1 2 #include <stdio.h> 3 #include <stdlib.h> 4 #include <assert.h> 5 #include <malloc.h> 6 7 typedef unsigned char UChar; 8 typedef unsigned int UInt; 9 typedef unsigned long int UWord; 10 typedef unsigned long long int ULong; 11 12 #define IS_32_ALIGNED(_ptr) (0 == (0x1F & (UWord)(_ptr))) 13 14 typedef union { UChar u8[32]; UInt u32[8]; } YMM; 15 16 typedef struct { YMM a1; YMM a2; YMM a3; YMM a4; ULong u64; } Block; 17 18 void showYMM ( YMM* vec ) 19 { 20 int i; 21 assert(IS_32_ALIGNED(vec)); 22 for (i = 31; i >= 0; i--) { 23 printf("%02x", (UInt)vec->u8[i]); 24 if (i > 0 && 0 == ((i+0) & 7)) printf("."); 25 } 26 } 27 28 void showBlock ( char* msg, Block* block ) 29 { 30 printf(" %s\n", msg); 31 printf(" "); showYMM(&block->a1); printf("\n"); 32 printf(" "); showYMM(&block->a2); printf("\n"); 33 printf(" "); showYMM(&block->a3); printf("\n"); 34 printf(" "); showYMM(&block->a4); printf("\n"); 35 printf(" %016llx\n", block->u64); 36 } 37 38 UChar randUChar ( void ) 39 { 40 static UInt seed = 80021; 41 seed = 1103515245 * seed + 12345; 42 return (seed >> 17) & 0xFF; 43 } 44 45 void randBlock ( Block* b ) 46 { 47 int i; 48 UChar* p = (UChar*)b; 49 for (i = 0; i < sizeof(Block); i++) 50 p[i] = randUChar(); 51 } 52 53 54 /* Generate a function test_NAME, that tests the given insn, in both 55 its mem and reg forms. The reg form of the insn may mention, as 56 operands only %ymm6, %ymm7, %ymm8, %ymm9 and %r14. The mem form of 57 the insn may mention as operands only (%rax), %ymm7, %ymm8, %ymm9 58 and %r14. It's OK for the insn to clobber ymm0, as this is needed 59 for testing PCMPxSTRx, and ymm6, as this is needed for testing 60 MOVMASK variants. */ 61 62 #define GEN_test_RandM(_name, _reg_form, _mem_form) \ 63 \ 64 __attribute__ ((noinline)) static void test_##_name ( void ) \ 65 { \ 66 Block* b = memalign(32, sizeof(Block)); \ 67 randBlock(b); \ 68 printf("%s(reg)\n", #_name); \ 69 showBlock("before", b); \ 70 __asm__ __volatile__( \ 71 "vmovdqa 0(%0),%%ymm7" "\n\t" \ 72 "vmovdqa 32(%0),%%ymm8" "\n\t" \ 73 "vmovdqa 64(%0),%%ymm6" "\n\t" \ 74 "vmovdqa 96(%0),%%ymm9" "\n\t" \ 75 "movq 128(%0),%%r14" "\n\t" \ 76 _reg_form "\n\t" \ 77 "vmovdqa %%ymm7, 0(%0)" "\n\t" \ 78 "vmovdqa %%ymm8, 32(%0)" "\n\t" \ 79 "vmovdqa %%ymm6, 64(%0)" "\n\t" \ 80 "vmovdqa %%ymm9, 96(%0)" "\n\t" \ 81 "movq %%r14, 128(%0)" "\n\t" \ 82 : /*OUT*/ \ 83 : /*IN*/"r"(b) \ 84 : /*TRASH*/"xmm0","xmm7","xmm8","xmm6","xmm9","r14","memory","cc" \ 85 ); \ 86 showBlock("after", b); \ 87 randBlock(b); \ 88 printf("%s(mem)\n", #_name); \ 89 showBlock("before", b); \ 90 __asm__ __volatile__( \ 91 "leaq 0(%0),%%rax" "\n\t" \ 92 "vmovdqa 32(%0),%%ymm8" "\n\t" \ 93 "vmovdqa 64(%0),%%ymm7" "\n\t" \ 94 "vmovdqa 96(%0),%%ymm9" "\n\t" \ 95 "movq 128(%0),%%r14" "\n\t" \ 96 _mem_form "\n\t" \ 97 "vmovdqa %%ymm8, 32(%0)" "\n\t" \ 98 "vmovdqa %%ymm7, 64(%0)" "\n\t" \ 99 "vmovdqa %%ymm9, 96(%0)" "\n\t" \ 100 "movq %%r14, 128(%0)" "\n\t" \ 101 : /*OUT*/ \ 102 : /*IN*/"r"(b) \ 103 : /*TRASH*/"xmm6", \ 104 "xmm0","xmm8","xmm7","xmm9","r14","rax","memory","cc" \ 105 ); \ 106 showBlock("after", b); \ 107 printf("\n"); \ 108 free(b); \ 109 } 110 111 #define GEN_test_Ronly(_name, _reg_form) \ 112 GEN_test_RandM(_name, _reg_form, "") 113 #define GEN_test_Monly(_name, _mem_form) \ 114 GEN_test_RandM(_name, "", _mem_form) 115 116 117 GEN_test_RandM(VPOR_128, 118 "vpor %%xmm6, %%xmm8, %%xmm7", 119 "vpor (%%rax), %%xmm8, %%xmm7") 120 121 GEN_test_RandM(VPXOR_128, 122 "vpxor %%xmm6, %%xmm8, %%xmm7", 123 "vpxor (%%rax), %%xmm8, %%xmm7") 124 125 GEN_test_RandM(VPSUBB_128, 126 "vpsubb %%xmm6, %%xmm8, %%xmm7", 127 "vpsubb (%%rax), %%xmm8, %%xmm7") 128 129 GEN_test_RandM(VPSUBD_128, 130 "vpsubd %%xmm6, %%xmm8, %%xmm7", 131 "vpsubd (%%rax), %%xmm8, %%xmm7") 132 133 GEN_test_RandM(VPADDD_128, 134 "vpaddd %%xmm6, %%xmm8, %%xmm7", 135 "vpaddd (%%rax), %%xmm8, %%xmm7") 136 137 GEN_test_RandM(VPMOVZXWD_128, 138 "vpmovzxwd %%xmm6, %%xmm8", 139 "vpmovzxwd (%%rax), %%xmm8") 140 141 GEN_test_RandM(VPMOVZXBW_128, 142 "vpmovzxbw %%xmm6, %%xmm8", 143 "vpmovzxbw (%%rax), %%xmm8") 144 145 GEN_test_RandM(VPBLENDVB_128, 146 "vpblendvb %%xmm9, %%xmm6, %%xmm8, %%xmm7", 147 "vpblendvb %%xmm9, (%%rax), %%xmm8, %%xmm7") 148 149 GEN_test_RandM(VPMINSD_128, 150 "vpminsd %%xmm6, %%xmm8, %%xmm7", 151 "vpminsd (%%rax), %%xmm8, %%xmm7") 152 153 GEN_test_RandM(VPMAXSD_128, 154 "vpmaxsd %%xmm6, %%xmm8, %%xmm7", 155 "vpmaxsd (%%rax), %%xmm8, %%xmm7") 156 157 GEN_test_RandM(VANDPD_128, 158 "vandpd %%xmm6, %%xmm8, %%xmm7", 159 "vandpd (%%rax), %%xmm8, %%xmm7") 160 161 GEN_test_RandM(VCVTSI2SD_32, 162 "vcvtsi2sdl %%r14d, %%xmm8, %%xmm7", 163 "vcvtsi2sdl (%%rax), %%xmm8, %%xmm7") 164 165 GEN_test_RandM(VCVTSI2SD_64, 166 "vcvtsi2sdq %%r14, %%xmm8, %%xmm7", 167 "vcvtsi2sdq (%%rax), %%xmm8, %%xmm7") 168 169 GEN_test_RandM(VCVTSI2SS_64, 170 "vcvtsi2ssq %%r14, %%xmm8, %%xmm7", 171 "vcvtsi2ssq (%%rax), %%xmm8, %%xmm7") 172 173 GEN_test_RandM(VCVTTSD2SI_32, 174 "vcvttsd2si %%xmm8, %%r14d", 175 "vcvttsd2si (%%rax), %%r14d") 176 177 GEN_test_RandM(VCVTTSD2SI_64, 178 "vcvttsd2si %%xmm8, %%r14", 179 "vcvttsd2si (%%rax), %%r14") 180 181 GEN_test_RandM(VCVTSD2SI_32, 182 "vcvtsd2si %%xmm8, %%r14d", 183 "vcvtsd2si (%%rax), %%r14d") 184 185 GEN_test_RandM(VCVTSD2SI_64, 186 "vcvtsd2si %%xmm8, %%r14", 187 "vcvtsd2si (%%rax), %%r14") 188 189 GEN_test_RandM(VPSHUFB_128, 190 "vpshufb %%xmm6, %%xmm8, %%xmm7", 191 "vpshufb (%%rax), %%xmm8, %%xmm7") 192 193 GEN_test_RandM(VCMPSD_128_0x0, 194 "vcmpsd $0, %%xmm6, %%xmm8, %%xmm7", 195 "vcmpsd $0, (%%rax), %%xmm8, %%xmm7") 196 GEN_test_RandM(VCMPSD_128_0x1, 197 "vcmpsd $1, %%xmm6, %%xmm8, %%xmm7", 198 "vcmpsd $1, (%%rax), %%xmm8, %%xmm7") 199 GEN_test_RandM(VCMPSD_128_0x2, 200 "vcmpsd $2, %%xmm6, %%xmm8, %%xmm7", 201 "vcmpsd $2, (%%rax), %%xmm8, %%xmm7") 202 GEN_test_RandM(VCMPSD_128_0x3, 203 "vcmpsd $3, %%xmm6, %%xmm8, %%xmm7", 204 "vcmpsd $3, (%%rax), %%xmm8, %%xmm7") 205 GEN_test_RandM(VCMPSD_128_0x4, 206 "vcmpsd $4, %%xmm6, %%xmm8, %%xmm7", 207 "vcmpsd $4, (%%rax), %%xmm8, %%xmm7") 208 GEN_test_RandM(VCMPSD_128_0x5, 209 "vcmpsd $5, %%xmm6, %%xmm8, %%xmm7", 210 "vcmpsd $5, (%%rax), %%xmm8, %%xmm7") 211 GEN_test_RandM(VCMPSD_128_0x6, 212 "vcmpsd $6, %%xmm6, %%xmm8, %%xmm7", 213 "vcmpsd $6, (%%rax), %%xmm8, %%xmm7") 214 GEN_test_RandM(VCMPSD_128_0x7, 215 "vcmpsd $7, %%xmm6, %%xmm8, %%xmm7", 216 "vcmpsd $7, (%%rax), %%xmm8, %%xmm7") 217 GEN_test_RandM(VCMPSD_128_0x8, 218 "vcmpsd $8, %%xmm6, %%xmm8, %%xmm7", 219 "vcmpsd $8, (%%rax), %%xmm8, %%xmm7") 220 GEN_test_RandM(VCMPSD_128_0xA, 221 "vcmpsd $0xA, %%xmm6, %%xmm8, %%xmm7", 222 "vcmpsd $0xA, (%%rax), %%xmm8, %%xmm7") 223 GEN_test_RandM(VCMPSD_128_0xC, 224 "vcmpsd $0xC, %%xmm6, %%xmm8, %%xmm7", 225 "vcmpsd $0xC, (%%rax), %%xmm8, %%xmm7") 226 GEN_test_RandM(VCMPSD_128_0xD, 227 "vcmpsd $0xD, %%xmm6, %%xmm8, %%xmm7", 228 "vcmpsd $0xD, (%%rax), %%xmm8, %%xmm7") 229 GEN_test_RandM(VCMPSD_128_0xE, 230 "vcmpsd $0xE, %%xmm6, %%xmm8, %%xmm7", 231 "vcmpsd $0xE, (%%rax), %%xmm8, %%xmm7") 232 GEN_test_RandM(VCMPSD_128_0x11, 233 "vcmpsd $0x11, %%xmm6, %%xmm8, %%xmm7", 234 "vcmpsd $0x11, (%%rax), %%xmm8, %%xmm7") 235 GEN_test_RandM(VCMPSD_128_0x12, 236 "vcmpsd $0x12, %%xmm6, %%xmm8, %%xmm7", 237 "vcmpsd $0x12, (%%rax), %%xmm8, %%xmm7") 238 GEN_test_RandM(VCMPSD_128_0x16, 239 "vcmpsd $0x16, %%xmm6, %%xmm8, %%xmm7", 240 "vcmpsd $0x16, (%%rax), %%xmm8, %%xmm7") 241 GEN_test_RandM(VCMPSD_128_0x1E, 242 "vcmpsd $0x1E, %%xmm6, %%xmm8, %%xmm7", 243 "vcmpsd $0x1E, (%%rax), %%xmm8, %%xmm7") 244 245 GEN_test_RandM(VSQRTSD_128, 246 "vsqrtsd %%xmm6, %%xmm8, %%xmm7", 247 "vsqrtsd (%%rax), %%xmm8, %%xmm7") 248 249 GEN_test_RandM(VORPS_128, 250 "vorps %%xmm6, %%xmm8, %%xmm7", 251 "vorps (%%rax), %%xmm8, %%xmm7") 252 253 GEN_test_RandM(VANDNPS_128, 254 "vandnps %%xmm6, %%xmm8, %%xmm7", 255 "vandnps (%%rax), %%xmm8, %%xmm7") 256 257 GEN_test_RandM(VMAXSS_128, 258 "vmaxss %%xmm6, %%xmm8, %%xmm7", 259 "vmaxss (%%rax), %%xmm8, %%xmm7") 260 261 GEN_test_RandM(VMINSS_128, 262 "vminss %%xmm6, %%xmm8, %%xmm7", 263 "vminss (%%rax), %%xmm8, %%xmm7") 264 265 GEN_test_RandM(VANDPS_128, 266 "vandps %%xmm6, %%xmm8, %%xmm7", 267 "vandps (%%rax), %%xmm8, %%xmm7") 268 269 GEN_test_RandM(VCVTSI2SS_128, 270 "vcvtsi2ssl %%r14d, %%xmm8, %%xmm7", 271 "vcvtsi2ssl (%%rax), %%xmm8, %%xmm7") 272 273 GEN_test_RandM(VUNPCKLPS_128, 274 "vunpcklps %%xmm6, %%xmm8, %%xmm7", 275 "vunpcklps (%%rax), %%xmm8, %%xmm7") 276 277 GEN_test_RandM(VDIVSS_128, 278 "vdivss %%xmm6, %%xmm8, %%xmm7", 279 "vdivss (%%rax), %%xmm8, %%xmm7") 280 281 GEN_test_RandM(VADDSS_128, 282 "vaddss %%xmm6, %%xmm8, %%xmm7", 283 "vaddss (%%rax), %%xmm8, %%xmm7") 284 285 GEN_test_RandM(VSUBSS_128, 286 "vsubss %%xmm6, %%xmm8, %%xmm7", 287 "vsubss (%%rax), %%xmm8, %%xmm7") 288 289 GEN_test_RandM(VMULSS_128, 290 "vmulss %%xmm6, %%xmm8, %%xmm7", 291 "vmulss (%%rax), %%xmm8, %%xmm7") 292 293 GEN_test_RandM(VPUNPCKLBW_128, 294 "vpunpcklbw %%xmm6, %%xmm8, %%xmm7", 295 "vpunpcklbw (%%rax), %%xmm8, %%xmm7") 296 297 GEN_test_RandM(VPUNPCKHBW_128, 298 "vpunpckhbw %%xmm6, %%xmm8, %%xmm7", 299 "vpunpckhbw (%%rax), %%xmm8, %%xmm7") 300 301 GEN_test_RandM(VCVTTSS2SI_32, 302 "vcvttss2si %%xmm8, %%r14d", 303 "vcvttss2si (%%rax), %%r14d") 304 305 GEN_test_RandM(VCVTSS2SI_32, 306 "vcvtss2si %%xmm8, %%r14d", 307 "vcvtss2si (%%rax), %%r14d") 308 309 GEN_test_RandM(VMOVQ_XMMorMEM64_to_XMM, 310 "vmovq %%xmm7, %%xmm8", 311 "vmovq (%%rax), %%xmm8") 312 313 /* NB tests the reg form only */ 314 GEN_test_Ronly(VMOVQ_XMM_to_IREG64, 315 "vmovq %%xmm7, %%r14") 316 317 /* This insn only exists in the reg-reg-reg form. */ 318 GEN_test_Ronly(VMOVHLPS_128, 319 "vmovhlps %%xmm6, %%xmm8, %%xmm7") 320 321 GEN_test_RandM(VPABSD_128, 322 "vpabsd %%xmm6, %%xmm8", 323 "vpabsd (%%rax), %%xmm8") 324 325 /* This insn only exists in the reg-reg-reg form. */ 326 GEN_test_Ronly(VMOVLHPS_128, 327 "vmovlhps %%xmm6, %%xmm8, %%xmm7") 328 329 GEN_test_Monly(VMOVNTDQ_128, 330 "vmovntdq %%xmm8, (%%rax)") 331 332 GEN_test_Monly(VMOVNTDQ_256, 333 "vmovntdq %%ymm8, (%%rax)") 334 335 GEN_test_RandM(VMOVUPS_XMM_to_XMMorMEM, 336 "vmovups %%xmm8, %%xmm7", 337 "vmovups %%xmm9, (%%rax)") 338 339 GEN_test_RandM(VMOVQ_IREGorMEM64_to_XMM, 340 "vmovq %%r14, %%xmm7", 341 "vmovq (%%rax), %%xmm9") 342 343 GEN_test_RandM(VPCMPESTRM_0x45_128, 344 "vpcmpestrm $0x45, %%xmm7, %%xmm8; movapd %%xmm0, %%xmm9", 345 "vpcmpestrm $0x45, (%%rax), %%xmm8; movapd %%xmm0, %%xmm9") 346 347 /* NB tests the reg form only */ 348 GEN_test_Ronly(VMOVD_XMM_to_IREG32, 349 "vmovd %%xmm7, %%r14d") 350 351 GEN_test_RandM(VCVTSD2SS_128, 352 "vcvtsd2ss %%xmm9, %%xmm8, %%xmm7", 353 "vcvtsd2ss (%%rax), %%xmm8, %%xmm7") 354 355 GEN_test_RandM(VCVTSS2SD_128, 356 "vcvtss2sd %%xmm9, %%xmm8, %%xmm7", 357 "vcvtss2sd (%%rax), %%xmm8, %%xmm7") 358 359 GEN_test_RandM(VPACKUSWB_128, 360 "vpackuswb %%xmm9, %%xmm8, %%xmm7", 361 "vpackuswb (%%rax), %%xmm8, %%xmm7") 362 363 GEN_test_RandM(VCVTTSS2SI_64, 364 "vcvttss2si %%xmm8, %%r14", 365 "vcvttss2si (%%rax), %%r14") 366 367 GEN_test_RandM(VCVTSS2SI_64, 368 "vcvtss2si %%xmm8, %%r14", 369 "vcvtss2si (%%rax), %%r14") 370 371 GEN_test_Ronly(VPMOVMSKB_128, 372 "vpmovmskb %%xmm8, %%r14") 373 374 GEN_test_RandM(VPAND_128, 375 "vpand %%xmm9, %%xmm8, %%xmm7", 376 "vpand (%%rax), %%xmm8, %%xmm7") 377 378 GEN_test_Monly(VMOVHPD_128_StoreForm, 379 "vmovhpd %%xmm8, (%%rax)") 380 381 GEN_test_Monly(VMOVHPS_128_StoreForm, 382 "vmovhps %%xmm8, (%%rax)") 383 384 GEN_test_RandM(VPCMPEQB_128, 385 "vpcmpeqb %%xmm9, %%xmm8, %%xmm7", 386 "vpcmpeqb (%%rax), %%xmm8, %%xmm7") 387 388 GEN_test_RandM(VSHUFPS_0x39_128, 389 "vshufps $0x39, %%xmm9, %%xmm8, %%xmm7", 390 "vshufps $0xC6, (%%rax), %%xmm8, %%xmm7") 391 392 GEN_test_RandM(VMULPS_128, 393 "vmulps %%xmm9, %%xmm8, %%xmm7", 394 "vmulps (%%rax), %%xmm8, %%xmm7") 395 396 GEN_test_RandM(VSUBPS_128, 397 "vsubps %%xmm9, %%xmm8, %%xmm7", 398 "vsubps (%%rax), %%xmm8, %%xmm7") 399 400 GEN_test_RandM(VADDPS_128, 401 "vaddps %%xmm9, %%xmm8, %%xmm7", 402 "vaddps (%%rax), %%xmm8, %%xmm7") 403 404 GEN_test_RandM(VMAXPS_128, 405 "vmaxps %%xmm9, %%xmm8, %%xmm7", 406 "vmaxps (%%rax), %%xmm8, %%xmm7") 407 408 GEN_test_RandM(VMAXPS_256, 409 "vmaxps %%ymm9, %%ymm8, %%ymm7", 410 "vmaxps (%%rax), %%ymm8, %%ymm7") 411 412 GEN_test_RandM(VMAXPD_128, 413 "vmaxpd %%xmm9, %%xmm8, %%xmm7", 414 "vmaxpd (%%rax), %%xmm8, %%xmm7") 415 416 GEN_test_RandM(VMAXPD_256, 417 "vmaxpd %%ymm9, %%ymm8, %%ymm7", 418 "vmaxpd (%%rax), %%ymm8, %%ymm7") 419 420 GEN_test_RandM(VMINPS_128, 421 "vminps %%xmm9, %%xmm8, %%xmm7", 422 "vminps (%%rax), %%xmm8, %%xmm7") 423 424 GEN_test_RandM(VMINPS_256, 425 "vminps %%ymm9, %%ymm8, %%ymm7", 426 "vminps (%%rax), %%ymm8, %%ymm7") 427 428 GEN_test_RandM(VMINPD_128, 429 "vminpd %%xmm9, %%xmm8, %%xmm7", 430 "vminpd (%%rax), %%xmm8, %%xmm7") 431 432 GEN_test_RandM(VMINPD_256, 433 "vminpd %%ymm9, %%ymm8, %%ymm7", 434 "vminpd (%%rax), %%ymm8, %%ymm7") 435 436 GEN_test_RandM(VCVTPS2DQ_128, 437 "vcvtps2dq %%xmm8, %%xmm7", 438 "vcvtps2dq (%%rax), %%xmm8") 439 440 GEN_test_RandM(VPSHUFLW_0x39_128, 441 "vpshuflw $0x39, %%xmm9, %%xmm7", 442 "vpshuflw $0xC6, (%%rax), %%xmm8") 443 444 GEN_test_RandM(VPSHUFHW_0x39_128, 445 "vpshufhw $0x39, %%xmm9, %%xmm7", 446 "vpshufhw $0xC6, (%%rax), %%xmm8") 447 448 GEN_test_RandM(VPMULLW_128, 449 "vpmullw %%xmm9, %%xmm8, %%xmm7", 450 "vpmullw (%%rax), %%xmm8, %%xmm7") 451 452 GEN_test_RandM(VPADDUSW_128, 453 "vpaddusw %%xmm9, %%xmm8, %%xmm7", 454 "vpaddusw (%%rax), %%xmm8, %%xmm7") 455 456 GEN_test_RandM(VPMULHUW_128, 457 "vpmulhuw %%xmm9, %%xmm8, %%xmm7", 458 "vpmulhuw (%%rax), %%xmm8, %%xmm7") 459 460 GEN_test_RandM(VPADDUSB_128, 461 "vpaddusb %%xmm9, %%xmm8, %%xmm7", 462 "vpaddusb (%%rax), %%xmm8, %%xmm7") 463 464 GEN_test_RandM(VPUNPCKLWD_128, 465 "vpunpcklwd %%xmm6, %%xmm8, %%xmm7", 466 "vpunpcklwd (%%rax), %%xmm8, %%xmm7") 467 468 GEN_test_RandM(VPUNPCKHWD_128, 469 "vpunpckhwd %%xmm6, %%xmm8, %%xmm7", 470 "vpunpckhwd (%%rax), %%xmm8, %%xmm7") 471 472 GEN_test_Ronly(VPSLLD_0x05_128, 473 "vpslld $0x5, %%xmm9, %%xmm7") 474 475 GEN_test_Ronly(VPSRLD_0x05_128, 476 "vpsrld $0x5, %%xmm9, %%xmm7") 477 478 GEN_test_Ronly(VPSRAD_0x05_128, 479 "vpsrad $0x5, %%xmm9, %%xmm7") 480 481 GEN_test_RandM(VPSUBUSB_128, 482 "vpsubusb %%xmm9, %%xmm8, %%xmm7", 483 "vpsubusb (%%rax), %%xmm8, %%xmm7") 484 485 GEN_test_RandM(VPSUBSB_128, 486 "vpsubsb %%xmm9, %%xmm8, %%xmm7", 487 "vpsubsb (%%rax), %%xmm8, %%xmm7") 488 489 GEN_test_Ronly(VPSRLDQ_0x05_128, 490 "vpsrldq $0x5, %%xmm9, %%xmm7") 491 492 GEN_test_Ronly(VPSLLDQ_0x05_128, 493 "vpslldq $0x5, %%xmm9, %%xmm7") 494 495 GEN_test_RandM(VPANDN_128, 496 "vpandn %%xmm9, %%xmm8, %%xmm7", 497 "vpandn (%%rax), %%xmm8, %%xmm7") 498 499 /* NB tests the mem form only */ 500 GEN_test_Monly(VMOVD_XMM_to_MEM32, 501 "vmovd %%xmm7, (%%rax)") 502 503 GEN_test_RandM(VPINSRD_128, 504 "vpinsrd $0, %%r14d, %%xmm8, %%xmm7", 505 "vpinsrd $3, (%%rax), %%xmm8, %%xmm7") 506 507 GEN_test_RandM(VPUNPCKLQDQ_128, 508 "vpunpcklqdq %%xmm6, %%xmm8, %%xmm7", 509 "vpunpcklqdq (%%rax), %%xmm8, %%xmm7") 510 511 GEN_test_Ronly(VPSRLW_0x05_128, 512 "vpsrlw $0x5, %%xmm9, %%xmm7") 513 514 GEN_test_Ronly(VPSLLW_0x05_128, 515 "vpsllw $0x5, %%xmm9, %%xmm7") 516 517 GEN_test_RandM(VPADDW_128, 518 "vpaddw %%xmm6, %%xmm8, %%xmm7", 519 "vpaddw (%%rax), %%xmm8, %%xmm7") 520 521 GEN_test_RandM(VPACKSSDW_128, 522 "vpackssdw %%xmm9, %%xmm8, %%xmm7", 523 "vpackssdw (%%rax), %%xmm8, %%xmm7") 524 525 GEN_test_RandM(VPUNPCKLDQ_128, 526 "vpunpckldq %%xmm6, %%xmm8, %%xmm7", 527 "vpunpckldq (%%rax), %%xmm8, %%xmm7") 528 529 GEN_test_RandM(VINSERTPS_0x39_128, 530 "vinsertps $0x39, %%xmm6, %%xmm8, %%xmm7", 531 "vinsertps $0xC6, (%%rax), %%xmm8, %%xmm7") 532 533 GEN_test_Monly(VMOVSD_M64_XMM, "vmovsd (%%rax), %%xmm8") 534 535 GEN_test_Monly(VMOVSS_M64_XMM, "vmovss (%%rax), %%xmm8") 536 537 GEN_test_Monly(VMOVSD_XMM_M64, "vmovsd %%xmm8, (%%rax)") 538 539 GEN_test_Monly(VMOVSS_XMM_M32, "vmovss %%xmm8, (%%rax)") 540 541 GEN_test_RandM(VMOVUPD_GtoE_128, 542 "vmovupd %%xmm9, %%xmm6", 543 "vmovupd %%xmm7, (%%rax)") 544 545 GEN_test_RandM(VMOVAPD_EtoG_128, 546 "vmovapd %%xmm6, %%xmm8", 547 "vmovapd (%%rax), %%xmm9") 548 549 GEN_test_RandM(VMOVAPD_EtoG_256, 550 "vmovapd %%ymm6, %%ymm8", 551 "vmovapd (%%rax), %%ymm9") 552 553 GEN_test_RandM(VMOVAPS_EtoG_128, 554 "vmovaps %%xmm6, %%xmm8", 555 "vmovaps (%%rax), %%xmm9") 556 557 GEN_test_RandM(VMOVAPS_GtoE_128, 558 "vmovaps %%xmm9, %%xmm6", 559 "vmovaps %%xmm7, (%%rax)") 560 561 GEN_test_RandM(VMOVAPS_GtoE_256, 562 "vmovaps %%ymm9, %%ymm6", 563 "vmovaps %%ymm7, (%%rax)") 564 565 GEN_test_RandM(VMOVAPD_GtoE_128, 566 "vmovapd %%xmm9, %%xmm6", 567 "vmovapd %%xmm7, (%%rax)") 568 569 GEN_test_RandM(VMOVAPD_GtoE_256, 570 "vmovapd %%ymm9, %%ymm6", 571 "vmovapd %%ymm7, (%%rax)") 572 573 GEN_test_RandM(VMOVDQU_EtoG_128, 574 "vmovdqu %%xmm6, %%xmm8", 575 "vmovdqu (%%rax), %%xmm9") 576 577 GEN_test_RandM(VMOVDQA_EtoG_128, 578 "vmovdqa %%xmm6, %%xmm8", 579 "vmovdqa (%%rax), %%xmm9") 580 581 GEN_test_RandM(VMOVDQA_EtoG_256, 582 "vmovdqa %%ymm6, %%ymm8", 583 "vmovdqa (%%rax), %%ymm9") 584 585 GEN_test_RandM(VMOVDQU_GtoE_128, 586 "vmovdqu %%xmm9, %%xmm6", 587 "vmovdqu %%xmm7, (%%rax)") 588 589 GEN_test_RandM(VMOVDQA_GtoE_128, 590 "vmovdqa %%xmm9, %%xmm6", 591 "vmovdqa %%xmm7, (%%rax)") 592 593 GEN_test_RandM(VMOVDQA_GtoE_256, 594 "vmovdqa %%ymm9, %%ymm6", 595 "vmovdqa %%ymm7, (%%rax)") 596 597 GEN_test_Monly(VMOVQ_XMM_MEM64, "vmovq %%xmm8, (%%rax)") 598 599 GEN_test_RandM(VMOVD_IREGorMEM32_to_XMM, 600 "vmovd %%r14d, %%xmm7", 601 "vmovd (%%rax), %%xmm9") 602 603 GEN_test_RandM(VMOVDDUP_XMMorMEM64_to_XMM, 604 "vmovddup %%xmm8, %%xmm7", 605 "vmovddup (%%rax), %%xmm9") 606 607 GEN_test_RandM(VCMPSS_128_0x0, 608 "vcmpss $0, %%xmm6, %%xmm8, %%xmm7", 609 "vcmpss $0, (%%rax), %%xmm8, %%xmm7") 610 GEN_test_RandM(VCMPSS_128_0x1, 611 "vcmpss $1, %%xmm6, %%xmm8, %%xmm7", 612 "vcmpss $1, (%%rax), %%xmm8, %%xmm7") 613 GEN_test_RandM(VCMPSS_128_0x2, 614 "vcmpss $2, %%xmm6, %%xmm8, %%xmm7", 615 "vcmpss $2, (%%rax), %%xmm8, %%xmm7") 616 GEN_test_RandM(VCMPSS_128_0x3, 617 "vcmpss $3, %%xmm6, %%xmm8, %%xmm7", 618 "vcmpss $3, (%%rax), %%xmm8, %%xmm7") 619 GEN_test_RandM(VCMPSS_128_0x4, 620 "vcmpss $4, %%xmm6, %%xmm8, %%xmm7", 621 "vcmpss $4, (%%rax), %%xmm8, %%xmm7") 622 GEN_test_RandM(VCMPSS_128_0x5, 623 "vcmpss $5, %%xmm6, %%xmm8, %%xmm7", 624 "vcmpss $5, (%%rax), %%xmm8, %%xmm7") 625 GEN_test_RandM(VCMPSS_128_0x6, 626 "vcmpss $6, %%xmm6, %%xmm8, %%xmm7", 627 "vcmpss $6, (%%rax), %%xmm8, %%xmm7") 628 GEN_test_RandM(VCMPSS_128_0x7, 629 "vcmpss $7, %%xmm6, %%xmm8, %%xmm7", 630 "vcmpss $7, (%%rax), %%xmm8, %%xmm7") 631 GEN_test_RandM(VCMPSS_128_0x8, 632 "vcmpss $8, %%xmm6, %%xmm8, %%xmm7", 633 "vcmpss $8, (%%rax), %%xmm8, %%xmm7") 634 GEN_test_RandM(VCMPSS_128_0xA, 635 "vcmpss $0xA, %%xmm6, %%xmm8, %%xmm7", 636 "vcmpss $0xA, (%%rax), %%xmm8, %%xmm7") 637 GEN_test_RandM(VCMPSS_128_0xC, 638 "vcmpss $0xC, %%xmm6, %%xmm8, %%xmm7", 639 "vcmpss $0xC, (%%rax), %%xmm8, %%xmm7") 640 GEN_test_RandM(VCMPSS_128_0xD, 641 "vcmpss $0xD, %%xmm6, %%xmm8, %%xmm7", 642 "vcmpss $0xD, (%%rax), %%xmm8, %%xmm7") 643 GEN_test_RandM(VCMPSS_128_0xE, 644 "vcmpss $0xE, %%xmm6, %%xmm8, %%xmm7", 645 "vcmpss $0xE, (%%rax), %%xmm8, %%xmm7") 646 GEN_test_RandM(VCMPSS_128_0x11, 647 "vcmpss $0x11, %%xmm6, %%xmm8, %%xmm7", 648 "vcmpss $0x11, (%%rax), %%xmm8, %%xmm7") 649 GEN_test_RandM(VCMPSS_128_0x12, 650 "vcmpss $0x12, %%xmm6, %%xmm8, %%xmm7", 651 "vcmpss $0x12, (%%rax), %%xmm8, %%xmm7") 652 GEN_test_RandM(VCMPSS_128_0x16, 653 "vcmpss $0x16, %%xmm6, %%xmm8, %%xmm7", 654 "vcmpss $0x16, (%%rax), %%xmm8, %%xmm7") 655 GEN_test_RandM(VCMPSS_128_0x1E, 656 "vcmpss $0x1E, %%xmm6, %%xmm8, %%xmm7", 657 "vcmpss $0x1E, (%%rax), %%xmm8, %%xmm7") 658 659 // The x suffix denotes a 128 -> 64 operation 660 GEN_test_RandM(VCVTPD2PS_128, 661 "vcvtpd2psx %%xmm8, %%xmm7", 662 "vcvtpd2psx (%%rax), %%xmm9") 663 664 GEN_test_RandM(VEXTRACTF128_0x0, 665 "vextractf128 $0x0, %%ymm7, %%xmm9", 666 "vextractf128 $0x0, %%ymm7, (%%rax)") 667 668 GEN_test_RandM(VEXTRACTF128_0x1, 669 "vextractf128 $0x1, %%ymm7, %%xmm9", 670 "vextractf128 $0x1, %%ymm7, (%%rax)") 671 672 GEN_test_RandM(VINSERTF128_0x0, 673 "vinsertf128 $0x0, %%xmm9, %%ymm7, %%ymm8", 674 "vinsertf128 $0x0, (%%rax), %%ymm7, %%ymm8") 675 676 GEN_test_RandM(VINSERTF128_0x1, 677 "vinsertf128 $0x1, %%xmm9, %%ymm7, %%ymm8", 678 "vinsertf128 $0x1, (%%rax), %%ymm7, %%ymm8") 679 680 GEN_test_RandM(VPEXTRD_128_0x0, 681 "vpextrd $0x0, %%xmm7, %%r14d", 682 "vpextrd $0x0, %%xmm7, (%%rax)") 683 684 GEN_test_RandM(VPEXTRD_128_0x3, 685 "vpextrd $0x3, %%xmm7, %%r14d", 686 "vpextrd $0x3, %%xmm7, (%%rax)") 687 688 GEN_test_RandM(VPCMPEQD_128, 689 "vpcmpeqd %%xmm6, %%xmm8, %%xmm7", 690 "vpcmpeqd (%%rax), %%xmm8, %%xmm7") 691 692 GEN_test_RandM(VPSHUFD_0x39_128, 693 "vpshufd $0x39, %%xmm9, %%xmm8", 694 "vpshufd $0xC6, (%%rax), %%xmm7") 695 696 GEN_test_RandM(VMAXSD_128, 697 "vmaxsd %%xmm6, %%xmm8, %%xmm7", 698 "vmaxsd (%%rax), %%xmm8, %%xmm7") 699 700 GEN_test_RandM(VDIVSD_128, 701 "vdivsd %%xmm6, %%xmm8, %%xmm7", 702 "vdivsd (%%rax), %%xmm8, %%xmm7") 703 704 GEN_test_RandM(VMINSD_128, 705 "vminsd %%xmm6, %%xmm8, %%xmm7", 706 "vminsd (%%rax), %%xmm8, %%xmm7") 707 708 GEN_test_RandM(VSUBSD_128, 709 "vsubsd %%xmm6, %%xmm8, %%xmm7", 710 "vsubsd (%%rax), %%xmm8, %%xmm7") 711 712 GEN_test_RandM(VADDSD_128, 713 "vaddsd %%xmm6, %%xmm8, %%xmm7", 714 "vaddsd (%%rax), %%xmm8, %%xmm7") 715 716 GEN_test_RandM(VMULSD_128, 717 "vmulsd %%xmm6, %%xmm8, %%xmm7", 718 "vmulsd (%%rax), %%xmm8, %%xmm7") 719 720 GEN_test_RandM(VXORPS_128, 721 "vxorps %%xmm6, %%xmm8, %%xmm7", 722 "vxorps (%%rax), %%xmm8, %%xmm7") 723 724 GEN_test_RandM(VXORPD_128, 725 "vxorpd %%xmm6, %%xmm8, %%xmm7", 726 "vxorpd (%%rax), %%xmm8, %%xmm7") 727 728 GEN_test_RandM(VORPD_128, 729 "vorpd %%xmm6, %%xmm8, %%xmm7", 730 "vorpd (%%rax), %%xmm8, %%xmm7") 731 732 GEN_test_RandM(VANDNPD_128, 733 "vandnpd %%xmm6, %%xmm8, %%xmm7", 734 "vandnpd (%%rax), %%xmm8, %%xmm7") 735 736 GEN_test_RandM(VCVTPS2PD_128, 737 "vcvtps2pd %%xmm6, %%xmm8", 738 "vcvtps2pd (%%rax), %%xmm8") 739 740 GEN_test_RandM(VUCOMISD_128, 741 "vucomisd %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 742 "vucomisd (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 743 744 GEN_test_RandM(VUCOMISS_128, 745 "vucomiss %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 746 "vucomiss (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 747 748 GEN_test_RandM(VPINSRQ_128, 749 "vpinsrq $0, %%r14, %%xmm8, %%xmm7", 750 "vpinsrq $1, (%%rax), %%xmm8, %%xmm7") 751 752 GEN_test_RandM(VPADDQ_128, 753 "vpaddq %%xmm6, %%xmm8, %%xmm7", 754 "vpaddq (%%rax), %%xmm8, %%xmm7") 755 756 GEN_test_RandM(VPSUBQ_128, 757 "vpsubq %%xmm6, %%xmm8, %%xmm7", 758 "vpsubq (%%rax), %%xmm8, %%xmm7") 759 760 GEN_test_RandM(VPSUBW_128, 761 "vpsubw %%xmm6, %%xmm8, %%xmm7", 762 "vpsubw (%%rax), %%xmm8, %%xmm7") 763 764 GEN_test_RandM(VMOVUPD_GtoE_256, 765 "vmovupd %%ymm9, %%ymm6", 766 "vmovupd %%ymm7, (%%rax)") 767 768 GEN_test_RandM(VMOVUPD_EtoG_256, 769 "vmovupd %%ymm6, %%ymm9", 770 "vmovupd (%%rax), %%ymm7") 771 772 GEN_test_RandM(VMULPD_256, 773 "vmulpd %%ymm6, %%ymm8, %%ymm7", 774 "vmulpd (%%rax), %%ymm8, %%ymm7") 775 776 GEN_test_RandM(VMOVUPD_EtoG_128, 777 "vmovupd %%xmm6, %%xmm9", 778 "vmovupd (%%rax), %%xmm7") 779 780 GEN_test_RandM(VADDPD_256, 781 "vaddpd %%ymm6, %%ymm8, %%ymm7", 782 "vaddpd (%%rax), %%ymm8, %%ymm7") 783 784 GEN_test_RandM(VSUBPD_256, 785 "vsubpd %%ymm6, %%ymm8, %%ymm7", 786 "vsubpd (%%rax), %%ymm8, %%ymm7") 787 788 GEN_test_RandM(VDIVPD_256, 789 "vdivpd %%ymm6, %%ymm8, %%ymm7", 790 "vdivpd (%%rax), %%ymm8, %%ymm7") 791 792 GEN_test_RandM(VPCMPEQQ_128, 793 "vpcmpeqq %%xmm6, %%xmm8, %%xmm7", 794 "vpcmpeqq (%%rax), %%xmm8, %%xmm7") 795 796 GEN_test_RandM(VSUBPD_128, 797 "vsubpd %%xmm6, %%xmm8, %%xmm7", 798 "vsubpd (%%rax), %%xmm8, %%xmm7") 799 800 GEN_test_RandM(VADDPD_128, 801 "vaddpd %%xmm6, %%xmm8, %%xmm7", 802 "vaddpd (%%rax), %%xmm8, %%xmm7") 803 804 GEN_test_RandM(VUNPCKLPD_128, 805 "vunpcklpd %%xmm6, %%xmm8, %%xmm7", 806 "vunpcklpd (%%rax), %%xmm8, %%xmm7") 807 808 GEN_test_RandM(VUNPCKHPD_128, 809 "vunpckhpd %%xmm6, %%xmm8, %%xmm7", 810 "vunpckhpd (%%rax), %%xmm8, %%xmm7") 811 812 GEN_test_RandM(VUNPCKHPS_128, 813 "vunpckhps %%xmm6, %%xmm8, %%xmm7", 814 "vunpckhps (%%rax), %%xmm8, %%xmm7") 815 816 GEN_test_RandM(VMOVUPS_EtoG_128, 817 "vmovups %%xmm6, %%xmm8", 818 "vmovups (%%rax), %%xmm9") 819 820 GEN_test_RandM(VADDPS_256, 821 "vaddps %%ymm6, %%ymm8, %%ymm7", 822 "vaddps (%%rax), %%ymm8, %%ymm7") 823 824 GEN_test_RandM(VSUBPS_256, 825 "vsubps %%ymm6, %%ymm8, %%ymm7", 826 "vsubps (%%rax), %%ymm8, %%ymm7") 827 828 GEN_test_RandM(VMULPS_256, 829 "vmulps %%ymm6, %%ymm8, %%ymm7", 830 "vmulps (%%rax), %%ymm8, %%ymm7") 831 832 GEN_test_RandM(VDIVPS_256, 833 "vdivps %%ymm6, %%ymm8, %%ymm7", 834 "vdivps (%%rax), %%ymm8, %%ymm7") 835 836 GEN_test_RandM(VPCMPGTQ_128, 837 "vpcmpgtq %%xmm6, %%xmm8, %%xmm7", 838 "vpcmpgtq (%%rax), %%xmm8, %%xmm7") 839 840 GEN_test_RandM(VPEXTRQ_128_0x0, 841 "vpextrq $0x0, %%xmm7, %%r14", 842 "vpextrq $0x0, %%xmm7, (%%rax)") 843 844 GEN_test_RandM(VPEXTRQ_128_0x1, 845 "vpextrq $0x1, %%xmm7, %%r14", 846 "vpextrq $0x1, %%xmm7, (%%rax)") 847 848 GEN_test_Ronly(VPSRLQ_0x05_128, 849 "vpsrlq $0x5, %%xmm9, %%xmm7") 850 851 GEN_test_RandM(VPMULUDQ_128, 852 "vpmuludq %%xmm6, %%xmm8, %%xmm7", 853 "vpmuludq (%%rax), %%xmm8, %%xmm7") 854 855 GEN_test_RandM(VPMULDQ_128, 856 "vpmuldq %%xmm6, %%xmm8, %%xmm7", 857 "vpmuldq (%%rax), %%xmm8, %%xmm7") 858 859 GEN_test_Ronly(VPSLLQ_0x05_128, 860 "vpsllq $0x5, %%xmm9, %%xmm7") 861 862 GEN_test_RandM(VPMAXUD_128, 863 "vpmaxud %%xmm6, %%xmm8, %%xmm7", 864 "vpmaxud (%%rax), %%xmm8, %%xmm7") 865 866 GEN_test_RandM(VPMINUD_128, 867 "vpminud %%xmm6, %%xmm8, %%xmm7", 868 "vpminud (%%rax), %%xmm8, %%xmm7") 869 870 GEN_test_RandM(VPMULLD_128, 871 "vpmulld %%xmm6, %%xmm8, %%xmm7", 872 "vpmulld (%%rax), %%xmm8, %%xmm7") 873 874 GEN_test_RandM(VPMAXUW_128, 875 "vpmaxuw %%xmm6, %%xmm8, %%xmm7", 876 "vpmaxuw (%%rax), %%xmm8, %%xmm7") 877 878 GEN_test_Ronly(VPEXTRW_128_EregOnly_toG_0x0, 879 "vpextrw $0x0, %%xmm7, %%r14d") 880 881 GEN_test_Ronly(VPEXTRW_128_EregOnly_toG_0x7, 882 "vpextrw $0x7, %%xmm7, %%r14d") 883 884 GEN_test_RandM(VPMINUW_128, 885 "vpminuw %%xmm6, %%xmm8, %%xmm7", 886 "vpminuw (%%rax), %%xmm8, %%xmm7") 887 888 GEN_test_RandM(VPHMINPOSUW_128, 889 "vphminposuw %%xmm6, %%xmm8", 890 "vphminposuw (%%rax), %%xmm7") 891 892 GEN_test_RandM(VPMAXSW_128, 893 "vpmaxsw %%xmm6, %%xmm8, %%xmm7", 894 "vpmaxsw (%%rax), %%xmm8, %%xmm7") 895 896 GEN_test_RandM(VPMINSW_128, 897 "vpminsw %%xmm6, %%xmm8, %%xmm7", 898 "vpminsw (%%rax), %%xmm8, %%xmm7") 899 900 GEN_test_RandM(VPMAXUB_128, 901 "vpmaxub %%xmm6, %%xmm8, %%xmm7", 902 "vpmaxub (%%rax), %%xmm8, %%xmm7") 903 904 GEN_test_RandM(VPEXTRB_GtoE_128_0x0, 905 "vpextrb $0x0, %%xmm8, %%r14", 906 "vpextrb $0x0, %%xmm8, (%%rax)") 907 908 GEN_test_RandM(VPEXTRB_GtoE_128_0x1, 909 "vpextrb $0x1, %%xmm8, %%r14", 910 "vpextrb $0x1, %%xmm8, (%%rax)") 911 912 GEN_test_RandM(VPEXTRB_GtoE_128_0x2, 913 "vpextrb $0x2, %%xmm8, %%r14", 914 "vpextrb $0x2, %%xmm8, (%%rax)") 915 916 GEN_test_RandM(VPEXTRB_GtoE_128_0x3, 917 "vpextrb $0x3, %%xmm8, %%r14", 918 "vpextrb $0x3, %%xmm8, (%%rax)") 919 920 GEN_test_RandM(VPEXTRB_GtoE_128_0x4, 921 "vpextrb $0x4, %%xmm8, %%r14", 922 "vpextrb $0x4, %%xmm8, (%%rax)") 923 924 GEN_test_RandM(VPEXTRB_GtoE_128_0x9, 925 "vpextrb $0x9, %%xmm8, %%r14", 926 "vpextrb $0x9, %%xmm8, (%%rax)") 927 928 GEN_test_RandM(VPEXTRB_GtoE_128_0xE, 929 "vpextrb $0xE, %%xmm8, %%r14", 930 "vpextrb $0xE, %%xmm8, (%%rax)") 931 932 GEN_test_RandM(VPEXTRB_GtoE_128_0xF, 933 "vpextrb $0xF, %%xmm8, %%r14", 934 "vpextrb $0xF, %%xmm8, (%%rax)") 935 936 GEN_test_RandM(VPMINUB_128, 937 "vpminub %%xmm6, %%xmm8, %%xmm7", 938 "vpminub (%%rax), %%xmm8, %%xmm7") 939 940 GEN_test_RandM(VPMAXSB_128, 941 "vpmaxsb %%xmm6, %%xmm8, %%xmm7", 942 "vpmaxsb (%%rax), %%xmm8, %%xmm7") 943 944 GEN_test_RandM(VPMINSB_128, 945 "vpminsb %%xmm6, %%xmm8, %%xmm7", 946 "vpminsb (%%rax), %%xmm8, %%xmm7") 947 948 GEN_test_RandM(VPERM2F128_0x00, 949 "vperm2f128 $0x00, %%ymm6, %%ymm8, %%ymm7", 950 "vperm2f128 $0x00, (%%rax), %%ymm8, %%ymm7") 951 GEN_test_RandM(VPERM2F128_0xFF, 952 "vperm2f128 $0xFF, %%ymm6, %%ymm8, %%ymm7", 953 "vperm2f128 $0xFF, (%%rax), %%ymm8, %%ymm7") 954 GEN_test_RandM(VPERM2F128_0x30, 955 "vperm2f128 $0x30, %%ymm6, %%ymm8, %%ymm7", 956 "vperm2f128 $0x30, (%%rax), %%ymm8, %%ymm7") 957 GEN_test_RandM(VPERM2F128_0x21, 958 "vperm2f128 $0x21, %%ymm6, %%ymm8, %%ymm7", 959 "vperm2f128 $0x21, (%%rax), %%ymm8, %%ymm7") 960 GEN_test_RandM(VPERM2F128_0x12, 961 "vperm2f128 $0x12, %%ymm6, %%ymm8, %%ymm7", 962 "vperm2f128 $0x12, (%%rax), %%ymm8, %%ymm7") 963 GEN_test_RandM(VPERM2F128_0x03, 964 "vperm2f128 $0x03, %%ymm6, %%ymm8, %%ymm7", 965 "vperm2f128 $0x03, (%%rax), %%ymm8, %%ymm7") 966 GEN_test_RandM(VPERM2F128_0x85, 967 "vperm2f128 $0x85, %%ymm6, %%ymm8, %%ymm7", 968 "vperm2f128 $0x85, (%%rax), %%ymm8, %%ymm7") 969 GEN_test_RandM(VPERM2F128_0x5A, 970 "vperm2f128 $0x5A, %%ymm6, %%ymm8, %%ymm7", 971 "vperm2f128 $0x5A, (%%rax), %%ymm8, %%ymm7") 972 973 GEN_test_RandM(VPERMILPD_256_0x0, 974 "vpermilpd $0x0, %%ymm6, %%ymm8", 975 "vpermilpd $0x1, (%%rax), %%ymm8") 976 GEN_test_RandM(VPERMILPD_256_0xF, 977 "vpermilpd $0xF, %%ymm6, %%ymm8", 978 "vpermilpd $0xE, (%%rax), %%ymm8") 979 GEN_test_RandM(VPERMILPD_256_0xA, 980 "vpermilpd $0xA, %%ymm6, %%ymm8", 981 "vpermilpd $0xB, (%%rax), %%ymm8") 982 GEN_test_RandM(VPERMILPD_256_0x5, 983 "vpermilpd $0x5, %%ymm6, %%ymm8", 984 "vpermilpd $0x4, (%%rax), %%ymm8") 985 986 GEN_test_RandM(VPERMILPD_128_0x0, 987 "vpermilpd $0x0, %%xmm6, %%xmm8", 988 "vpermilpd $0x1, (%%rax), %%xmm8") 989 GEN_test_RandM(VPERMILPD_128_0x3, 990 "vpermilpd $0x3, %%xmm6, %%xmm8", 991 "vpermilpd $0x2, (%%rax), %%xmm8") 992 993 GEN_test_RandM(VUNPCKLPD_256, 994 "vunpcklpd %%ymm6, %%ymm8, %%ymm7", 995 "vunpcklpd (%%rax), %%ymm8, %%ymm7") 996 997 GEN_test_RandM(VUNPCKHPD_256, 998 "vunpckhpd %%ymm6, %%ymm8, %%ymm7", 999 "vunpckhpd (%%rax), %%ymm8, %%ymm7") 1000 1001 GEN_test_RandM(VSHUFPS_0x39_256, 1002 "vshufps $0x39, %%ymm9, %%ymm8, %%ymm7", 1003 "vshufps $0xC6, (%%rax), %%ymm8, %%ymm7") 1004 1005 GEN_test_RandM(VUNPCKLPS_256, 1006 "vunpcklps %%ymm6, %%ymm8, %%ymm7", 1007 "vunpcklps (%%rax), %%ymm8, %%ymm7") 1008 1009 GEN_test_RandM(VUNPCKHPS_256, 1010 "vunpckhps %%ymm6, %%ymm8, %%ymm7", 1011 "vunpckhps (%%rax), %%ymm8, %%ymm7") 1012 1013 GEN_test_RandM(VXORPD_256, 1014 "vxorpd %%ymm6, %%ymm8, %%ymm7", 1015 "vxorpd (%%rax), %%ymm8, %%ymm7") 1016 1017 GEN_test_Monly(VBROADCASTSD_256, 1018 "vbroadcastsd (%%rax), %%ymm8") 1019 1020 GEN_test_RandM(VCMPPD_128_0x4, 1021 "vcmppd $4, %%xmm6, %%xmm8, %%xmm7", 1022 "vcmppd $4, (%%rax), %%xmm8, %%xmm7") 1023 1024 GEN_test_RandM(VCMPPD_256_0x4, 1025 "vcmppd $4, %%ymm6, %%ymm8, %%ymm7", 1026 "vcmppd $4, (%%rax), %%ymm8, %%ymm7") 1027 1028 GEN_test_RandM(VCMPPS_128_0x4, 1029 "vcmpps $4, %%xmm6, %%xmm8, %%xmm7", 1030 "vcmpps $4, (%%rax), %%xmm8, %%xmm7") 1031 1032 GEN_test_RandM(VCMPPS_256_0x4, 1033 "vcmpps $4, %%ymm6, %%ymm8, %%ymm7", 1034 "vcmpps $4, (%%rax), %%ymm8, %%ymm7") 1035 1036 GEN_test_RandM(VCVTDQ2PD_128, 1037 "vcvtdq2pd %%xmm6, %%xmm8", 1038 "vcvtdq2pd (%%rax), %%xmm8") 1039 1040 GEN_test_RandM(VDIVPD_128, 1041 "vdivpd %%xmm6, %%xmm8, %%xmm7", 1042 "vdivpd (%%rax), %%xmm8, %%xmm7") 1043 1044 GEN_test_RandM(VANDPD_256, 1045 "vandpd %%ymm6, %%ymm8, %%ymm7", 1046 "vandpd (%%rax), %%ymm8, %%ymm7") 1047 1048 GEN_test_RandM(VPMOVSXBW_128, 1049 "vpmovsxbw %%xmm6, %%xmm8", 1050 "vpmovsxbw (%%rax), %%xmm8") 1051 1052 GEN_test_RandM(VPSUBUSW_128, 1053 "vpsubusw %%xmm9, %%xmm8, %%xmm7", 1054 "vpsubusw (%%rax), %%xmm8, %%xmm7") 1055 1056 GEN_test_RandM(VPSUBSW_128, 1057 "vpsubsw %%xmm9, %%xmm8, %%xmm7", 1058 "vpsubsw (%%rax), %%xmm8, %%xmm7") 1059 1060 GEN_test_RandM(VPCMPEQW_128, 1061 "vpcmpeqw %%xmm6, %%xmm8, %%xmm7", 1062 "vpcmpeqw (%%rax), %%xmm8, %%xmm7") 1063 1064 GEN_test_RandM(VPADDB_128, 1065 "vpaddb %%xmm6, %%xmm8, %%xmm7", 1066 "vpaddb (%%rax), %%xmm8, %%xmm7") 1067 1068 GEN_test_RandM(VMOVAPS_EtoG_256, 1069 "vmovaps %%ymm6, %%ymm8", 1070 "vmovaps (%%rax), %%ymm9") 1071 1072 GEN_test_RandM(VCVTDQ2PD_256, 1073 "vcvtdq2pd %%xmm6, %%ymm8", 1074 "vcvtdq2pd (%%rax), %%ymm8") 1075 1076 GEN_test_Monly(VMOVHPD_128_LoadForm, 1077 "vmovhpd (%%rax), %%xmm8, %%xmm7") 1078 1079 GEN_test_Monly(VMOVHPS_128_LoadForm, 1080 "vmovhps (%%rax), %%xmm8, %%xmm7") 1081 1082 // The y suffix denotes a 256 -> 128 operation 1083 GEN_test_RandM(VCVTPD2PS_256, 1084 "vcvtpd2psy %%ymm8, %%xmm7", 1085 "vcvtpd2psy (%%rax), %%xmm9") 1086 1087 GEN_test_RandM(VPUNPCKHDQ_128, 1088 "vpunpckhdq %%xmm6, %%xmm8, %%xmm7", 1089 "vpunpckhdq (%%rax), %%xmm8, %%xmm7") 1090 1091 GEN_test_Monly(VBROADCASTSS_128, 1092 "vbroadcastss (%%rax), %%xmm8") 1093 1094 GEN_test_RandM(VPMOVSXDQ_128, 1095 "vpmovsxdq %%xmm6, %%xmm8", 1096 "vpmovsxdq (%%rax), %%xmm8") 1097 1098 GEN_test_RandM(VPMOVSXWD_128, 1099 "vpmovsxwd %%xmm6, %%xmm8", 1100 "vpmovsxwd (%%rax), %%xmm8") 1101 1102 GEN_test_RandM(VDIVPS_128, 1103 "vdivps %%xmm9, %%xmm8, %%xmm7", 1104 "vdivps (%%rax), %%xmm8, %%xmm7") 1105 1106 GEN_test_RandM(VANDPS_256, 1107 "vandps %%ymm6, %%ymm8, %%ymm7", 1108 "vandps (%%rax), %%ymm8, %%ymm7") 1109 1110 GEN_test_RandM(VXORPS_256, 1111 "vxorps %%ymm6, %%ymm8, %%ymm7", 1112 "vxorps (%%rax), %%ymm8, %%ymm7") 1113 1114 GEN_test_RandM(VORPS_256, 1115 "vorps %%ymm6, %%ymm8, %%ymm7", 1116 "vorps (%%rax), %%ymm8, %%ymm7") 1117 1118 GEN_test_RandM(VANDNPD_256, 1119 "vandnpd %%ymm6, %%ymm8, %%ymm7", 1120 "vandnpd (%%rax), %%ymm8, %%ymm7") 1121 1122 GEN_test_RandM(VANDNPS_256, 1123 "vandnps %%ymm6, %%ymm8, %%ymm7", 1124 "vandnps (%%rax), %%ymm8, %%ymm7") 1125 1126 GEN_test_RandM(VORPD_256, 1127 "vorpd %%ymm6, %%ymm8, %%ymm7", 1128 "vorpd (%%rax), %%ymm8, %%ymm7") 1129 1130 GEN_test_RandM(VPERMILPS_256_0x0F, 1131 "vpermilps $0x0F, %%ymm6, %%ymm8", 1132 "vpermilps $0x1E, (%%rax), %%ymm8") 1133 GEN_test_RandM(VPERMILPS_256_0xFA, 1134 "vpermilps $0xFA, %%ymm6, %%ymm8", 1135 "vpermilps $0xE5, (%%rax), %%ymm8") 1136 GEN_test_RandM(VPERMILPS_256_0xA3, 1137 "vpermilps $0xA3, %%ymm6, %%ymm8", 1138 "vpermilps $0xB4, (%%rax), %%ymm8") 1139 GEN_test_RandM(VPERMILPS_256_0x5A, 1140 "vpermilps $0x5A, %%ymm6, %%ymm8", 1141 "vpermilps $0x45, (%%rax), %%ymm8") 1142 1143 GEN_test_RandM(VPMULHW_128, 1144 "vpmulhw %%xmm9, %%xmm8, %%xmm7", 1145 "vpmulhw (%%rax), %%xmm8, %%xmm7") 1146 1147 GEN_test_RandM(VPUNPCKHQDQ_128, 1148 "vpunpckhqdq %%xmm6, %%xmm8, %%xmm7", 1149 "vpunpckhqdq (%%rax), %%xmm8, %%xmm7") 1150 1151 GEN_test_Ronly(VPSRAW_0x05_128, 1152 "vpsraw $0x5, %%xmm9, %%xmm7") 1153 1154 GEN_test_RandM(VPCMPGTB_128, 1155 "vpcmpgtb %%xmm6, %%xmm8, %%xmm7", 1156 "vpcmpgtb (%%rax), %%xmm8, %%xmm7") 1157 1158 GEN_test_RandM(VPCMPGTW_128, 1159 "vpcmpgtw %%xmm6, %%xmm8, %%xmm7", 1160 "vpcmpgtw (%%rax), %%xmm8, %%xmm7") 1161 1162 GEN_test_RandM(VPCMPGTD_128, 1163 "vpcmpgtd %%xmm6, %%xmm8, %%xmm7", 1164 "vpcmpgtd (%%rax), %%xmm8, %%xmm7") 1165 1166 GEN_test_RandM(VPMOVZXBD_128, 1167 "vpmovzxbd %%xmm6, %%xmm8", 1168 "vpmovzxbd (%%rax), %%xmm8") 1169 1170 GEN_test_RandM(VPMOVSXBD_128, 1171 "vpmovsxbd %%xmm6, %%xmm8", 1172 "vpmovsxbd (%%rax), %%xmm8") 1173 1174 GEN_test_RandM(VPINSRB_128_1of3, 1175 "vpinsrb $0, %%r14d, %%xmm8, %%xmm7", 1176 "vpinsrb $3, (%%rax), %%xmm8, %%xmm7") 1177 GEN_test_RandM(VPINSRB_128_2of3, 1178 "vpinsrb $6, %%r14d, %%xmm8, %%xmm7", 1179 "vpinsrb $9, (%%rax), %%xmm8, %%xmm7") 1180 GEN_test_RandM(VPINSRB_128_3of3, 1181 "vpinsrb $12, %%r14d, %%xmm8, %%xmm7", 1182 "vpinsrb $15, (%%rax), %%xmm8, %%xmm7") 1183 1184 GEN_test_RandM(VPINSRW_128_1of4, 1185 "vpinsrw $0, %%r14d, %%xmm8, %%xmm7", 1186 "vpinsrw $3, (%%rax), %%xmm8, %%xmm7") 1187 GEN_test_RandM(VPINSRW_128_2of4, 1188 "vpinsrw $2, %%r14d, %%xmm8, %%xmm7", 1189 "vpinsrw $3, (%%rax), %%xmm8, %%xmm7") 1190 GEN_test_RandM(VPINSRW_128_3of4, 1191 "vpinsrw $4, %%r14d, %%xmm8, %%xmm7", 1192 "vpinsrw $5, (%%rax), %%xmm8, %%xmm7") 1193 GEN_test_RandM(VPINSRW_128_4of4, 1194 "vpinsrw $6, %%r14d, %%xmm8, %%xmm7", 1195 "vpinsrw $7, (%%rax), %%xmm8, %%xmm7") 1196 1197 GEN_test_RandM(VCOMISD_128, 1198 "vcomisd %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 1199 "vcomisd (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 1200 1201 GEN_test_RandM(VCOMISS_128, 1202 "vcomiss %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 1203 "vcomiss (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 1204 1205 GEN_test_RandM(VMOVUPS_YMM_to_YMMorMEM, 1206 "vmovups %%ymm8, %%ymm7", 1207 "vmovups %%ymm9, (%%rax)") 1208 1209 GEN_test_RandM(VDPPD_128_1of4, 1210 "vdppd $0x00, %%xmm6, %%xmm8, %%xmm7", 1211 "vdppd $0xA5, (%%rax), %%xmm9, %%xmm6") 1212 GEN_test_RandM(VDPPD_128_2of4, 1213 "vdppd $0x5A, %%xmm6, %%xmm8, %%xmm7", 1214 "vdppd $0xFF, (%%rax), %%xmm9, %%xmm6") 1215 GEN_test_RandM(VDPPD_128_3of4, 1216 "vdppd $0x0F, %%xmm6, %%xmm8, %%xmm7", 1217 "vdppd $0x37, (%%rax), %%xmm9, %%xmm6") 1218 GEN_test_RandM(VDPPD_128_4of4, 1219 "vdppd $0xF0, %%xmm6, %%xmm8, %%xmm7", 1220 "vdppd $0x73, (%%rax), %%xmm9, %%xmm6") 1221 1222 GEN_test_RandM(VDPPS_128_1of4, 1223 "vdpps $0x00, %%xmm6, %%xmm8, %%xmm7", 1224 "vdpps $0xA5, (%%rax), %%xmm9, %%xmm6") 1225 GEN_test_RandM(VDPPS_128_2of4, 1226 "vdpps $0x5A, %%xmm6, %%xmm8, %%xmm7", 1227 "vdpps $0xFF, (%%rax), %%xmm9, %%xmm6") 1228 GEN_test_RandM(VDPPS_128_3of4, 1229 "vdpps $0x0F, %%xmm6, %%xmm8, %%xmm7", 1230 "vdpps $0x37, (%%rax), %%xmm9, %%xmm6") 1231 GEN_test_RandM(VDPPS_128_4of4, 1232 "vdpps $0xF0, %%xmm6, %%xmm8, %%xmm7", 1233 "vdpps $0x73, (%%rax), %%xmm9, %%xmm6") 1234 1235 GEN_test_RandM(VDPPS_256_1of4, 1236 "vdpps $0x00, %%ymm6, %%ymm8, %%ymm7", 1237 "vdpps $0xA5, (%%rax), %%ymm9, %%ymm6") 1238 GEN_test_RandM(VDPPS_256_2of4, 1239 "vdpps $0x5A, %%ymm6, %%ymm8, %%ymm7", 1240 "vdpps $0xFF, (%%rax), %%ymm9, %%ymm6") 1241 GEN_test_RandM(VDPPS_256_3of4, 1242 "vdpps $0x0F, %%ymm6, %%ymm8, %%ymm7", 1243 "vdpps $0x37, (%%rax), %%ymm9, %%ymm6") 1244 GEN_test_RandM(VDPPS_256_4of4, 1245 "vdpps $0xF0, %%ymm6, %%ymm8, %%ymm7", 1246 "vdpps $0x73, (%%rax), %%ymm9, %%ymm6") 1247 1248 GEN_test_Monly(VBROADCASTSS_256, 1249 "vbroadcastss (%%rax), %%ymm8") 1250 1251 GEN_test_RandM(VPALIGNR_128_1of3, 1252 "vpalignr $0, %%xmm6, %%xmm8, %%xmm7", 1253 "vpalignr $3, (%%rax), %%xmm8, %%xmm7") 1254 GEN_test_RandM(VPALIGNR_128_2of3, 1255 "vpalignr $6, %%xmm6, %%xmm8, %%xmm7", 1256 "vpalignr $9, (%%rax), %%xmm8, %%xmm7") 1257 GEN_test_RandM(VPALIGNR_128_3of3, 1258 "vpalignr $12, %%xmm6, %%xmm8, %%xmm7", 1259 "vpalignr $15, (%%rax), %%xmm8, %%xmm7") 1260 1261 GEN_test_Ronly(VMOVSD_REG_XMM, "vmovsd %%xmm9, %%xmm7, %%xmm8") 1262 1263 GEN_test_Ronly(VMOVSS_REG_XMM, "vmovss %%xmm9, %%xmm7, %%xmm8") 1264 1265 GEN_test_Monly(VMOVLPD_128_M64_XMM_XMM, "vmovlpd (%%rax), %%xmm8, %%xmm7") 1266 1267 GEN_test_Monly(VMOVLPD_128_XMM_M64, "vmovlpd %%xmm7, (%%rax)") 1268 1269 GEN_test_RandM(VSHUFPD_128_1of2, 1270 "vshufpd $0, %%xmm9, %%xmm8, %%xmm7", 1271 "vshufpd $1, (%%rax), %%xmm8, %%xmm7") 1272 GEN_test_RandM(VSHUFPD_128_2of2, 1273 "vshufpd $2, %%xmm9, %%xmm8, %%xmm7", 1274 "vshufpd $3, (%%rax), %%xmm8, %%xmm7") 1275 1276 GEN_test_RandM(VSHUFPD_256_1of2, 1277 "vshufpd $0x00, %%ymm9, %%ymm8, %%ymm7", 1278 "vshufpd $0xFF, (%%rax), %%ymm8, %%ymm7") 1279 GEN_test_RandM(VSHUFPD_256_2of2, 1280 "vshufpd $0x5A, %%ymm9, %%ymm8, %%ymm7", 1281 "vshufpd $0xA5, (%%rax), %%ymm8, %%ymm7") 1282 1283 GEN_test_RandM(VPERMILPS_128_0x00, 1284 "vpermilps $0x00, %%xmm6, %%xmm8", 1285 "vpermilps $0x01, (%%rax), %%xmm8") 1286 GEN_test_RandM(VPERMILPS_128_0xFE, 1287 "vpermilps $0xFE, %%xmm6, %%xmm8", 1288 "vpermilps $0xFF, (%%rax), %%xmm8") 1289 GEN_test_RandM(VPERMILPS_128_0x30, 1290 "vpermilps $0x30, %%xmm6, %%xmm8", 1291 "vpermilps $0x03, (%%rax), %%xmm8") 1292 GEN_test_RandM(VPERMILPS_128_0x21, 1293 "vpermilps $0x21, %%xmm6, %%xmm8", 1294 "vpermilps $0x12, (%%rax), %%xmm8") 1295 GEN_test_RandM(VPERMILPS_128_0xD7, 1296 "vpermilps $0xD7, %%xmm6, %%xmm8", 1297 "vpermilps $0x6C, (%%rax), %%xmm8") 1298 GEN_test_RandM(VPERMILPS_128_0xB5, 1299 "vpermilps $0xB5, %%xmm6, %%xmm8", 1300 "vpermilps $0x4A, (%%rax), %%xmm8") 1301 GEN_test_RandM(VPERMILPS_128_0x85, 1302 "vpermilps $0x85, %%xmm6, %%xmm8", 1303 "vpermilps $0xDC, (%%rax), %%xmm8") 1304 GEN_test_RandM(VPERMILPS_128_0x29, 1305 "vpermilps $0x29, %%xmm6, %%xmm8", 1306 "vpermilps $0x92, (%%rax), %%xmm8") 1307 1308 GEN_test_RandM(VBLENDPS_128_1of3, 1309 "vblendps $0, %%xmm6, %%xmm8, %%xmm7", 1310 "vblendps $3, (%%rax), %%xmm8, %%xmm7") 1311 GEN_test_RandM(VBLENDPS_128_2of3, 1312 "vblendps $6, %%xmm6, %%xmm8, %%xmm7", 1313 "vblendps $9, (%%rax), %%xmm8, %%xmm7") 1314 GEN_test_RandM(VBLENDPS_128_3of3, 1315 "vblendps $12, %%xmm6, %%xmm8, %%xmm7", 1316 "vblendps $15, (%%rax), %%xmm8, %%xmm7") 1317 1318 GEN_test_RandM(VBLENDPD_128_1of2, 1319 "vblendpd $0, %%xmm6, %%xmm8, %%xmm7", 1320 "vblendpd $1, (%%rax), %%xmm8, %%xmm7") 1321 GEN_test_RandM(VBLENDPD_128_2of2, 1322 "vblendpd $2, %%xmm6, %%xmm8, %%xmm7", 1323 "vblendpd $3, (%%rax), %%xmm8, %%xmm7") 1324 1325 GEN_test_RandM(VBLENDPD_256_1of3, 1326 "vblendpd $0, %%ymm6, %%ymm8, %%ymm7", 1327 "vblendpd $3, (%%rax), %%ymm8, %%ymm7") 1328 GEN_test_RandM(VBLENDPD_256_2of3, 1329 "vblendpd $6, %%ymm6, %%ymm8, %%ymm7", 1330 "vblendpd $9, (%%rax), %%ymm8, %%ymm7") 1331 GEN_test_RandM(VBLENDPD_256_3of3, 1332 "vblendpd $12, %%ymm6, %%ymm8, %%ymm7", 1333 "vblendpd $15, (%%rax), %%ymm8, %%ymm7") 1334 1335 GEN_test_RandM(VPBLENDW_128_0x00, 1336 "vpblendw $0x00, %%xmm6, %%xmm8, %%xmm7", 1337 "vpblendw $0x01, (%%rax), %%xmm8, %%xmm7") 1338 GEN_test_RandM(VPBLENDW_128_0xFE, 1339 "vpblendw $0xFE, %%xmm6, %%xmm8, %%xmm7", 1340 "vpblendw $0xFF, (%%rax), %%xmm8, %%xmm7") 1341 GEN_test_RandM(VPBLENDW_128_0x30, 1342 "vpblendw $0x30, %%xmm6, %%xmm8, %%xmm7", 1343 "vpblendw $0x03, (%%rax), %%xmm8, %%xmm7") 1344 GEN_test_RandM(VPBLENDW_128_0x21, 1345 "vpblendw $0x21, %%xmm6, %%xmm8, %%xmm7", 1346 "vpblendw $0x12, (%%rax), %%xmm8, %%xmm7") 1347 GEN_test_RandM(VPBLENDW_128_0xD7, 1348 "vpblendw $0xD7, %%xmm6, %%xmm8, %%xmm7", 1349 "vpblendw $0x6C, (%%rax), %%xmm8, %%xmm7") 1350 GEN_test_RandM(VPBLENDW_128_0xB5, 1351 "vpblendw $0xB5, %%xmm6, %%xmm8, %%xmm7", 1352 "vpblendw $0x4A, (%%rax), %%xmm8, %%xmm7") 1353 GEN_test_RandM(VPBLENDW_128_0x85, 1354 "vpblendw $0x85, %%xmm6, %%xmm8, %%xmm7", 1355 "vpblendw $0xDC, (%%rax), %%xmm8, %%xmm7") 1356 GEN_test_RandM(VPBLENDW_128_0x29, 1357 "vpblendw $0x29, %%xmm6, %%xmm8, %%xmm7", 1358 "vpblendw $0x92, (%%rax), %%xmm8, %%xmm7") 1359 1360 GEN_test_RandM(VMOVUPS_EtoG_256, 1361 "vmovups %%ymm6, %%ymm9", 1362 "vmovups (%%rax), %%ymm7") 1363 1364 GEN_test_RandM(VSQRTSS_128, 1365 "vsqrtss %%xmm6, %%xmm8, %%xmm7", 1366 "vsqrtss (%%rax), %%xmm8, %%xmm7") 1367 1368 GEN_test_RandM(VSQRTPS_128, 1369 "vsqrtps %%xmm6, %%xmm8", 1370 "vsqrtps (%%rax), %%xmm8") 1371 1372 GEN_test_RandM(VSQRTPS_256, 1373 "vsqrtps %%ymm6, %%ymm8", 1374 "vsqrtps (%%rax), %%ymm8") 1375 1376 GEN_test_RandM(VSQRTPD_128, 1377 "vsqrtpd %%xmm6, %%xmm8", 1378 "vsqrtpd (%%rax), %%xmm8") 1379 1380 GEN_test_RandM(VSQRTPD_256, 1381 "vsqrtpd %%ymm6, %%ymm8", 1382 "vsqrtpd (%%rax), %%ymm8") 1383 1384 GEN_test_RandM(VRSQRTSS_128, 1385 "vrsqrtss %%xmm6, %%xmm8, %%xmm7", 1386 "vrsqrtss (%%rax), %%xmm8, %%xmm7") 1387 1388 GEN_test_RandM(VRSQRTPS_128, 1389 "vrsqrtps %%xmm6, %%xmm8", 1390 "vrsqrtps (%%rax), %%xmm8") 1391 1392 GEN_test_RandM(VRSQRTPS_256, 1393 "vrsqrtps %%ymm6, %%ymm8", 1394 "vrsqrtps (%%rax), %%ymm8") 1395 1396 GEN_test_RandM(VMOVDQU_GtoE_256, 1397 "vmovdqu %%ymm9, %%ymm6", 1398 "vmovdqu %%ymm7, (%%rax)") 1399 1400 GEN_test_RandM(VCVTPS2PD_256, 1401 "vcvtps2pd %%xmm9, %%ymm6", 1402 "vcvtps2pd (%%rax), %%ymm7") 1403 1404 GEN_test_RandM(VCVTTPS2DQ_128, 1405 "vcvttps2dq %%xmm9, %%xmm6", 1406 "vcvttps2dq (%%rax), %%xmm7") 1407 1408 GEN_test_RandM(VCVTTPS2DQ_256, 1409 "vcvttps2dq %%ymm9, %%ymm6", 1410 "vcvttps2dq (%%rax), %%ymm7") 1411 1412 GEN_test_RandM(VCVTDQ2PS_128, 1413 "vcvtdq2ps %%xmm9, %%xmm6", 1414 "vcvtdq2ps (%%rax), %%xmm7") 1415 1416 GEN_test_RandM(VCVTDQ2PS_256, 1417 "vcvtdq2ps %%ymm9, %%ymm6", 1418 "vcvtdq2ps (%%rax), %%ymm7") 1419 1420 GEN_test_RandM(VCVTTPD2DQ_128, 1421 "vcvttpd2dqx %%xmm9, %%xmm6", 1422 "vcvttpd2dqx (%%rax), %%xmm7") 1423 1424 GEN_test_RandM(VCVTTPD2DQ_256, 1425 "vcvttpd2dqy %%ymm9, %%xmm6", 1426 "vcvttpd2dqy (%%rax), %%xmm7") 1427 1428 GEN_test_RandM(VCVTPD2DQ_128, 1429 "vcvtpd2dqx %%xmm9, %%xmm6", 1430 "vcvtpd2dqx (%%rax), %%xmm7") 1431 1432 GEN_test_RandM(VCVTPD2DQ_256, 1433 "vcvtpd2dqy %%ymm9, %%xmm6", 1434 "vcvtpd2dqy (%%rax), %%xmm7") 1435 1436 GEN_test_RandM(VMOVSLDUP_128, 1437 "vmovsldup %%xmm9, %%xmm6", 1438 "vmovsldup (%%rax), %%xmm7") 1439 1440 GEN_test_RandM(VMOVSLDUP_256, 1441 "vmovsldup %%ymm9, %%ymm6", 1442 "vmovsldup (%%rax), %%ymm7") 1443 1444 GEN_test_RandM(VMOVSHDUP_128, 1445 "vmovshdup %%xmm9, %%xmm6", 1446 "vmovshdup (%%rax), %%xmm7") 1447 1448 GEN_test_RandM(VMOVSHDUP_256, 1449 "vmovshdup %%ymm9, %%ymm6", 1450 "vmovshdup (%%rax), %%ymm7") 1451 1452 GEN_test_RandM(VPERMILPS_VAR_128, 1453 "vpermilps %%xmm6, %%xmm8, %%xmm7", 1454 "vpermilps (%%rax), %%xmm8, %%xmm7") 1455 1456 GEN_test_RandM(VPERMILPD_VAR_128, 1457 "vpermilpd %%xmm6, %%xmm8, %%xmm7", 1458 "vpermilpd (%%rax), %%xmm8, %%xmm7") 1459 1460 GEN_test_RandM(VPERMILPS_VAR_256, 1461 "vpermilps %%ymm6, %%ymm8, %%ymm7", 1462 "vpermilps (%%rax), %%ymm8, %%ymm7") 1463 1464 GEN_test_RandM(VPERMILPD_VAR_256, 1465 "vpermilpd %%ymm6, %%ymm8, %%ymm7", 1466 "vpermilpd (%%rax), %%ymm8, %%ymm7") 1467 1468 GEN_test_RandM(VPSLLW_128, 1469 "andl $15, %%r14d;" 1470 "vmovd %%r14d, %%xmm6;" 1471 "vpsllw %%xmm6, %%xmm8, %%xmm9", 1472 "andq $15, 128(%%rax);" 1473 "vpsllw 128(%%rax), %%xmm8, %%xmm9") 1474 1475 GEN_test_RandM(VPSRLW_128, 1476 "andl $15, %%r14d;" 1477 "vmovd %%r14d, %%xmm6;" 1478 "vpsrlw %%xmm6, %%xmm8, %%xmm9", 1479 "andq $15, 128(%%rax);" 1480 "vpsrlw 128(%%rax), %%xmm8, %%xmm9") 1481 1482 GEN_test_RandM(VPSRAW_128, 1483 "andl $31, %%r14d;" 1484 "vmovd %%r14d, %%xmm6;" 1485 "vpsraw %%xmm6, %%xmm8, %%xmm9", 1486 "andq $15, 128(%%rax);" 1487 "vpsraw 128(%%rax), %%xmm8, %%xmm9") 1488 1489 GEN_test_RandM(VPSLLD_128, 1490 "andl $31, %%r14d;" 1491 "vmovd %%r14d, %%xmm6;" 1492 "vpslld %%xmm6, %%xmm8, %%xmm9", 1493 "andq $31, 128(%%rax);" 1494 "vpslld 128(%%rax), %%xmm8, %%xmm9") 1495 1496 GEN_test_RandM(VPSRLD_128, 1497 "andl $31, %%r14d;" 1498 "vmovd %%r14d, %%xmm6;" 1499 "vpsrld %%xmm6, %%xmm8, %%xmm9", 1500 "andq $31, 128(%%rax);" 1501 "vpsrld 128(%%rax), %%xmm8, %%xmm9") 1502 1503 GEN_test_RandM(VPSRAD_128, 1504 "andl $31, %%r14d;" 1505 "vmovd %%r14d, %%xmm6;" 1506 "vpsrad %%xmm6, %%xmm8, %%xmm9", 1507 "andq $31, 128(%%rax);" 1508 "vpsrad 128(%%rax), %%xmm8, %%xmm9") 1509 1510 GEN_test_RandM(VPSLLQ_128, 1511 "andl $63, %%r14d;" 1512 "vmovd %%r14d, %%xmm6;" 1513 "vpsllq %%xmm6, %%xmm8, %%xmm9", 1514 "andq $63, 128(%%rax);" 1515 "vpsllq 128(%%rax), %%xmm8, %%xmm9") 1516 1517 GEN_test_RandM(VPSRLQ_128, 1518 "andl $63, %%r14d;" 1519 "vmovd %%r14d, %%xmm6;" 1520 "vpsrlq %%xmm6, %%xmm8, %%xmm9", 1521 "andq $63, 128(%%rax);" 1522 "vpsrlq 128(%%rax), %%xmm8, %%xmm9") 1523 1524 GEN_test_RandM(VROUNDPS_128_0x0, 1525 "vroundps $0x0, %%xmm8, %%xmm9", 1526 "vroundps $0x0, (%%rax), %%xmm9") 1527 GEN_test_RandM(VROUNDPS_128_0x1, 1528 "vroundps $0x1, %%xmm8, %%xmm9", 1529 "vroundps $0x1, (%%rax), %%xmm9") 1530 GEN_test_RandM(VROUNDPS_128_0x2, 1531 "vroundps $0x2, %%xmm8, %%xmm9", 1532 "vroundps $0x2, (%%rax), %%xmm9") 1533 GEN_test_RandM(VROUNDPS_128_0x3, 1534 "vroundps $0x3, %%xmm8, %%xmm9", 1535 "vroundps $0x3, (%%rax), %%xmm9") 1536 GEN_test_RandM(VROUNDPS_128_0x4, 1537 "vroundps $0x4, %%xmm8, %%xmm9", 1538 "vroundps $0x4, (%%rax), %%xmm9") 1539 1540 GEN_test_RandM(VROUNDPS_256_0x0, 1541 "vroundps $0x0, %%ymm8, %%ymm9", 1542 "vroundps $0x0, (%%rax), %%ymm9") 1543 GEN_test_RandM(VROUNDPS_256_0x1, 1544 "vroundps $0x1, %%ymm8, %%ymm9", 1545 "vroundps $0x1, (%%rax), %%ymm9") 1546 GEN_test_RandM(VROUNDPS_256_0x2, 1547 "vroundps $0x2, %%ymm8, %%ymm9", 1548 "vroundps $0x2, (%%rax), %%ymm9") 1549 GEN_test_RandM(VROUNDPS_256_0x3, 1550 "vroundps $0x3, %%ymm8, %%ymm9", 1551 "vroundps $0x3, (%%rax), %%ymm9") 1552 GEN_test_RandM(VROUNDPS_256_0x4, 1553 "vroundps $0x4, %%ymm8, %%ymm9", 1554 "vroundps $0x4, (%%rax), %%ymm9") 1555 1556 GEN_test_RandM(VROUNDPD_128_0x0, 1557 "vroundpd $0x0, %%xmm8, %%xmm9", 1558 "vroundpd $0x0, (%%rax), %%xmm9") 1559 GEN_test_RandM(VROUNDPD_128_0x1, 1560 "vroundpd $0x1, %%xmm8, %%xmm9", 1561 "vroundpd $0x1, (%%rax), %%xmm9") 1562 GEN_test_RandM(VROUNDPD_128_0x2, 1563 "vroundpd $0x2, %%xmm8, %%xmm9", 1564 "vroundpd $0x2, (%%rax), %%xmm9") 1565 GEN_test_RandM(VROUNDPD_128_0x3, 1566 "vroundpd $0x3, %%xmm8, %%xmm9", 1567 "vroundpd $0x3, (%%rax), %%xmm9") 1568 GEN_test_RandM(VROUNDPD_128_0x4, 1569 "vroundpd $0x4, %%xmm8, %%xmm9", 1570 "vroundpd $0x4, (%%rax), %%xmm9") 1571 1572 GEN_test_RandM(VROUNDPD_256_0x0, 1573 "vroundpd $0x0, %%ymm8, %%ymm9", 1574 "vroundpd $0x0, (%%rax), %%ymm9") 1575 GEN_test_RandM(VROUNDPD_256_0x1, 1576 "vroundpd $0x1, %%ymm8, %%ymm9", 1577 "vroundpd $0x1, (%%rax), %%ymm9") 1578 GEN_test_RandM(VROUNDPD_256_0x2, 1579 "vroundpd $0x2, %%ymm8, %%ymm9", 1580 "vroundpd $0x2, (%%rax), %%ymm9") 1581 GEN_test_RandM(VROUNDPD_256_0x3, 1582 "vroundpd $0x3, %%ymm8, %%ymm9", 1583 "vroundpd $0x3, (%%rax), %%ymm9") 1584 GEN_test_RandM(VROUNDPD_256_0x4, 1585 "vroundpd $0x4, %%ymm8, %%ymm9", 1586 "vroundpd $0x4, (%%rax), %%ymm9") 1587 1588 GEN_test_RandM(VPMADDWD_128, 1589 "vpmaddwd %%xmm6, %%xmm8, %%xmm7", 1590 "vpmaddwd (%%rax), %%xmm8, %%xmm7") 1591 1592 GEN_test_RandM(VADDSUBPS_128, 1593 "vaddsubps %%xmm6, %%xmm8, %%xmm7", 1594 "vaddsubps (%%rax), %%xmm8, %%xmm7") 1595 1596 GEN_test_RandM(VADDSUBPS_256, 1597 "vaddsubps %%ymm6, %%ymm8, %%ymm7", 1598 "vaddsubps (%%rax), %%ymm8, %%ymm7") 1599 1600 GEN_test_RandM(VADDSUBPD_128, 1601 "vaddsubpd %%xmm6, %%xmm8, %%xmm7", 1602 "vaddsubpd (%%rax), %%xmm8, %%xmm7") 1603 1604 GEN_test_RandM(VADDSUBPD_256, 1605 "vaddsubpd %%ymm6, %%ymm8, %%ymm7", 1606 "vaddsubpd (%%rax), %%ymm8, %%ymm7") 1607 1608 GEN_test_RandM(VROUNDSS_0x0, 1609 "vroundss $0x0, %%xmm8, %%xmm6, %%xmm9", 1610 "vroundss $0x0, (%%rax), %%xmm6, %%xmm9") 1611 GEN_test_RandM(VROUNDSS_0x1, 1612 "vroundss $0x1, %%xmm8, %%xmm6, %%xmm9", 1613 "vroundss $0x1, (%%rax), %%xmm6, %%xmm9") 1614 GEN_test_RandM(VROUNDSS_0x2, 1615 "vroundss $0x2, %%xmm8, %%xmm6, %%xmm9", 1616 "vroundss $0x2, (%%rax), %%xmm6, %%xmm9") 1617 GEN_test_RandM(VROUNDSS_0x3, 1618 "vroundss $0x3, %%xmm8, %%xmm6, %%xmm9", 1619 "vroundss $0x3, (%%rax), %%xmm6, %%xmm9") 1620 GEN_test_RandM(VROUNDSS_0x4, 1621 "vroundss $0x4, %%xmm8, %%xmm6, %%xmm9", 1622 "vroundss $0x4, (%%rax), %%xmm6, %%xmm9") 1623 GEN_test_RandM(VROUNDSS_0x5, 1624 "vroundss $0x5, %%xmm8, %%xmm6, %%xmm9", 1625 "vroundss $0x5, (%%rax), %%xmm6, %%xmm9") 1626 1627 GEN_test_RandM(VROUNDSD_0x0, 1628 "vroundsd $0x0, %%xmm8, %%xmm6, %%xmm9", 1629 "vroundsd $0x0, (%%rax), %%xmm6, %%xmm9") 1630 GEN_test_RandM(VROUNDSD_0x1, 1631 "vroundsd $0x1, %%xmm8, %%xmm6, %%xmm9", 1632 "vroundsd $0x1, (%%rax), %%xmm6, %%xmm9") 1633 GEN_test_RandM(VROUNDSD_0x2, 1634 "vroundsd $0x2, %%xmm8, %%xmm6, %%xmm9", 1635 "vroundsd $0x2, (%%rax), %%xmm6, %%xmm9") 1636 GEN_test_RandM(VROUNDSD_0x3, 1637 "vroundsd $0x3, %%xmm8, %%xmm6, %%xmm9", 1638 "vroundsd $0x3, (%%rax), %%xmm6, %%xmm9") 1639 GEN_test_RandM(VROUNDSD_0x4, 1640 "vroundsd $0x4, %%xmm8, %%xmm6, %%xmm9", 1641 "vroundsd $0x4, (%%rax), %%xmm6, %%xmm9") 1642 GEN_test_RandM(VROUNDSD_0x5, 1643 "vroundsd $0x5, %%xmm8, %%xmm6, %%xmm9", 1644 "vroundsd $0x5, (%%rax), %%xmm6, %%xmm9") 1645 1646 GEN_test_RandM(VPTEST_128_1, 1647 "vptest %%xmm6, %%xmm8; " 1648 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1649 "vptest (%%rax), %%xmm8; " 1650 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1651 1652 /* Here we ignore the boilerplate-supplied data and try to do 1653 x AND x and x AND NOT x. Not a great test but better 1654 than nothing. */ 1655 GEN_test_RandM(VPTEST_128_2, 1656 "vmovups %%xmm6, %%xmm8;" 1657 "vptest %%xmm6, %%xmm8; " 1658 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1659 "vmovups (%%rax), %%xmm8;" 1660 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1661 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1662 "vptest (%%rax), %%xmm8; " 1663 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1664 1665 GEN_test_RandM(VPTEST_256_1, 1666 "vptest %%ymm6, %%ymm8; " 1667 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1668 "vptest (%%rax), %%ymm8; " 1669 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1670 1671 /* Here we ignore the boilerplate-supplied data and try to do 1672 x AND x and x AND NOT x. Not a great test but better 1673 than nothing. */ 1674 GEN_test_RandM(VPTEST_256_2, 1675 "vmovups %%ymm6, %%ymm8;" 1676 "vptest %%ymm6, %%ymm8; " 1677 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1678 "vmovups (%%rax), %%ymm8;" 1679 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1680 "subq $1024, %%rsp;" 1681 "vmovups %%xmm7,512(%%rsp);" 1682 "vmovups %%xmm7,528(%%rsp);" 1683 "vmovups 512(%%rsp), %%ymm7;" 1684 "addq $1024, %%rsp;" 1685 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1686 "vptest (%%rax), %%ymm8; " 1687 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1688 1689 1690 /* VTESTPS/VTESTPD: test once with all-0 operands, once with 1691 one all-0s and one all 1s, and once with random data. */ 1692 1693 GEN_test_RandM(VTESTPS_128_1, 1694 "vtestps %%xmm6, %%xmm8; " 1695 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1696 "vtestps (%%rax), %%xmm8; " 1697 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1698 1699 /* Here we ignore the boilerplate-supplied data and try to do 1700 x AND x and x AND NOT x. Not a great test but better 1701 than nothing. */ 1702 GEN_test_RandM(VTESTPS_128_2, 1703 "vmovups %%xmm6, %%xmm8;" 1704 "vtestps %%xmm6, %%xmm8; " 1705 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1706 "vmovups (%%rax), %%xmm8;" 1707 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1708 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1709 "vtestps (%%rax), %%xmm8; " 1710 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1711 1712 GEN_test_RandM(VTESTPS_128_3, 1713 "vtestps %%xmm8, %%xmm9; " 1714 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1715 "vtestps (%%rax), %%xmm9; " 1716 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1717 1718 1719 1720 1721 GEN_test_RandM(VTESTPS_256_1, 1722 "vtestps %%ymm6, %%ymm8; " 1723 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1724 "vtestps (%%rax), %%ymm8; " 1725 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1726 1727 /* Here we ignore the boilerplate-supplied data and try to do 1728 x AND x and x AND NOT x. Not a great test but better 1729 than nothing. */ 1730 GEN_test_RandM(VTESTPS_256_2, 1731 "vmovups %%ymm6, %%ymm8;" 1732 "vtestps %%ymm6, %%ymm8; " 1733 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1734 "vmovups (%%rax), %%ymm8;" 1735 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1736 "subq $1024, %%rsp;" 1737 "vmovups %%xmm7,512(%%rsp);" 1738 "vmovups %%xmm7,528(%%rsp);" 1739 "vmovups 512(%%rsp), %%ymm7;" 1740 "addq $1024, %%rsp;" 1741 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1742 "vtestps (%%rax), %%ymm8; " 1743 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1744 1745 GEN_test_RandM(VTESTPS_256_3, 1746 "vtestps %%ymm8, %%ymm9; " 1747 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1748 "vtestps (%%rax), %%ymm9; " 1749 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1750 1751 1752 1753 GEN_test_RandM(VTESTPD_128_1, 1754 "vtestpd %%xmm6, %%xmm8; " 1755 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1756 "vtestpd (%%rax), %%xmm8; " 1757 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1758 1759 /* Here we ignore the boilerplate-supplied data and try to do 1760 x AND x and x AND NOT x. Not a great test but better 1761 than nothing. */ 1762 GEN_test_RandM(VTESTPD_128_2, 1763 "vmovups %%xmm6, %%xmm8;" 1764 "vtestpd %%xmm6, %%xmm8; " 1765 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1766 "vmovups (%%rax), %%xmm8;" 1767 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1768 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1769 "vtestpd (%%rax), %%xmm8; " 1770 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1771 1772 GEN_test_RandM(VTESTPD_128_3, 1773 "vtestpd %%xmm8, %%xmm9; " 1774 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1775 "vtestpd (%%rax), %%xmm9; " 1776 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1777 1778 1779 1780 1781 GEN_test_RandM(VTESTPD_256_1, 1782 "vtestpd %%ymm6, %%ymm8; " 1783 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1784 "vtestpd (%%rax), %%ymm8; " 1785 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1786 1787 /* Here we ignore the boilerplate-supplied data and try to do 1788 x AND x and x AND NOT x. Not a great test but better 1789 than nothing. */ 1790 GEN_test_RandM(VTESTPD_256_2, 1791 "vmovups %%ymm6, %%ymm8;" 1792 "vtestpd %%ymm6, %%ymm8; " 1793 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1794 "vmovups (%%rax), %%ymm8;" 1795 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1796 "subq $1024, %%rsp;" 1797 "vmovups %%xmm7,512(%%rsp);" 1798 "vmovups %%xmm7,528(%%rsp);" 1799 "vmovups 512(%%rsp), %%ymm7;" 1800 "addq $1024, %%rsp;" 1801 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1802 "vtestpd (%%rax), %%ymm8; " 1803 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1804 1805 GEN_test_RandM(VTESTPD_256_3, 1806 "vtestpd %%ymm8, %%ymm9; " 1807 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1808 "vtestpd (%%rax), %%ymm9; " 1809 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1810 1811 GEN_test_RandM(VBLENDVPS_128, 1812 "vblendvps %%xmm9, %%xmm6, %%xmm8, %%xmm7", 1813 "vblendvps %%xmm9, (%%rax), %%xmm8, %%xmm7") 1814 1815 GEN_test_RandM(VBLENDVPS_256, 1816 "vblendvps %%ymm9, %%ymm6, %%ymm8, %%ymm7", 1817 "vblendvps %%ymm9, (%%rax), %%ymm8, %%ymm7") 1818 1819 GEN_test_RandM(VBLENDVPD_128, 1820 "vblendvpd %%xmm9, %%xmm6, %%xmm8, %%xmm7", 1821 "vblendvpd %%xmm9, (%%rax), %%xmm8, %%xmm7") 1822 1823 GEN_test_RandM(VBLENDVPD_256, 1824 "vblendvpd %%ymm9, %%ymm6, %%ymm8, %%ymm7", 1825 "vblendvpd %%ymm9, (%%rax), %%ymm8, %%ymm7") 1826 1827 1828 GEN_test_RandM(VHADDPS_128, 1829 "vhaddps %%xmm6, %%xmm8, %%xmm7", 1830 "vhaddps (%%rax), %%xmm8, %%xmm7") 1831 1832 GEN_test_RandM(VHADDPS_256, 1833 "vhaddps %%ymm6, %%ymm8, %%ymm7", 1834 "vhaddps (%%rax), %%ymm8, %%ymm7") 1835 1836 GEN_test_RandM(VHADDPD_128, 1837 "vhaddpd %%xmm6, %%xmm8, %%xmm7", 1838 "vhaddpd (%%rax), %%xmm8, %%xmm7") 1839 1840 GEN_test_RandM(VHADDPD_256, 1841 "vhaddpd %%ymm6, %%ymm8, %%ymm7", 1842 "vhaddpd (%%rax), %%ymm8, %%ymm7") 1843 1844 GEN_test_RandM(VHSUBPS_128, 1845 "vhsubps %%xmm6, %%xmm8, %%xmm7", 1846 "vhsubps (%%rax), %%xmm8, %%xmm7") 1847 1848 GEN_test_RandM(VHSUBPS_256, 1849 "vhsubps %%ymm6, %%ymm8, %%ymm7", 1850 "vhsubps (%%rax), %%ymm8, %%ymm7") 1851 1852 GEN_test_RandM(VHSUBPD_128, 1853 "vhsubpd %%xmm6, %%xmm8, %%xmm7", 1854 "vhsubpd (%%rax), %%xmm8, %%xmm7") 1855 1856 GEN_test_RandM(VHSUBPD_256, 1857 "vhsubpd %%ymm6, %%ymm8, %%ymm7", 1858 "vhsubpd (%%rax), %%ymm8, %%ymm7") 1859 1860 GEN_test_RandM(VEXTRACTPS_0x0, 1861 "vextractps $0, %%xmm8, %%r14d", 1862 "vextractps $0, %%xmm8, (%%rax)") 1863 1864 GEN_test_RandM(VEXTRACTPS_0x1, 1865 "vextractps $1, %%xmm8, %%r14d", 1866 "vextractps $1, %%xmm8, (%%rax)") 1867 1868 GEN_test_RandM(VEXTRACTPS_0x2, 1869 "vextractps $2, %%xmm8, %%r14d", 1870 "vextractps $2, %%xmm8, (%%rax)") 1871 1872 GEN_test_RandM(VEXTRACTPS_0x3, 1873 "vextractps $3, %%xmm8, %%r14d", 1874 "vextractps $3, %%xmm8, (%%rax)") 1875 1876 GEN_test_Monly(VLDDQU_128, 1877 "vlddqu 1(%%rax), %%xmm8") 1878 1879 GEN_test_Monly(VLDDQU_256, 1880 "vlddqu 1(%%rax), %%ymm8") 1881 1882 GEN_test_Monly(VMOVNTDQA_128, 1883 "vmovntdqa (%%rax), %%xmm9") 1884 1885 GEN_test_Monly(VMASKMOVDQU_128, 1886 "xchgq %%rax, %%rdi;" 1887 "vmaskmovdqu %%xmm8, %%xmm9;" 1888 "xchgq %%rax, %%rdi") 1889 1890 GEN_test_Ronly(VMOVMSKPD_128, 1891 "vmovmskpd %%xmm9, %%r14d") 1892 1893 GEN_test_Ronly(VMOVMSKPD_256, 1894 "vmovmskpd %%ymm9, %%r14d") 1895 1896 GEN_test_Ronly(VMOVMSKPS_128, 1897 "vmovmskps %%xmm9, %%r14d") 1898 1899 GEN_test_Ronly(VMOVMSKPS_256, 1900 "vmovmskps %%ymm9, %%r14d") 1901 1902 GEN_test_Monly(VMOVNTPD_128, 1903 "vmovntpd %%xmm9, (%%rax)") 1904 1905 GEN_test_Monly(VMOVNTPD_256, 1906 "vmovntpd %%ymm9, (%%rax)") 1907 1908 GEN_test_Monly(VMOVNTPS_128, 1909 "vmovntps %%xmm9, (%%rax)") 1910 1911 GEN_test_Monly(VMOVNTPS_256, 1912 "vmovntps %%ymm9, (%%rax)") 1913 1914 GEN_test_RandM(VPACKSSWB_128, 1915 "vpacksswb %%xmm6, %%xmm8, %%xmm7", 1916 "vpacksswb (%%rax), %%xmm8, %%xmm7") 1917 1918 GEN_test_RandM(VPAVGB_128, 1919 "vpavgb %%xmm6, %%xmm8, %%xmm7", 1920 "vpavgb (%%rax), %%xmm8, %%xmm7") 1921 1922 GEN_test_RandM(VPAVGW_128, 1923 "vpavgw %%xmm6, %%xmm8, %%xmm7", 1924 "vpavgw (%%rax), %%xmm8, %%xmm7") 1925 1926 GEN_test_RandM(VPADDSB_128, 1927 "vpaddsb %%xmm6, %%xmm8, %%xmm7", 1928 "vpaddsb (%%rax), %%xmm8, %%xmm7") 1929 1930 GEN_test_RandM(VPADDSW_128, 1931 "vpaddsw %%xmm6, %%xmm8, %%xmm7", 1932 "vpaddsw (%%rax), %%xmm8, %%xmm7") 1933 1934 GEN_test_RandM(VPHADDW_128, 1935 "vphaddw %%xmm6, %%xmm8, %%xmm7", 1936 "vphaddw (%%rax), %%xmm8, %%xmm7") 1937 1938 GEN_test_RandM(VPHADDD_128, 1939 "vphaddd %%xmm6, %%xmm8, %%xmm7", 1940 "vphaddd (%%rax), %%xmm8, %%xmm7") 1941 1942 GEN_test_RandM(VPHADDSW_128, 1943 "vphaddsw %%xmm6, %%xmm8, %%xmm7", 1944 "vphaddsw (%%rax), %%xmm8, %%xmm7") 1945 1946 GEN_test_RandM(VPMADDUBSW_128, 1947 "vpmaddubsw %%xmm6, %%xmm8, %%xmm7", 1948 "vpmaddubsw (%%rax), %%xmm8, %%xmm7") 1949 1950 GEN_test_RandM(VPHSUBW_128, 1951 "vphsubw %%xmm6, %%xmm8, %%xmm7", 1952 "vphsubw (%%rax), %%xmm8, %%xmm7") 1953 1954 GEN_test_RandM(VPHSUBD_128, 1955 "vphsubd %%xmm6, %%xmm8, %%xmm7", 1956 "vphsubd (%%rax), %%xmm8, %%xmm7") 1957 1958 GEN_test_RandM(VPHSUBSW_128, 1959 "vphsubsw %%xmm6, %%xmm8, %%xmm7", 1960 "vphsubsw (%%rax), %%xmm8, %%xmm7") 1961 1962 GEN_test_RandM(VPABSB_128, 1963 "vpabsb %%xmm6, %%xmm7", 1964 "vpabsb (%%rax), %%xmm7") 1965 1966 GEN_test_RandM(VPABSW_128, 1967 "vpabsw %%xmm6, %%xmm7", 1968 "vpabsw (%%rax), %%xmm7") 1969 1970 GEN_test_RandM(VPMOVSXBQ_128, 1971 "vpmovsxbq %%xmm6, %%xmm8", 1972 "vpmovsxbq (%%rax), %%xmm8") 1973 1974 GEN_test_RandM(VPMOVSXWQ_128, 1975 "vpmovsxwq %%xmm6, %%xmm8", 1976 "vpmovsxwq (%%rax), %%xmm8") 1977 1978 GEN_test_RandM(VPACKUSDW_128, 1979 "vpackusdw %%xmm6, %%xmm8, %%xmm7", 1980 "vpackusdw (%%rax), %%xmm8, %%xmm7") 1981 1982 GEN_test_RandM(VPMOVZXBQ_128, 1983 "vpmovzxbq %%xmm6, %%xmm8", 1984 "vpmovzxbq (%%rax), %%xmm8") 1985 1986 GEN_test_RandM(VPMOVZXWQ_128, 1987 "vpmovzxwq %%xmm6, %%xmm8", 1988 "vpmovzxwq (%%rax), %%xmm8") 1989 1990 GEN_test_RandM(VPMOVZXDQ_128, 1991 "vpmovzxdq %%xmm6, %%xmm8", 1992 "vpmovzxdq (%%rax), %%xmm8") 1993 1994 GEN_test_RandM(VMPSADBW_128_0x0, 1995 "vmpsadbw $0, %%xmm6, %%xmm8, %%xmm7", 1996 "vmpsadbw $0, (%%rax), %%xmm8, %%xmm7") 1997 GEN_test_RandM(VMPSADBW_128_0x1, 1998 "vmpsadbw $1, %%xmm6, %%xmm8, %%xmm7", 1999 "vmpsadbw $1, (%%rax), %%xmm8, %%xmm7") 2000 GEN_test_RandM(VMPSADBW_128_0x2, 2001 "vmpsadbw $2, %%xmm6, %%xmm8, %%xmm7", 2002 "vmpsadbw $2, (%%rax), %%xmm8, %%xmm7") 2003 GEN_test_RandM(VMPSADBW_128_0x3, 2004 "vmpsadbw $3, %%xmm6, %%xmm8, %%xmm7", 2005 "vmpsadbw $3, (%%rax), %%xmm8, %%xmm7") 2006 GEN_test_RandM(VMPSADBW_128_0x4, 2007 "vmpsadbw $4, %%xmm6, %%xmm8, %%xmm7", 2008 "vmpsadbw $4, (%%rax), %%xmm8, %%xmm7") 2009 GEN_test_RandM(VMPSADBW_128_0x5, 2010 "vmpsadbw $5, %%xmm6, %%xmm8, %%xmm7", 2011 "vmpsadbw $5, (%%rax), %%xmm8, %%xmm7") 2012 GEN_test_RandM(VMPSADBW_128_0x6, 2013 "vmpsadbw $6, %%xmm6, %%xmm8, %%xmm7", 2014 "vmpsadbw $6, (%%rax), %%xmm8, %%xmm7") 2015 GEN_test_RandM(VMPSADBW_128_0x7, 2016 "vmpsadbw $7, %%xmm6, %%xmm8, %%xmm7", 2017 "vmpsadbw $7, (%%rax), %%xmm8, %%xmm7") 2018 2019 GEN_test_RandM(VMOVDDUP_YMMorMEM256_to_YMM, 2020 "vmovddup %%ymm8, %%ymm7", 2021 "vmovddup (%%rax), %%ymm9") 2022 2023 GEN_test_Monly(VMOVLPS_128_M64_XMM_XMM, "vmovlps (%%rax), %%xmm8, %%xmm7") 2024 2025 GEN_test_Monly(VMOVLPS_128_XMM_M64, "vmovlps %%xmm7, (%%rax)") 2026 2027 GEN_test_RandM(VRCPSS_128, 2028 "vrcpss %%xmm6, %%xmm8, %%xmm7", 2029 "vrcpss (%%rax), %%xmm8, %%xmm7") 2030 2031 GEN_test_RandM(VRCPPS_128, 2032 "vrcpps %%xmm6, %%xmm8", 2033 "vrcpps (%%rax), %%xmm8") 2034 2035 GEN_test_RandM(VRCPPS_256, 2036 "vrcpps %%ymm6, %%ymm8", 2037 "vrcpps (%%rax), %%ymm8") 2038 2039 GEN_test_RandM(VPSADBW_128, 2040 "vpsadbw %%xmm6, %%xmm8, %%xmm7", 2041 "vpsadbw (%%rax), %%xmm8, %%xmm7") 2042 2043 GEN_test_RandM(VPSIGNB_128, 2044 "vpsignb %%xmm6, %%xmm8, %%xmm7", 2045 "vpsignb (%%rax), %%xmm8, %%xmm7") 2046 2047 GEN_test_RandM(VPSIGNW_128, 2048 "vpsignw %%xmm6, %%xmm8, %%xmm7", 2049 "vpsignw (%%rax), %%xmm8, %%xmm7") 2050 2051 GEN_test_RandM(VPSIGND_128, 2052 "vpsignd %%xmm6, %%xmm8, %%xmm7", 2053 "vpsignd (%%rax), %%xmm8, %%xmm7") 2054 2055 GEN_test_RandM(VPMULHRSW_128, 2056 "vpmulhrsw %%xmm6, %%xmm8, %%xmm7", 2057 "vpmulhrsw (%%rax), %%xmm8, %%xmm7") 2058 2059 GEN_test_Monly(VBROADCASTF128, 2060 "vbroadcastf128 (%%rax), %%ymm9") 2061 2062 GEN_test_RandM(VPEXTRW_128_0x0, 2063 "vpextrw $0x0, %%xmm7, %%r14d", 2064 "vpextrw $0x0, %%xmm7, (%%rax)") 2065 GEN_test_RandM(VPEXTRW_128_0x1, 2066 "vpextrw $0x1, %%xmm7, %%r14d", 2067 "vpextrw $0x1, %%xmm7, (%%rax)") 2068 GEN_test_RandM(VPEXTRW_128_0x2, 2069 "vpextrw $0x2, %%xmm7, %%r14d", 2070 "vpextrw $0x2, %%xmm7, (%%rax)") 2071 GEN_test_RandM(VPEXTRW_128_0x3, 2072 "vpextrw $0x3, %%xmm7, %%r14d", 2073 "vpextrw $0x3, %%xmm7, (%%rax)") 2074 GEN_test_RandM(VPEXTRW_128_0x4, 2075 "vpextrw $0x4, %%xmm7, %%r14d", 2076 "vpextrw $0x4, %%xmm7, (%%rax)") 2077 GEN_test_RandM(VPEXTRW_128_0x5, 2078 "vpextrw $0x5, %%xmm7, %%r14d", 2079 "vpextrw $0x5, %%xmm7, (%%rax)") 2080 GEN_test_RandM(VPEXTRW_128_0x6, 2081 "vpextrw $0x6, %%xmm7, %%r14d", 2082 "vpextrw $0x6, %%xmm7, (%%rax)") 2083 GEN_test_RandM(VPEXTRW_128_0x7, 2084 "vpextrw $0x7, %%xmm7, %%r14d", 2085 "vpextrw $0x7, %%xmm7, (%%rax)") 2086 2087 GEN_test_RandM(VAESENC, 2088 "vaesenc %%xmm6, %%xmm8, %%xmm7", 2089 "vaesenc (%%rax), %%xmm8, %%xmm7") 2090 2091 GEN_test_RandM(VAESENCLAST, 2092 "vaesenclast %%xmm6, %%xmm8, %%xmm7", 2093 "vaesenclast (%%rax), %%xmm8, %%xmm7") 2094 2095 GEN_test_RandM(VAESDEC, 2096 "vaesdec %%xmm6, %%xmm8, %%xmm7", 2097 "vaesdec (%%rax), %%xmm8, %%xmm7") 2098 2099 GEN_test_RandM(VAESDECLAST, 2100 "vaesdeclast %%xmm6, %%xmm8, %%xmm7", 2101 "vaesdeclast (%%rax), %%xmm8, %%xmm7") 2102 2103 GEN_test_RandM(VAESIMC, 2104 "vaesimc %%xmm6, %%xmm7", 2105 "vaesimc (%%rax), %%xmm7") 2106 2107 GEN_test_RandM(VAESKEYGENASSIST_0x00, 2108 "vaeskeygenassist $0x00, %%xmm6, %%xmm7", 2109 "vaeskeygenassist $0x00, (%%rax), %%xmm7") 2110 GEN_test_RandM(VAESKEYGENASSIST_0x31, 2111 "vaeskeygenassist $0x31, %%xmm6, %%xmm7", 2112 "vaeskeygenassist $0x31, (%%rax), %%xmm7") 2113 GEN_test_RandM(VAESKEYGENASSIST_0xB2, 2114 "vaeskeygenassist $0xb2, %%xmm6, %%xmm7", 2115 "vaeskeygenassist $0xb2, (%%rax), %%xmm7") 2116 GEN_test_RandM(VAESKEYGENASSIST_0xFF, 2117 "vaeskeygenassist $0xFF, %%xmm6, %%xmm7", 2118 "vaeskeygenassist $0xFF, (%%rax), %%xmm7") 2119 2120 GEN_test_RandM(VPCLMULQDQ_0x00, 2121 "vpclmulqdq $0x00, %%xmm6, %%xmm8, %%xmm7", 2122 "vpclmulqdq $0x00, (%%rax), %%xmm8, %%xmm7") 2123 GEN_test_RandM(VPCLMULQDQ_0x01, 2124 "vpclmulqdq $0x01, %%xmm6, %%xmm8, %%xmm7", 2125 "vpclmulqdq $0x01, (%%rax), %%xmm8, %%xmm7") 2126 GEN_test_RandM(VPCLMULQDQ_0x10, 2127 "vpclmulqdq $0x10, %%xmm6, %%xmm8, %%xmm7", 2128 "vpclmulqdq $0x10, (%%rax), %%xmm8, %%xmm7") 2129 GEN_test_RandM(VPCLMULQDQ_0x11, 2130 "vpclmulqdq $0x11, %%xmm6, %%xmm8, %%xmm7", 2131 "vpclmulqdq $0x11, (%%rax), %%xmm8, %%xmm7") 2132 GEN_test_RandM(VPCLMULQDQ_0xFF, 2133 "vpclmulqdq $0xFF, %%xmm6, %%xmm8, %%xmm7", 2134 "vpclmulqdq $0xFF, (%%rax), %%xmm8, %%xmm7") 2135 2136 GEN_test_RandM(VCMPSS_128_0x9, 2137 "vcmpss $0x9, %%xmm6, %%xmm8, %%xmm7", 2138 "vcmpss $0x9, (%%rax), %%xmm8, %%xmm7") 2139 2140 GEN_test_Monly(VMASKMOVPS_128_LoadForm, 2141 "vmaskmovps (%%rax), %%xmm8, %%xmm7;" 2142 "vxorps %%xmm6, %%xmm6, %%xmm6;" 2143 "vmaskmovps (%%rax,%%rax,4), %%xmm6, %%xmm9") 2144 2145 GEN_test_Monly(VMASKMOVPS_256_LoadForm, 2146 "vmaskmovps (%%rax), %%ymm8, %%ymm7;" 2147 "vxorps %%ymm6, %%ymm6, %%ymm6;" 2148 "vmaskmovps (%%rax,%%rax,4), %%ymm6, %%ymm9") 2149 2150 GEN_test_Monly(VMASKMOVPD_128_LoadForm, 2151 "vmaskmovpd (%%rax), %%xmm8, %%xmm7;" 2152 "vxorpd %%xmm6, %%xmm6, %%xmm6;" 2153 "vmaskmovpd (%%rax,%%rax,4), %%xmm6, %%xmm9") 2154 2155 GEN_test_Monly(VMASKMOVPD_256_LoadForm, 2156 "vmaskmovpd (%%rax), %%ymm8, %%ymm7;" 2157 "vxorpd %%ymm6, %%ymm6, %%ymm6;" 2158 "vmaskmovpd (%%rax,%%rax,4), %%ymm6, %%ymm9") 2159 2160 GEN_test_Monly(VMASKMOVPS_128_StoreForm, 2161 "vmaskmovps %%xmm8, %%xmm7, (%%rax);" 2162 "vxorps %%xmm6, %%xmm6, %%xmm6;" 2163 "vmaskmovps %%xmm9, %%xmm6, (%%rax,%%rax,4)") 2164 2165 GEN_test_Monly(VMASKMOVPS_256_StoreForm, 2166 "vmaskmovps %%ymm8, %%ymm7, (%%rax);" 2167 "vxorps %%ymm6, %%ymm6, %%ymm6;" 2168 "vmaskmovps %%ymm9, %%ymm6, (%%rax,%%rax,4)") 2169 2170 GEN_test_Monly(VMASKMOVPD_128_StoreForm, 2171 "vmaskmovpd %%xmm8, %%xmm7, (%%rax);" 2172 "vxorpd %%xmm6, %%xmm6, %%xmm6;" 2173 "vmaskmovpd %%xmm9, %%xmm6, (%%rax,%%rax,4)") 2174 2175 GEN_test_Monly(VMASKMOVPD_256_StoreForm, 2176 "vmaskmovpd %%ymm8, %%ymm7, (%%rax);" 2177 "vxorpd %%ymm6, %%ymm6, %%ymm6;" 2178 "vmaskmovpd %%ymm9, %%ymm6, (%%rax,%%rax,4)") 2179 2180 /* Comment duplicated above, for convenient reference: 2181 Allowed operands in test insns: 2182 Reg form: %ymm6, %ymm7, %ymm8, %ymm9 and %r14. 2183 Mem form: (%rax), %ymm7, %ymm8, %ymm9 and %r14. 2184 Imm8 etc fields are also allowed, where they make sense. 2185 Both forms may use ymm0 as scratch. Mem form may also use 2186 ymm6 as scratch. 2187 */ 2188 2189 #define N_DEFAULT_ITERS 3 2190 2191 // Do the specified test some number of times 2192 #define DO_N(_iters, _testfn) \ 2193 do { int i; for (i = 0; i < (_iters); i++) { test_##_testfn(); } } while (0) 2194 2195 // Do the specified test the default number of times 2196 #define DO_D(_testfn) DO_N(N_DEFAULT_ITERS, _testfn) 2197 2198 2199 int main ( void ) 2200 { 2201 DO_D( VMOVUPD_EtoG_256 ); 2202 DO_D( VMOVUPD_GtoE_256 ); 2203 DO_D( VPSUBW_128 ); 2204 DO_D( VPSUBQ_128 ); 2205 DO_D( VPADDQ_128 ); 2206 DO_D( VPINSRQ_128 ); 2207 DO_D( VUCOMISS_128 ); 2208 DO_D( VUCOMISD_128 ); 2209 DO_D( VCVTPS2PD_128 ); 2210 DO_D( VANDNPD_128 ); 2211 DO_D( VORPD_128 ); 2212 DO_D( VXORPD_128 ); 2213 DO_D( VXORPS_128 ); 2214 DO_D( VMULSD_128 ); 2215 DO_D( VADDSD_128 ); 2216 DO_D( VMINSD_128 ); 2217 DO_D( VSUBSD_128 ); 2218 DO_D( VDIVSD_128 ); 2219 DO_D( VMAXSD_128 ); 2220 DO_D( VPSHUFD_0x39_128 ); 2221 DO_D( VPCMPEQD_128 ); 2222 DO_D( VPEXTRD_128_0x3 ); 2223 DO_D( VPEXTRD_128_0x0 ); 2224 DO_D( VINSERTF128_0x0 ); 2225 DO_D( VINSERTF128_0x1 ); 2226 DO_D( VEXTRACTF128_0x0 ); 2227 DO_D( VEXTRACTF128_0x1 ); 2228 DO_D( VCVTPD2PS_128 ); 2229 /* Test all CMPSS variants; this code is tricky. */ 2230 DO_D( VCMPSS_128_0x0 ); 2231 DO_D( VCMPSS_128_0x1 ); 2232 DO_D( VCMPSS_128_0x2 ); 2233 DO_D( VCMPSS_128_0x3 ); 2234 DO_D( VCMPSS_128_0x4 ); 2235 DO_D( VCMPSS_128_0x5 ); 2236 DO_D( VCMPSS_128_0x6 ); 2237 DO_D( VCMPSS_128_0x7 ); 2238 DO_D( VCMPSS_128_0x8 ); 2239 DO_D( VCMPSS_128_0xA ); 2240 DO_D( VCMPSS_128_0xC ); 2241 DO_D( VCMPSS_128_0xC ); 2242 DO_D( VCMPSS_128_0xD ); 2243 DO_D( VCMPSS_128_0xE ); 2244 DO_D( VCMPSS_128_0x11 ); 2245 DO_D( VCMPSS_128_0x12); 2246 DO_D( VCMPSS_128_0x16 ); 2247 DO_D( VCMPSS_128_0x1E ); 2248 DO_D( VMOVDDUP_XMMorMEM64_to_XMM ); 2249 DO_D( VMOVD_IREGorMEM32_to_XMM ); 2250 DO_D( VMOVQ_XMM_MEM64 ); 2251 DO_D( VMOVDQA_GtoE_256 ); 2252 DO_D( VMOVDQA_GtoE_128 ); 2253 DO_D( VMOVDQU_GtoE_128 ); 2254 DO_D( VMOVDQA_EtoG_256 ); 2255 DO_D( VMOVDQA_EtoG_128 ); 2256 DO_D( VMOVDQU_EtoG_128 ); 2257 DO_D( VMOVAPD_GtoE_128 ); 2258 DO_D( VMOVAPD_GtoE_256 ); 2259 DO_D( VMOVAPS_GtoE_128 ); 2260 DO_D( VMOVAPS_GtoE_256 ); 2261 DO_D( VMOVAPS_EtoG_128 ); 2262 DO_D( VMOVAPD_EtoG_256 ); 2263 DO_D( VMOVAPD_EtoG_128 ); 2264 DO_D( VMOVUPD_GtoE_128 ); 2265 DO_D( VMOVSS_XMM_M32 ); 2266 DO_D( VMOVSD_XMM_M64 ); 2267 DO_D( VMOVSS_M64_XMM ); 2268 DO_D( VMOVSD_M64_XMM ); 2269 DO_D( VINSERTPS_0x39_128 ); 2270 DO_D( VPUNPCKLDQ_128 ); 2271 DO_D( VPACKSSDW_128 ); 2272 DO_D( VPADDW_128 ); 2273 DO_D( VPSRLW_0x05_128 ); 2274 DO_D( VPSLLW_0x05_128 ); 2275 DO_D( VPUNPCKLQDQ_128 ); 2276 DO_D( VPINSRD_128 ); 2277 DO_D( VMOVD_XMM_to_MEM32 ); 2278 DO_D( VPANDN_128 ); 2279 DO_D( VPSLLDQ_0x05_128 ); 2280 DO_D( VPSRLDQ_0x05_128 ); 2281 DO_D( VPSUBUSB_128 ); 2282 DO_D( VPSUBSB_128 ); 2283 DO_D( VPSLLD_0x05_128 ); 2284 DO_D( VPSRLD_0x05_128 ); 2285 DO_D( VPSRAD_0x05_128 ); 2286 DO_D( VPUNPCKLWD_128 ); 2287 DO_D( VPUNPCKHWD_128 ); 2288 DO_D( VPADDUSB_128 ); 2289 DO_D( VPMULHUW_128 ); 2290 DO_D( VPADDUSW_128 ); 2291 DO_D( VPMULLW_128 ); 2292 DO_D( VPSHUFHW_0x39_128 ); 2293 DO_D( VPSHUFLW_0x39_128 ); 2294 DO_D( VCVTPS2DQ_128 ); 2295 DO_D( VSUBPS_128 ); 2296 DO_D( VADDPS_128 ); 2297 DO_D( VMULPS_128 ); 2298 DO_D( VMAXPS_128 ); 2299 DO_D( VMINPS_128 ); 2300 DO_D( VSHUFPS_0x39_128 ); 2301 DO_D( VPCMPEQB_128 ); 2302 DO_D( VMOVHPD_128_StoreForm ); 2303 DO_D( VPAND_128 ); 2304 DO_D( VPMOVMSKB_128 ); 2305 DO_D( VCVTTSS2SI_64 ); 2306 DO_D( VPACKUSWB_128 ); 2307 DO_D( VCVTSS2SD_128 ); 2308 DO_D( VCVTSD2SS_128 ); 2309 DO_D( VMOVD_XMM_to_IREG32 ); 2310 DO_D( VPCMPESTRM_0x45_128 ); 2311 DO_D( VMOVQ_IREGorMEM64_to_XMM ); 2312 DO_D( VMOVUPS_XMM_to_XMMorMEM ); 2313 DO_D( VMOVNTDQ_128 ); 2314 DO_D( VMOVLHPS_128 ); 2315 DO_D( VPABSD_128 ); 2316 DO_D( VMOVHLPS_128 ); 2317 DO_D( VMOVQ_XMM_to_IREG64 ); 2318 DO_D( VMOVQ_XMMorMEM64_to_XMM ); 2319 DO_D( VCVTTSS2SI_32 ); 2320 DO_D( VPUNPCKLBW_128 ); 2321 DO_D( VPUNPCKHBW_128 ); 2322 DO_D( VMULSS_128 ); 2323 DO_D( VSUBSS_128 ); 2324 DO_D( VADDSS_128 ); 2325 DO_D( VDIVSS_128 ); 2326 DO_D( VUNPCKLPS_128 ); 2327 DO_D( VCVTSI2SS_128 ); 2328 DO_D( VANDPS_128 ); 2329 DO_D( VMINSS_128 ); 2330 DO_D( VMAXSS_128 ); 2331 DO_D( VANDNPS_128 ); 2332 DO_D( VORPS_128 ); 2333 DO_D( VSQRTSD_128 ); 2334 /* Test all CMPSS variants; this code is tricky. */ 2335 DO_D( VCMPSD_128_0x0 ); 2336 DO_D( VCMPSD_128_0x1 ); 2337 DO_D( VCMPSD_128_0x2 ); 2338 DO_D( VCMPSD_128_0x3 ); 2339 DO_D( VCMPSD_128_0x4 ); 2340 DO_D( VCMPSD_128_0x5 ); 2341 DO_D( VCMPSD_128_0x6 ); 2342 DO_D( VCMPSD_128_0x7 ); 2343 DO_D( VCMPSD_128_0x8 ); 2344 DO_D( VCMPSD_128_0xA ); 2345 DO_D( VCMPSD_128_0xC ); 2346 DO_D( VCMPSD_128_0xD ); 2347 DO_D( VCMPSD_128_0xE ); 2348 DO_D( VCMPSD_128_0x11 ); 2349 DO_D( VCMPSD_128_0x12 ); 2350 DO_D( VCMPSD_128_0x16 ); 2351 DO_D( VCMPSD_128_0x1E ); 2352 DO_D( VPSHUFB_128 ); 2353 DO_D( VCVTTSD2SI_32 ); 2354 DO_D( VCVTTSD2SI_64 ); 2355 DO_D( VCVTSI2SS_64 ); 2356 DO_D( VCVTSI2SD_64 ); 2357 DO_D( VCVTSI2SD_32 ); 2358 DO_D( VPOR_128 ); 2359 DO_D( VPXOR_128 ); 2360 DO_D( VPSUBB_128 ); 2361 DO_D( VPSUBD_128 ); 2362 DO_D( VPADDD_128 ); 2363 DO_D( VPMOVZXBW_128 ); 2364 DO_D( VPMOVZXWD_128 ); 2365 DO_D( VPBLENDVB_128 ); 2366 DO_D( VPMINSD_128 ); 2367 DO_D( VPMAXSD_128 ); 2368 DO_D( VANDPD_128 ); 2369 DO_D( VMULPD_256 ); 2370 DO_D( VMOVUPD_EtoG_128 ); 2371 DO_D( VADDPD_256 ); 2372 DO_D( VSUBPD_256 ); 2373 DO_D( VDIVPD_256 ); 2374 DO_D( VPCMPEQQ_128 ); 2375 DO_D( VSUBPD_128 ); 2376 DO_D( VADDPD_128 ); 2377 DO_D( VUNPCKLPD_128 ); 2378 DO_D( VUNPCKHPD_128 ); 2379 DO_D( VUNPCKHPS_128 ); 2380 DO_D( VMOVUPS_EtoG_128 ); 2381 DO_D( VADDPS_256 ); 2382 DO_D( VSUBPS_256 ); 2383 DO_D( VMULPS_256 ); 2384 DO_D( VDIVPS_256 ); 2385 DO_D( VPCMPGTQ_128 ); 2386 DO_D( VPEXTRQ_128_0x0 ); 2387 DO_D( VPEXTRQ_128_0x1 ); 2388 DO_D( VPSRLQ_0x05_128 ); 2389 DO_D( VPMULUDQ_128 ); 2390 DO_D( VPSLLQ_0x05_128 ); 2391 DO_D( VPMAXUD_128 ); 2392 DO_D( VPMINUD_128 ); 2393 DO_D( VPMULLD_128 ); 2394 DO_D( VPMAXUW_128 ); 2395 DO_D( VPEXTRW_128_EregOnly_toG_0x0 ); 2396 DO_D( VPEXTRW_128_EregOnly_toG_0x7 ); 2397 DO_D( VPMINUW_128 ); 2398 DO_D( VPHMINPOSUW_128 ); 2399 DO_D( VPMAXSW_128 ); 2400 DO_D( VPMINSW_128 ); 2401 DO_D( VPMAXUB_128 ); 2402 DO_D( VPEXTRB_GtoE_128_0x0 ); 2403 DO_D( VPEXTRB_GtoE_128_0x1 ); 2404 DO_D( VPEXTRB_GtoE_128_0x2 ); 2405 DO_D( VPEXTRB_GtoE_128_0x3 ); 2406 DO_D( VPEXTRB_GtoE_128_0x4 ); 2407 DO_D( VPEXTRB_GtoE_128_0x9 ); 2408 DO_D( VPEXTRB_GtoE_128_0xE ); 2409 DO_D( VPEXTRB_GtoE_128_0xF ); 2410 DO_D( VPMINUB_128 ); 2411 DO_D( VPMAXSB_128 ); 2412 DO_D( VPMINSB_128 ); 2413 DO_D( VPERM2F128_0x00 ); 2414 DO_D( VPERM2F128_0xFF ); 2415 DO_D( VPERM2F128_0x30 ); 2416 DO_D( VPERM2F128_0x21 ); 2417 DO_D( VPERM2F128_0x12 ); 2418 DO_D( VPERM2F128_0x03 ); 2419 DO_D( VPERM2F128_0x85 ); 2420 DO_D( VPERM2F128_0x5A ); 2421 DO_D( VPERMILPD_256_0x0 ); 2422 DO_D( VPERMILPD_256_0xF ); 2423 DO_D( VPERMILPD_256_0xA ); 2424 DO_D( VPERMILPD_256_0x5 ); 2425 DO_D( VPERMILPD_128_0x0 ); 2426 DO_D( VPERMILPD_128_0x3 ); 2427 DO_D( VUNPCKLPD_256 ); 2428 DO_D( VUNPCKHPD_256 ); 2429 DO_D( VSHUFPS_0x39_256 ); 2430 DO_D( VUNPCKLPS_256 ); 2431 DO_D( VUNPCKHPS_256 ); 2432 DO_D( VXORPD_256 ); 2433 DO_D( VBROADCASTSD_256 ); 2434 DO_D( VCMPPD_128_0x4 ); 2435 DO_D( VCVTDQ2PD_128 ); 2436 DO_D( VDIVPD_128 ); 2437 DO_D( VANDPD_256 ); 2438 DO_D( VPMOVSXBW_128 ); 2439 DO_D( VPSUBUSW_128 ); 2440 DO_D( VPSUBSW_128 ); 2441 DO_D( VPCMPEQW_128 ); 2442 DO_D( VPADDB_128 ); 2443 DO_D( VMOVAPS_EtoG_256 ); 2444 DO_D( VCVTDQ2PD_256 ); 2445 DO_D( VMOVHPD_128_LoadForm ); 2446 DO_D( VCVTPD2PS_256 ); 2447 DO_D( VPUNPCKHDQ_128 ); 2448 DO_D( VBROADCASTSS_128 ); 2449 DO_D( VPMOVSXDQ_128 ); 2450 DO_D( VPMOVSXWD_128 ); 2451 DO_D( VDIVPS_128 ); 2452 DO_D( VANDPS_256 ); 2453 DO_D( VXORPS_256 ); 2454 DO_D( VORPS_256 ); 2455 DO_D( VANDNPD_256 ); 2456 DO_D( VANDNPS_256 ); 2457 DO_D( VORPD_256 ); 2458 DO_D( VPERMILPS_256_0x0F ); 2459 DO_D( VPERMILPS_256_0xFA ); 2460 DO_D( VPERMILPS_256_0xA3 ); 2461 DO_D( VPERMILPS_256_0x5A ); 2462 DO_D( VPMULHW_128 ); 2463 DO_D( VPUNPCKHQDQ_128 ); 2464 DO_D( VPSRAW_0x05_128 ); 2465 DO_D( VPCMPGTD_128 ); 2466 DO_D( VPMOVZXBD_128 ); 2467 DO_D( VPMOVSXBD_128 ); 2468 DO_D( VPINSRB_128_1of3 ); 2469 DO_D( VPINSRB_128_2of3 ); 2470 DO_D( VPINSRB_128_3of3 ); 2471 DO_D( VCOMISD_128 ); 2472 DO_D( VCOMISS_128 ); 2473 DO_D( VMOVUPS_YMM_to_YMMorMEM ); 2474 DO_D( VDPPD_128_1of4 ); 2475 DO_D( VDPPD_128_2of4 ); 2476 DO_D( VDPPD_128_3of4 ); 2477 DO_D( VDPPD_128_4of4 ); 2478 DO_D( VPINSRW_128_1of4 ); 2479 DO_D( VPINSRW_128_2of4 ); 2480 DO_D( VPINSRW_128_3of4 ); 2481 DO_D( VPINSRW_128_4of4 ); 2482 DO_D( VBROADCASTSS_256 ); 2483 DO_D( VPALIGNR_128_1of3 ); 2484 DO_D( VPALIGNR_128_2of3 ); 2485 DO_D( VPALIGNR_128_3of3 ); 2486 DO_D( VMOVSD_REG_XMM ); 2487 DO_D( VMOVSS_REG_XMM ); 2488 DO_D( VMOVLPD_128_M64_XMM_XMM ); 2489 DO_D( VMOVLPD_128_XMM_M64 ); 2490 DO_D( VSHUFPD_128_1of2 ); 2491 DO_D( VSHUFPD_128_2of2 ); 2492 DO_D( VSHUFPD_256_1of2 ); 2493 DO_D( VSHUFPD_256_2of2 ); 2494 DO_D( VPERMILPS_128_0x00 ); 2495 DO_D( VPERMILPS_128_0xFE ); 2496 DO_D( VPERMILPS_128_0x30 ); 2497 DO_D( VPERMILPS_128_0x21 ); 2498 DO_D( VPERMILPS_128_0xD7 ); 2499 DO_D( VPERMILPS_128_0xB5 ); 2500 DO_D( VPERMILPS_128_0x85 ); 2501 DO_D( VPERMILPS_128_0x29 ); 2502 DO_D( VBLENDPS_128_1of3 ); 2503 DO_D( VBLENDPS_128_2of3 ); 2504 DO_D( VBLENDPS_128_3of3 ); 2505 DO_D( VBLENDPD_128_1of2 ); 2506 DO_D( VBLENDPD_128_2of2 ); 2507 DO_D( VBLENDPD_256_1of3 ); 2508 DO_D( VBLENDPD_256_2of3 ); 2509 DO_D( VBLENDPD_256_3of3 ); 2510 DO_D( VPBLENDW_128_0x00 ); 2511 DO_D( VPBLENDW_128_0xFE ); 2512 DO_D( VPBLENDW_128_0x30 ); 2513 DO_D( VPBLENDW_128_0x21 ); 2514 DO_D( VPBLENDW_128_0xD7 ); 2515 DO_D( VPBLENDW_128_0xB5 ); 2516 DO_D( VPBLENDW_128_0x85 ); 2517 DO_D( VPBLENDW_128_0x29 ); 2518 DO_D( VMOVUPS_EtoG_256 ); 2519 DO_D( VSQRTSS_128 ); 2520 DO_D( VSQRTPS_128 ); 2521 DO_D( VSQRTPS_256 ); 2522 DO_D( VSQRTPD_128 ); 2523 DO_D( VSQRTPD_256 ); 2524 DO_D( VRSQRTSS_128 ); 2525 DO_D( VRSQRTPS_128 ); 2526 DO_D( VRSQRTPS_256 ); 2527 DO_D( VMOVDQU_GtoE_256 ); 2528 DO_D( VCVTPS2PD_256 ); 2529 DO_D( VCVTTPS2DQ_128 ); 2530 DO_D( VCVTTPS2DQ_256 ); 2531 DO_D( VCVTDQ2PS_128 ); 2532 DO_D( VCVTDQ2PS_256 ); 2533 DO_D( VCVTTPD2DQ_128 ); 2534 DO_D( VCVTTPD2DQ_256 ); 2535 DO_D( VCVTPD2DQ_128 ); 2536 DO_D( VCVTPD2DQ_256 ); 2537 DO_D( VMOVSLDUP_128 ); 2538 DO_D( VMOVSLDUP_256 ); 2539 DO_D( VMOVSHDUP_128 ); 2540 DO_D( VMOVSHDUP_256 ); 2541 DO_D( VPERMILPS_VAR_128 ); 2542 DO_D( VPERMILPD_VAR_128 ); 2543 DO_D( VPERMILPS_VAR_256 ); 2544 DO_D( VPERMILPD_VAR_256 ); 2545 DO_D( VPSLLW_128 ); 2546 DO_D( VPSRLW_128 ); 2547 DO_D( VPSRAW_128 ); 2548 DO_D( VPSLLD_128 ); 2549 DO_D( VPSRLD_128 ); 2550 DO_D( VPSRAD_128 ); 2551 DO_D( VPSLLQ_128 ); 2552 DO_D( VPSRLQ_128 ); 2553 DO_D( VROUNDPS_128_0x0 ); 2554 DO_D( VROUNDPS_128_0x1 ); 2555 DO_D( VROUNDPS_128_0x2 ); 2556 DO_D( VROUNDPS_128_0x3 ); 2557 DO_D( VROUNDPS_128_0x4 ); 2558 DO_D( VROUNDPS_256_0x0 ); 2559 DO_D( VROUNDPS_256_0x1 ); 2560 DO_D( VROUNDPS_256_0x2 ); 2561 DO_D( VROUNDPS_256_0x3 ); 2562 DO_D( VROUNDPS_256_0x4 ); 2563 DO_D( VROUNDPD_128_0x0 ); 2564 DO_D( VROUNDPD_128_0x1 ); 2565 DO_D( VROUNDPD_128_0x2 ); 2566 DO_D( VROUNDPD_128_0x3 ); 2567 DO_D( VROUNDPD_128_0x4 ); 2568 DO_D( VROUNDPD_256_0x0 ); 2569 DO_D( VROUNDPD_256_0x1 ); 2570 DO_D( VROUNDPD_256_0x2 ); 2571 DO_D( VROUNDPD_256_0x3 ); 2572 DO_D( VROUNDPD_256_0x4 ); 2573 DO_D( VROUNDSS_0x0 ); 2574 DO_D( VROUNDSS_0x1 ); 2575 DO_D( VROUNDSS_0x2 ); 2576 DO_D( VROUNDSS_0x3 ); 2577 DO_D( VROUNDSS_0x4 ); 2578 DO_D( VROUNDSS_0x5 ); 2579 DO_D( VROUNDSD_0x0 ); 2580 DO_D( VROUNDSD_0x1 ); 2581 DO_D( VROUNDSD_0x2 ); 2582 DO_D( VROUNDSD_0x3 ); 2583 DO_D( VROUNDSD_0x4 ); 2584 DO_D( VROUNDSD_0x5 ); 2585 DO_D( VPTEST_128_1 ); 2586 DO_D( VPTEST_128_2 ); 2587 DO_D( VPTEST_256_1 ); 2588 DO_D( VPTEST_256_2 ); 2589 DO_D( VTESTPS_128_1 ); 2590 DO_D( VTESTPS_128_2 ); 2591 DO_N( 10, VTESTPS_128_3 ); 2592 DO_D( VTESTPS_256_1 ); 2593 DO_D( VTESTPS_256_2 ); 2594 DO_N( 10, VTESTPS_256_3 ); 2595 DO_D( VTESTPD_128_1 ); 2596 DO_D( VTESTPD_128_2 ); 2597 DO_N( 10, VTESTPD_128_3 ); 2598 DO_D( VTESTPD_256_1 ); 2599 DO_D( VTESTPD_256_2 ); 2600 DO_N( 10, VTESTPD_256_3 ); 2601 DO_D( VBLENDVPS_128 ); 2602 DO_D( VBLENDVPS_256 ); 2603 DO_D( VBLENDVPD_128 ); 2604 DO_D( VBLENDVPD_256 ); 2605 DO_D( VPMULDQ_128 ); 2606 DO_D( VCMPPD_256_0x4 ); 2607 DO_D( VCMPPS_128_0x4 ); 2608 DO_D( VCMPPS_256_0x4 ); 2609 DO_D( VPCMPGTB_128 ); 2610 DO_D( VPCMPGTW_128 ); 2611 DO_D( VPMADDWD_128 ); 2612 DO_D( VADDSUBPS_128 ); 2613 DO_D( VADDSUBPS_256 ); 2614 DO_D( VADDSUBPD_128 ); 2615 DO_D( VADDSUBPD_256 ); 2616 DO_D( VCVTSS2SI_64 ); 2617 DO_D( VCVTSS2SI_32 ); 2618 DO_D( VCVTSD2SI_32 ); 2619 DO_D( VCVTSD2SI_64 ); 2620 DO_D( VDPPS_128_1of4 ); 2621 DO_D( VDPPS_128_2of4 ); 2622 DO_D( VDPPS_128_3of4 ); 2623 DO_D( VDPPS_128_4of4 ); 2624 DO_D( VDPPS_256_1of4 ); 2625 DO_D( VDPPS_256_2of4 ); 2626 DO_D( VDPPS_256_3of4 ); 2627 DO_D( VDPPS_256_4of4 ); 2628 DO_D( VHADDPS_128 ); 2629 DO_D( VHADDPS_256 ); 2630 DO_D( VHADDPD_128 ); 2631 DO_D( VHADDPD_256 ); 2632 DO_D( VHSUBPS_128 ); 2633 DO_D( VHSUBPS_256 ); 2634 DO_D( VHSUBPD_128 ); 2635 DO_D( VHSUBPD_256 ); 2636 DO_D( VEXTRACTPS_0x0 ); 2637 DO_D( VEXTRACTPS_0x1 ); 2638 DO_D( VEXTRACTPS_0x2 ); 2639 DO_D( VEXTRACTPS_0x3 ); 2640 DO_D( VLDDQU_128 ); 2641 DO_D( VLDDQU_256 ); 2642 DO_D( VMAXPS_256 ); 2643 DO_D( VMAXPD_128 ); 2644 DO_D( VMAXPD_256 ); 2645 DO_D( VMINPS_256 ); 2646 DO_D( VMINPD_128 ); 2647 DO_D( VMINPD_256 ); 2648 DO_D( VMOVHPS_128_StoreForm ); 2649 DO_D( VMOVNTDQ_256 ); 2650 DO_D( VMOVHPS_128_LoadForm ); 2651 DO_D( VMOVNTDQA_128 ); 2652 DO_D( VMASKMOVDQU_128 ); 2653 DO_D( VMOVMSKPD_128 ); 2654 DO_D( VMOVMSKPD_256 ); 2655 DO_D( VMOVMSKPS_128 ); 2656 DO_D( VMOVMSKPS_256 ); 2657 DO_D( VMOVNTPD_128 ); 2658 DO_D( VMOVNTPD_256 ); 2659 DO_D( VMOVNTPS_128 ); 2660 DO_D( VMOVNTPS_256 ); 2661 DO_D( VPACKSSWB_128 ); 2662 DO_D( VPAVGB_128 ); 2663 DO_D( VPAVGW_128 ); 2664 DO_D( VPADDSB_128 ); 2665 DO_D( VPADDSW_128 ); 2666 DO_D( VPHADDW_128 ); 2667 DO_D( VPHADDD_128 ); 2668 DO_D( VPHADDSW_128 ); 2669 DO_D( VPMADDUBSW_128 ); 2670 DO_D( VPHSUBW_128 ); 2671 DO_D( VPHSUBD_128 ); 2672 DO_D( VPHSUBSW_128 ); 2673 DO_D( VPABSB_128 ); 2674 DO_D( VPABSW_128 ); 2675 DO_D( VPMOVSXBQ_128 ); 2676 DO_D( VPMOVSXWQ_128 ); 2677 DO_D( VPACKUSDW_128 ); 2678 DO_D( VPMOVZXBQ_128 ); 2679 DO_D( VPMOVZXWQ_128 ); 2680 DO_D( VPMOVZXDQ_128 ); 2681 DO_D( VMPSADBW_128_0x0 ); 2682 DO_D( VMPSADBW_128_0x1 ); 2683 DO_D( VMPSADBW_128_0x2 ); 2684 DO_D( VMPSADBW_128_0x3 ); 2685 DO_D( VMPSADBW_128_0x4 ); 2686 DO_D( VMPSADBW_128_0x5 ); 2687 DO_D( VMPSADBW_128_0x6 ); 2688 DO_D( VMPSADBW_128_0x7 ); 2689 DO_D( VMOVDDUP_YMMorMEM256_to_YMM ); 2690 DO_D( VMOVLPS_128_M64_XMM_XMM ); 2691 DO_D( VMOVLPS_128_XMM_M64 ); 2692 DO_D( VRCPSS_128 ); 2693 DO_D( VRCPPS_128 ); 2694 DO_D( VRCPPS_256 ); 2695 DO_D( VPSADBW_128 ); 2696 DO_D( VPSIGNB_128 ); 2697 DO_D( VPSIGNW_128 ); 2698 DO_D( VPSIGND_128 ); 2699 DO_D( VPMULHRSW_128 ); 2700 DO_D( VBROADCASTF128 ); 2701 DO_D( VPEXTRW_128_0x0 ); 2702 DO_D( VPEXTRW_128_0x1 ); 2703 DO_D( VPEXTRW_128_0x2 ); 2704 DO_D( VPEXTRW_128_0x3 ); 2705 DO_D( VPEXTRW_128_0x4 ); 2706 DO_D( VPEXTRW_128_0x5 ); 2707 DO_D( VPEXTRW_128_0x6 ); 2708 DO_D( VPEXTRW_128_0x7 ); 2709 DO_D( VAESENC ); 2710 DO_D( VAESENCLAST ); 2711 DO_D( VAESDEC ); 2712 DO_D( VAESDECLAST ); 2713 DO_D( VAESIMC ); 2714 DO_D( VAESKEYGENASSIST_0x00 ); 2715 DO_D( VAESKEYGENASSIST_0x31 ); 2716 DO_D( VAESKEYGENASSIST_0xB2 ); 2717 DO_D( VAESKEYGENASSIST_0xFF ); 2718 DO_D( VPCLMULQDQ_0x00 ); 2719 DO_D( VPCLMULQDQ_0x01 ); 2720 DO_D( VPCLMULQDQ_0x10 ); 2721 DO_D( VPCLMULQDQ_0x11 ); 2722 DO_D( VPCLMULQDQ_0xFF ); 2723 DO_D( VCMPSS_128_0x9 ); 2724 DO_D( VMASKMOVPS_128_LoadForm ); 2725 DO_D( VMASKMOVPS_256_LoadForm ); 2726 DO_D( VMASKMOVPD_128_LoadForm ); 2727 DO_D( VMASKMOVPD_256_LoadForm ); 2728 DO_D( VMASKMOVPS_128_StoreForm ); 2729 DO_D( VMASKMOVPS_256_StoreForm ); 2730 DO_D( VMASKMOVPD_128_StoreForm ); 2731 DO_D( VMASKMOVPD_256_StoreForm ); 2732 return 0; 2733 } 2734 2735