mmintrin.h revision 272461
1208562Srwatson/*===---- mmintrin.h - MMX intrinsics --------------------------------------=== 2208562Srwatson * 3208562Srwatson * Permission is hereby granted, free of charge, to any person obtaining a copy 4208562Srwatson * of this software and associated documentation files (the "Software"), to deal 5208562Srwatson * in the Software without restriction, including without limitation the rights 6208562Srwatson * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 7208562Srwatson * copies of the Software, and to permit persons to whom the Software is 8208562Srwatson * furnished to do so, subject to the following conditions: 9208562Srwatson * 10208562Srwatson * The above copyright notice and this permission notice shall be included in 11208562Srwatson * all copies or substantial portions of the Software. 12208562Srwatson * 13208562Srwatson * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14208562Srwatson * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15208562Srwatson * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16208562Srwatson * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17208562Srwatson * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 18208562Srwatson * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 19208562Srwatson * THE SOFTWARE. 20208562Srwatson * 21208562Srwatson *===-----------------------------------------------------------------------=== 22208562Srwatson */ 23208562Srwatson 24208562Srwatson#ifndef __MMINTRIN_H 25208562Srwatson#define __MMINTRIN_H 26208562Srwatson 27208562Srwatson#ifndef __MMX__ 28208562Srwatson#error "MMX instruction set not enabled" 29208562Srwatson#else 30208562Srwatson 31208562Srwatsontypedef long long __m64 __attribute__((__vector_size__(8))); 32208562Srwatson 33208562Srwatsontypedef int __v2si __attribute__((__vector_size__(8))); 34208562Srwatsontypedef short __v4hi __attribute__((__vector_size__(8))); 35208562Srwatsontypedef char __v8qi __attribute__((__vector_size__(8))); 36208562Srwatson 37208562Srwatsonstatic __inline__ void __attribute__((__always_inline__, __nodebug__)) 38208562Srwatson_mm_empty(void) 39208562Srwatson{ 40208562Srwatson __builtin_ia32_emms(); 41208562Srwatson} 42208562Srwatson 43208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 44208562Srwatson_mm_cvtsi32_si64(int __i) 45208562Srwatson{ 46208562Srwatson return (__m64)__builtin_ia32_vec_init_v2si(__i, 0); 47208562Srwatson} 48208562Srwatson 49208562Srwatsonstatic __inline__ int __attribute__((__always_inline__, __nodebug__)) 50208562Srwatson_mm_cvtsi64_si32(__m64 __m) 51208562Srwatson{ 52208562Srwatson return __builtin_ia32_vec_ext_v2si((__v2si)__m, 0); 53208562Srwatson} 54208562Srwatson 55208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 56208562Srwatson_mm_cvtsi64_m64(long long __i) 57208562Srwatson{ 58208562Srwatson return (__m64)__i; 59208562Srwatson} 60208562Srwatson 61208562Srwatsonstatic __inline__ long long __attribute__((__always_inline__, __nodebug__)) 62208562Srwatson_mm_cvtm64_si64(__m64 __m) 63208562Srwatson{ 64208562Srwatson return (long long)__m; 65208562Srwatson} 66208562Srwatson 67208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 68208562Srwatson_mm_packs_pi16(__m64 __m1, __m64 __m2) 69208562Srwatson{ 70208562Srwatson return (__m64)__builtin_ia32_packsswb((__v4hi)__m1, (__v4hi)__m2); 71208562Srwatson} 72208562Srwatson 73208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 74208562Srwatson_mm_packs_pi32(__m64 __m1, __m64 __m2) 75208562Srwatson{ 76208562Srwatson return (__m64)__builtin_ia32_packssdw((__v2si)__m1, (__v2si)__m2); 77208562Srwatson} 78208602Srwatson 79208602Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 80208562Srwatson_mm_packs_pu16(__m64 __m1, __m64 __m2) 81208562Srwatson{ 82208562Srwatson return (__m64)__builtin_ia32_packuswb((__v4hi)__m1, (__v4hi)__m2); 83208562Srwatson} 84208562Srwatson 85208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 86208562Srwatson_mm_unpackhi_pi8(__m64 __m1, __m64 __m2) 87208562Srwatson{ 88208562Srwatson return (__m64)__builtin_ia32_punpckhbw((__v8qi)__m1, (__v8qi)__m2); 89208562Srwatson} 90208562Srwatson 91208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 92208562Srwatson_mm_unpackhi_pi16(__m64 __m1, __m64 __m2) 93208562Srwatson{ 94208562Srwatson return (__m64)__builtin_ia32_punpckhwd((__v4hi)__m1, (__v4hi)__m2); 95208562Srwatson} 96208562Srwatson 97208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 98208562Srwatson_mm_unpackhi_pi32(__m64 __m1, __m64 __m2) 99208562Srwatson{ 100208562Srwatson return (__m64)__builtin_ia32_punpckhdq((__v2si)__m1, (__v2si)__m2); 101208562Srwatson} 102208562Srwatson 103208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 104208562Srwatson_mm_unpacklo_pi8(__m64 __m1, __m64 __m2) 105208562Srwatson{ 106208562Srwatson return (__m64)__builtin_ia32_punpcklbw((__v8qi)__m1, (__v8qi)__m2); 107208602Srwatson} 108208602Srwatson 109208602Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 110208602Srwatson_mm_unpacklo_pi16(__m64 __m1, __m64 __m2) 111208562Srwatson{ 112208602Srwatson return (__m64)__builtin_ia32_punpcklwd((__v4hi)__m1, (__v4hi)__m2); 113208602Srwatson} 114208602Srwatson 115208602Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 116208602Srwatson_mm_unpacklo_pi32(__m64 __m1, __m64 __m2) 117208602Srwatson{ 118208602Srwatson return (__m64)__builtin_ia32_punpckldq((__v2si)__m1, (__v2si)__m2); 119208602Srwatson} 120208562Srwatson 121208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 122208562Srwatson_mm_add_pi8(__m64 __m1, __m64 __m2) 123208562Srwatson{ 124208562Srwatson return (__m64)__builtin_ia32_paddb((__v8qi)__m1, (__v8qi)__m2); 125208562Srwatson} 126208562Srwatson 127208562Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 128208562Srwatson_mm_add_pi16(__m64 __m1, __m64 __m2) 129208562Srwatson{ 130208562Srwatson return (__m64)__builtin_ia32_paddw((__v4hi)__m1, (__v4hi)__m2); 131208602Srwatson} 132208562Srwatson 133208602Srwatsonstatic __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 134208562Srwatson_mm_add_pi32(__m64 __m1, __m64 __m2) 135208562Srwatson{ 136208562Srwatson return (__m64)__builtin_ia32_paddd((__v2si)__m1, (__v2si)__m2); 137208562Srwatson} 138208562Srwatson 139static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 140_mm_adds_pi8(__m64 __m1, __m64 __m2) 141{ 142 return (__m64)__builtin_ia32_paddsb((__v8qi)__m1, (__v8qi)__m2); 143} 144 145static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 146_mm_adds_pi16(__m64 __m1, __m64 __m2) 147{ 148 return (__m64)__builtin_ia32_paddsw((__v4hi)__m1, (__v4hi)__m2); 149} 150 151static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 152_mm_adds_pu8(__m64 __m1, __m64 __m2) 153{ 154 return (__m64)__builtin_ia32_paddusb((__v8qi)__m1, (__v8qi)__m2); 155} 156 157static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 158_mm_adds_pu16(__m64 __m1, __m64 __m2) 159{ 160 return (__m64)__builtin_ia32_paddusw((__v4hi)__m1, (__v4hi)__m2); 161} 162 163static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 164_mm_sub_pi8(__m64 __m1, __m64 __m2) 165{ 166 return (__m64)__builtin_ia32_psubb((__v8qi)__m1, (__v8qi)__m2); 167} 168 169static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 170_mm_sub_pi16(__m64 __m1, __m64 __m2) 171{ 172 return (__m64)__builtin_ia32_psubw((__v4hi)__m1, (__v4hi)__m2); 173} 174 175static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 176_mm_sub_pi32(__m64 __m1, __m64 __m2) 177{ 178 return (__m64)__builtin_ia32_psubd((__v2si)__m1, (__v2si)__m2); 179} 180 181static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 182_mm_subs_pi8(__m64 __m1, __m64 __m2) 183{ 184 return (__m64)__builtin_ia32_psubsb((__v8qi)__m1, (__v8qi)__m2); 185} 186 187static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 188_mm_subs_pi16(__m64 __m1, __m64 __m2) 189{ 190 return (__m64)__builtin_ia32_psubsw((__v4hi)__m1, (__v4hi)__m2); 191} 192 193static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 194_mm_subs_pu8(__m64 __m1, __m64 __m2) 195{ 196 return (__m64)__builtin_ia32_psubusb((__v8qi)__m1, (__v8qi)__m2); 197} 198 199static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 200_mm_subs_pu16(__m64 __m1, __m64 __m2) 201{ 202 return (__m64)__builtin_ia32_psubusw((__v4hi)__m1, (__v4hi)__m2); 203} 204 205static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 206_mm_madd_pi16(__m64 __m1, __m64 __m2) 207{ 208 return (__m64)__builtin_ia32_pmaddwd((__v4hi)__m1, (__v4hi)__m2); 209} 210 211static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 212_mm_mulhi_pi16(__m64 __m1, __m64 __m2) 213{ 214 return (__m64)__builtin_ia32_pmulhw((__v4hi)__m1, (__v4hi)__m2); 215} 216 217static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 218_mm_mullo_pi16(__m64 __m1, __m64 __m2) 219{ 220 return (__m64)__builtin_ia32_pmullw((__v4hi)__m1, (__v4hi)__m2); 221} 222 223static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 224_mm_sll_pi16(__m64 __m, __m64 __count) 225{ 226 return (__m64)__builtin_ia32_psllw((__v4hi)__m, __count); 227} 228 229static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 230_mm_slli_pi16(__m64 __m, int __count) 231{ 232 return (__m64)__builtin_ia32_psllwi((__v4hi)__m, __count); 233} 234 235static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 236_mm_sll_pi32(__m64 __m, __m64 __count) 237{ 238 return (__m64)__builtin_ia32_pslld((__v2si)__m, __count); 239} 240 241static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 242_mm_slli_pi32(__m64 __m, int __count) 243{ 244 return (__m64)__builtin_ia32_pslldi((__v2si)__m, __count); 245} 246 247static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 248_mm_sll_si64(__m64 __m, __m64 __count) 249{ 250 return (__m64)__builtin_ia32_psllq(__m, __count); 251} 252 253static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 254_mm_slli_si64(__m64 __m, int __count) 255{ 256 return (__m64)__builtin_ia32_psllqi(__m, __count); 257} 258 259static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 260_mm_sra_pi16(__m64 __m, __m64 __count) 261{ 262 return (__m64)__builtin_ia32_psraw((__v4hi)__m, __count); 263} 264 265static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 266_mm_srai_pi16(__m64 __m, int __count) 267{ 268 return (__m64)__builtin_ia32_psrawi((__v4hi)__m, __count); 269} 270 271static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 272_mm_sra_pi32(__m64 __m, __m64 __count) 273{ 274 return (__m64)__builtin_ia32_psrad((__v2si)__m, __count); 275} 276 277static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 278_mm_srai_pi32(__m64 __m, int __count) 279{ 280 return (__m64)__builtin_ia32_psradi((__v2si)__m, __count); 281} 282 283static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 284_mm_srl_pi16(__m64 __m, __m64 __count) 285{ 286 return (__m64)__builtin_ia32_psrlw((__v4hi)__m, __count); 287} 288 289static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 290_mm_srli_pi16(__m64 __m, int __count) 291{ 292 return (__m64)__builtin_ia32_psrlwi((__v4hi)__m, __count); 293} 294 295static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 296_mm_srl_pi32(__m64 __m, __m64 __count) 297{ 298 return (__m64)__builtin_ia32_psrld((__v2si)__m, __count); 299} 300 301static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 302_mm_srli_pi32(__m64 __m, int __count) 303{ 304 return (__m64)__builtin_ia32_psrldi((__v2si)__m, __count); 305} 306 307static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 308_mm_srl_si64(__m64 __m, __m64 __count) 309{ 310 return (__m64)__builtin_ia32_psrlq(__m, __count); 311} 312 313static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 314_mm_srli_si64(__m64 __m, int __count) 315{ 316 return (__m64)__builtin_ia32_psrlqi(__m, __count); 317} 318 319static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 320_mm_and_si64(__m64 __m1, __m64 __m2) 321{ 322 return __builtin_ia32_pand(__m1, __m2); 323} 324 325static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 326_mm_andnot_si64(__m64 __m1, __m64 __m2) 327{ 328 return __builtin_ia32_pandn(__m1, __m2); 329} 330 331static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 332_mm_or_si64(__m64 __m1, __m64 __m2) 333{ 334 return __builtin_ia32_por(__m1, __m2); 335} 336 337static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 338_mm_xor_si64(__m64 __m1, __m64 __m2) 339{ 340 return __builtin_ia32_pxor(__m1, __m2); 341} 342 343static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 344_mm_cmpeq_pi8(__m64 __m1, __m64 __m2) 345{ 346 return (__m64)__builtin_ia32_pcmpeqb((__v8qi)__m1, (__v8qi)__m2); 347} 348 349static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 350_mm_cmpeq_pi16(__m64 __m1, __m64 __m2) 351{ 352 return (__m64)__builtin_ia32_pcmpeqw((__v4hi)__m1, (__v4hi)__m2); 353} 354 355static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 356_mm_cmpeq_pi32(__m64 __m1, __m64 __m2) 357{ 358 return (__m64)__builtin_ia32_pcmpeqd((__v2si)__m1, (__v2si)__m2); 359} 360 361static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 362_mm_cmpgt_pi8(__m64 __m1, __m64 __m2) 363{ 364 return (__m64)__builtin_ia32_pcmpgtb((__v8qi)__m1, (__v8qi)__m2); 365} 366 367static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 368_mm_cmpgt_pi16(__m64 __m1, __m64 __m2) 369{ 370 return (__m64)__builtin_ia32_pcmpgtw((__v4hi)__m1, (__v4hi)__m2); 371} 372 373static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 374_mm_cmpgt_pi32(__m64 __m1, __m64 __m2) 375{ 376 return (__m64)__builtin_ia32_pcmpgtd((__v2si)__m1, (__v2si)__m2); 377} 378 379static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 380_mm_setzero_si64(void) 381{ 382 return (__m64){ 0LL }; 383} 384 385static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 386_mm_set_pi32(int __i1, int __i0) 387{ 388 return (__m64)__builtin_ia32_vec_init_v2si(__i0, __i1); 389} 390 391static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 392_mm_set_pi16(short __s3, short __s2, short __s1, short __s0) 393{ 394 return (__m64)__builtin_ia32_vec_init_v4hi(__s0, __s1, __s2, __s3); 395} 396 397static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 398_mm_set_pi8(char __b7, char __b6, char __b5, char __b4, char __b3, char __b2, 399 char __b1, char __b0) 400{ 401 return (__m64)__builtin_ia32_vec_init_v8qi(__b0, __b1, __b2, __b3, 402 __b4, __b5, __b6, __b7); 403} 404 405static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 406_mm_set1_pi32(int __i) 407{ 408 return _mm_set_pi32(__i, __i); 409} 410 411static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 412_mm_set1_pi16(short __w) 413{ 414 return _mm_set_pi16(__w, __w, __w, __w); 415} 416 417static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 418_mm_set1_pi8(char __b) 419{ 420 return _mm_set_pi8(__b, __b, __b, __b, __b, __b, __b, __b); 421} 422 423static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 424_mm_setr_pi32(int __i0, int __i1) 425{ 426 return _mm_set_pi32(__i1, __i0); 427} 428 429static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 430_mm_setr_pi16(short __w0, short __w1, short __w2, short __w3) 431{ 432 return _mm_set_pi16(__w3, __w2, __w1, __w0); 433} 434 435static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) 436_mm_setr_pi8(char __b0, char __b1, char __b2, char __b3, char __b4, char __b5, 437 char __b6, char __b7) 438{ 439 return _mm_set_pi8(__b7, __b6, __b5, __b4, __b3, __b2, __b1, __b0); 440} 441 442 443/* Aliases for compatibility. */ 444#define _m_empty _mm_empty 445#define _m_from_int _mm_cvtsi32_si64 446#define _m_to_int _mm_cvtsi64_si32 447#define _m_packsswb _mm_packs_pi16 448#define _m_packssdw _mm_packs_pi32 449#define _m_packuswb _mm_packs_pu16 450#define _m_punpckhbw _mm_unpackhi_pi8 451#define _m_punpckhwd _mm_unpackhi_pi16 452#define _m_punpckhdq _mm_unpackhi_pi32 453#define _m_punpcklbw _mm_unpacklo_pi8 454#define _m_punpcklwd _mm_unpacklo_pi16 455#define _m_punpckldq _mm_unpacklo_pi32 456#define _m_paddb _mm_add_pi8 457#define _m_paddw _mm_add_pi16 458#define _m_paddd _mm_add_pi32 459#define _m_paddsb _mm_adds_pi8 460#define _m_paddsw _mm_adds_pi16 461#define _m_paddusb _mm_adds_pu8 462#define _m_paddusw _mm_adds_pu16 463#define _m_psubb _mm_sub_pi8 464#define _m_psubw _mm_sub_pi16 465#define _m_psubd _mm_sub_pi32 466#define _m_psubsb _mm_subs_pi8 467#define _m_psubsw _mm_subs_pi16 468#define _m_psubusb _mm_subs_pu8 469#define _m_psubusw _mm_subs_pu16 470#define _m_pmaddwd _mm_madd_pi16 471#define _m_pmulhw _mm_mulhi_pi16 472#define _m_pmullw _mm_mullo_pi16 473#define _m_psllw _mm_sll_pi16 474#define _m_psllwi _mm_slli_pi16 475#define _m_pslld _mm_sll_pi32 476#define _m_pslldi _mm_slli_pi32 477#define _m_psllq _mm_sll_si64 478#define _m_psllqi _mm_slli_si64 479#define _m_psraw _mm_sra_pi16 480#define _m_psrawi _mm_srai_pi16 481#define _m_psrad _mm_sra_pi32 482#define _m_psradi _mm_srai_pi32 483#define _m_psrlw _mm_srl_pi16 484#define _m_psrlwi _mm_srli_pi16 485#define _m_psrld _mm_srl_pi32 486#define _m_psrldi _mm_srli_pi32 487#define _m_psrlq _mm_srl_si64 488#define _m_psrlqi _mm_srli_si64 489#define _m_pand _mm_and_si64 490#define _m_pandn _mm_andnot_si64 491#define _m_por _mm_or_si64 492#define _m_pxor _mm_xor_si64 493#define _m_pcmpeqb _mm_cmpeq_pi8 494#define _m_pcmpeqw _mm_cmpeq_pi16 495#define _m_pcmpeqd _mm_cmpeq_pi32 496#define _m_pcmpgtb _mm_cmpgt_pi8 497#define _m_pcmpgtw _mm_cmpgt_pi16 498#define _m_pcmpgtd _mm_cmpgt_pi32 499 500#endif /* __MMX__ */ 501 502#endif /* __MMINTRIN_H */ 503 504