mmx-builtins.c revision da500caf3ccd04214465e3c457c83e3a4adf8175
1// RUN: %clang_cc1 %s -O3 -triple=x86_64-apple-darwin -target-feature +ssse3 -S -o - | FileCheck %s 2 3#include <tmmintrin.h> 4 5__m64 test1(__m64 a, __m64 b) { 6 // CHECK: phaddw 7 return _mm_hadd_pi16(a, b); 8} 9 10__m64 test2(__m64 a, __m64 b) { 11 // CHECK: phaddd 12 return _mm_hadd_pi32(a, b); 13} 14 15__m64 test3(__m64 a, __m64 b) { 16 // CHECK: phaddsw 17 return _mm_hadds_pi16(a, b); 18} 19 20__m64 test4(__m64 a, __m64 b) { 21 // CHECK: phsubw 22 return _mm_hsub_pi16(a, b); 23} 24 25__m64 test5(__m64 a, __m64 b) { 26 // CHECK: phsubd 27 return _mm_hsub_pi32(a, b); 28} 29 30__m64 test6(__m64 a, __m64 b) { 31 // CHECK: phsubsw 32 return _mm_hsubs_pi16(a, b); 33} 34 35__m64 test7(__m64 a, __m64 b) { 36 // CHECK: pmaddubsw 37 return _mm_maddubs_pi16(a, b); 38} 39 40__m64 test8(__m64 a, __m64 b) { 41 // CHECK: pmulhrsw 42 return _mm_mulhrs_pi16(a, b); 43} 44 45__m64 test9(__m64 a, __m64 b) { 46 // CHECK: pshufb 47 return _mm_shuffle_pi8(a, b); 48} 49 50__m64 test10(__m64 a, __m64 b) { 51 // CHECK: psignb 52 return _mm_sign_pi8(a, b); 53} 54 55__m64 test11(__m64 a, __m64 b) { 56 // CHECK: psignw 57 return _mm_sign_pi16(a, b); 58} 59 60__m64 test12(__m64 a, __m64 b) { 61 // CHECK: psignd 62 return _mm_sign_pi32(a, b); 63} 64 65__m64 test13(__m64 a) { 66 // CHECK: pabsb 67 return _mm_abs_pi8(a); 68} 69 70__m64 test14(__m64 a) { 71 // CHECK: pabsw 72 return _mm_abs_pi16(a); 73} 74 75__m64 test15(__m64 a) { 76 // CHECK: pabsd 77 return _mm_abs_pi32(a); 78} 79 80__m64 test16(__m64 a, __m64 b) { 81 // CHECK: palignr 82 return _mm_alignr_pi8(a, b, 2); 83} 84 85__m64 test17(__m128d a) { 86 // CHECK: cvtpd2pi 87 return _mm_cvtpd_pi32(a); 88} 89 90__m64 test18(__m128d a) { 91 // CHECK: cvttpd2pi 92 return _mm_cvttpd_pi32(a); 93} 94 95__m128d test19(__m64 a) { 96 // CHECK: cvtpi2pd 97 return _mm_cvtpi32_pd(a); 98} 99 100__m64 test20(__m64 a, __m64 b) { 101 // CHECK: pmuludq 102 return _mm_mul_su32(a, b); 103} 104 105__m64 test21(__m64 a) { 106 // CHECK: pshufw 107 return _mm_shuffle_pi16(a, 3); 108} 109 110__m64 test22(__m64 a, __m64 b) { 111 // CHECK: pmulhuw 112 return _mm_mulhi_pu16(a, b); 113} 114 115void test23(__m64 d, __m64 n, char *p) { 116 // CHECK: maskmovq 117 _mm_maskmove_si64(d, n, p); 118} 119 120int test24(__m64 a) { 121 // CHECK: pmovmskb 122 return _mm_movemask_pi8(a); 123} 124 125void test25(__m64 *p, __m64 a) { 126 // CHECK: movntq 127 _mm_stream_pi(p, a); 128} 129 130__m64 test26(__m64 a, __m64 b) { 131 // CHECK: pavgb 132 return _mm_avg_pu8(a, b); 133} 134 135__m64 test27(__m64 a, __m64 b) { 136 // CHECK: pavgw 137 return _mm_avg_pu16(a, b); 138} 139 140__m64 test28(__m64 a, __m64 b) { 141 // CHECK: pmaxub 142 return _mm_max_pu8(a, b); 143} 144 145__m64 test29(__m64 a, __m64 b) { 146 // CHECK: pmaxsw 147 return _mm_max_pi16(a, b); 148} 149 150__m64 test30(__m64 a, __m64 b) { 151 // CHECK: pminub 152 return _mm_min_pu8(a, b); 153} 154 155__m64 test31(__m64 a, __m64 b) { 156 // CHECK: pminsw 157 return _mm_min_pi16(a, b); 158} 159 160__m64 test32(__m64 a, __m64 b) { 161 // CHECK: psadbw 162 return _mm_sad_pu8(a, b); 163} 164 165__m64 test33(__m64 a, __m64 b) { 166 // CHECK: paddb 167 return _mm_add_pi8(a, b); 168} 169 170__m64 test34(__m64 a, __m64 b) { 171 // CHECK: paddw 172 return _mm_add_pi16(a, b); 173} 174 175__m64 test35(__m64 a, __m64 b) { 176 // CHECK: paddd 177 return _mm_add_pi32(a, b); 178} 179 180__m64 test36(__m64 a, __m64 b) { 181 // CHECK: paddq 182 return __builtin_ia32_paddq(a, b); 183} 184 185__m64 test37(__m64 a, __m64 b) { 186 // CHECK: paddsb 187 return _mm_adds_pi8(a, b); 188} 189 190__m64 test38(__m64 a, __m64 b) { 191 // CHECK: paddsw 192 return _mm_adds_pi16(a, b); 193} 194 195__m64 test39(__m64 a, __m64 b) { 196 // CHECK: paddusb 197 return _mm_adds_pu8(a, b); 198} 199 200__m64 test40(__m64 a, __m64 b) { 201 // CHECK: paddusw 202 return _mm_adds_pu16(a, b); 203} 204 205__m64 test41(__m64 a, __m64 b) { 206 // CHECK: psubb 207 return _mm_sub_pi8(a, b); 208} 209 210__m64 test42(__m64 a, __m64 b) { 211 // CHECK: psubw 212 return _mm_sub_pi16(a, b); 213} 214 215__m64 test43(__m64 a, __m64 b) { 216 // CHECK: psubd 217 return _mm_sub_pi32(a, b); 218} 219 220__m64 test44(__m64 a, __m64 b) { 221 // CHECK: psubq 222 return __builtin_ia32_psubq(a, b); 223} 224 225__m64 test45(__m64 a, __m64 b) { 226 // CHECK: psubsb 227 return _mm_subs_pi8(a, b); 228} 229 230__m64 test46(__m64 a, __m64 b) { 231 // CHECK: psubsw 232 return _mm_subs_pi16(a, b); 233} 234 235__m64 test47(__m64 a, __m64 b) { 236 // CHECK: psubusb 237 return _mm_subs_pu8(a, b); 238} 239 240__m64 test48(__m64 a, __m64 b) { 241 // CHECK: psubusw 242 return _mm_subs_pu16(a, b); 243} 244 245__m64 test49(__m64 a, __m64 b) { 246 // CHECK: pmaddwd 247 return _mm_madd_pi16(a, b); 248} 249 250__m64 test50(__m64 a, __m64 b) { 251 // CHECK: pmulhw 252 return _mm_mulhi_pi16(a, b); 253} 254 255__m64 test51(__m64 a, __m64 b) { 256 // CHECK: pmullw 257 return _mm_mullo_pi16(a, b); 258} 259 260__m64 test52(__m64 a, __m64 b) { 261 // CHECK: pmullw 262 return _mm_mullo_pi16(a, b); 263} 264 265__m64 test53(__m64 a, __m64 b) { 266 // CHECK: pand 267 return _mm_and_si64(a, b); 268} 269 270__m64 test54(__m64 a, __m64 b) { 271 // CHECK: pandn 272 return _mm_andnot_si64(a, b); 273} 274 275__m64 test55(__m64 a, __m64 b) { 276 // CHECK: por 277 return _mm_or_si64(a, b); 278} 279 280__m64 test56(__m64 a, __m64 b) { 281 // CHECK: pxor 282 return _mm_xor_si64(a, b); 283} 284 285__m64 test57(__m64 a, __m64 b) { 286 // CHECK: pavgb 287 return _mm_avg_pu8(a, b); 288} 289 290__m64 test58(__m64 a, __m64 b) { 291 // CHECK: pavgw 292 return _mm_avg_pu16(a, b); 293} 294 295__m64 test59(__m64 a, __m64 b) { 296 // CHECK: psllw 297 return _mm_sll_pi16(a, b); 298} 299 300__m64 test60(__m64 a, __m64 b) { 301 // CHECK: pslld 302 return _mm_sll_pi32(a, b); 303} 304 305__m64 test61(__m64 a, __m64 b) { 306 // CHECK: psllq 307 return _mm_sll_si64(a, b); 308} 309 310__m64 test62(__m64 a, __m64 b) { 311 // CHECK: psrlw 312 return _mm_srl_pi16(a, b); 313} 314 315__m64 test63(__m64 a, __m64 b) { 316 // CHECK: psrld 317 return _mm_srl_pi32(a, b); 318} 319 320__m64 test64(__m64 a, __m64 b) { 321 // CHECK: psrlq 322 return _mm_srl_si64(a, b); 323} 324 325__m64 test65(__m64 a, __m64 b) { 326 // CHECK: psraw 327 return _mm_sra_pi16(a, b); 328} 329 330__m64 test66(__m64 a, __m64 b) { 331 // CHECK: psrad 332 return _mm_sra_pi32(a, b); 333} 334 335__m64 test67(__m64 a) { 336 // CHECK: psllw 337 return _mm_slli_pi16(a, 3); 338} 339 340__m64 test68(__m64 a) { 341 // CHECK: pslld 342 return _mm_slli_pi32(a, 3); 343} 344 345__m64 test69(__m64 a) { 346 // CHECK: psllq 347 return _mm_slli_si64(a, 3); 348} 349 350__m64 test70(__m64 a) { 351 // CHECK: psrlw 352 return _mm_srli_pi16(a, 3); 353} 354 355__m64 test71(__m64 a) { 356 // CHECK: psrld 357 return _mm_srli_pi32(a, 3); 358} 359 360__m64 test72(__m64 a) { 361 // CHECK: psrlq 362 return _mm_srli_si64(a, 3); 363} 364 365__m64 test73(__m64 a) { 366 // CHECK: psraw 367 return _mm_srai_pi16(a, 3); 368} 369 370__m64 test74(__m64 a) { 371 // CHECK: psrad 372 return _mm_srai_pi32(a, 3); 373} 374 375__m64 test75(__m64 a, __m64 b) { 376 // CHECK: packsswb 377 return _mm_packs_pi16(a, b); 378} 379 380__m64 test76(__m64 a, __m64 b) { 381 // CHECK: packssdw 382 return _mm_packs_pi32(a, b); 383} 384 385__m64 test77(__m64 a, __m64 b) { 386 // CHECK: packuswb 387 return _mm_packs_pu16(a, b); 388} 389 390__m64 test78(__m64 a, __m64 b) { 391 // CHECK: punpckhbw 392 return _mm_unpackhi_pi8(a, b); 393} 394 395__m64 test79(__m64 a, __m64 b) { 396 // CHECK: punpckhwd 397 return _mm_unpackhi_pi16(a, b); 398} 399 400__m64 test80(__m64 a, __m64 b) { 401 // CHECK: punpckhdq 402 return _mm_unpackhi_pi32(a, b); 403} 404 405__m64 test81(__m64 a, __m64 b) { 406 // CHECK: punpcklbw 407 return _mm_unpacklo_pi8(a, b); 408} 409 410__m64 test82(__m64 a, __m64 b) { 411 // CHECK: punpcklwd 412 return _mm_unpacklo_pi16(a, b); 413} 414 415__m64 test83(__m64 a, __m64 b) { 416 // CHECK: punpckldq 417 return _mm_unpacklo_pi32(a, b); 418} 419 420__m64 test84(__m64 a, __m64 b) { 421 // CHECK: pcmpeqb 422 return _mm_cmpeq_pi8(a, b); 423} 424 425__m64 test85(__m64 a, __m64 b) { 426 // CHECK: pcmpeqw 427 return _mm_cmpeq_pi16(a, b); 428} 429 430__m64 test86(__m64 a, __m64 b) { 431 // CHECK: pcmpeqd 432 return _mm_cmpeq_pi32(a, b); 433} 434 435__m64 test87(__m64 a, __m64 b) { 436 // CHECK: pcmpgtb 437 return _mm_cmpgt_pi8(a, b); 438} 439 440__m64 test88(__m64 a, __m64 b) { 441 // CHECK: pcmpgtw 442 return _mm_cmpgt_pi16(a, b); 443} 444 445__m64 test89(__m64 a, __m64 b) { 446 // CHECK: pcmpgtd 447 return _mm_cmpgt_pi32(a, b); 448} 449