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