xmmintrin.h revision 296373
1/* Copyright (C) 2002, 2003, 2004, 2005, 2006, 2007 2 Free Software Foundation, Inc. 3 4 This file is part of GCC. 5 6 GCC is free software; you can redistribute it and/or modify 7 it under the terms of the GNU General Public License as published by 8 the Free Software Foundation; either version 2, or (at your option) 9 any later version. 10 11 GCC is distributed in the hope that it will be useful, 12 but WITHOUT ANY WARRANTY; without even the implied warranty of 13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 GNU General Public License for more details. 15 16 You should have received a copy of the GNU General Public License 17 along with GCC; see the file COPYING. If not, write to 18 the Free Software Foundation, 51 Franklin Street, Fifth Floor, 19 Boston, MA 02110-1301, USA. */ 20 21/* As a special exception, if you include this header file into source 22 files compiled by GCC, this header file does not by itself cause 23 the resulting executable to be covered by the GNU General Public 24 License. This exception does not however invalidate any other 25 reasons why the executable file might be covered by the GNU General 26 Public License. */ 27 28/* Implemented from the specification included in the Intel C++ Compiler 29 User Guide and Reference, version 9.0. */ 30 31#ifndef _XMMINTRIN_H_INCLUDED 32#define _XMMINTRIN_H_INCLUDED 33 34#ifndef __SSE__ 35# error "SSE instruction set not enabled" 36#else 37 38/* We need type definitions from the MMX header file. */ 39#include <mmintrin.h> 40 41/* Get _mm_malloc () and _mm_free (). */ 42#if __STDC_HOSTED__ 43#include <mm_malloc.h> 44#endif 45 46/* The Intel API is flexible enough that we must allow aliasing with other 47 vector types, and their scalar components. */ 48typedef float __m128 __attribute__ ((__vector_size__ (16), __may_alias__)); 49 50/* Internal data types for implementing the intrinsics. */ 51typedef float __v4sf __attribute__ ((__vector_size__ (16))); 52 53/* Create a selector for use with the SHUFPS instruction. */ 54#define _MM_SHUFFLE(fp3,fp2,fp1,fp0) \ 55 (((fp3) << 6) | ((fp2) << 4) | ((fp1) << 2) | (fp0)) 56 57/* Constants for use with _mm_prefetch. */ 58enum _mm_hint 59{ 60 _MM_HINT_T0 = 3, 61 _MM_HINT_T1 = 2, 62 _MM_HINT_T2 = 1, 63 _MM_HINT_NTA = 0 64}; 65 66/* Bits in the MXCSR. */ 67#define _MM_EXCEPT_MASK 0x003f 68#define _MM_EXCEPT_INVALID 0x0001 69#define _MM_EXCEPT_DENORM 0x0002 70#define _MM_EXCEPT_DIV_ZERO 0x0004 71#define _MM_EXCEPT_OVERFLOW 0x0008 72#define _MM_EXCEPT_UNDERFLOW 0x0010 73#define _MM_EXCEPT_INEXACT 0x0020 74 75#define _MM_MASK_MASK 0x1f80 76#define _MM_MASK_INVALID 0x0080 77#define _MM_MASK_DENORM 0x0100 78#define _MM_MASK_DIV_ZERO 0x0200 79#define _MM_MASK_OVERFLOW 0x0400 80#define _MM_MASK_UNDERFLOW 0x0800 81#define _MM_MASK_INEXACT 0x1000 82 83#define _MM_ROUND_MASK 0x6000 84#define _MM_ROUND_NEAREST 0x0000 85#define _MM_ROUND_DOWN 0x2000 86#define _MM_ROUND_UP 0x4000 87#define _MM_ROUND_TOWARD_ZERO 0x6000 88 89#define _MM_FLUSH_ZERO_MASK 0x8000 90#define _MM_FLUSH_ZERO_ON 0x8000 91#define _MM_FLUSH_ZERO_OFF 0x0000 92 93/* Create a vector of zeros. */ 94static __inline __m128 __attribute__((__always_inline__)) 95_mm_setzero_ps (void) 96{ 97 return __extension__ (__m128){ 0.0f, 0.0f, 0.0f, 0.0f }; 98} 99 100/* Perform the respective operation on the lower SPFP (single-precision 101 floating-point) values of A and B; the upper three SPFP values are 102 passed through from A. */ 103 104static __inline __m128 __attribute__((__always_inline__)) 105_mm_add_ss (__m128 __A, __m128 __B) 106{ 107 return (__m128) __builtin_ia32_addss ((__v4sf)__A, (__v4sf)__B); 108} 109 110static __inline __m128 __attribute__((__always_inline__)) 111_mm_sub_ss (__m128 __A, __m128 __B) 112{ 113 return (__m128) __builtin_ia32_subss ((__v4sf)__A, (__v4sf)__B); 114} 115 116static __inline __m128 __attribute__((__always_inline__)) 117_mm_mul_ss (__m128 __A, __m128 __B) 118{ 119 return (__m128) __builtin_ia32_mulss ((__v4sf)__A, (__v4sf)__B); 120} 121 122static __inline __m128 __attribute__((__always_inline__)) 123_mm_div_ss (__m128 __A, __m128 __B) 124{ 125 return (__m128) __builtin_ia32_divss ((__v4sf)__A, (__v4sf)__B); 126} 127 128static __inline __m128 __attribute__((__always_inline__)) 129_mm_sqrt_ss (__m128 __A) 130{ 131 return (__m128) __builtin_ia32_sqrtss ((__v4sf)__A); 132} 133 134static __inline __m128 __attribute__((__always_inline__)) 135_mm_rcp_ss (__m128 __A) 136{ 137 return (__m128) __builtin_ia32_rcpss ((__v4sf)__A); 138} 139 140static __inline __m128 __attribute__((__always_inline__)) 141_mm_rsqrt_ss (__m128 __A) 142{ 143 return (__m128) __builtin_ia32_rsqrtss ((__v4sf)__A); 144} 145 146static __inline __m128 __attribute__((__always_inline__)) 147_mm_min_ss (__m128 __A, __m128 __B) 148{ 149 return (__m128) __builtin_ia32_minss ((__v4sf)__A, (__v4sf)__B); 150} 151 152static __inline __m128 __attribute__((__always_inline__)) 153_mm_max_ss (__m128 __A, __m128 __B) 154{ 155 return (__m128) __builtin_ia32_maxss ((__v4sf)__A, (__v4sf)__B); 156} 157 158/* Perform the respective operation on the four SPFP values in A and B. */ 159 160static __inline __m128 __attribute__((__always_inline__)) 161_mm_add_ps (__m128 __A, __m128 __B) 162{ 163 return (__m128) __builtin_ia32_addps ((__v4sf)__A, (__v4sf)__B); 164} 165 166static __inline __m128 __attribute__((__always_inline__)) 167_mm_sub_ps (__m128 __A, __m128 __B) 168{ 169 return (__m128) __builtin_ia32_subps ((__v4sf)__A, (__v4sf)__B); 170} 171 172static __inline __m128 __attribute__((__always_inline__)) 173_mm_mul_ps (__m128 __A, __m128 __B) 174{ 175 return (__m128) __builtin_ia32_mulps ((__v4sf)__A, (__v4sf)__B); 176} 177 178static __inline __m128 __attribute__((__always_inline__)) 179_mm_div_ps (__m128 __A, __m128 __B) 180{ 181 return (__m128) __builtin_ia32_divps ((__v4sf)__A, (__v4sf)__B); 182} 183 184static __inline __m128 __attribute__((__always_inline__)) 185_mm_sqrt_ps (__m128 __A) 186{ 187 return (__m128) __builtin_ia32_sqrtps ((__v4sf)__A); 188} 189 190static __inline __m128 __attribute__((__always_inline__)) 191_mm_rcp_ps (__m128 __A) 192{ 193 return (__m128) __builtin_ia32_rcpps ((__v4sf)__A); 194} 195 196static __inline __m128 __attribute__((__always_inline__)) 197_mm_rsqrt_ps (__m128 __A) 198{ 199 return (__m128) __builtin_ia32_rsqrtps ((__v4sf)__A); 200} 201 202static __inline __m128 __attribute__((__always_inline__)) 203_mm_min_ps (__m128 __A, __m128 __B) 204{ 205 return (__m128) __builtin_ia32_minps ((__v4sf)__A, (__v4sf)__B); 206} 207 208static __inline __m128 __attribute__((__always_inline__)) 209_mm_max_ps (__m128 __A, __m128 __B) 210{ 211 return (__m128) __builtin_ia32_maxps ((__v4sf)__A, (__v4sf)__B); 212} 213 214/* Perform logical bit-wise operations on 128-bit values. */ 215 216static __inline __m128 __attribute__((__always_inline__)) 217_mm_and_ps (__m128 __A, __m128 __B) 218{ 219 return __builtin_ia32_andps (__A, __B); 220} 221 222static __inline __m128 __attribute__((__always_inline__)) 223_mm_andnot_ps (__m128 __A, __m128 __B) 224{ 225 return __builtin_ia32_andnps (__A, __B); 226} 227 228static __inline __m128 __attribute__((__always_inline__)) 229_mm_or_ps (__m128 __A, __m128 __B) 230{ 231 return __builtin_ia32_orps (__A, __B); 232} 233 234static __inline __m128 __attribute__((__always_inline__)) 235_mm_xor_ps (__m128 __A, __m128 __B) 236{ 237 return __builtin_ia32_xorps (__A, __B); 238} 239 240/* Perform a comparison on the lower SPFP values of A and B. If the 241 comparison is true, place a mask of all ones in the result, otherwise a 242 mask of zeros. The upper three SPFP values are passed through from A. */ 243 244static __inline __m128 __attribute__((__always_inline__)) 245_mm_cmpeq_ss (__m128 __A, __m128 __B) 246{ 247 return (__m128) __builtin_ia32_cmpeqss ((__v4sf)__A, (__v4sf)__B); 248} 249 250static __inline __m128 __attribute__((__always_inline__)) 251_mm_cmplt_ss (__m128 __A, __m128 __B) 252{ 253 return (__m128) __builtin_ia32_cmpltss ((__v4sf)__A, (__v4sf)__B); 254} 255 256static __inline __m128 __attribute__((__always_inline__)) 257_mm_cmple_ss (__m128 __A, __m128 __B) 258{ 259 return (__m128) __builtin_ia32_cmpless ((__v4sf)__A, (__v4sf)__B); 260} 261 262static __inline __m128 __attribute__((__always_inline__)) 263_mm_cmpgt_ss (__m128 __A, __m128 __B) 264{ 265 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 266 (__v4sf) 267 __builtin_ia32_cmpltss ((__v4sf) __B, 268 (__v4sf) 269 __A)); 270} 271 272static __inline __m128 __attribute__((__always_inline__)) 273_mm_cmpge_ss (__m128 __A, __m128 __B) 274{ 275 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 276 (__v4sf) 277 __builtin_ia32_cmpless ((__v4sf) __B, 278 (__v4sf) 279 __A)); 280} 281 282static __inline __m128 __attribute__((__always_inline__)) 283_mm_cmpneq_ss (__m128 __A, __m128 __B) 284{ 285 return (__m128) __builtin_ia32_cmpneqss ((__v4sf)__A, (__v4sf)__B); 286} 287 288static __inline __m128 __attribute__((__always_inline__)) 289_mm_cmpnlt_ss (__m128 __A, __m128 __B) 290{ 291 return (__m128) __builtin_ia32_cmpnltss ((__v4sf)__A, (__v4sf)__B); 292} 293 294static __inline __m128 __attribute__((__always_inline__)) 295_mm_cmpnle_ss (__m128 __A, __m128 __B) 296{ 297 return (__m128) __builtin_ia32_cmpnless ((__v4sf)__A, (__v4sf)__B); 298} 299 300static __inline __m128 __attribute__((__always_inline__)) 301_mm_cmpngt_ss (__m128 __A, __m128 __B) 302{ 303 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 304 (__v4sf) 305 __builtin_ia32_cmpnltss ((__v4sf) __B, 306 (__v4sf) 307 __A)); 308} 309 310static __inline __m128 __attribute__((__always_inline__)) 311_mm_cmpnge_ss (__m128 __A, __m128 __B) 312{ 313 return (__m128) __builtin_ia32_movss ((__v4sf) __A, 314 (__v4sf) 315 __builtin_ia32_cmpnless ((__v4sf) __B, 316 (__v4sf) 317 __A)); 318} 319 320static __inline __m128 __attribute__((__always_inline__)) 321_mm_cmpord_ss (__m128 __A, __m128 __B) 322{ 323 return (__m128) __builtin_ia32_cmpordss ((__v4sf)__A, (__v4sf)__B); 324} 325 326static __inline __m128 __attribute__((__always_inline__)) 327_mm_cmpunord_ss (__m128 __A, __m128 __B) 328{ 329 return (__m128) __builtin_ia32_cmpunordss ((__v4sf)__A, (__v4sf)__B); 330} 331 332/* Perform a comparison on the four SPFP values of A and B. For each 333 element, if the comparison is true, place a mask of all ones in the 334 result, otherwise a mask of zeros. */ 335 336static __inline __m128 __attribute__((__always_inline__)) 337_mm_cmpeq_ps (__m128 __A, __m128 __B) 338{ 339 return (__m128) __builtin_ia32_cmpeqps ((__v4sf)__A, (__v4sf)__B); 340} 341 342static __inline __m128 __attribute__((__always_inline__)) 343_mm_cmplt_ps (__m128 __A, __m128 __B) 344{ 345 return (__m128) __builtin_ia32_cmpltps ((__v4sf)__A, (__v4sf)__B); 346} 347 348static __inline __m128 __attribute__((__always_inline__)) 349_mm_cmple_ps (__m128 __A, __m128 __B) 350{ 351 return (__m128) __builtin_ia32_cmpleps ((__v4sf)__A, (__v4sf)__B); 352} 353 354static __inline __m128 __attribute__((__always_inline__)) 355_mm_cmpgt_ps (__m128 __A, __m128 __B) 356{ 357 return (__m128) __builtin_ia32_cmpgtps ((__v4sf)__A, (__v4sf)__B); 358} 359 360static __inline __m128 __attribute__((__always_inline__)) 361_mm_cmpge_ps (__m128 __A, __m128 __B) 362{ 363 return (__m128) __builtin_ia32_cmpgeps ((__v4sf)__A, (__v4sf)__B); 364} 365 366static __inline __m128 __attribute__((__always_inline__)) 367_mm_cmpneq_ps (__m128 __A, __m128 __B) 368{ 369 return (__m128) __builtin_ia32_cmpneqps ((__v4sf)__A, (__v4sf)__B); 370} 371 372static __inline __m128 __attribute__((__always_inline__)) 373_mm_cmpnlt_ps (__m128 __A, __m128 __B) 374{ 375 return (__m128) __builtin_ia32_cmpnltps ((__v4sf)__A, (__v4sf)__B); 376} 377 378static __inline __m128 __attribute__((__always_inline__)) 379_mm_cmpnle_ps (__m128 __A, __m128 __B) 380{ 381 return (__m128) __builtin_ia32_cmpnleps ((__v4sf)__A, (__v4sf)__B); 382} 383 384static __inline __m128 __attribute__((__always_inline__)) 385_mm_cmpngt_ps (__m128 __A, __m128 __B) 386{ 387 return (__m128) __builtin_ia32_cmpngtps ((__v4sf)__A, (__v4sf)__B); 388} 389 390static __inline __m128 __attribute__((__always_inline__)) 391_mm_cmpnge_ps (__m128 __A, __m128 __B) 392{ 393 return (__m128) __builtin_ia32_cmpngeps ((__v4sf)__A, (__v4sf)__B); 394} 395 396static __inline __m128 __attribute__((__always_inline__)) 397_mm_cmpord_ps (__m128 __A, __m128 __B) 398{ 399 return (__m128) __builtin_ia32_cmpordps ((__v4sf)__A, (__v4sf)__B); 400} 401 402static __inline __m128 __attribute__((__always_inline__)) 403_mm_cmpunord_ps (__m128 __A, __m128 __B) 404{ 405 return (__m128) __builtin_ia32_cmpunordps ((__v4sf)__A, (__v4sf)__B); 406} 407 408/* Compare the lower SPFP values of A and B and return 1 if true 409 and 0 if false. */ 410 411static __inline int __attribute__((__always_inline__)) 412_mm_comieq_ss (__m128 __A, __m128 __B) 413{ 414 return __builtin_ia32_comieq ((__v4sf)__A, (__v4sf)__B); 415} 416 417static __inline int __attribute__((__always_inline__)) 418_mm_comilt_ss (__m128 __A, __m128 __B) 419{ 420 return __builtin_ia32_comilt ((__v4sf)__A, (__v4sf)__B); 421} 422 423static __inline int __attribute__((__always_inline__)) 424_mm_comile_ss (__m128 __A, __m128 __B) 425{ 426 return __builtin_ia32_comile ((__v4sf)__A, (__v4sf)__B); 427} 428 429static __inline int __attribute__((__always_inline__)) 430_mm_comigt_ss (__m128 __A, __m128 __B) 431{ 432 return __builtin_ia32_comigt ((__v4sf)__A, (__v4sf)__B); 433} 434 435static __inline int __attribute__((__always_inline__)) 436_mm_comige_ss (__m128 __A, __m128 __B) 437{ 438 return __builtin_ia32_comige ((__v4sf)__A, (__v4sf)__B); 439} 440 441static __inline int __attribute__((__always_inline__)) 442_mm_comineq_ss (__m128 __A, __m128 __B) 443{ 444 return __builtin_ia32_comineq ((__v4sf)__A, (__v4sf)__B); 445} 446 447static __inline int __attribute__((__always_inline__)) 448_mm_ucomieq_ss (__m128 __A, __m128 __B) 449{ 450 return __builtin_ia32_ucomieq ((__v4sf)__A, (__v4sf)__B); 451} 452 453static __inline int __attribute__((__always_inline__)) 454_mm_ucomilt_ss (__m128 __A, __m128 __B) 455{ 456 return __builtin_ia32_ucomilt ((__v4sf)__A, (__v4sf)__B); 457} 458 459static __inline int __attribute__((__always_inline__)) 460_mm_ucomile_ss (__m128 __A, __m128 __B) 461{ 462 return __builtin_ia32_ucomile ((__v4sf)__A, (__v4sf)__B); 463} 464 465static __inline int __attribute__((__always_inline__)) 466_mm_ucomigt_ss (__m128 __A, __m128 __B) 467{ 468 return __builtin_ia32_ucomigt ((__v4sf)__A, (__v4sf)__B); 469} 470 471static __inline int __attribute__((__always_inline__)) 472_mm_ucomige_ss (__m128 __A, __m128 __B) 473{ 474 return __builtin_ia32_ucomige ((__v4sf)__A, (__v4sf)__B); 475} 476 477static __inline int __attribute__((__always_inline__)) 478_mm_ucomineq_ss (__m128 __A, __m128 __B) 479{ 480 return __builtin_ia32_ucomineq ((__v4sf)__A, (__v4sf)__B); 481} 482 483/* Convert the lower SPFP value to a 32-bit integer according to the current 484 rounding mode. */ 485static __inline int __attribute__((__always_inline__)) 486_mm_cvtss_si32 (__m128 __A) 487{ 488 return __builtin_ia32_cvtss2si ((__v4sf) __A); 489} 490 491static __inline int __attribute__((__always_inline__)) 492_mm_cvt_ss2si (__m128 __A) 493{ 494 return _mm_cvtss_si32 (__A); 495} 496 497#ifdef __x86_64__ 498/* Convert the lower SPFP value to a 32-bit integer according to the 499 current rounding mode. */ 500 501/* Intel intrinsic. */ 502static __inline long long __attribute__((__always_inline__)) 503_mm_cvtss_si64 (__m128 __A) 504{ 505 return __builtin_ia32_cvtss2si64 ((__v4sf) __A); 506} 507 508/* Microsoft intrinsic. */ 509static __inline long long __attribute__((__always_inline__)) 510_mm_cvtss_si64x (__m128 __A) 511{ 512 return __builtin_ia32_cvtss2si64 ((__v4sf) __A); 513} 514#endif 515 516/* Convert the two lower SPFP values to 32-bit integers according to the 517 current rounding mode. Return the integers in packed form. */ 518static __inline __m64 __attribute__((__always_inline__)) 519_mm_cvtps_pi32 (__m128 __A) 520{ 521 return (__m64) __builtin_ia32_cvtps2pi ((__v4sf) __A); 522} 523 524static __inline __m64 __attribute__((__always_inline__)) 525_mm_cvt_ps2pi (__m128 __A) 526{ 527 return _mm_cvtps_pi32 (__A); 528} 529 530/* Truncate the lower SPFP value to a 32-bit integer. */ 531static __inline int __attribute__((__always_inline__)) 532_mm_cvttss_si32 (__m128 __A) 533{ 534 return __builtin_ia32_cvttss2si ((__v4sf) __A); 535} 536 537static __inline int __attribute__((__always_inline__)) 538_mm_cvtt_ss2si (__m128 __A) 539{ 540 return _mm_cvttss_si32 (__A); 541} 542 543#ifdef __x86_64__ 544/* Truncate the lower SPFP value to a 32-bit integer. */ 545 546/* Intel intrinsic. */ 547static __inline long long __attribute__((__always_inline__)) 548_mm_cvttss_si64 (__m128 __A) 549{ 550 return __builtin_ia32_cvttss2si64 ((__v4sf) __A); 551} 552 553/* Microsoft intrinsic. */ 554static __inline long long __attribute__((__always_inline__)) 555_mm_cvttss_si64x (__m128 __A) 556{ 557 return __builtin_ia32_cvttss2si64 ((__v4sf) __A); 558} 559#endif 560 561/* Truncate the two lower SPFP values to 32-bit integers. Return the 562 integers in packed form. */ 563static __inline __m64 __attribute__((__always_inline__)) 564_mm_cvttps_pi32 (__m128 __A) 565{ 566 return (__m64) __builtin_ia32_cvttps2pi ((__v4sf) __A); 567} 568 569static __inline __m64 __attribute__((__always_inline__)) 570_mm_cvtt_ps2pi (__m128 __A) 571{ 572 return _mm_cvttps_pi32 (__A); 573} 574 575/* Convert B to a SPFP value and insert it as element zero in A. */ 576static __inline __m128 __attribute__((__always_inline__)) 577_mm_cvtsi32_ss (__m128 __A, int __B) 578{ 579 return (__m128) __builtin_ia32_cvtsi2ss ((__v4sf) __A, __B); 580} 581 582static __inline __m128 __attribute__((__always_inline__)) 583_mm_cvt_si2ss (__m128 __A, int __B) 584{ 585 return _mm_cvtsi32_ss (__A, __B); 586} 587 588#ifdef __x86_64__ 589/* Convert B to a SPFP value and insert it as element zero in A. */ 590 591/* Intel intrinsic. */ 592static __inline __m128 __attribute__((__always_inline__)) 593_mm_cvtsi64_ss (__m128 __A, long long __B) 594{ 595 return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B); 596} 597 598/* Microsoft intrinsic. */ 599static __inline __m128 __attribute__((__always_inline__)) 600_mm_cvtsi64x_ss (__m128 __A, long long __B) 601{ 602 return (__m128) __builtin_ia32_cvtsi642ss ((__v4sf) __A, __B); 603} 604#endif 605 606/* Convert the two 32-bit values in B to SPFP form and insert them 607 as the two lower elements in A. */ 608static __inline __m128 __attribute__((__always_inline__)) 609_mm_cvtpi32_ps (__m128 __A, __m64 __B) 610{ 611 return (__m128) __builtin_ia32_cvtpi2ps ((__v4sf) __A, (__v2si)__B); 612} 613 614static __inline __m128 __attribute__((__always_inline__)) 615_mm_cvt_pi2ps (__m128 __A, __m64 __B) 616{ 617 return _mm_cvtpi32_ps (__A, __B); 618} 619 620/* Convert the four signed 16-bit values in A to SPFP form. */ 621static __inline __m128 __attribute__((__always_inline__)) 622_mm_cvtpi16_ps (__m64 __A) 623{ 624 __v4hi __sign; 625 __v2si __hisi, __losi; 626 __v4sf __r; 627 628 /* This comparison against zero gives us a mask that can be used to 629 fill in the missing sign bits in the unpack operations below, so 630 that we get signed values after unpacking. */ 631 __sign = __builtin_ia32_pcmpgtw ((__v4hi)0LL, (__v4hi)__A); 632 633 /* Convert the four words to doublewords. */ 634 __hisi = (__v2si) __builtin_ia32_punpckhwd ((__v4hi)__A, __sign); 635 __losi = (__v2si) __builtin_ia32_punpcklwd ((__v4hi)__A, __sign); 636 637 /* Convert the doublewords to floating point two at a time. */ 638 __r = (__v4sf) _mm_setzero_ps (); 639 __r = __builtin_ia32_cvtpi2ps (__r, __hisi); 640 __r = __builtin_ia32_movlhps (__r, __r); 641 __r = __builtin_ia32_cvtpi2ps (__r, __losi); 642 643 return (__m128) __r; 644} 645 646/* Convert the four unsigned 16-bit values in A to SPFP form. */ 647static __inline __m128 __attribute__((__always_inline__)) 648_mm_cvtpu16_ps (__m64 __A) 649{ 650 __v2si __hisi, __losi; 651 __v4sf __r; 652 653 /* Convert the four words to doublewords. */ 654 __hisi = (__v2si) __builtin_ia32_punpckhwd ((__v4hi)__A, (__v4hi)0LL); 655 __losi = (__v2si) __builtin_ia32_punpcklwd ((__v4hi)__A, (__v4hi)0LL); 656 657 /* Convert the doublewords to floating point two at a time. */ 658 __r = (__v4sf) _mm_setzero_ps (); 659 __r = __builtin_ia32_cvtpi2ps (__r, __hisi); 660 __r = __builtin_ia32_movlhps (__r, __r); 661 __r = __builtin_ia32_cvtpi2ps (__r, __losi); 662 663 return (__m128) __r; 664} 665 666/* Convert the low four signed 8-bit values in A to SPFP form. */ 667static __inline __m128 __attribute__((__always_inline__)) 668_mm_cvtpi8_ps (__m64 __A) 669{ 670 __v8qi __sign; 671 672 /* This comparison against zero gives us a mask that can be used to 673 fill in the missing sign bits in the unpack operations below, so 674 that we get signed values after unpacking. */ 675 __sign = __builtin_ia32_pcmpgtb ((__v8qi)0LL, (__v8qi)__A); 676 677 /* Convert the four low bytes to words. */ 678 __A = (__m64) __builtin_ia32_punpcklbw ((__v8qi)__A, __sign); 679 680 return _mm_cvtpi16_ps(__A); 681} 682 683/* Convert the low four unsigned 8-bit values in A to SPFP form. */ 684static __inline __m128 __attribute__((__always_inline__)) 685_mm_cvtpu8_ps(__m64 __A) 686{ 687 __A = (__m64) __builtin_ia32_punpcklbw ((__v8qi)__A, (__v8qi)0LL); 688 return _mm_cvtpu16_ps(__A); 689} 690 691/* Convert the four signed 32-bit values in A and B to SPFP form. */ 692static __inline __m128 __attribute__((__always_inline__)) 693_mm_cvtpi32x2_ps(__m64 __A, __m64 __B) 694{ 695 __v4sf __zero = (__v4sf) _mm_setzero_ps (); 696 __v4sf __sfa = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__A); 697 __v4sf __sfb = __builtin_ia32_cvtpi2ps (__zero, (__v2si)__B); 698 return (__m128) __builtin_ia32_movlhps (__sfa, __sfb); 699} 700 701/* Convert the four SPFP values in A to four signed 16-bit integers. */ 702static __inline __m64 __attribute__((__always_inline__)) 703_mm_cvtps_pi16(__m128 __A) 704{ 705 __v4sf __hisf = (__v4sf)__A; 706 __v4sf __losf = __builtin_ia32_movhlps (__hisf, __hisf); 707 __v2si __hisi = __builtin_ia32_cvtps2pi (__hisf); 708 __v2si __losi = __builtin_ia32_cvtps2pi (__losf); 709 return (__m64) __builtin_ia32_packssdw (__hisi, __losi); 710} 711 712/* Convert the four SPFP values in A to four signed 8-bit integers. */ 713static __inline __m64 __attribute__((__always_inline__)) 714_mm_cvtps_pi8(__m128 __A) 715{ 716 __v4hi __tmp = (__v4hi) _mm_cvtps_pi16 (__A); 717 return (__m64) __builtin_ia32_packsswb (__tmp, (__v4hi)0LL); 718} 719 720/* Selects four specific SPFP values from A and B based on MASK. */ 721#if 0 722static __inline __m128 __attribute__((__always_inline__)) 723_mm_shuffle_ps (__m128 __A, __m128 __B, int __mask) 724{ 725 return (__m128) __builtin_ia32_shufps ((__v4sf)__A, (__v4sf)__B, __mask); 726} 727#else 728#define _mm_shuffle_ps(A, B, MASK) \ 729 ((__m128) __builtin_ia32_shufps ((__v4sf)(A), (__v4sf)(B), (MASK))) 730#endif 731 732 733/* Selects and interleaves the upper two SPFP values from A and B. */ 734static __inline __m128 __attribute__((__always_inline__)) 735_mm_unpackhi_ps (__m128 __A, __m128 __B) 736{ 737 return (__m128) __builtin_ia32_unpckhps ((__v4sf)__A, (__v4sf)__B); 738} 739 740/* Selects and interleaves the lower two SPFP values from A and B. */ 741static __inline __m128 __attribute__((__always_inline__)) 742_mm_unpacklo_ps (__m128 __A, __m128 __B) 743{ 744 return (__m128) __builtin_ia32_unpcklps ((__v4sf)__A, (__v4sf)__B); 745} 746 747/* Sets the upper two SPFP values with 64-bits of data loaded from P; 748 the lower two values are passed through from A. */ 749static __inline __m128 __attribute__((__always_inline__)) 750_mm_loadh_pi (__m128 __A, __m64 const *__P) 751{ 752 return (__m128) __builtin_ia32_loadhps ((__v4sf)__A, (__v2si *)__P); 753} 754 755/* Stores the upper two SPFP values of A into P. */ 756static __inline void __attribute__((__always_inline__)) 757_mm_storeh_pi (__m64 *__P, __m128 __A) 758{ 759 __builtin_ia32_storehps ((__v2si *)__P, (__v4sf)__A); 760} 761 762/* Moves the upper two values of B into the lower two values of A. */ 763static __inline __m128 __attribute__((__always_inline__)) 764_mm_movehl_ps (__m128 __A, __m128 __B) 765{ 766 return (__m128) __builtin_ia32_movhlps ((__v4sf)__A, (__v4sf)__B); 767} 768 769/* Moves the lower two values of B into the upper two values of A. */ 770static __inline __m128 __attribute__((__always_inline__)) 771_mm_movelh_ps (__m128 __A, __m128 __B) 772{ 773 return (__m128) __builtin_ia32_movlhps ((__v4sf)__A, (__v4sf)__B); 774} 775 776/* Sets the lower two SPFP values with 64-bits of data loaded from P; 777 the upper two values are passed through from A. */ 778static __inline __m128 __attribute__((__always_inline__)) 779_mm_loadl_pi (__m128 __A, __m64 const *__P) 780{ 781 return (__m128) __builtin_ia32_loadlps ((__v4sf)__A, (__v2si *)__P); 782} 783 784/* Stores the lower two SPFP values of A into P. */ 785static __inline void __attribute__((__always_inline__)) 786_mm_storel_pi (__m64 *__P, __m128 __A) 787{ 788 __builtin_ia32_storelps ((__v2si *)__P, (__v4sf)__A); 789} 790 791/* Creates a 4-bit mask from the most significant bits of the SPFP values. */ 792static __inline int __attribute__((__always_inline__)) 793_mm_movemask_ps (__m128 __A) 794{ 795 return __builtin_ia32_movmskps ((__v4sf)__A); 796} 797 798/* Return the contents of the control register. */ 799static __inline unsigned int __attribute__((__always_inline__)) 800_mm_getcsr (void) 801{ 802 return __builtin_ia32_stmxcsr (); 803} 804 805/* Read exception bits from the control register. */ 806static __inline unsigned int __attribute__((__always_inline__)) 807_MM_GET_EXCEPTION_STATE (void) 808{ 809 return _mm_getcsr() & _MM_EXCEPT_MASK; 810} 811 812static __inline unsigned int __attribute__((__always_inline__)) 813_MM_GET_EXCEPTION_MASK (void) 814{ 815 return _mm_getcsr() & _MM_MASK_MASK; 816} 817 818static __inline unsigned int __attribute__((__always_inline__)) 819_MM_GET_ROUNDING_MODE (void) 820{ 821 return _mm_getcsr() & _MM_ROUND_MASK; 822} 823 824static __inline unsigned int __attribute__((__always_inline__)) 825_MM_GET_FLUSH_ZERO_MODE (void) 826{ 827 return _mm_getcsr() & _MM_FLUSH_ZERO_MASK; 828} 829 830/* Set the control register to I. */ 831static __inline void __attribute__((__always_inline__)) 832_mm_setcsr (unsigned int __I) 833{ 834 __builtin_ia32_ldmxcsr (__I); 835} 836 837/* Set exception bits in the control register. */ 838static __inline void __attribute__((__always_inline__)) 839_MM_SET_EXCEPTION_STATE(unsigned int __mask) 840{ 841 _mm_setcsr((_mm_getcsr() & ~_MM_EXCEPT_MASK) | __mask); 842} 843 844static __inline void __attribute__((__always_inline__)) 845_MM_SET_EXCEPTION_MASK (unsigned int __mask) 846{ 847 _mm_setcsr((_mm_getcsr() & ~_MM_MASK_MASK) | __mask); 848} 849 850static __inline void __attribute__((__always_inline__)) 851_MM_SET_ROUNDING_MODE (unsigned int __mode) 852{ 853 _mm_setcsr((_mm_getcsr() & ~_MM_ROUND_MASK) | __mode); 854} 855 856static __inline void __attribute__((__always_inline__)) 857_MM_SET_FLUSH_ZERO_MODE (unsigned int __mode) 858{ 859 _mm_setcsr((_mm_getcsr() & ~_MM_FLUSH_ZERO_MASK) | __mode); 860} 861 862/* Create a vector with element 0 as F and the rest zero. */ 863static __inline __m128 __attribute__((__always_inline__)) 864_mm_set_ss (float __F) 865{ 866 return __extension__ (__m128)(__v4sf){ __F, 0, 0, 0 }; 867} 868 869/* Create a vector with all four elements equal to F. */ 870static __inline __m128 __attribute__((__always_inline__)) 871_mm_set1_ps (float __F) 872{ 873 return __extension__ (__m128)(__v4sf){ __F, __F, __F, __F }; 874} 875 876static __inline __m128 __attribute__((__always_inline__)) 877_mm_set_ps1 (float __F) 878{ 879 return _mm_set1_ps (__F); 880} 881 882/* Create a vector with element 0 as *P and the rest zero. */ 883static __inline __m128 __attribute__((__always_inline__)) 884_mm_load_ss (float const *__P) 885{ 886 return _mm_set_ss (*__P); 887} 888 889/* Create a vector with all four elements equal to *P. */ 890static __inline __m128 __attribute__((__always_inline__)) 891_mm_load1_ps (float const *__P) 892{ 893 return _mm_set1_ps (*__P); 894} 895 896static __inline __m128 __attribute__((__always_inline__)) 897_mm_load_ps1 (float const *__P) 898{ 899 return _mm_load1_ps (__P); 900} 901 902/* Load four SPFP values from P. The address must be 16-byte aligned. */ 903static __inline __m128 __attribute__((__always_inline__)) 904_mm_load_ps (float const *__P) 905{ 906 return (__m128) *(__v4sf *)__P; 907} 908 909/* Load four SPFP values from P. The address need not be 16-byte aligned. */ 910static __inline __m128 __attribute__((__always_inline__)) 911_mm_loadu_ps (float const *__P) 912{ 913 return (__m128) __builtin_ia32_loadups (__P); 914} 915 916/* Load four SPFP values in reverse order. The address must be aligned. */ 917static __inline __m128 __attribute__((__always_inline__)) 918_mm_loadr_ps (float const *__P) 919{ 920 __v4sf __tmp = *(__v4sf *)__P; 921 return (__m128) __builtin_ia32_shufps (__tmp, __tmp, _MM_SHUFFLE (0,1,2,3)); 922} 923 924/* Create the vector [Z Y X W]. */ 925static __inline __m128 __attribute__((__always_inline__)) 926_mm_set_ps (const float __Z, const float __Y, const float __X, const float __W) 927{ 928 return __extension__ (__m128)(__v4sf){ __W, __X, __Y, __Z }; 929} 930 931/* Create the vector [W X Y Z]. */ 932static __inline __m128 __attribute__((__always_inline__)) 933_mm_setr_ps (float __Z, float __Y, float __X, float __W) 934{ 935 return __extension__ (__m128)(__v4sf){ __Z, __Y, __X, __W }; 936} 937 938/* Stores the lower SPFP value. */ 939static __inline void __attribute__((__always_inline__)) 940_mm_store_ss (float *__P, __m128 __A) 941{ 942 *__P = __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0); 943} 944 945static __inline float __attribute__((__always_inline__)) 946_mm_cvtss_f32 (__m128 __A) 947{ 948 return __builtin_ia32_vec_ext_v4sf ((__v4sf)__A, 0); 949} 950 951/* Store four SPFP values. The address must be 16-byte aligned. */ 952static __inline void __attribute__((__always_inline__)) 953_mm_store_ps (float *__P, __m128 __A) 954{ 955 *(__v4sf *)__P = (__v4sf)__A; 956} 957 958/* Store four SPFP values. The address need not be 16-byte aligned. */ 959static __inline void __attribute__((__always_inline__)) 960_mm_storeu_ps (float *__P, __m128 __A) 961{ 962 __builtin_ia32_storeups (__P, (__v4sf)__A); 963} 964 965/* Store the lower SPFP value across four words. */ 966static __inline void __attribute__((__always_inline__)) 967_mm_store1_ps (float *__P, __m128 __A) 968{ 969 __v4sf __va = (__v4sf)__A; 970 __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,0,0,0)); 971 _mm_storeu_ps (__P, __tmp); 972} 973 974static __inline void __attribute__((__always_inline__)) 975_mm_store_ps1 (float *__P, __m128 __A) 976{ 977 _mm_store1_ps (__P, __A); 978} 979 980/* Store four SPFP values in reverse order. The address must be aligned. */ 981static __inline void __attribute__((__always_inline__)) 982_mm_storer_ps (float *__P, __m128 __A) 983{ 984 __v4sf __va = (__v4sf)__A; 985 __v4sf __tmp = __builtin_ia32_shufps (__va, __va, _MM_SHUFFLE (0,1,2,3)); 986 _mm_store_ps (__P, __tmp); 987} 988 989/* Sets the low SPFP value of A from the low value of B. */ 990static __inline __m128 __attribute__((__always_inline__)) 991_mm_move_ss (__m128 __A, __m128 __B) 992{ 993 return (__m128) __builtin_ia32_movss ((__v4sf)__A, (__v4sf)__B); 994} 995 996/* Extracts one of the four words of A. The selector N must be immediate. */ 997#if 0 998static __inline int __attribute__((__always_inline__)) 999_mm_extract_pi16 (__m64 const __A, int const __N) 1000{ 1001 return __builtin_ia32_vec_ext_v4hi ((__v4hi)__A, __N); 1002} 1003 1004static __inline int __attribute__((__always_inline__)) 1005_m_pextrw (__m64 const __A, int const __N) 1006{ 1007 return _mm_extract_pi16 (__A, __N); 1008} 1009#else 1010#define _mm_extract_pi16(A, N) __builtin_ia32_vec_ext_v4hi ((__v4hi)(A), (N)) 1011#define _m_pextrw(A, N) _mm_extract_pi16((A), (N)) 1012#endif 1013 1014/* Inserts word D into one of four words of A. The selector N must be 1015 immediate. */ 1016#if 0 1017static __inline __m64 __attribute__((__always_inline__)) 1018_mm_insert_pi16 (__m64 const __A, int const __D, int const __N) 1019{ 1020 return (__m64) __builtin_ia32_vec_set_v4hi ((__v4hi)__A, __D, __N); 1021} 1022 1023static __inline __m64 __attribute__((__always_inline__)) 1024_m_pinsrw (__m64 const __A, int const __D, int const __N) 1025{ 1026 return _mm_insert_pi16 (__A, __D, __N); 1027} 1028#else 1029#define _mm_insert_pi16(A, D, N) \ 1030 ((__m64) __builtin_ia32_vec_set_v4hi ((__v4hi)(A), (D), (N))) 1031#define _m_pinsrw(A, D, N) _mm_insert_pi16((A), (D), (N)) 1032#endif 1033 1034/* Compute the element-wise maximum of signed 16-bit values. */ 1035static __inline __m64 __attribute__((__always_inline__)) 1036_mm_max_pi16 (__m64 __A, __m64 __B) 1037{ 1038 return (__m64) __builtin_ia32_pmaxsw ((__v4hi)__A, (__v4hi)__B); 1039} 1040 1041static __inline __m64 __attribute__((__always_inline__)) 1042_m_pmaxsw (__m64 __A, __m64 __B) 1043{ 1044 return _mm_max_pi16 (__A, __B); 1045} 1046 1047/* Compute the element-wise maximum of unsigned 8-bit values. */ 1048static __inline __m64 __attribute__((__always_inline__)) 1049_mm_max_pu8 (__m64 __A, __m64 __B) 1050{ 1051 return (__m64) __builtin_ia32_pmaxub ((__v8qi)__A, (__v8qi)__B); 1052} 1053 1054static __inline __m64 __attribute__((__always_inline__)) 1055_m_pmaxub (__m64 __A, __m64 __B) 1056{ 1057 return _mm_max_pu8 (__A, __B); 1058} 1059 1060/* Compute the element-wise minimum of signed 16-bit values. */ 1061static __inline __m64 __attribute__((__always_inline__)) 1062_mm_min_pi16 (__m64 __A, __m64 __B) 1063{ 1064 return (__m64) __builtin_ia32_pminsw ((__v4hi)__A, (__v4hi)__B); 1065} 1066 1067static __inline __m64 __attribute__((__always_inline__)) 1068_m_pminsw (__m64 __A, __m64 __B) 1069{ 1070 return _mm_min_pi16 (__A, __B); 1071} 1072 1073/* Compute the element-wise minimum of unsigned 8-bit values. */ 1074static __inline __m64 __attribute__((__always_inline__)) 1075_mm_min_pu8 (__m64 __A, __m64 __B) 1076{ 1077 return (__m64) __builtin_ia32_pminub ((__v8qi)__A, (__v8qi)__B); 1078} 1079 1080static __inline __m64 __attribute__((__always_inline__)) 1081_m_pminub (__m64 __A, __m64 __B) 1082{ 1083 return _mm_min_pu8 (__A, __B); 1084} 1085 1086/* Create an 8-bit mask of the signs of 8-bit values. */ 1087static __inline int __attribute__((__always_inline__)) 1088_mm_movemask_pi8 (__m64 __A) 1089{ 1090 return __builtin_ia32_pmovmskb ((__v8qi)__A); 1091} 1092 1093static __inline int __attribute__((__always_inline__)) 1094_m_pmovmskb (__m64 __A) 1095{ 1096 return _mm_movemask_pi8 (__A); 1097} 1098 1099/* Multiply four unsigned 16-bit values in A by four unsigned 16-bit values 1100 in B and produce the high 16 bits of the 32-bit results. */ 1101static __inline __m64 __attribute__((__always_inline__)) 1102_mm_mulhi_pu16 (__m64 __A, __m64 __B) 1103{ 1104 return (__m64) __builtin_ia32_pmulhuw ((__v4hi)__A, (__v4hi)__B); 1105} 1106 1107static __inline __m64 __attribute__((__always_inline__)) 1108_m_pmulhuw (__m64 __A, __m64 __B) 1109{ 1110 return _mm_mulhi_pu16 (__A, __B); 1111} 1112 1113/* Return a combination of the four 16-bit values in A. The selector 1114 must be an immediate. */ 1115#if 0 1116static __inline __m64 __attribute__((__always_inline__)) 1117_mm_shuffle_pi16 (__m64 __A, int __N) 1118{ 1119 return (__m64) __builtin_ia32_pshufw ((__v4hi)__A, __N); 1120} 1121 1122static __inline __m64 __attribute__((__always_inline__)) 1123_m_pshufw (__m64 __A, int __N) 1124{ 1125 return _mm_shuffle_pi16 (__A, __N); 1126} 1127#else 1128#define _mm_shuffle_pi16(A, N) \ 1129 ((__m64) __builtin_ia32_pshufw ((__v4hi)(A), (N))) 1130#define _m_pshufw(A, N) _mm_shuffle_pi16 ((A), (N)) 1131#endif 1132 1133/* Conditionally store byte elements of A into P. The high bit of each 1134 byte in the selector N determines whether the corresponding byte from 1135 A is stored. */ 1136static __inline void __attribute__((__always_inline__)) 1137_mm_maskmove_si64 (__m64 __A, __m64 __N, char *__P) 1138{ 1139 __builtin_ia32_maskmovq ((__v8qi)__A, (__v8qi)__N, __P); 1140} 1141 1142static __inline void __attribute__((__always_inline__)) 1143_m_maskmovq (__m64 __A, __m64 __N, char *__P) 1144{ 1145 _mm_maskmove_si64 (__A, __N, __P); 1146} 1147 1148/* Compute the rounded averages of the unsigned 8-bit values in A and B. */ 1149static __inline __m64 __attribute__((__always_inline__)) 1150_mm_avg_pu8 (__m64 __A, __m64 __B) 1151{ 1152 return (__m64) __builtin_ia32_pavgb ((__v8qi)__A, (__v8qi)__B); 1153} 1154 1155static __inline __m64 __attribute__((__always_inline__)) 1156_m_pavgb (__m64 __A, __m64 __B) 1157{ 1158 return _mm_avg_pu8 (__A, __B); 1159} 1160 1161/* Compute the rounded averages of the unsigned 16-bit values in A and B. */ 1162static __inline __m64 __attribute__((__always_inline__)) 1163_mm_avg_pu16 (__m64 __A, __m64 __B) 1164{ 1165 return (__m64) __builtin_ia32_pavgw ((__v4hi)__A, (__v4hi)__B); 1166} 1167 1168static __inline __m64 __attribute__((__always_inline__)) 1169_m_pavgw (__m64 __A, __m64 __B) 1170{ 1171 return _mm_avg_pu16 (__A, __B); 1172} 1173 1174/* Compute the sum of the absolute differences of the unsigned 8-bit 1175 values in A and B. Return the value in the lower 16-bit word; the 1176 upper words are cleared. */ 1177static __inline __m64 __attribute__((__always_inline__)) 1178_mm_sad_pu8 (__m64 __A, __m64 __B) 1179{ 1180 return (__m64) __builtin_ia32_psadbw ((__v8qi)__A, (__v8qi)__B); 1181} 1182 1183static __inline __m64 __attribute__((__always_inline__)) 1184_m_psadbw (__m64 __A, __m64 __B) 1185{ 1186 return _mm_sad_pu8 (__A, __B); 1187} 1188 1189/* Loads one cache line from address P to a location "closer" to the 1190 processor. The selector I specifies the type of prefetch operation. */ 1191#if 0 1192static __inline void __attribute__((__always_inline__)) 1193_mm_prefetch (void *__P, enum _mm_hint __I) 1194{ 1195 __builtin_prefetch (__P, 0, __I); 1196} 1197#else 1198#define _mm_prefetch(P, I) \ 1199 __builtin_prefetch ((P), 0, (I)) 1200#endif 1201 1202/* Stores the data in A to the address P without polluting the caches. */ 1203static __inline void __attribute__((__always_inline__)) 1204_mm_stream_pi (__m64 *__P, __m64 __A) 1205{ 1206 __builtin_ia32_movntq ((unsigned long long *)__P, (unsigned long long)__A); 1207} 1208 1209/* Likewise. The address must be 16-byte aligned. */ 1210static __inline void __attribute__((__always_inline__)) 1211_mm_stream_ps (float *__P, __m128 __A) 1212{ 1213 __builtin_ia32_movntps (__P, (__v4sf)__A); 1214} 1215 1216/* Guarantees that every preceding store is globally visible before 1217 any subsequent store. */ 1218static __inline void __attribute__((__always_inline__)) 1219_mm_sfence (void) 1220{ 1221 __builtin_ia32_sfence (); 1222} 1223 1224/* The execution of the next instruction is delayed by an implementation 1225 specific amount of time. The instruction does not modify the 1226 architectural state. */ 1227static __inline void __attribute__((__always_inline__)) 1228_mm_pause (void) 1229{ 1230 __asm__ __volatile__ ("rep; nop" : : ); 1231} 1232 1233/* Transpose the 4x4 matrix composed of row[0-3]. */ 1234#define _MM_TRANSPOSE4_PS(row0, row1, row2, row3) \ 1235do { \ 1236 __v4sf __r0 = (row0), __r1 = (row1), __r2 = (row2), __r3 = (row3); \ 1237 __v4sf __t0 = __builtin_ia32_unpcklps (__r0, __r1); \ 1238 __v4sf __t1 = __builtin_ia32_unpcklps (__r2, __r3); \ 1239 __v4sf __t2 = __builtin_ia32_unpckhps (__r0, __r1); \ 1240 __v4sf __t3 = __builtin_ia32_unpckhps (__r2, __r3); \ 1241 (row0) = __builtin_ia32_movlhps (__t0, __t1); \ 1242 (row1) = __builtin_ia32_movhlps (__t1, __t0); \ 1243 (row2) = __builtin_ia32_movlhps (__t2, __t3); \ 1244 (row3) = __builtin_ia32_movhlps (__t3, __t2); \ 1245} while (0) 1246 1247/* For backward source compatibility. */ 1248#ifdef __SSE2__ 1249#include <emmintrin.h> 1250#endif 1251 1252#endif /* __SSE__ */ 1253#endif /* _XMMINTRIN_H_INCLUDED */ 1254