1/* Copyright (C) 2002-2014 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 _XMMINTRIN_H_INCLUDED 28#define _XMMINTRIN_H_INCLUDED 29 30/* We need type definitions from the MMX header file. */ 31#include <mmintrin.h> 32 33/* Get _mm_malloc () and _mm_free (). */ 34#include <mm_malloc.h> 35 36/* Constants for use with _mm_prefetch. */ 37enum _mm_hint 38{ 39 /* _MM_HINT_ET is _MM_HINT_T with set 3rd bit. */ 40 _MM_HINT_ET0 = 7, 41 _MM_HINT_ET1 = 6, 42 _MM_HINT_T0 = 3, 43 _MM_HINT_T1 = 2, 44 _MM_HINT_T2 = 1, 45 _MM_HINT_NTA = 0 46}; 47 48/* Loads one cache line from address P to a location "closer" to the 49 processor. The selector I specifies the type of prefetch operation. */ 50#ifdef __OPTIMIZE__ 51extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 52_mm_prefetch (const void *__P, enum _mm_hint __I) 53{ 54 __builtin_prefetch (__P, (__I & 0x4) >> 2, __I & 0x3); 55} 56#else 57#define _mm_prefetch(P, I) \ 58 __builtin_prefetch ((P), ((I & 0x4) >> 2), (I & 0x3)) 59#endif 60 61#ifndef __SSE__ 62#pragma GCC push_options 63#pragma GCC target("sse") 64#define __DISABLE_SSE__ 65#endif /* __SSE__ */ 66 67/* The Intel API is flexible enough that we must allow aliasing with other 68 vector types, and their scalar components. */ 69typedef float __m128 __attribute__ ((__vector_size__ (16), __may_alias__)); 70 71/* Internal data types for implementing the intrinsics. */ 72typedef float __v4sf __attribute__ ((__vector_size__ (16))); 73 74/* Create a selector for use with the SHUFPS instruction. */ 75#define _MM_SHUFFLE(fp3,fp2,fp1,fp0) \ 76 (((fp3) << 6) | ((fp2) << 4) | ((fp1) << 2) | (fp0)) 77 78/* Bits in the MXCSR. */ 79#define _MM_EXCEPT_MASK 0x003f 80#define _MM_EXCEPT_INVALID 0x0001 81#define _MM_EXCEPT_DENORM 0x0002 82#define _MM_EXCEPT_DIV_ZERO 0x0004 83#define _MM_EXCEPT_OVERFLOW 0x0008 84#define _MM_EXCEPT_UNDERFLOW 0x0010 85#define _MM_EXCEPT_INEXACT 0x0020 86 87#define _MM_MASK_MASK 0x1f80 88#define _MM_MASK_INVALID 0x0080 89#define _MM_MASK_DENORM 0x0100 90#define _MM_MASK_DIV_ZERO 0x0200 91#define _MM_MASK_OVERFLOW 0x0400 92#define _MM_MASK_UNDERFLOW 0x0800 93#define _MM_MASK_INEXACT 0x1000 94 95#define _MM_ROUND_MASK 0x6000 96#define _MM_ROUND_NEAREST 0x0000 97#define _MM_ROUND_DOWN 0x2000 98#define _MM_ROUND_UP 0x4000 99#define _MM_ROUND_TOWARD_ZERO 0x6000 100 101#define _MM_FLUSH_ZERO_MASK 0x8000 102#define _MM_FLUSH_ZERO_ON 0x8000 103#define _MM_FLUSH_ZERO_OFF 0x0000 104 105/* Create an undefined vector. */ 106extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 107_mm_undefined_ps (void) 108{ 109 __m128 __Y = __Y; 110 return __Y; 111} 112 113/* Create a vector of zeros. */ 114extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 115_mm_setzero_ps (void) 116{ 117 return __extension__ (__m128){ 0.0f, 0.0f, 0.0f, 0.0f }; 118} 119 120/* Perform the respective operation on the lower SPFP (single-precision 121 floating-point) values of A and B; the upper three SPFP values are 122 passed through from A. */ 123 124extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 125_mm_add_ss (__m128 __A, __m128 __B) 126{ 127 return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B); 128} 129 130extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 131_mm_sub_ss (__m128 __A, __m128 __B) 132{ 133 return (__m128) __builtin_ia32_subss ((__v4sf)__A, (__v4sf)__B); 134} 135 136extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 137_mm_mul_ss (__m128 __A, __m128 __B) 138{ 139 return (__m128) __builtin_ia32_mulss ((__v4sf)__A, (__v4sf)__B); 140} 141 142extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 143_mm_div_ss (__m128 __A, __m128 __B) 144{ 145 return (__m128) __builtin_ia32_divss ((__v4sf)__A, (__v4sf)__B); 146} 147 148extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 149_mm_sqrt_ss (__m128 __A) 150{ 151 return (__m128) __builtin_ia32_sqrtss ((__v4sf)__A); 152} 153 154extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 155_mm_rcp_ss (__m128 __A) 156{ 157 return (__m128) __builtin_ia32_rcpss ((__v4sf)__A); 158} 159 160extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 161_mm_rsqrt_ss (__m128 __A) 162{ 163 return (__m128) __builtin_ia32_rsqrtss ((__v4sf)__A); 164} 165 166extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 167_mm_min_ss (__m128 __A, __m128 __B) 168{ 169 return (__m128) __builtin_ia32_minss ((__v4sf)__A, (__v4sf)__B); 170} 171 172extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 173_mm_max_ss (__m128 __A, __m128 __B) 174{ 175 return (__m128) __builtin_ia32_maxss ((__v4sf)__A, (__v4sf)__B); 176} 177 178/* Perform the respective operation on the four SPFP values in A and B. */ 179 180extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 181_mm_add_ps (__m128 __A, __m128 __B) 182{ 183 return (__m128) __builtin_ia32_addps ((__v4sf)__A, (__v4sf)__B); 184} 185 186extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 187_mm_sub_ps (__m128 __A, __m128 __B) 188{ 189 return (__m128) __builtin_ia32_subps ((__v4sf)__A, (__v4sf)__B); 190} 191 192extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 193_mm_mul_ps (__m128 __A, __m128 __B) 194{ 195 return (__m128) __builtin_ia32_mulps ((__v4sf)__A, (__v4sf)__B); 196} 197 198extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 199_mm_div_ps (__m128 __A, __m128 __B) 200{ 201 return (__m128) __builtin_ia32_divps ((__v4sf)__A, (__v4sf)__B); 202} 203 204extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 205_mm_sqrt_ps (__m128 __A) 206{ 207 return (__m128) __builtin_ia32_sqrtps ((__v4sf)__A); 208} 209 210extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 211_mm_rcp_ps (__m128 __A) 212{ 213 return (__m128) __builtin_ia32_rcpps ((__v4sf)__A); 214} 215 216extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 217_mm_rsqrt_ps (__m128 __A) 218{ 219 return (__m128) __builtin_ia32_rsqrtps ((__v4sf)__A); 220} 221 222extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 223_mm_min_ps (__m128 __A, __m128 __B) 224{ 225 return (__m128) __builtin_ia32_minps ((__v4sf)__A, (__v4sf)__B); 226} 227 228extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 229_mm_max_ps (__m128 __A, __m128 __B) 230{ 231 return (__m128) __builtin_ia32_maxps ((__v4sf)__A, (__v4sf)__B); 232} 233 234/* Perform logical bit-wise operations on 128-bit values. */ 235 236extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 237_mm_and_ps (__m128 __A, __m128 __B) 238{ 239 return __builtin_ia32_andps (__A, __B); 240} 241 242extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 243_mm_andnot_ps (__m128 __A, __m128 __B) 244{ 245 return __builtin_ia32_andnps (__A, __B); 246} 247 248extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 249_mm_or_ps (__m128 __A, __m128 __B) 250{ 251 return __builtin_ia32_orps (__A, __B); 252} 253 254extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 255_mm_xor_ps (__m128 __A, __m128 __B) 256{ 257 return __builtin_ia32_xorps (__A, __B); 258} 259 260/* Perform a comparison on the lower SPFP values of A and B. If the 261 comparison is true, place a mask of all ones in the result, otherwise a 262 mask of zeros. The upper three SPFP values are passed through from A. */ 263 264extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 265_mm_cmpeq_ss (__m128 __A, __m128 __B) 266{ 267 return (__m128) __builtin_ia32_cmpeqss ((__v4sf)__A, (__v4sf)__B); 268} 269 270extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 271_mm_cmplt_ss (__m128 __A, __m128 __B) 272{ 273 return (__m128) __builtin_ia32_cmpltss ((__v4sf)__A, (__v4sf)__B); 274} 275 276extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 277_mm_cmple_ss (__m128 __A, __m128 __B) 278{ 279 return (__m128) __builtin_ia32_cmpless ((__v4sf)__A, (__v4sf)__B); 280} 281 282extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 283_mm_cmpgt_ss (__m128 __A, __m128 __B) 284{ 285 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 286 (__v4sf) 287 __builtin_ia32_cmpltss ((__v4sf) __B, 288 (__v4sf) 289 __A)); 290} 291 292extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 293_mm_cmpge_ss (__m128 __A, __m128 __B) 294{ 295 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 296 (__v4sf) 297 __builtin_ia32_cmpless ((__v4sf) __B, 298 (__v4sf) 299 __A)); 300} 301 302extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 303_mm_cmpneq_ss (__m128 __A, __m128 __B) 304{ 305 return (__m128) __builtin_ia32_cmpneqss ((__v4sf)__A, (__v4sf)__B); 306} 307 308extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 309_mm_cmpnlt_ss (__m128 __A, __m128 __B) 310{ 311 return (__m128) __builtin_ia32_cmpnltss ((__v4sf)__A, (__v4sf)__B); 312} 313 314extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 315_mm_cmpnle_ss (__m128 __A, __m128 __B) 316{ 317 return (__m128) __builtin_ia32_cmpnless ((__v4sf)__A, (__v4sf)__B); 318} 319 320extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 321_mm_cmpngt_ss (__m128 __A, __m128 __B) 322{ 323 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 324 (__v4sf) 325 __builtin_ia32_cmpnltss ((__v4sf) __B, 326 (__v4sf) 327 __A)); 328} 329 330extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 331_mm_cmpnge_ss (__m128 __A, __m128 __B) 332{ 333 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 334 (__v4sf) 335 __builtin_ia32_cmpnless ((__v4sf) __B, 336 (__v4sf) 337 __A)); 338} 339 340extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 341_mm_cmpord_ss (__m128 __A, __m128 __B) 342{ 343 return (__m128) __builtin_ia32_cmpordss ((__v4sf)__A, (__v4sf)__B); 344} 345 346extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 347_mm_cmpunord_ss (__m128 __A, __m128 __B) 348{ 349 return (__m128) __builtin_ia32_cmpunordss ((__v4sf)__A, (__v4sf)__B); 350} 351 352/* Perform a comparison on the four SPFP values of A and B. For each 353 element, if the comparison is true, place a mask of all ones in the 354 result, otherwise a mask of zeros. */ 355 356extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 357_mm_cmpeq_ps (__m128 __A, __m128 __B) 358{ 359 return (__m128) __builtin_ia32_cmpeqps ((__v4sf)__A, (__v4sf)__B); 360} 361 362extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 363_mm_cmplt_ps (__m128 __A, __m128 __B) 364{ 365 return (__m128) __builtin_ia32_cmpltps ((__v4sf)__A, (__v4sf)__B); 366} 367 368extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 369_mm_cmple_ps (__m128 __A, __m128 __B) 370{ 371 return (__m128) __builtin_ia32_cmpleps ((__v4sf)__A, (__v4sf)__B); 372} 373 374extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 375_mm_cmpgt_ps (__m128 __A, __m128 __B) 376{ 377 return (__m128) __builtin_ia32_cmpgtps ((__v4sf)__A, (__v4sf)__B); 378} 379 380extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 381_mm_cmpge_ps (__m128 __A, __m128 __B) 382{ 383 return (__m128) __builtin_ia32_cmpgeps ((__v4sf)__A, (__v4sf)__B); 384} 385 386extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 387_mm_cmpneq_ps (__m128 __A, __m128 __B) 388{ 389 return (__m128) __builtin_ia32_cmpneqps ((__v4sf)__A, (__v4sf)__B); 390} 391 392extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 393_mm_cmpnlt_ps (__m128 __A, __m128 __B) 394{ 395 return (__m128) __builtin_ia32_cmpnltps ((__v4sf)__A, (__v4sf)__B); 396} 397 398extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 399_mm_cmpnle_ps (__m128 __A, __m128 __B) 400{ 401 return (__m128) __builtin_ia32_cmpnleps ((__v4sf)__A, (__v4sf)__B); 402} 403 404extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 405_mm_cmpngt_ps (__m128 __A, __m128 __B) 406{ 407 return (__m128) __builtin_ia32_cmpngtps ((__v4sf)__A, (__v4sf)__B); 408} 409 410extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 411_mm_cmpnge_ps (__m128 __A, __m128 __B) 412{ 413 return (__m128) __builtin_ia32_cmpngeps ((__v4sf)__A, (__v4sf)__B); 414} 415 416extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 417_mm_cmpord_ps (__m128 __A, __m128 __B) 418{ 419 return (__m128) __builtin_ia32_cmpordps ((__v4sf)__A, (__v4sf)__B); 420} 421 422extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 423_mm_cmpunord_ps (__m128 __A, __m128 __B) 424{ 425 return (__m128) __builtin_ia32_cmpunordps ((__v4sf)__A, (__v4sf)__B); 426} 427 428/* Compare the lower SPFP values of A and B and return 1 if true 429 and 0 if false. */ 430 431extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 432_mm_comieq_ss (__m128 __A, __m128 __B) 433{ 434 return __builtin_ia32_comieq ((__v4sf)__A, (__v4sf)__B); 435} 436 437extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 438_mm_comilt_ss (__m128 __A, __m128 __B) 439{ 440 return __builtin_ia32_comilt ((__v4sf)__A, (__v4sf)__B); 441} 442 443extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 444_mm_comile_ss (__m128 __A, __m128 __B) 445{ 446 return __builtin_ia32_comile ((__v4sf)__A, (__v4sf)__B); 447} 448 449extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 450_mm_comigt_ss (__m128 __A, __m128 __B) 451{ 452 return __builtin_ia32_comigt ((__v4sf)__A, (__v4sf)__B); 453} 454 455extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 456_mm_comige_ss (__m128 __A, __m128 __B) 457{ 458 return __builtin_ia32_comige ((__v4sf)__A, (__v4sf)__B); 459} 460 461extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 462_mm_comineq_ss (__m128 __A, __m128 __B) 463{ 464 return __builtin_ia32_comineq ((__v4sf)__A, (__v4sf)__B); 465} 466 467extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 468_mm_ucomieq_ss (__m128 __A, __m128 __B) 469{ 470 return __builtin_ia32_ucomieq ((__v4sf)__A, (__v4sf)__B); 471} 472 473extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 474_mm_ucomilt_ss (__m128 __A, __m128 __B) 475{ 476 return __builtin_ia32_ucomilt ((__v4sf)__A, (__v4sf)__B); 477} 478 479extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 480_mm_ucomile_ss (__m128 __A, __m128 __B) 481{ 482 return __builtin_ia32_ucomile ((__v4sf)__A, (__v4sf)__B); 483} 484 485extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 486_mm_ucomigt_ss (__m128 __A, __m128 __B) 487{ 488 return __builtin_ia32_ucomigt ((__v4sf)__A, (__v4sf)__B); 489} 490 491extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 492_mm_ucomige_ss (__m128 __A, __m128 __B) 493{ 494 return __builtin_ia32_ucomige ((__v4sf)__A, (__v4sf)__B); 495} 496 497extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 498_mm_ucomineq_ss (__m128 __A, __m128 __B) 499{ 500 return __builtin_ia32_ucomineq ((__v4sf)__A, (__v4sf)__B); 501} 502 503/* Convert the lower SPFP value to a 32-bit integer according to the current 504 rounding mode. */ 505extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 506_mm_cvtss_si32 (__m128 __A) 507{ 508 return __builtin_ia32_cvtss2si ((__v4sf) __A); 509} 510 511extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 512_mm_cvt_ss2si (__m128 __A) 513{ 514 return _mm_cvtss_si32 (__A); 515} 516 517#ifdef __x86_64__ 518/* Convert the lower SPFP value to a 32-bit integer according to the 519 current rounding mode. */ 520 521/* Intel intrinsic. */ 522extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 523_mm_cvtss_si64 (__m128 __A) 524{ 525 return __builtin_ia32_cvtss2si64 ((__v4sf) __A); 526} 527 528/* Microsoft intrinsic. */ 529extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 530_mm_cvtss_si64x (__m128 __A) 531{ 532 return __builtin_ia32_cvtss2si64 ((__v4sf) __A); 533} 534#endif 535 536/* Convert the two lower SPFP values to 32-bit integers according to the 537 current rounding mode. Return the integers in packed form. */ 538extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 539_mm_cvtps_pi32 (__m128 __A) 540{ 541 return (__m64) __builtin_ia32_cvtps2pi ((__v4sf) __A); 542} 543 544extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 545_mm_cvt_ps2pi (__m128 __A) 546{ 547 return _mm_cvtps_pi32 (__A); 548} 549 550/* Truncate the lower SPFP value to a 32-bit integer. */ 551extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 552_mm_cvttss_si32 (__m128 __A) 553{ 554 return __builtin_ia32_cvttss2si ((__v4sf) __A); 555} 556 557extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 558_mm_cvtt_ss2si (__m128 __A) 559{ 560 return _mm_cvttss_si32 (__A); 561} 562 563#ifdef __x86_64__ 564/* Truncate the lower SPFP value to a 32-bit integer. */ 565 566/* Intel intrinsic. */ 567extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 568_mm_cvttss_si64 (__m128 __A) 569{ 570 return __builtin_ia32_cvttss2si64 ((__v4sf) __A); 571} 572 573/* Microsoft intrinsic. */ 574extern __inline long long __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 575_mm_cvttss_si64x (__m128 __A) 576{ 577 return __builtin_ia32_cvttss2si64 ((__v4sf) __A); 578} 579#endif 580 581/* Truncate the two lower SPFP values to 32-bit integers. Return the 582 integers in packed form. */ 583extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 584_mm_cvttps_pi32 (__m128 __A) 585{ 586 return (__m64) __builtin_ia32_cvttps2pi ((__v4sf) __A); 587} 588 589extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 590_mm_cvtt_ps2pi (__m128 __A) 591{ 592 return _mm_cvttps_pi32 (__A); 593} 594 595/* Convert B to a SPFP value and insert it as element zero in A. */ 596extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 597_mm_cvtsi32_ss (__m128 __A, int __B) 598{ 599 return (__m128) __builtin_ia32_cvtsi2ss ((__v4sf) __A, __B); 600} 601 602extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 603_mm_cvt_si2ss (__m128 __A, int __B) 604{ 605 return _mm_cvtsi32_ss (__A, __B); 606} 607 608#ifdef __x86_64__ 609/* Convert B to a SPFP value and insert it as element zero in A. */ 610 611/* Intel intrinsic. */ 612extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 613_mm_cvtsi64_ss (__m128 __A, long long __B) 614{ 615 return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B); 616} 617 618/* Microsoft intrinsic. */ 619extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 620_mm_cvtsi64x_ss (__m128 __A, long long __B) 621{ 622 return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B); 623} 624#endif 625 626/* Convert the two 32-bit values in B to SPFP form and insert them 627 as the two lower elements in A. */ 628extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 629_mm_cvtpi32_ps (__m128 __A, __m64 __B) 630{ 631 return (__m128) __builtin_ia32_cvtpi2ps ((__v4sf) __A, (__v2si)__B); 632} 633 634extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 635_mm_cvt_pi2ps (__m128 __A, __m64 __B) 636{ 637 return _mm_cvtpi32_ps (__A, __B); 638} 639 640/* Convert the four signed 16-bit values in A to SPFP form. */ 641extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 642_mm_cvtpi16_ps (__m64 __A) 643{ 644 __v4hi __sign; 645 __v2si __hisi, __losi; 646 __v4sf __zero, __ra, __rb; 647 648 /* This comparison against zero gives us a mask that can be used to 649 fill in the missing sign bits in the unpack operations below, so 650 that we get signed values after unpacking. */ 651 __sign = __builtin_ia32_pcmpgtw ((__v4hi)0LL, (__v4hi)__A); 652 653 /* Convert the four words to doublewords. */ 654 __losi = (__v2si) __builtin_ia32_punpcklwd ((__v4hi)__A, __sign); 655 __hisi = (__v2si) __builtin_ia32_punpckhwd ((__v4hi)__A, __sign); 656 657 /* Convert the doublewords to floating point two at a time. */ 658 __zero = (__v4sf) _mm_setzero_ps (); 659 __ra = __builtin_ia32_cvtpi2ps (__zero, __losi); 660 __rb = __builtin_ia32_cvtpi2ps (__ra, __hisi); 661 662 return (__m128) __builtin_ia32_movlhps (__ra, __rb); 663} 664 665/* Convert the four unsigned 16-bit values in A to SPFP form. */ 666extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 667_mm_cvtpu16_ps (__m64 __A) 668{ 669 __v2si __hisi, __losi; 670 __v4sf __zero, __ra, __rb; 671 672 /* Convert the four words to doublewords. */ 673 __losi = (__v2si) __builtin_ia32_punpcklwd ((__v4hi)__A, (__v4hi)0LL); 674 __hisi = (__v2si) __builtin_ia32_punpckhwd ((__v4hi)__A, (__v4hi)0LL); 675 676 /* Convert the doublewords to floating point two at a time. */ 677 __zero = (__v4sf) _mm_setzero_ps (); 678 __ra = __builtin_ia32_cvtpi2ps (__zero, __losi); 679 __rb = __builtin_ia32_cvtpi2ps (__ra, __hisi); 680 681 return (__m128) __builtin_ia32_movlhps (__ra, __rb); 682} 683 684/* Convert the low four signed 8-bit values in A to SPFP form. */ 685extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 686_mm_cvtpi8_ps (__m64 __A) 687{ 688 __v8qi __sign; 689 690 /* This comparison against zero gives us a mask that can be used to 691 fill in the missing sign bits in the unpack operations below, so 692 that we get signed values after unpacking. */ 693 __sign = __builtin_ia32_pcmpgtb ((__v8qi)0LL, (__v8qi)__A); 694 695 /* Convert the four low bytes to words. */ 696 __A = (__m64) __builtin_ia32_punpcklbw ((__v8qi)__A, __sign); 697 698 return _mm_cvtpi16_ps(__A); 699} 700 701/* Convert the low four unsigned 8-bit values in A to SPFP form. */ 702extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 703_mm_cvtpu8_ps(__m64 __A) 704{ 705 __A = (__m64) __builtin_ia32_punpcklbw ((__v8qi)__A, (__v8qi)0LL); 706 return _mm_cvtpu16_ps(__A); 707} 708 709/* Convert the four signed 32-bit values in A and B to SPFP form. */ 710extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 711_mm_cvtpi32x2_ps(__m64 __A, __m64 __B) 712{ 713 __v4sf __zero = (__v4sf) _mm_setzero_ps (); 714 __v4sf __sfa = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__A); 715 __v4sf __sfb = __builtin_ia32_cvtpi2ps (__sfa, (__v2si)__B); 716 return (__m128) __builtin_ia32_movlhps (__sfa, __sfb); 717} 718 719/* Convert the four SPFP values in A to four signed 16-bit integers. */ 720extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 721_mm_cvtps_pi16(__m128 __A) 722{ 723 __v4sf __hisf = (__v4sf)__A; 724 __v4sf __losf = __builtin_ia32_movhlps (__hisf, __hisf); 725 __v2si __hisi = __builtin_ia32_cvtps2pi (__hisf); 726 __v2si __losi = __builtin_ia32_cvtps2pi (__losf); 727 return (__m64) __builtin_ia32_packssdw (__hisi, __losi); 728} 729 730/* Convert the four SPFP values in A to four signed 8-bit integers. */ 731extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 732_mm_cvtps_pi8(__m128 __A) 733{ 734 __v4hi __tmp = (__v4hi) _mm_cvtps_pi16 (__A); 735 return (__m64) __builtin_ia32_packsswb (__tmp, (__v4hi)0LL); 736} 737 738/* Selects four specific SPFP values from A and B based on MASK. */ 739#ifdef __OPTIMIZE__ 740extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 741_mm_shuffle_ps (__m128 __A, __m128 __B, int const __mask) 742{ 743 return (__m128) __builtin_ia32_shufps ((__v4sf)__A, (__v4sf)__B, __mask); 744} 745#else 746#define _mm_shuffle_ps(A, B, MASK) \ 747 ((__m128) __builtin_ia32_shufps ((__v4sf)(__m128)(A), \ 748 (__v4sf)(__m128)(B), (int)(MASK))) 749#endif 750 751/* Selects and interleaves the upper two SPFP values from A and B. */ 752extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 753_mm_unpackhi_ps (__m128 __A, __m128 __B) 754{ 755 return (__m128) __builtin_ia32_unpckhps ((__v4sf)__A, (__v4sf)__B); 756} 757 758/* Selects and interleaves the lower two SPFP values from A and B. */ 759extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 760_mm_unpacklo_ps (__m128 __A, __m128 __B) 761{ 762 return (__m128) __builtin_ia32_unpcklps ((__v4sf)__A, (__v4sf)__B); 763} 764 765/* Sets the upper two SPFP values with 64-bits of data loaded from P; 766 the lower two values are passed through from A. */ 767extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 768_mm_loadh_pi (__m128 __A, __m64 const *__P) 769{ 770 return (__m128) __builtin_ia32_loadhps ((__v4sf)__A, (const __v2sf *)__P); 771} 772 773/* Stores the upper two SPFP values of A into P. */ 774extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 775_mm_storeh_pi (__m64 *__P, __m128 __A) 776{ 777 __builtin_ia32_storehps ((__v2sf *)__P, (__v4sf)__A); 778} 779 780/* Moves the upper two values of B into the lower two values of A. */ 781extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 782_mm_movehl_ps (__m128 __A, __m128 __B) 783{ 784 return (__m128) __builtin_ia32_movhlps ((__v4sf)__A, (__v4sf)__B); 785} 786 787/* Moves the lower two values of B into the upper two values of A. */ 788extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 789_mm_movelh_ps (__m128 __A, __m128 __B) 790{ 791 return (__m128) __builtin_ia32_movlhps ((__v4sf)__A, (__v4sf)__B); 792} 793 794/* Sets the lower two SPFP values with 64-bits of data loaded from P; 795 the upper two values are passed through from A. */ 796extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 797_mm_loadl_pi (__m128 __A, __m64 const *__P) 798{ 799 return (__m128) __builtin_ia32_loadlps ((__v4sf)__A, (const __v2sf *)__P); 800} 801 802/* Stores the lower two SPFP values of A into P. */ 803extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 804_mm_storel_pi (__m64 *__P, __m128 __A) 805{ 806 __builtin_ia32_storelps ((__v2sf *)__P, (__v4sf)__A); 807} 808 809/* Creates a 4-bit mask from the most significant bits of the SPFP values. */ 810extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 811_mm_movemask_ps (__m128 __A) 812{ 813 return __builtin_ia32_movmskps ((__v4sf)__A); 814} 815 816/* Return the contents of the control register. */ 817extern __inline unsigned int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 818_mm_getcsr (void) 819{ 820 return __builtin_ia32_stmxcsr (); 821} 822 823/* Read exception bits from the control register. */ 824extern __inline unsigned int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 825_MM_GET_EXCEPTION_STATE (void) 826{ 827 return _mm_getcsr() & _MM_EXCEPT_MASK; 828} 829 830extern __inline unsigned int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 831_MM_GET_EXCEPTION_MASK (void) 832{ 833 return _mm_getcsr() & _MM_MASK_MASK; 834} 835 836extern __inline unsigned int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 837_MM_GET_ROUNDING_MODE (void) 838{ 839 return _mm_getcsr() & _MM_ROUND_MASK; 840} 841 842extern __inline unsigned int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 843_MM_GET_FLUSH_ZERO_MODE (void) 844{ 845 return _mm_getcsr() & _MM_FLUSH_ZERO_MASK; 846} 847 848/* Set the control register to I. */ 849extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 850_mm_setcsr (unsigned int __I) 851{ 852 __builtin_ia32_ldmxcsr (__I); 853} 854 855/* Set exception bits in the control register. */ 856extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 857_MM_SET_EXCEPTION_STATE(unsigned int __mask) 858{ 859 _mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | __mask); 860} 861 862extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 863_MM_SET_EXCEPTION_MASK (unsigned int __mask) 864{ 865 _mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | __mask); 866} 867 868extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 869_MM_SET_ROUNDING_MODE (unsigned int __mode) 870{ 871 _mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | __mode); 872} 873 874extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 875_MM_SET_FLUSH_ZERO_MODE (unsigned int __mode) 876{ 877 _mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | __mode); 878} 879 880/* Create a vector with element 0 as F and the rest zero. */ 881extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 882_mm_set_ss (float __F) 883{ 884 return __extension__ (__m128)(__v4sf){ __F, 0.0f, 0.0f, 0.0f }; 885} 886 887/* Create a vector with all four elements equal to F. */ 888extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 889_mm_set1_ps (float __F) 890{ 891 return __extension__ (__m128)(__v4sf){ __F, __F, __F, __F }; 892} 893 894extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 895_mm_set_ps1 (float __F) 896{ 897 return _mm_set1_ps (__F); 898} 899 900/* Create a vector with element 0 as *P and the rest zero. */ 901extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 902_mm_load_ss (float const *__P) 903{ 904 return _mm_set_ss (*__P); 905} 906 907/* Create a vector with all four elements equal to *P. */ 908extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 909_mm_load1_ps (float const *__P) 910{ 911 return _mm_set1_ps (*__P); 912} 913 914extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 915_mm_load_ps1 (float const *__P) 916{ 917 return _mm_load1_ps (__P); 918} 919 920/* Load four SPFP values from P. The address must be 16-byte aligned. */ 921extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 922_mm_load_ps (float const *__P) 923{ 924 return (__m128) *(__v4sf *)__P; 925} 926 927/* Load four SPFP values from P. The address need not be 16-byte aligned. */ 928extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 929_mm_loadu_ps (float const *__P) 930{ 931 return (__m128) __builtin_ia32_loadups (__P); 932} 933 934/* Load four SPFP values in reverse order. The address must be aligned. */ 935extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 936_mm_loadr_ps (float const *__P) 937{ 938 __v4sf __tmp = *(__v4sf *)__P; 939 return (__m128) __builtin_ia32_shufps (__tmp, __tmp, _MM_SHUFFLE (0,1,2,3)); 940} 941 942/* Create the vector [Z Y X W]. */ 943extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 944_mm_set_ps (const float __Z, const float __Y, const float __X, const float __W) 945{ 946 return __extension__ (__m128)(__v4sf){ __W, __X, __Y, __Z }; 947} 948 949/* Create the vector [W X Y Z]. */ 950extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 951_mm_setr_ps (float __Z, float __Y, float __X, float __W) 952{ 953 return __extension__ (__m128)(__v4sf){ __Z, __Y, __X, __W }; 954} 955 956/* Stores the lower SPFP value. */ 957extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 958_mm_store_ss (float *__P, __m128 __A) 959{ 960 *__P = __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0); 961} 962 963extern __inline float __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 964_mm_cvtss_f32 (__m128 __A) 965{ 966 return __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0); 967} 968 969/* Store four SPFP values. The address must be 16-byte aligned. */ 970extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 971_mm_store_ps (float *__P, __m128 __A) 972{ 973 *(__v4sf *)__P = (__v4sf)__A; 974} 975 976/* Store four SPFP values. The address need not be 16-byte aligned. */ 977extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 978_mm_storeu_ps (float *__P, __m128 __A) 979{ 980 __builtin_ia32_storeups (__P, (__v4sf)__A); 981} 982 983/* Store the lower SPFP value across four words. */ 984extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 985_mm_store1_ps (float *__P, __m128 __A) 986{ 987 __v4sf __va = (__v4sf)__A; 988 __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,0,0,0)); 989 _mm_storeu_ps (__P, __tmp); 990} 991 992extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 993_mm_store_ps1 (float *__P, __m128 __A) 994{ 995 _mm_store1_ps (__P, __A); 996} 997 998/* Store four SPFP values in reverse order. The address must be aligned. */ 999extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1000_mm_storer_ps (float *__P, __m128 __A) 1001{ 1002 __v4sf __va = (__v4sf)__A; 1003 __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,1,2,3)); 1004 _mm_store_ps (__P, __tmp); 1005} 1006 1007/* Sets the low SPFP value of A from the low value of B. */ 1008extern __inline __m128 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1009_mm_move_ss (__m128 __A, __m128 __B) 1010{ 1011 return (__m128) __builtin_ia32_movss ((__v4sf)__A, (__v4sf)__B); 1012} 1013 1014/* Extracts one of the four words of A. The selector N must be immediate. */ 1015#ifdef __OPTIMIZE__ 1016extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1017_mm_extract_pi16 (__m64 const __A, int const __N) 1018{ 1019 return __builtin_ia32_vec_ext_v4hi ((__v4hi)__A, __N); 1020} 1021 1022extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1023_m_pextrw (__m64 const __A, int const __N) 1024{ 1025 return _mm_extract_pi16 (__A, __N); 1026} 1027#else 1028#define _mm_extract_pi16(A, N) \ 1029 ((int) __builtin_ia32_vec_ext_v4hi ((__v4hi)(__m64)(A), (int)(N))) 1030 1031#define _m_pextrw(A, N) _mm_extract_pi16(A, N) 1032#endif 1033 1034/* Inserts word D into one of four words of A. The selector N must be 1035 immediate. */ 1036#ifdef __OPTIMIZE__ 1037extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1038_mm_insert_pi16 (__m64 const __A, int const __D, int const __N) 1039{ 1040 return (__m64) __builtin_ia32_vec_set_v4hi ((__v4hi)__A, __D, __N); 1041} 1042 1043extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1044_m_pinsrw (__m64 const __A, int const __D, int const __N) 1045{ 1046 return _mm_insert_pi16 (__A, __D, __N); 1047} 1048#else 1049#define _mm_insert_pi16(A, D, N) \ 1050 ((__m64) __builtin_ia32_vec_set_v4hi ((__v4hi)(__m64)(A), \ 1051 (int)(D), (int)(N))) 1052 1053#define _m_pinsrw(A, D, N) _mm_insert_pi16(A, D, N) 1054#endif 1055 1056/* Compute the element-wise maximum of signed 16-bit values. */ 1057extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1058_mm_max_pi16 (__m64 __A, __m64 __B) 1059{ 1060 return (__m64) __builtin_ia32_pmaxsw ((__v4hi)__A, (__v4hi)__B); 1061} 1062 1063extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1064_m_pmaxsw (__m64 __A, __m64 __B) 1065{ 1066 return _mm_max_pi16 (__A, __B); 1067} 1068 1069/* Compute the element-wise maximum of unsigned 8-bit values. */ 1070extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1071_mm_max_pu8 (__m64 __A, __m64 __B) 1072{ 1073 return (__m64) __builtin_ia32_pmaxub ((__v8qi)__A, (__v8qi)__B); 1074} 1075 1076extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1077_m_pmaxub (__m64 __A, __m64 __B) 1078{ 1079 return _mm_max_pu8 (__A, __B); 1080} 1081 1082/* Compute the element-wise minimum of signed 16-bit values. */ 1083extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1084_mm_min_pi16 (__m64 __A, __m64 __B) 1085{ 1086 return (__m64) __builtin_ia32_pminsw ((__v4hi)__A, (__v4hi)__B); 1087} 1088 1089extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1090_m_pminsw (__m64 __A, __m64 __B) 1091{ 1092 return _mm_min_pi16 (__A, __B); 1093} 1094 1095/* Compute the element-wise minimum of unsigned 8-bit values. */ 1096extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1097_mm_min_pu8 (__m64 __A, __m64 __B) 1098{ 1099 return (__m64) __builtin_ia32_pminub ((__v8qi)__A, (__v8qi)__B); 1100} 1101 1102extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1103_m_pminub (__m64 __A, __m64 __B) 1104{ 1105 return _mm_min_pu8 (__A, __B); 1106} 1107 1108/* Create an 8-bit mask of the signs of 8-bit values. */ 1109extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1110_mm_movemask_pi8 (__m64 __A) 1111{ 1112 return __builtin_ia32_pmovmskb ((__v8qi)__A); 1113} 1114 1115extern __inline int __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1116_m_pmovmskb (__m64 __A) 1117{ 1118 return _mm_movemask_pi8 (__A); 1119} 1120 1121/* Multiply four unsigned 16-bit values in A by four unsigned 16-bit values 1122 in B and produce the high 16 bits of the 32-bit results. */ 1123extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1124_mm_mulhi_pu16 (__m64 __A, __m64 __B) 1125{ 1126 return (__m64) __builtin_ia32_pmulhuw ((__v4hi)__A, (__v4hi)__B); 1127} 1128 1129extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1130_m_pmulhuw (__m64 __A, __m64 __B) 1131{ 1132 return _mm_mulhi_pu16 (__A, __B); 1133} 1134 1135/* Return a combination of the four 16-bit values in A. The selector 1136 must be an immediate. */ 1137#ifdef __OPTIMIZE__ 1138extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1139_mm_shuffle_pi16 (__m64 __A, int const __N) 1140{ 1141 return (__m64) __builtin_ia32_pshufw ((__v4hi)__A, __N); 1142} 1143 1144extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1145_m_pshufw (__m64 __A, int const __N) 1146{ 1147 return _mm_shuffle_pi16 (__A, __N); 1148} 1149#else 1150#define _mm_shuffle_pi16(A, N) \ 1151 ((__m64) __builtin_ia32_pshufw ((__v4hi)(__m64)(A), (int)(N))) 1152 1153#define _m_pshufw(A, N) _mm_shuffle_pi16 (A, N) 1154#endif 1155 1156/* Conditionally store byte elements of A into P. The high bit of each 1157 byte in the selector N determines whether the corresponding byte from 1158 A is stored. */ 1159extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1160_mm_maskmove_si64 (__m64 __A, __m64 __N, char *__P) 1161{ 1162 __builtin_ia32_maskmovq ((__v8qi)__A, (__v8qi)__N, __P); 1163} 1164 1165extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1166_m_maskmovq (__m64 __A, __m64 __N, char *__P) 1167{ 1168 _mm_maskmove_si64 (__A, __N, __P); 1169} 1170 1171/* Compute the rounded averages of the unsigned 8-bit values in A and B. */ 1172extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1173_mm_avg_pu8 (__m64 __A, __m64 __B) 1174{ 1175 return (__m64) __builtin_ia32_pavgb ((__v8qi)__A, (__v8qi)__B); 1176} 1177 1178extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1179_m_pavgb (__m64 __A, __m64 __B) 1180{ 1181 return _mm_avg_pu8 (__A, __B); 1182} 1183 1184/* Compute the rounded averages of the unsigned 16-bit values in A and B. */ 1185extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1186_mm_avg_pu16 (__m64 __A, __m64 __B) 1187{ 1188 return (__m64) __builtin_ia32_pavgw ((__v4hi)__A, (__v4hi)__B); 1189} 1190 1191extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1192_m_pavgw (__m64 __A, __m64 __B) 1193{ 1194 return _mm_avg_pu16 (__A, __B); 1195} 1196 1197/* Compute the sum of the absolute differences of the unsigned 8-bit 1198 values in A and B. Return the value in the lower 16-bit word; the 1199 upper words are cleared. */ 1200extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1201_mm_sad_pu8 (__m64 __A, __m64 __B) 1202{ 1203 return (__m64) __builtin_ia32_psadbw ((__v8qi)__A, (__v8qi)__B); 1204} 1205 1206extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1207_m_psadbw (__m64 __A, __m64 __B) 1208{ 1209 return _mm_sad_pu8 (__A, __B); 1210} 1211 1212/* Stores the data in A to the address P without polluting the caches. */ 1213extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1214_mm_stream_pi (__m64 *__P, __m64 __A) 1215{ 1216 __builtin_ia32_movntq ((unsigned long long *)__P, (unsigned long long)__A); 1217} 1218 1219/* Likewise. The address must be 16-byte aligned. */ 1220extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1221_mm_stream_ps (float *__P, __m128 __A) 1222{ 1223 __builtin_ia32_movntps (__P, (__v4sf)__A); 1224} 1225 1226/* Guarantees that every preceding store is globally visible before 1227 any subsequent store. */ 1228extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1229_mm_sfence (void) 1230{ 1231 __builtin_ia32_sfence (); 1232} 1233 1234/* The execution of the next instruction is delayed by an implementation 1235 specific amount of time. The instruction does not modify the 1236 architectural state. */ 1237extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__)) 1238_mm_pause (void) 1239{ 1240 __builtin_ia32_pause (); 1241} 1242 1243/* Transpose the 4x4 matrix composed of row[0-3]. */ 1244#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ 1245do { \ 1246 __v4sf __r0 = (row0), __r1 = (row1), __r2 = (row2), __r3 = (row3); \ 1247 __v4sf __t0 = __builtin_ia32_unpcklps (__r0, __r1); \ 1248 __v4sf __t1 = __builtin_ia32_unpcklps (__r2, __r3); \ 1249 __v4sf __t2 = __builtin_ia32_unpckhps (__r0, __r1); \ 1250 __v4sf __t3 = __builtin_ia32_unpckhps (__r2, __r3); \ 1251 (row0) = __builtin_ia32_movlhps (__t0, __t1); \ 1252 (row1) = __builtin_ia32_movhlps (__t1, __t0); \ 1253 (row2) = __builtin_ia32_movlhps (__t2, __t3); \ 1254 (row3) = __builtin_ia32_movhlps (__t3, __t2); \ 1255} while (0) 1256 1257/* For backward source compatibility. */ 1258# include <emmintrin.h> 1259 1260#ifdef __DISABLE_SSE__ 1261#undef __DISABLE_SSE__ 1262#pragma GCC pop_options 1263#endif /* __DISABLE_SSE__ */ 1264 1265#endif /* _XMMINTRIN_H_INCLUDED */ 1266