnot_si128 (__m128i __A, __m128i __B) 1280 { 1281 return (__m128i)__builtin_ia32_pandn128 ((__v2di)__A, (__v2di)__B); 1282 } 1283 1284 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1285 _mm_or_si128 (__m128i __A, __m128i __B) 1286 { 1287 return (__m128i) ((__v2du)__A | (__v2du)__B); 1288 } 1289 1290 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1291 _mm_xor_si128 (__m128i __A, __m128i __B) 1292 { 1293 return (__m128i) ((__v2du)__A ^ (__v2du)__B); 1294 } 1295 1296 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1297 _mm_cmpeq_epi8 (__m128i __A, __m128i __B) 1298 { 1299 return (__m128i) ((__v16qs)__A == (__v16qs)__B); 1300 } 1301 1302 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1303 _mm_cmpeq_epi16 (__m128i __A, __m128i __B) 1304 { 1305 return (__m128i) ((__v8hi)__A == (__v8hi)__B); 1306 } 1307 1308 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1309 _mm_cmpeq_epi32 (__m128i __A, __m128i __B) 1310 { 1311 return (__m128i) ((__v4si)__A == (__v4si)__B); 1312 } 1313 1314 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1315 _mm_cmplt_epi8 (__m128i __A, __m128i __B) 1316 { 1317 return (__m128i) ((__v16qs)__A < (__v16qs)__B); 1318 } 1319 1320 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1321 _mm_cmplt_epi16 (__m128i __A, __m128i __B) 1322 { 1323 return (__m128i) ((__v8hi)__A < (__v8hi)__B); 1324 } 1325 1326 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1327 _mm_cmplt_epi32 (__m128i __A, __m128i __B) 1328 { 1329 return (__m128i) ((__v4si)__A < (__v4si)__B); 1330 } 1331 1332 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1333 _mm_cmpgt_epi8 (__m128i __A, __m128i __B) 1334 { 1335 return (__m128i) ((__v16qs)__A > (__v16qs)__B); 1336 } 1337 1338 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1339 _mm_cmpgt_epi16 (__m128i __A, __m128i __B) 1340 { 1341 return (__m128i) ((__v8hi)__A > (__v8hi)__B); 1342 } 1343 1344 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1345 _mm_cmpgt_epi32 (__m128i __A, __m128i __B) 1346 { 1347 return (__m128i) ((__v4si)__A > (__v4si)__B); 1348 } 1349 1350 #ifdef __OPTIMIZE__ 1351 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1352 _mm_extract_epi16 (__m128i const __A, int const __N) 1353 { 1354 return (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)__A, __N); 1355 } 1356 1357 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1358 _mm_insert_epi16 (__m128i const __A, int const __D, int const __N) 1359 { 1360 return (__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)__A, __D, __N); 1361 } 1362 #else 1363 #define _mm_extract_epi16(A, N) \ 1364 ((int) (unsigned short) __builtin_ia32_vec_ext_v8hi ((__v8hi)(__m128i)(A), (int)(N))) 1365 #define _mm_insert_epi16(A, D, N) \ 1366 ((__m128i) __builtin_ia32_vec_set_v8hi ((__v8hi)(__m128i)(A), \ 1367 (int)(D), (int)(N))) 1368 #endif 1369 1370 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1371 _mm_max_epi16 (__m128i __A, __m128i __B) 1372 { 1373 return (__m128i)__builtin_ia32_pmaxsw128 ((__v8hi)__A, (__v8hi)__B); 1374 } 1375 1376 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1377 _mm_max_epu8 (__m128i __A, __m128i __B) 1378 { 1379 return (__m128i)__builtin_ia32_pmaxub128 ((__v16qi)__A, (__v16qi)__B); 1380 } 1381 1382 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1383 _mm_min_epi16 (__m128i __A, __m128i __B) 1384 { 1385 return (__m128i)__builtin_ia32_pminsw128 ((__v8hi)__A, (__v8hi)__B); 1386 } 1387 1388 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1389 _mm_min_epu8 (__m128i __A, __m128i __B) 1390 { 1391 return (__m128i)__builtin_ia32_pminub128 ((__v16qi)__A, (__v16qi)__B); 1392 } 1393 1394 extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1395 _mm_movemask_epi8 (__m128i __A) 1396 { 1397 return __builtin_ia32_pmovmskb128 ((__v16qi)__A); 1398 } 1399 1400 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1401 _mm_mulhi_epu16 (__m128i __A, __m128i __B) 1402 { 1403 return (__m128i)__builtin_ia32_pmulhuw128 ((__v8hi)__A, (__v8hi)__B); 1404 } 1405 1406 #ifdef __OPTIMIZE__ 1407 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1408 _mm_shufflehi_epi16 (__m128i __A, const int __mask) 1409 { 1410 return (__m128i)__builtin_ia32_pshufhw ((__v8hi)__A, __mask); 1411 } 1412 1413 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1414 _mm_shufflelo_epi16 (__m128i __A, const int __mask) 1415 { 1416 return (__m128i)__builtin_ia32_pshuflw ((__v8hi)__A, __mask); 1417 } 1418 1419 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1420 _mm_shuffle_epi32 (__m128i __A, const int __mask) 1421 { 1422 return (__m128i)__builtin_ia32_pshufd ((__v4si)__A, __mask); 1423 } 1424 #else 1425 #define _mm_shufflehi_epi16(A, N) \ 1426 ((__m128i)__builtin_ia32_pshufhw ((__v8hi)(__m128i)(A), (int)(N))) 1427 #define _mm_shufflelo_epi16(A, N) \ 1428 ((__m128i)__builtin_ia32_pshuflw ((__v8hi)(__m128i)(A), (int)(N))) 1429 #define _mm_shuffle_epi32(A, N) \ 1430 ((__m128i)__builtin_ia32_pshufd ((__v4si)(__m128i)(A), (int)(N))) 1431 #endif 1432 1433 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1434 _mm_maskmoveu_si128 (__m128i __A, __m128i __B, char *__C) 1435 { 1436 __builtin_ia32_maskmovdqu ((__v16qi)__A, (__v16qi)__B, __C); 1437 } 1438 1439 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1440 _mm_avg_epu8 (__m128i __A, __m128i __B) 1441 { 1442 return (__m128i)__builtin_ia32_pavgb128 ((__v16qi)__A, (__v16qi)__B); 1443 } 1444 1445 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1446 _mm_avg_epu16 (__m128i __A, __m128i __B) 1447 { 1448 return (__m128i)__builtin_ia32_pavgw128 ((__v8hi)__A, (__v8hi)__B); 1449 } 1450 1451 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1452 _mm_sad_epu8 (__m128i __A, __m128i __B) 1453 { 1454 return (__m128i)__builtin_ia32_psadbw128 ((__v16qi)__A, (__v16qi)__B); 1455 } 1456 1457 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1458 _mm_stream_si32 (int *__A, int __B) 1459 { 1460 __builtin_ia32_movnti (__A, __B); 1461 } 1462 1463 #ifdef __x86_64__ 1464 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1465 _mm_stream_si64 (long long int *__A, long long int __B) 1466 { 1467 __builtin_ia32_movnti64 (__A, __B); 1468 } 1469 #endif 1470 1471 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1472 _mm_stream_si128 (__m128i *__A, __m128i __B) 1473 { 1474 __builtin_ia32_movntdq ((__v2di *)__A, (__v2di)__B); 1475 } 1476 1477 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1478 _mm_stream_pd (double *__A, __m128d __B) 1479 { 1480 __builtin_ia32_movntpd (__A, (__v2df)__B); 1481 } 1482 1483 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1484 _mm_clflush (void const *__A) 1485 { 1486 __builtin_ia32_clflush (__A); 1487 } 1488 1489 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1490 _mm_lfence (void) 1491 { 1492 __builtin_ia32_lfence (); 1493 } 1494 1495 extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1496 _mm_mfence (void) 1497 { 1498 __builtin_ia32_mfence (); 1499 } 1500 1501 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1502 _mm_cvtsi32_si128 (int __A) 1503 { 1504 return _mm_set_epi32 (0, 0, 0, __A); 1505 } 1506 1507 #ifdef __x86_64__ 1508 /* Intel intrinsic. */ 1509 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1510 _mm_cvtsi64_si128 (long long __A) 1511 { 1512 return _mm_set_epi64x (0, __A); 1513 } 1514 1515 /* Microsoft intrinsic. */ 1516 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1517 _mm_cvtsi64x_si128 (long long __A) 1518 { 1519 return _mm_set_epi64x (0, __A); 1520 } 1521 #endif 1522 1523 /* Casts between various SP, DP, INT vector types. Note that these do no 1524 conversion of values, they just change the type. */ 1525 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1526 _mm_castpd_ps(__m128d __A) 1527 { 1528 return (__m128) __A; 1529 } 1530 1531 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1532 _mm_castpd_si128(__m128d __A) 1533 { 1534 return (__m128i) __A; 1535 } 1536 1537 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1538 _mm_castps_pd(__m128 __A) 1539 { 1540 return (__m128d) __A; 1541 } 1542 1543 extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1544 _mm_castps_si128(__m128 __A) 1545 { 1546 return (__m128i) __A; 1547 } 1548 1549 extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1550 _mm_castsi128_ps(__m128i __A) 1551 { 1552 return (__m128) __A; 1553 } 1554 1555 extern __inline __m128d __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1556 _mm_castsi128_pd(__m128i __A) 1557 { 1558 return (__m128d) __A; 1559 } 1560 1561 #ifdef __DISABLE_SSE2__ 1562 #undef __DISABLE_SSE2__ 1563 #pragma GCC pop_options 1564 #endif /* __DISABLE_SSE2__ */ 1565 1566 #endif /* _EMMINTRIN_H_INCLUDED */ 1567