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_0xA, 216 "vcmpsd $0xA, %%xmm6, %%xmm8, %%xmm7", 217 "vcmpsd $0xA, (%%rax), %%xmm8, %%xmm7") 218 GEN_test_RandM(VCMPSD_128_0xC, 219 "vcmpsd $0xC, %%xmm6, %%xmm8, %%xmm7", 220 "vcmpsd $0xC, (%%rax), %%xmm8, %%xmm7") 221 GEN_test_RandM(VCMPSD_128_0xD, 222 "vcmpsd $0xD, %%xmm6, %%xmm8, %%xmm7", 223 "vcmpsd $0xD, (%%rax), %%xmm8, %%xmm7") 224 GEN_test_RandM(VCMPSD_128_0xE, 225 "vcmpsd $0xE, %%xmm6, %%xmm8, %%xmm7", 226 "vcmpsd $0xE, (%%rax), %%xmm8, %%xmm7") 227 GEN_test_RandM(VCMPSD_128_0x11, 228 "vcmpsd $0x11, %%xmm6, %%xmm8, %%xmm7", 229 "vcmpsd $0x11, (%%rax), %%xmm8, %%xmm7") 230 GEN_test_RandM(VCMPSD_128_0x12, 231 "vcmpsd $0x12, %%xmm6, %%xmm8, %%xmm7", 232 "vcmpsd $0x12, (%%rax), %%xmm8, %%xmm7") 233 GEN_test_RandM(VCMPSD_128_0x16, 234 "vcmpsd $0x16, %%xmm6, %%xmm8, %%xmm7", 235 "vcmpsd $0x16, (%%rax), %%xmm8, %%xmm7") 236 GEN_test_RandM(VCMPSD_128_0x1E, 237 "vcmpsd $0x1E, %%xmm6, %%xmm8, %%xmm7", 238 "vcmpsd $0x1E, (%%rax), %%xmm8, %%xmm7") 239 240 GEN_test_RandM(VSQRTSD_128, 241 "vsqrtsd %%xmm6, %%xmm8, %%xmm7", 242 "vsqrtsd (%%rax), %%xmm8, %%xmm7") 243 244 GEN_test_RandM(VORPS_128, 245 "vorps %%xmm6, %%xmm8, %%xmm7", 246 "vorps (%%rax), %%xmm8, %%xmm7") 247 248 GEN_test_RandM(VANDNPS_128, 249 "vandnps %%xmm6, %%xmm8, %%xmm7", 250 "vandnps (%%rax), %%xmm8, %%xmm7") 251 252 GEN_test_RandM(VMAXSS_128, 253 "vmaxss %%xmm6, %%xmm8, %%xmm7", 254 "vmaxss (%%rax), %%xmm8, %%xmm7") 255 256 GEN_test_RandM(VMINSS_128, 257 "vminss %%xmm6, %%xmm8, %%xmm7", 258 "vminss (%%rax), %%xmm8, %%xmm7") 259 260 GEN_test_RandM(VANDPS_128, 261 "vandps %%xmm6, %%xmm8, %%xmm7", 262 "vandps (%%rax), %%xmm8, %%xmm7") 263 264 GEN_test_RandM(VCVTSI2SS_128, 265 "vcvtsi2ssl %%r14d, %%xmm8, %%xmm7", 266 "vcvtsi2ssl (%%rax), %%xmm8, %%xmm7") 267 268 GEN_test_RandM(VUNPCKLPS_128, 269 "vunpcklps %%xmm6, %%xmm8, %%xmm7", 270 "vunpcklps (%%rax), %%xmm8, %%xmm7") 271 272 GEN_test_RandM(VDIVSS_128, 273 "vdivss %%xmm6, %%xmm8, %%xmm7", 274 "vdivss (%%rax), %%xmm8, %%xmm7") 275 276 GEN_test_RandM(VADDSS_128, 277 "vaddss %%xmm6, %%xmm8, %%xmm7", 278 "vaddss (%%rax), %%xmm8, %%xmm7") 279 280 GEN_test_RandM(VSUBSS_128, 281 "vsubss %%xmm6, %%xmm8, %%xmm7", 282 "vsubss (%%rax), %%xmm8, %%xmm7") 283 284 GEN_test_RandM(VMULSS_128, 285 "vmulss %%xmm6, %%xmm8, %%xmm7", 286 "vmulss (%%rax), %%xmm8, %%xmm7") 287 288 GEN_test_RandM(VPUNPCKLBW_128, 289 "vpunpcklbw %%xmm6, %%xmm8, %%xmm7", 290 "vpunpcklbw (%%rax), %%xmm8, %%xmm7") 291 292 GEN_test_RandM(VPUNPCKHBW_128, 293 "vpunpckhbw %%xmm6, %%xmm8, %%xmm7", 294 "vpunpckhbw (%%rax), %%xmm8, %%xmm7") 295 296 GEN_test_RandM(VCVTTSS2SI_32, 297 "vcvttss2si %%xmm8, %%r14d", 298 "vcvttss2si (%%rax), %%r14d") 299 300 GEN_test_RandM(VCVTSS2SI_32, 301 "vcvtss2si %%xmm8, %%r14d", 302 "vcvtss2si (%%rax), %%r14d") 303 304 GEN_test_RandM(VMOVQ_XMMorMEM64_to_XMM, 305 "vmovq %%xmm7, %%xmm8", 306 "vmovq (%%rax), %%xmm8") 307 308 /* NB tests the reg form only */ 309 GEN_test_Ronly(VMOVQ_XMM_to_IREG64, 310 "vmovq %%xmm7, %%r14") 311 312 /* This insn only exists in the reg-reg-reg form. */ 313 GEN_test_Ronly(VMOVHLPS_128, 314 "vmovhlps %%xmm6, %%xmm8, %%xmm7") 315 316 GEN_test_RandM(VPABSD_128, 317 "vpabsd %%xmm6, %%xmm8", 318 "vpabsd (%%rax), %%xmm8") 319 320 /* This insn only exists in the reg-reg-reg form. */ 321 GEN_test_Ronly(VMOVLHPS_128, 322 "vmovlhps %%xmm6, %%xmm8, %%xmm7") 323 324 GEN_test_Monly(VMOVNTDQ_128, 325 "vmovntdq %%xmm8, (%%rax)") 326 327 GEN_test_Monly(VMOVNTDQ_256, 328 "vmovntdq %%ymm8, (%%rax)") 329 330 GEN_test_RandM(VMOVUPS_XMM_to_XMMorMEM, 331 "vmovups %%xmm8, %%xmm7", 332 "vmovups %%xmm9, (%%rax)") 333 334 GEN_test_RandM(VMOVQ_IREGorMEM64_to_XMM, 335 "vmovq %%r14, %%xmm7", 336 "vmovq (%%rax), %%xmm9") 337 338 GEN_test_RandM(VPCMPESTRM_0x45_128, 339 "vpcmpestrm $0x45, %%xmm7, %%xmm8; movapd %%xmm0, %%xmm9", 340 "vpcmpestrm $0x45, (%%rax), %%xmm8; movapd %%xmm0, %%xmm9") 341 342 /* NB tests the reg form only */ 343 GEN_test_Ronly(VMOVD_XMM_to_IREG32, 344 "vmovd %%xmm7, %%r14d") 345 346 GEN_test_RandM(VCVTSD2SS_128, 347 "vcvtsd2ss %%xmm9, %%xmm8, %%xmm7", 348 "vcvtsd2ss (%%rax), %%xmm8, %%xmm7") 349 350 GEN_test_RandM(VCVTSS2SD_128, 351 "vcvtss2sd %%xmm9, %%xmm8, %%xmm7", 352 "vcvtss2sd (%%rax), %%xmm8, %%xmm7") 353 354 GEN_test_RandM(VPACKUSWB_128, 355 "vpackuswb %%xmm9, %%xmm8, %%xmm7", 356 "vpackuswb (%%rax), %%xmm8, %%xmm7") 357 358 GEN_test_RandM(VCVTTSS2SI_64, 359 "vcvttss2si %%xmm8, %%r14", 360 "vcvttss2si (%%rax), %%r14") 361 362 GEN_test_RandM(VCVTSS2SI_64, 363 "vcvtss2si %%xmm8, %%r14", 364 "vcvtss2si (%%rax), %%r14") 365 366 GEN_test_Ronly(VPMOVMSKB_128, 367 "vpmovmskb %%xmm8, %%r14") 368 369 GEN_test_RandM(VPAND_128, 370 "vpand %%xmm9, %%xmm8, %%xmm7", 371 "vpand (%%rax), %%xmm8, %%xmm7") 372 373 GEN_test_Monly(VMOVHPD_128_StoreForm, 374 "vmovhpd %%xmm8, (%%rax)") 375 376 GEN_test_Monly(VMOVHPS_128_StoreForm, 377 "vmovhps %%xmm8, (%%rax)") 378 379 GEN_test_RandM(VPCMPEQB_128, 380 "vpcmpeqb %%xmm9, %%xmm8, %%xmm7", 381 "vpcmpeqb (%%rax), %%xmm8, %%xmm7") 382 383 GEN_test_RandM(VSHUFPS_0x39_128, 384 "vshufps $0x39, %%xmm9, %%xmm8, %%xmm7", 385 "vshufps $0xC6, (%%rax), %%xmm8, %%xmm7") 386 387 GEN_test_RandM(VMULPS_128, 388 "vmulps %%xmm9, %%xmm8, %%xmm7", 389 "vmulps (%%rax), %%xmm8, %%xmm7") 390 391 GEN_test_RandM(VSUBPS_128, 392 "vsubps %%xmm9, %%xmm8, %%xmm7", 393 "vsubps (%%rax), %%xmm8, %%xmm7") 394 395 GEN_test_RandM(VADDPS_128, 396 "vaddps %%xmm9, %%xmm8, %%xmm7", 397 "vaddps (%%rax), %%xmm8, %%xmm7") 398 399 GEN_test_RandM(VMAXPS_128, 400 "vmaxps %%xmm9, %%xmm8, %%xmm7", 401 "vmaxps (%%rax), %%xmm8, %%xmm7") 402 403 GEN_test_RandM(VMAXPS_256, 404 "vmaxps %%ymm9, %%ymm8, %%ymm7", 405 "vmaxps (%%rax), %%ymm8, %%ymm7") 406 407 GEN_test_RandM(VMAXPD_128, 408 "vmaxpd %%xmm9, %%xmm8, %%xmm7", 409 "vmaxpd (%%rax), %%xmm8, %%xmm7") 410 411 GEN_test_RandM(VMAXPD_256, 412 "vmaxpd %%ymm9, %%ymm8, %%ymm7", 413 "vmaxpd (%%rax), %%ymm8, %%ymm7") 414 415 GEN_test_RandM(VMINPS_128, 416 "vminps %%xmm9, %%xmm8, %%xmm7", 417 "vminps (%%rax), %%xmm8, %%xmm7") 418 419 GEN_test_RandM(VMINPS_256, 420 "vminps %%ymm9, %%ymm8, %%ymm7", 421 "vminps (%%rax), %%ymm8, %%ymm7") 422 423 GEN_test_RandM(VMINPD_128, 424 "vminpd %%xmm9, %%xmm8, %%xmm7", 425 "vminpd (%%rax), %%xmm8, %%xmm7") 426 427 GEN_test_RandM(VMINPD_256, 428 "vminpd %%ymm9, %%ymm8, %%ymm7", 429 "vminpd (%%rax), %%ymm8, %%ymm7") 430 431 GEN_test_RandM(VCVTPS2DQ_128, 432 "vcvtps2dq %%xmm8, %%xmm7", 433 "vcvtps2dq (%%rax), %%xmm8") 434 435 GEN_test_RandM(VPSHUFLW_0x39_128, 436 "vpshuflw $0x39, %%xmm9, %%xmm7", 437 "vpshuflw $0xC6, (%%rax), %%xmm8") 438 439 GEN_test_RandM(VPSHUFHW_0x39_128, 440 "vpshufhw $0x39, %%xmm9, %%xmm7", 441 "vpshufhw $0xC6, (%%rax), %%xmm8") 442 443 GEN_test_RandM(VPMULLW_128, 444 "vpmullw %%xmm9, %%xmm8, %%xmm7", 445 "vpmullw (%%rax), %%xmm8, %%xmm7") 446 447 GEN_test_RandM(VPADDUSW_128, 448 "vpaddusw %%xmm9, %%xmm8, %%xmm7", 449 "vpaddusw (%%rax), %%xmm8, %%xmm7") 450 451 GEN_test_RandM(VPMULHUW_128, 452 "vpmulhuw %%xmm9, %%xmm8, %%xmm7", 453 "vpmulhuw (%%rax), %%xmm8, %%xmm7") 454 455 GEN_test_RandM(VPADDUSB_128, 456 "vpaddusb %%xmm9, %%xmm8, %%xmm7", 457 "vpaddusb (%%rax), %%xmm8, %%xmm7") 458 459 GEN_test_RandM(VPUNPCKLWD_128, 460 "vpunpcklwd %%xmm6, %%xmm8, %%xmm7", 461 "vpunpcklwd (%%rax), %%xmm8, %%xmm7") 462 463 GEN_test_RandM(VPUNPCKHWD_128, 464 "vpunpckhwd %%xmm6, %%xmm8, %%xmm7", 465 "vpunpckhwd (%%rax), %%xmm8, %%xmm7") 466 467 GEN_test_Ronly(VPSLLD_0x05_128, 468 "vpslld $0x5, %%xmm9, %%xmm7") 469 470 GEN_test_Ronly(VPSRLD_0x05_128, 471 "vpsrld $0x5, %%xmm9, %%xmm7") 472 473 GEN_test_Ronly(VPSRAD_0x05_128, 474 "vpsrad $0x5, %%xmm9, %%xmm7") 475 476 GEN_test_RandM(VPSUBUSB_128, 477 "vpsubusb %%xmm9, %%xmm8, %%xmm7", 478 "vpsubusb (%%rax), %%xmm8, %%xmm7") 479 480 GEN_test_RandM(VPSUBSB_128, 481 "vpsubsb %%xmm9, %%xmm8, %%xmm7", 482 "vpsubsb (%%rax), %%xmm8, %%xmm7") 483 484 GEN_test_Ronly(VPSRLDQ_0x05_128, 485 "vpsrldq $0x5, %%xmm9, %%xmm7") 486 487 GEN_test_Ronly(VPSLLDQ_0x05_128, 488 "vpslldq $0x5, %%xmm9, %%xmm7") 489 490 GEN_test_RandM(VPANDN_128, 491 "vpandn %%xmm9, %%xmm8, %%xmm7", 492 "vpandn (%%rax), %%xmm8, %%xmm7") 493 494 /* NB tests the mem form only */ 495 GEN_test_Monly(VMOVD_XMM_to_MEM32, 496 "vmovd %%xmm7, (%%rax)") 497 498 GEN_test_RandM(VPINSRD_128, 499 "vpinsrd $0, %%r14d, %%xmm8, %%xmm7", 500 "vpinsrd $3, (%%rax), %%xmm8, %%xmm7") 501 502 GEN_test_RandM(VPUNPCKLQDQ_128, 503 "vpunpcklqdq %%xmm6, %%xmm8, %%xmm7", 504 "vpunpcklqdq (%%rax), %%xmm8, %%xmm7") 505 506 GEN_test_Ronly(VPSRLW_0x05_128, 507 "vpsrlw $0x5, %%xmm9, %%xmm7") 508 509 GEN_test_Ronly(VPSLLW_0x05_128, 510 "vpsllw $0x5, %%xmm9, %%xmm7") 511 512 GEN_test_RandM(VPADDW_128, 513 "vpaddw %%xmm6, %%xmm8, %%xmm7", 514 "vpaddw (%%rax), %%xmm8, %%xmm7") 515 516 GEN_test_RandM(VPACKSSDW_128, 517 "vpackssdw %%xmm9, %%xmm8, %%xmm7", 518 "vpackssdw (%%rax), %%xmm8, %%xmm7") 519 520 GEN_test_RandM(VPUNPCKLDQ_128, 521 "vpunpckldq %%xmm6, %%xmm8, %%xmm7", 522 "vpunpckldq (%%rax), %%xmm8, %%xmm7") 523 524 GEN_test_RandM(VINSERTPS_0x39_128, 525 "vinsertps $0x39, %%xmm6, %%xmm8, %%xmm7", 526 "vinsertps $0xC6, (%%rax), %%xmm8, %%xmm7") 527 528 GEN_test_Monly(VMOVSD_M64_XMM, "vmovsd (%%rax), %%xmm8") 529 530 GEN_test_Monly(VMOVSS_M64_XMM, "vmovss (%%rax), %%xmm8") 531 532 GEN_test_Monly(VMOVSD_XMM_M64, "vmovsd %%xmm8, (%%rax)") 533 534 GEN_test_Monly(VMOVSS_XMM_M32, "vmovss %%xmm8, (%%rax)") 535 536 GEN_test_RandM(VMOVUPD_GtoE_128, 537 "vmovupd %%xmm9, %%xmm6", 538 "vmovupd %%xmm7, (%%rax)") 539 540 GEN_test_RandM(VMOVAPD_EtoG_128, 541 "vmovapd %%xmm6, %%xmm8", 542 "vmovapd (%%rax), %%xmm9") 543 544 GEN_test_RandM(VMOVAPD_EtoG_256, 545 "vmovapd %%ymm6, %%ymm8", 546 "vmovapd (%%rax), %%ymm9") 547 548 GEN_test_RandM(VMOVAPS_EtoG_128, 549 "vmovaps %%xmm6, %%xmm8", 550 "vmovaps (%%rax), %%xmm9") 551 552 GEN_test_RandM(VMOVAPS_GtoE_128, 553 "vmovaps %%xmm9, %%xmm6", 554 "vmovaps %%xmm7, (%%rax)") 555 556 GEN_test_RandM(VMOVAPS_GtoE_256, 557 "vmovaps %%ymm9, %%ymm6", 558 "vmovaps %%ymm7, (%%rax)") 559 560 GEN_test_RandM(VMOVAPD_GtoE_128, 561 "vmovapd %%xmm9, %%xmm6", 562 "vmovapd %%xmm7, (%%rax)") 563 564 GEN_test_RandM(VMOVAPD_GtoE_256, 565 "vmovapd %%ymm9, %%ymm6", 566 "vmovapd %%ymm7, (%%rax)") 567 568 GEN_test_RandM(VMOVDQU_EtoG_128, 569 "vmovdqu %%xmm6, %%xmm8", 570 "vmovdqu (%%rax), %%xmm9") 571 572 GEN_test_RandM(VMOVDQA_EtoG_128, 573 "vmovdqa %%xmm6, %%xmm8", 574 "vmovdqa (%%rax), %%xmm9") 575 576 GEN_test_RandM(VMOVDQA_EtoG_256, 577 "vmovdqa %%ymm6, %%ymm8", 578 "vmovdqa (%%rax), %%ymm9") 579 580 GEN_test_RandM(VMOVDQU_GtoE_128, 581 "vmovdqu %%xmm9, %%xmm6", 582 "vmovdqu %%xmm7, (%%rax)") 583 584 GEN_test_RandM(VMOVDQA_GtoE_128, 585 "vmovdqa %%xmm9, %%xmm6", 586 "vmovdqa %%xmm7, (%%rax)") 587 588 GEN_test_RandM(VMOVDQA_GtoE_256, 589 "vmovdqa %%ymm9, %%ymm6", 590 "vmovdqa %%ymm7, (%%rax)") 591 592 GEN_test_Monly(VMOVQ_XMM_MEM64, "vmovq %%xmm8, (%%rax)") 593 594 GEN_test_RandM(VMOVD_IREGorMEM32_to_XMM, 595 "vmovd %%r14d, %%xmm7", 596 "vmovd (%%rax), %%xmm9") 597 598 GEN_test_RandM(VMOVDDUP_XMMorMEM64_to_XMM, 599 "vmovddup %%xmm8, %%xmm7", 600 "vmovddup (%%rax), %%xmm9") 601 602 GEN_test_RandM(VCMPSS_128_0x0, 603 "vcmpss $0, %%xmm6, %%xmm8, %%xmm7", 604 "vcmpss $0, (%%rax), %%xmm8, %%xmm7") 605 GEN_test_RandM(VCMPSS_128_0x1, 606 "vcmpss $1, %%xmm6, %%xmm8, %%xmm7", 607 "vcmpss $1, (%%rax), %%xmm8, %%xmm7") 608 GEN_test_RandM(VCMPSS_128_0x2, 609 "vcmpss $2, %%xmm6, %%xmm8, %%xmm7", 610 "vcmpss $2, (%%rax), %%xmm8, %%xmm7") 611 GEN_test_RandM(VCMPSS_128_0x3, 612 "vcmpss $3, %%xmm6, %%xmm8, %%xmm7", 613 "vcmpss $3, (%%rax), %%xmm8, %%xmm7") 614 GEN_test_RandM(VCMPSS_128_0x4, 615 "vcmpss $4, %%xmm6, %%xmm8, %%xmm7", 616 "vcmpss $4, (%%rax), %%xmm8, %%xmm7") 617 GEN_test_RandM(VCMPSS_128_0x5, 618 "vcmpss $5, %%xmm6, %%xmm8, %%xmm7", 619 "vcmpss $5, (%%rax), %%xmm8, %%xmm7") 620 GEN_test_RandM(VCMPSS_128_0x6, 621 "vcmpss $6, %%xmm6, %%xmm8, %%xmm7", 622 "vcmpss $6, (%%rax), %%xmm8, %%xmm7") 623 GEN_test_RandM(VCMPSS_128_0x7, 624 "vcmpss $7, %%xmm6, %%xmm8, %%xmm7", 625 "vcmpss $7, (%%rax), %%xmm8, %%xmm7") 626 GEN_test_RandM(VCMPSS_128_0xA, 627 "vcmpss $0xA, %%xmm6, %%xmm8, %%xmm7", 628 "vcmpss $0xA, (%%rax), %%xmm8, %%xmm7") 629 GEN_test_RandM(VCMPSS_128_0xC, 630 "vcmpss $0xC, %%xmm6, %%xmm8, %%xmm7", 631 "vcmpss $0xC, (%%rax), %%xmm8, %%xmm7") 632 GEN_test_RandM(VCMPSS_128_0xD, 633 "vcmpss $0xD, %%xmm6, %%xmm8, %%xmm7", 634 "vcmpss $0xD, (%%rax), %%xmm8, %%xmm7") 635 GEN_test_RandM(VCMPSS_128_0xE, 636 "vcmpss $0xE, %%xmm6, %%xmm8, %%xmm7", 637 "vcmpss $0xE, (%%rax), %%xmm8, %%xmm7") 638 GEN_test_RandM(VCMPSS_128_0x11, 639 "vcmpss $0x11, %%xmm6, %%xmm8, %%xmm7", 640 "vcmpss $0x11, (%%rax), %%xmm8, %%xmm7") 641 GEN_test_RandM(VCMPSS_128_0x12, 642 "vcmpss $0x12, %%xmm6, %%xmm8, %%xmm7", 643 "vcmpss $0x12, (%%rax), %%xmm8, %%xmm7") 644 GEN_test_RandM(VCMPSS_128_0x16, 645 "vcmpss $0x16, %%xmm6, %%xmm8, %%xmm7", 646 "vcmpss $0x16, (%%rax), %%xmm8, %%xmm7") 647 GEN_test_RandM(VCMPSS_128_0x1E, 648 "vcmpss $0x1E, %%xmm6, %%xmm8, %%xmm7", 649 "vcmpss $0x1E, (%%rax), %%xmm8, %%xmm7") 650 651 // The x suffix denotes a 128 -> 64 operation 652 GEN_test_RandM(VCVTPD2PS_128, 653 "vcvtpd2psx %%xmm8, %%xmm7", 654 "vcvtpd2psx (%%rax), %%xmm9") 655 656 GEN_test_RandM(VEXTRACTF128_0x0, 657 "vextractf128 $0x0, %%ymm7, %%xmm9", 658 "vextractf128 $0x0, %%ymm7, (%%rax)") 659 660 GEN_test_RandM(VEXTRACTF128_0x1, 661 "vextractf128 $0x1, %%ymm7, %%xmm9", 662 "vextractf128 $0x1, %%ymm7, (%%rax)") 663 664 GEN_test_RandM(VINSERTF128_0x0, 665 "vinsertf128 $0x0, %%xmm9, %%ymm7, %%ymm8", 666 "vinsertf128 $0x0, (%%rax), %%ymm7, %%ymm8") 667 668 GEN_test_RandM(VINSERTF128_0x1, 669 "vinsertf128 $0x1, %%xmm9, %%ymm7, %%ymm8", 670 "vinsertf128 $0x1, (%%rax), %%ymm7, %%ymm8") 671 672 GEN_test_RandM(VPEXTRD_128_0x0, 673 "vpextrd $0x0, %%xmm7, %%r14d", 674 "vpextrd $0x0, %%xmm7, (%%rax)") 675 676 GEN_test_RandM(VPEXTRD_128_0x3, 677 "vpextrd $0x3, %%xmm7, %%r14d", 678 "vpextrd $0x3, %%xmm7, (%%rax)") 679 680 GEN_test_RandM(VPCMPEQD_128, 681 "vpcmpeqd %%xmm6, %%xmm8, %%xmm7", 682 "vpcmpeqd (%%rax), %%xmm8, %%xmm7") 683 684 GEN_test_RandM(VPSHUFD_0x39_128, 685 "vpshufd $0x39, %%xmm9, %%xmm8", 686 "vpshufd $0xC6, (%%rax), %%xmm7") 687 688 GEN_test_RandM(VMAXSD_128, 689 "vmaxsd %%xmm6, %%xmm8, %%xmm7", 690 "vmaxsd (%%rax), %%xmm8, %%xmm7") 691 692 GEN_test_RandM(VDIVSD_128, 693 "vdivsd %%xmm6, %%xmm8, %%xmm7", 694 "vdivsd (%%rax), %%xmm8, %%xmm7") 695 696 GEN_test_RandM(VMINSD_128, 697 "vminsd %%xmm6, %%xmm8, %%xmm7", 698 "vminsd (%%rax), %%xmm8, %%xmm7") 699 700 GEN_test_RandM(VSUBSD_128, 701 "vsubsd %%xmm6, %%xmm8, %%xmm7", 702 "vsubsd (%%rax), %%xmm8, %%xmm7") 703 704 GEN_test_RandM(VADDSD_128, 705 "vaddsd %%xmm6, %%xmm8, %%xmm7", 706 "vaddsd (%%rax), %%xmm8, %%xmm7") 707 708 GEN_test_RandM(VMULSD_128, 709 "vmulsd %%xmm6, %%xmm8, %%xmm7", 710 "vmulsd (%%rax), %%xmm8, %%xmm7") 711 712 GEN_test_RandM(VXORPS_128, 713 "vxorps %%xmm6, %%xmm8, %%xmm7", 714 "vxorps (%%rax), %%xmm8, %%xmm7") 715 716 GEN_test_RandM(VXORPD_128, 717 "vxorpd %%xmm6, %%xmm8, %%xmm7", 718 "vxorpd (%%rax), %%xmm8, %%xmm7") 719 720 GEN_test_RandM(VORPD_128, 721 "vorpd %%xmm6, %%xmm8, %%xmm7", 722 "vorpd (%%rax), %%xmm8, %%xmm7") 723 724 GEN_test_RandM(VANDNPD_128, 725 "vandnpd %%xmm6, %%xmm8, %%xmm7", 726 "vandnpd (%%rax), %%xmm8, %%xmm7") 727 728 GEN_test_RandM(VCVTPS2PD_128, 729 "vcvtps2pd %%xmm6, %%xmm8", 730 "vcvtps2pd (%%rax), %%xmm8") 731 732 GEN_test_RandM(VUCOMISD_128, 733 "vucomisd %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 734 "vucomisd (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 735 736 GEN_test_RandM(VUCOMISS_128, 737 "vucomiss %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 738 "vucomiss (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 739 740 GEN_test_RandM(VPINSRQ_128, 741 "vpinsrq $0, %%r14, %%xmm8, %%xmm7", 742 "vpinsrq $1, (%%rax), %%xmm8, %%xmm7") 743 744 GEN_test_RandM(VPADDQ_128, 745 "vpaddq %%xmm6, %%xmm8, %%xmm7", 746 "vpaddq (%%rax), %%xmm8, %%xmm7") 747 748 GEN_test_RandM(VPSUBQ_128, 749 "vpsubq %%xmm6, %%xmm8, %%xmm7", 750 "vpsubq (%%rax), %%xmm8, %%xmm7") 751 752 GEN_test_RandM(VPSUBW_128, 753 "vpsubw %%xmm6, %%xmm8, %%xmm7", 754 "vpsubw (%%rax), %%xmm8, %%xmm7") 755 756 GEN_test_RandM(VMOVUPD_GtoE_256, 757 "vmovupd %%ymm9, %%ymm6", 758 "vmovupd %%ymm7, (%%rax)") 759 760 GEN_test_RandM(VMOVUPD_EtoG_256, 761 "vmovupd %%ymm6, %%ymm9", 762 "vmovupd (%%rax), %%ymm7") 763 764 GEN_test_RandM(VMULPD_256, 765 "vmulpd %%ymm6, %%ymm8, %%ymm7", 766 "vmulpd (%%rax), %%ymm8, %%ymm7") 767 768 GEN_test_RandM(VMOVUPD_EtoG_128, 769 "vmovupd %%xmm6, %%xmm9", 770 "vmovupd (%%rax), %%xmm7") 771 772 GEN_test_RandM(VADDPD_256, 773 "vaddpd %%ymm6, %%ymm8, %%ymm7", 774 "vaddpd (%%rax), %%ymm8, %%ymm7") 775 776 GEN_test_RandM(VSUBPD_256, 777 "vsubpd %%ymm6, %%ymm8, %%ymm7", 778 "vsubpd (%%rax), %%ymm8, %%ymm7") 779 780 GEN_test_RandM(VDIVPD_256, 781 "vdivpd %%ymm6, %%ymm8, %%ymm7", 782 "vdivpd (%%rax), %%ymm8, %%ymm7") 783 784 GEN_test_RandM(VPCMPEQQ_128, 785 "vpcmpeqq %%xmm6, %%xmm8, %%xmm7", 786 "vpcmpeqq (%%rax), %%xmm8, %%xmm7") 787 788 GEN_test_RandM(VSUBPD_128, 789 "vsubpd %%xmm6, %%xmm8, %%xmm7", 790 "vsubpd (%%rax), %%xmm8, %%xmm7") 791 792 GEN_test_RandM(VADDPD_128, 793 "vaddpd %%xmm6, %%xmm8, %%xmm7", 794 "vaddpd (%%rax), %%xmm8, %%xmm7") 795 796 GEN_test_RandM(VUNPCKLPD_128, 797 "vunpcklpd %%xmm6, %%xmm8, %%xmm7", 798 "vunpcklpd (%%rax), %%xmm8, %%xmm7") 799 800 GEN_test_RandM(VUNPCKHPD_128, 801 "vunpckhpd %%xmm6, %%xmm8, %%xmm7", 802 "vunpckhpd (%%rax), %%xmm8, %%xmm7") 803 804 GEN_test_RandM(VUNPCKHPS_128, 805 "vunpckhps %%xmm6, %%xmm8, %%xmm7", 806 "vunpckhps (%%rax), %%xmm8, %%xmm7") 807 808 GEN_test_RandM(VMOVUPS_EtoG_128, 809 "vmovups %%xmm6, %%xmm8", 810 "vmovups (%%rax), %%xmm9") 811 812 GEN_test_RandM(VADDPS_256, 813 "vaddps %%ymm6, %%ymm8, %%ymm7", 814 "vaddps (%%rax), %%ymm8, %%ymm7") 815 816 GEN_test_RandM(VSUBPS_256, 817 "vsubps %%ymm6, %%ymm8, %%ymm7", 818 "vsubps (%%rax), %%ymm8, %%ymm7") 819 820 GEN_test_RandM(VMULPS_256, 821 "vmulps %%ymm6, %%ymm8, %%ymm7", 822 "vmulps (%%rax), %%ymm8, %%ymm7") 823 824 GEN_test_RandM(VDIVPS_256, 825 "vdivps %%ymm6, %%ymm8, %%ymm7", 826 "vdivps (%%rax), %%ymm8, %%ymm7") 827 828 GEN_test_RandM(VPCMPGTQ_128, 829 "vpcmpgtq %%xmm6, %%xmm8, %%xmm7", 830 "vpcmpgtq (%%rax), %%xmm8, %%xmm7") 831 832 GEN_test_RandM(VPEXTRQ_128_0x0, 833 "vpextrq $0x0, %%xmm7, %%r14", 834 "vpextrq $0x0, %%xmm7, (%%rax)") 835 836 GEN_test_RandM(VPEXTRQ_128_0x1, 837 "vpextrq $0x1, %%xmm7, %%r14", 838 "vpextrq $0x1, %%xmm7, (%%rax)") 839 840 GEN_test_Ronly(VPSRLQ_0x05_128, 841 "vpsrlq $0x5, %%xmm9, %%xmm7") 842 843 GEN_test_RandM(VPMULUDQ_128, 844 "vpmuludq %%xmm6, %%xmm8, %%xmm7", 845 "vpmuludq (%%rax), %%xmm8, %%xmm7") 846 847 GEN_test_RandM(VPMULDQ_128, 848 "vpmuldq %%xmm6, %%xmm8, %%xmm7", 849 "vpmuldq (%%rax), %%xmm8, %%xmm7") 850 851 GEN_test_Ronly(VPSLLQ_0x05_128, 852 "vpsllq $0x5, %%xmm9, %%xmm7") 853 854 GEN_test_RandM(VPMAXUD_128, 855 "vpmaxud %%xmm6, %%xmm8, %%xmm7", 856 "vpmaxud (%%rax), %%xmm8, %%xmm7") 857 858 GEN_test_RandM(VPMINUD_128, 859 "vpminud %%xmm6, %%xmm8, %%xmm7", 860 "vpminud (%%rax), %%xmm8, %%xmm7") 861 862 GEN_test_RandM(VPMULLD_128, 863 "vpmulld %%xmm6, %%xmm8, %%xmm7", 864 "vpmulld (%%rax), %%xmm8, %%xmm7") 865 866 GEN_test_RandM(VPMAXUW_128, 867 "vpmaxuw %%xmm6, %%xmm8, %%xmm7", 868 "vpmaxuw (%%rax), %%xmm8, %%xmm7") 869 870 GEN_test_Ronly(VPEXTRW_128_EregOnly_toG_0x0, 871 "vpextrw $0x0, %%xmm7, %%r14d") 872 873 GEN_test_Ronly(VPEXTRW_128_EregOnly_toG_0x7, 874 "vpextrw $0x7, %%xmm7, %%r14d") 875 876 GEN_test_RandM(VPMINUW_128, 877 "vpminuw %%xmm6, %%xmm8, %%xmm7", 878 "vpminuw (%%rax), %%xmm8, %%xmm7") 879 880 GEN_test_RandM(VPHMINPOSUW_128, 881 "vphminposuw %%xmm6, %%xmm8", 882 "vphminposuw (%%rax), %%xmm7") 883 884 GEN_test_RandM(VPMAXSW_128, 885 "vpmaxsw %%xmm6, %%xmm8, %%xmm7", 886 "vpmaxsw (%%rax), %%xmm8, %%xmm7") 887 888 GEN_test_RandM(VPMINSW_128, 889 "vpminsw %%xmm6, %%xmm8, %%xmm7", 890 "vpminsw (%%rax), %%xmm8, %%xmm7") 891 892 GEN_test_RandM(VPMAXUB_128, 893 "vpmaxub %%xmm6, %%xmm8, %%xmm7", 894 "vpmaxub (%%rax), %%xmm8, %%xmm7") 895 896 GEN_test_RandM(VPEXTRB_GtoE_128_0x0, 897 "vpextrb $0x0, %%xmm8, %%r14", 898 "vpextrb $0x0, %%xmm8, (%%rax)") 899 900 GEN_test_RandM(VPEXTRB_GtoE_128_0x1, 901 "vpextrb $0x1, %%xmm8, %%r14", 902 "vpextrb $0x1, %%xmm8, (%%rax)") 903 904 GEN_test_RandM(VPEXTRB_GtoE_128_0x2, 905 "vpextrb $0x2, %%xmm8, %%r14", 906 "vpextrb $0x2, %%xmm8, (%%rax)") 907 908 GEN_test_RandM(VPEXTRB_GtoE_128_0x3, 909 "vpextrb $0x3, %%xmm8, %%r14", 910 "vpextrb $0x3, %%xmm8, (%%rax)") 911 912 GEN_test_RandM(VPEXTRB_GtoE_128_0x4, 913 "vpextrb $0x4, %%xmm8, %%r14", 914 "vpextrb $0x4, %%xmm8, (%%rax)") 915 916 GEN_test_RandM(VPEXTRB_GtoE_128_0x9, 917 "vpextrb $0x9, %%xmm8, %%r14", 918 "vpextrb $0x9, %%xmm8, (%%rax)") 919 920 GEN_test_RandM(VPEXTRB_GtoE_128_0xE, 921 "vpextrb $0xE, %%xmm8, %%r14", 922 "vpextrb $0xE, %%xmm8, (%%rax)") 923 924 GEN_test_RandM(VPEXTRB_GtoE_128_0xF, 925 "vpextrb $0xF, %%xmm8, %%r14", 926 "vpextrb $0xF, %%xmm8, (%%rax)") 927 928 GEN_test_RandM(VPMINUB_128, 929 "vpminub %%xmm6, %%xmm8, %%xmm7", 930 "vpminub (%%rax), %%xmm8, %%xmm7") 931 932 GEN_test_RandM(VPMAXSB_128, 933 "vpmaxsb %%xmm6, %%xmm8, %%xmm7", 934 "vpmaxsb (%%rax), %%xmm8, %%xmm7") 935 936 GEN_test_RandM(VPMINSB_128, 937 "vpminsb %%xmm6, %%xmm8, %%xmm7", 938 "vpminsb (%%rax), %%xmm8, %%xmm7") 939 940 GEN_test_RandM(VPERM2F128_0x00, 941 "vperm2f128 $0x00, %%ymm6, %%ymm8, %%ymm7", 942 "vperm2f128 $0x00, (%%rax), %%ymm8, %%ymm7") 943 GEN_test_RandM(VPERM2F128_0xFF, 944 "vperm2f128 $0xFF, %%ymm6, %%ymm8, %%ymm7", 945 "vperm2f128 $0xFF, (%%rax), %%ymm8, %%ymm7") 946 GEN_test_RandM(VPERM2F128_0x30, 947 "vperm2f128 $0x30, %%ymm6, %%ymm8, %%ymm7", 948 "vperm2f128 $0x30, (%%rax), %%ymm8, %%ymm7") 949 GEN_test_RandM(VPERM2F128_0x21, 950 "vperm2f128 $0x21, %%ymm6, %%ymm8, %%ymm7", 951 "vperm2f128 $0x21, (%%rax), %%ymm8, %%ymm7") 952 GEN_test_RandM(VPERM2F128_0x12, 953 "vperm2f128 $0x12, %%ymm6, %%ymm8, %%ymm7", 954 "vperm2f128 $0x12, (%%rax), %%ymm8, %%ymm7") 955 GEN_test_RandM(VPERM2F128_0x03, 956 "vperm2f128 $0x03, %%ymm6, %%ymm8, %%ymm7", 957 "vperm2f128 $0x03, (%%rax), %%ymm8, %%ymm7") 958 GEN_test_RandM(VPERM2F128_0x85, 959 "vperm2f128 $0x85, %%ymm6, %%ymm8, %%ymm7", 960 "vperm2f128 $0x85, (%%rax), %%ymm8, %%ymm7") 961 GEN_test_RandM(VPERM2F128_0x5A, 962 "vperm2f128 $0x5A, %%ymm6, %%ymm8, %%ymm7", 963 "vperm2f128 $0x5A, (%%rax), %%ymm8, %%ymm7") 964 965 GEN_test_RandM(VPERMILPD_256_0x0, 966 "vpermilpd $0x0, %%ymm6, %%ymm8", 967 "vpermilpd $0x1, (%%rax), %%ymm8") 968 GEN_test_RandM(VPERMILPD_256_0xF, 969 "vpermilpd $0xF, %%ymm6, %%ymm8", 970 "vpermilpd $0xE, (%%rax), %%ymm8") 971 GEN_test_RandM(VPERMILPD_256_0xA, 972 "vpermilpd $0xA, %%ymm6, %%ymm8", 973 "vpermilpd $0xB, (%%rax), %%ymm8") 974 GEN_test_RandM(VPERMILPD_256_0x5, 975 "vpermilpd $0x5, %%ymm6, %%ymm8", 976 "vpermilpd $0x4, (%%rax), %%ymm8") 977 978 GEN_test_RandM(VPERMILPD_128_0x0, 979 "vpermilpd $0x0, %%xmm6, %%xmm8", 980 "vpermilpd $0x1, (%%rax), %%xmm8") 981 GEN_test_RandM(VPERMILPD_128_0x3, 982 "vpermilpd $0x3, %%xmm6, %%xmm8", 983 "vpermilpd $0x2, (%%rax), %%xmm8") 984 985 GEN_test_RandM(VUNPCKLPD_256, 986 "vunpcklpd %%ymm6, %%ymm8, %%ymm7", 987 "vunpcklpd (%%rax), %%ymm8, %%ymm7") 988 989 GEN_test_RandM(VUNPCKHPD_256, 990 "vunpckhpd %%ymm6, %%ymm8, %%ymm7", 991 "vunpckhpd (%%rax), %%ymm8, %%ymm7") 992 993 GEN_test_RandM(VSHUFPS_0x39_256, 994 "vshufps $0x39, %%ymm9, %%ymm8, %%ymm7", 995 "vshufps $0xC6, (%%rax), %%ymm8, %%ymm7") 996 997 GEN_test_RandM(VUNPCKLPS_256, 998 "vunpcklps %%ymm6, %%ymm8, %%ymm7", 999 "vunpcklps (%%rax), %%ymm8, %%ymm7") 1000 1001 GEN_test_RandM(VUNPCKHPS_256, 1002 "vunpckhps %%ymm6, %%ymm8, %%ymm7", 1003 "vunpckhps (%%rax), %%ymm8, %%ymm7") 1004 1005 GEN_test_RandM(VXORPD_256, 1006 "vxorpd %%ymm6, %%ymm8, %%ymm7", 1007 "vxorpd (%%rax), %%ymm8, %%ymm7") 1008 1009 GEN_test_Monly(VBROADCASTSD_256, 1010 "vbroadcastsd (%%rax), %%ymm8") 1011 1012 GEN_test_RandM(VCMPPD_128_0x4, 1013 "vcmppd $4, %%xmm6, %%xmm8, %%xmm7", 1014 "vcmppd $4, (%%rax), %%xmm8, %%xmm7") 1015 1016 GEN_test_RandM(VCMPPD_256_0x4, 1017 "vcmppd $4, %%ymm6, %%ymm8, %%ymm7", 1018 "vcmppd $4, (%%rax), %%ymm8, %%ymm7") 1019 1020 GEN_test_RandM(VCMPPS_128_0x4, 1021 "vcmpps $4, %%xmm6, %%xmm8, %%xmm7", 1022 "vcmpps $4, (%%rax), %%xmm8, %%xmm7") 1023 1024 GEN_test_RandM(VCMPPS_256_0x4, 1025 "vcmpps $4, %%ymm6, %%ymm8, %%ymm7", 1026 "vcmpps $4, (%%rax), %%ymm8, %%ymm7") 1027 1028 GEN_test_RandM(VCVTDQ2PD_128, 1029 "vcvtdq2pd %%xmm6, %%xmm8", 1030 "vcvtdq2pd (%%rax), %%xmm8") 1031 1032 GEN_test_RandM(VDIVPD_128, 1033 "vdivpd %%xmm6, %%xmm8, %%xmm7", 1034 "vdivpd (%%rax), %%xmm8, %%xmm7") 1035 1036 GEN_test_RandM(VANDPD_256, 1037 "vandpd %%ymm6, %%ymm8, %%ymm7", 1038 "vandpd (%%rax), %%ymm8, %%ymm7") 1039 1040 GEN_test_RandM(VPMOVSXBW_128, 1041 "vpmovsxbw %%xmm6, %%xmm8", 1042 "vpmovsxbw (%%rax), %%xmm8") 1043 1044 GEN_test_RandM(VPSUBUSW_128, 1045 "vpsubusw %%xmm9, %%xmm8, %%xmm7", 1046 "vpsubusw (%%rax), %%xmm8, %%xmm7") 1047 1048 GEN_test_RandM(VPSUBSW_128, 1049 "vpsubsw %%xmm9, %%xmm8, %%xmm7", 1050 "vpsubsw (%%rax), %%xmm8, %%xmm7") 1051 1052 GEN_test_RandM(VPCMPEQW_128, 1053 "vpcmpeqw %%xmm6, %%xmm8, %%xmm7", 1054 "vpcmpeqw (%%rax), %%xmm8, %%xmm7") 1055 1056 GEN_test_RandM(VPADDB_128, 1057 "vpaddb %%xmm6, %%xmm8, %%xmm7", 1058 "vpaddb (%%rax), %%xmm8, %%xmm7") 1059 1060 GEN_test_RandM(VMOVAPS_EtoG_256, 1061 "vmovaps %%ymm6, %%ymm8", 1062 "vmovaps (%%rax), %%ymm9") 1063 1064 GEN_test_RandM(VCVTDQ2PD_256, 1065 "vcvtdq2pd %%xmm6, %%ymm8", 1066 "vcvtdq2pd (%%rax), %%ymm8") 1067 1068 GEN_test_Monly(VMOVHPD_128_LoadForm, 1069 "vmovhpd (%%rax), %%xmm8, %%xmm7") 1070 1071 GEN_test_Monly(VMOVHPS_128_LoadForm, 1072 "vmovhps (%%rax), %%xmm8, %%xmm7") 1073 1074 // The y suffix denotes a 256 -> 128 operation 1075 GEN_test_RandM(VCVTPD2PS_256, 1076 "vcvtpd2psy %%ymm8, %%xmm7", 1077 "vcvtpd2psy (%%rax), %%xmm9") 1078 1079 GEN_test_RandM(VPUNPCKHDQ_128, 1080 "vpunpckhdq %%xmm6, %%xmm8, %%xmm7", 1081 "vpunpckhdq (%%rax), %%xmm8, %%xmm7") 1082 1083 GEN_test_Monly(VBROADCASTSS_128, 1084 "vbroadcastss (%%rax), %%xmm8") 1085 1086 GEN_test_RandM(VPMOVSXDQ_128, 1087 "vpmovsxdq %%xmm6, %%xmm8", 1088 "vpmovsxdq (%%rax), %%xmm8") 1089 1090 GEN_test_RandM(VPMOVSXWD_128, 1091 "vpmovsxwd %%xmm6, %%xmm8", 1092 "vpmovsxwd (%%rax), %%xmm8") 1093 1094 GEN_test_RandM(VDIVPS_128, 1095 "vdivps %%xmm9, %%xmm8, %%xmm7", 1096 "vdivps (%%rax), %%xmm8, %%xmm7") 1097 1098 GEN_test_RandM(VANDPS_256, 1099 "vandps %%ymm6, %%ymm8, %%ymm7", 1100 "vandps (%%rax), %%ymm8, %%ymm7") 1101 1102 GEN_test_RandM(VXORPS_256, 1103 "vxorps %%ymm6, %%ymm8, %%ymm7", 1104 "vxorps (%%rax), %%ymm8, %%ymm7") 1105 1106 GEN_test_RandM(VORPS_256, 1107 "vorps %%ymm6, %%ymm8, %%ymm7", 1108 "vorps (%%rax), %%ymm8, %%ymm7") 1109 1110 GEN_test_RandM(VANDNPD_256, 1111 "vandnpd %%ymm6, %%ymm8, %%ymm7", 1112 "vandnpd (%%rax), %%ymm8, %%ymm7") 1113 1114 GEN_test_RandM(VANDNPS_256, 1115 "vandnps %%ymm6, %%ymm8, %%ymm7", 1116 "vandnps (%%rax), %%ymm8, %%ymm7") 1117 1118 GEN_test_RandM(VORPD_256, 1119 "vorpd %%ymm6, %%ymm8, %%ymm7", 1120 "vorpd (%%rax), %%ymm8, %%ymm7") 1121 1122 GEN_test_RandM(VPERMILPS_256_0x0F, 1123 "vpermilps $0x0F, %%ymm6, %%ymm8", 1124 "vpermilps $0x1E, (%%rax), %%ymm8") 1125 GEN_test_RandM(VPERMILPS_256_0xFA, 1126 "vpermilps $0xFA, %%ymm6, %%ymm8", 1127 "vpermilps $0xE5, (%%rax), %%ymm8") 1128 GEN_test_RandM(VPERMILPS_256_0xA3, 1129 "vpermilps $0xA3, %%ymm6, %%ymm8", 1130 "vpermilps $0xB4, (%%rax), %%ymm8") 1131 GEN_test_RandM(VPERMILPS_256_0x5A, 1132 "vpermilps $0x5A, %%ymm6, %%ymm8", 1133 "vpermilps $0x45, (%%rax), %%ymm8") 1134 1135 GEN_test_RandM(VPMULHW_128, 1136 "vpmulhw %%xmm9, %%xmm8, %%xmm7", 1137 "vpmulhw (%%rax), %%xmm8, %%xmm7") 1138 1139 GEN_test_RandM(VPUNPCKHQDQ_128, 1140 "vpunpckhqdq %%xmm6, %%xmm8, %%xmm7", 1141 "vpunpckhqdq (%%rax), %%xmm8, %%xmm7") 1142 1143 GEN_test_Ronly(VPSRAW_0x05_128, 1144 "vpsraw $0x5, %%xmm9, %%xmm7") 1145 1146 GEN_test_RandM(VPCMPGTB_128, 1147 "vpcmpgtb %%xmm6, %%xmm8, %%xmm7", 1148 "vpcmpgtb (%%rax), %%xmm8, %%xmm7") 1149 1150 GEN_test_RandM(VPCMPGTW_128, 1151 "vpcmpgtw %%xmm6, %%xmm8, %%xmm7", 1152 "vpcmpgtw (%%rax), %%xmm8, %%xmm7") 1153 1154 GEN_test_RandM(VPCMPGTD_128, 1155 "vpcmpgtd %%xmm6, %%xmm8, %%xmm7", 1156 "vpcmpgtd (%%rax), %%xmm8, %%xmm7") 1157 1158 GEN_test_RandM(VPMOVZXBD_128, 1159 "vpmovzxbd %%xmm6, %%xmm8", 1160 "vpmovzxbd (%%rax), %%xmm8") 1161 1162 GEN_test_RandM(VPMOVSXBD_128, 1163 "vpmovsxbd %%xmm6, %%xmm8", 1164 "vpmovsxbd (%%rax), %%xmm8") 1165 1166 GEN_test_RandM(VPINSRB_128_1of3, 1167 "vpinsrb $0, %%r14d, %%xmm8, %%xmm7", 1168 "vpinsrb $3, (%%rax), %%xmm8, %%xmm7") 1169 GEN_test_RandM(VPINSRB_128_2of3, 1170 "vpinsrb $6, %%r14d, %%xmm8, %%xmm7", 1171 "vpinsrb $9, (%%rax), %%xmm8, %%xmm7") 1172 GEN_test_RandM(VPINSRB_128_3of3, 1173 "vpinsrb $12, %%r14d, %%xmm8, %%xmm7", 1174 "vpinsrb $15, (%%rax), %%xmm8, %%xmm7") 1175 1176 GEN_test_RandM(VPINSRW_128_1of4, 1177 "vpinsrw $0, %%r14d, %%xmm8, %%xmm7", 1178 "vpinsrw $3, (%%rax), %%xmm8, %%xmm7") 1179 GEN_test_RandM(VPINSRW_128_2of4, 1180 "vpinsrw $2, %%r14d, %%xmm8, %%xmm7", 1181 "vpinsrw $3, (%%rax), %%xmm8, %%xmm7") 1182 GEN_test_RandM(VPINSRW_128_3of4, 1183 "vpinsrw $4, %%r14d, %%xmm8, %%xmm7", 1184 "vpinsrw $5, (%%rax), %%xmm8, %%xmm7") 1185 GEN_test_RandM(VPINSRW_128_4of4, 1186 "vpinsrw $6, %%r14d, %%xmm8, %%xmm7", 1187 "vpinsrw $7, (%%rax), %%xmm8, %%xmm7") 1188 1189 GEN_test_RandM(VCOMISD_128, 1190 "vcomisd %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 1191 "vcomisd (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 1192 1193 GEN_test_RandM(VCOMISS_128, 1194 "vcomiss %%xmm6, %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14", 1195 "vcomiss (%%rax), %%xmm8; pushfq; popq %%r14; andq $0x8D5, %%r14") 1196 1197 GEN_test_RandM(VMOVUPS_YMM_to_YMMorMEM, 1198 "vmovups %%ymm8, %%ymm7", 1199 "vmovups %%ymm9, (%%rax)") 1200 1201 GEN_test_RandM(VDPPD_128_1of4, 1202 "vdppd $0x00, %%xmm6, %%xmm8, %%xmm7", 1203 "vdppd $0xA5, (%%rax), %%xmm9, %%xmm6") 1204 GEN_test_RandM(VDPPD_128_2of4, 1205 "vdppd $0x5A, %%xmm6, %%xmm8, %%xmm7", 1206 "vdppd $0xFF, (%%rax), %%xmm9, %%xmm6") 1207 GEN_test_RandM(VDPPD_128_3of4, 1208 "vdppd $0x0F, %%xmm6, %%xmm8, %%xmm7", 1209 "vdppd $0x37, (%%rax), %%xmm9, %%xmm6") 1210 GEN_test_RandM(VDPPD_128_4of4, 1211 "vdppd $0xF0, %%xmm6, %%xmm8, %%xmm7", 1212 "vdppd $0x73, (%%rax), %%xmm9, %%xmm6") 1213 1214 GEN_test_RandM(VDPPS_128_1of4, 1215 "vdpps $0x00, %%xmm6, %%xmm8, %%xmm7", 1216 "vdpps $0xA5, (%%rax), %%xmm9, %%xmm6") 1217 GEN_test_RandM(VDPPS_128_2of4, 1218 "vdpps $0x5A, %%xmm6, %%xmm8, %%xmm7", 1219 "vdpps $0xFF, (%%rax), %%xmm9, %%xmm6") 1220 GEN_test_RandM(VDPPS_128_3of4, 1221 "vdpps $0x0F, %%xmm6, %%xmm8, %%xmm7", 1222 "vdpps $0x37, (%%rax), %%xmm9, %%xmm6") 1223 GEN_test_RandM(VDPPS_128_4of4, 1224 "vdpps $0xF0, %%xmm6, %%xmm8, %%xmm7", 1225 "vdpps $0x73, (%%rax), %%xmm9, %%xmm6") 1226 1227 GEN_test_RandM(VDPPS_256_1of4, 1228 "vdpps $0x00, %%ymm6, %%ymm8, %%ymm7", 1229 "vdpps $0xA5, (%%rax), %%ymm9, %%ymm6") 1230 GEN_test_RandM(VDPPS_256_2of4, 1231 "vdpps $0x5A, %%ymm6, %%ymm8, %%ymm7", 1232 "vdpps $0xFF, (%%rax), %%ymm9, %%ymm6") 1233 GEN_test_RandM(VDPPS_256_3of4, 1234 "vdpps $0x0F, %%ymm6, %%ymm8, %%ymm7", 1235 "vdpps $0x37, (%%rax), %%ymm9, %%ymm6") 1236 GEN_test_RandM(VDPPS_256_4of4, 1237 "vdpps $0xF0, %%ymm6, %%ymm8, %%ymm7", 1238 "vdpps $0x73, (%%rax), %%ymm9, %%ymm6") 1239 1240 GEN_test_Monly(VBROADCASTSS_256, 1241 "vbroadcastss (%%rax), %%ymm8") 1242 1243 GEN_test_RandM(VPALIGNR_128_1of3, 1244 "vpalignr $0, %%xmm6, %%xmm8, %%xmm7", 1245 "vpalignr $3, (%%rax), %%xmm8, %%xmm7") 1246 GEN_test_RandM(VPALIGNR_128_2of3, 1247 "vpalignr $6, %%xmm6, %%xmm8, %%xmm7", 1248 "vpalignr $9, (%%rax), %%xmm8, %%xmm7") 1249 GEN_test_RandM(VPALIGNR_128_3of3, 1250 "vpalignr $12, %%xmm6, %%xmm8, %%xmm7", 1251 "vpalignr $15, (%%rax), %%xmm8, %%xmm7") 1252 1253 GEN_test_Ronly(VMOVSD_REG_XMM, "vmovsd %%xmm9, %%xmm7, %%xmm8") 1254 1255 GEN_test_Ronly(VMOVSS_REG_XMM, "vmovss %%xmm9, %%xmm7, %%xmm8") 1256 1257 GEN_test_Monly(VMOVLPD_128_M64_XMM_XMM, "vmovlpd (%%rax), %%xmm8, %%xmm7") 1258 1259 GEN_test_Monly(VMOVLPD_128_XMM_M64, "vmovlpd %%xmm7, (%%rax)") 1260 1261 GEN_test_RandM(VSHUFPD_128_1of2, 1262 "vshufpd $0, %%xmm9, %%xmm8, %%xmm7", 1263 "vshufpd $1, (%%rax), %%xmm8, %%xmm7") 1264 GEN_test_RandM(VSHUFPD_128_2of2, 1265 "vshufpd $2, %%xmm9, %%xmm8, %%xmm7", 1266 "vshufpd $3, (%%rax), %%xmm8, %%xmm7") 1267 1268 GEN_test_RandM(VSHUFPD_256_1of2, 1269 "vshufpd $0x00, %%ymm9, %%ymm8, %%ymm7", 1270 "vshufpd $0xFF, (%%rax), %%ymm8, %%ymm7") 1271 GEN_test_RandM(VSHUFPD_256_2of2, 1272 "vshufpd $0x5A, %%ymm9, %%ymm8, %%ymm7", 1273 "vshufpd $0xA5, (%%rax), %%ymm8, %%ymm7") 1274 1275 GEN_test_RandM(VPERMILPS_128_0x00, 1276 "vpermilps $0x00, %%xmm6, %%xmm8", 1277 "vpermilps $0x01, (%%rax), %%xmm8") 1278 GEN_test_RandM(VPERMILPS_128_0xFE, 1279 "vpermilps $0xFE, %%xmm6, %%xmm8", 1280 "vpermilps $0xFF, (%%rax), %%xmm8") 1281 GEN_test_RandM(VPERMILPS_128_0x30, 1282 "vpermilps $0x30, %%xmm6, %%xmm8", 1283 "vpermilps $0x03, (%%rax), %%xmm8") 1284 GEN_test_RandM(VPERMILPS_128_0x21, 1285 "vpermilps $0x21, %%xmm6, %%xmm8", 1286 "vpermilps $0x12, (%%rax), %%xmm8") 1287 GEN_test_RandM(VPERMILPS_128_0xD7, 1288 "vpermilps $0xD7, %%xmm6, %%xmm8", 1289 "vpermilps $0x6C, (%%rax), %%xmm8") 1290 GEN_test_RandM(VPERMILPS_128_0xB5, 1291 "vpermilps $0xB5, %%xmm6, %%xmm8", 1292 "vpermilps $0x4A, (%%rax), %%xmm8") 1293 GEN_test_RandM(VPERMILPS_128_0x85, 1294 "vpermilps $0x85, %%xmm6, %%xmm8", 1295 "vpermilps $0xDC, (%%rax), %%xmm8") 1296 GEN_test_RandM(VPERMILPS_128_0x29, 1297 "vpermilps $0x29, %%xmm6, %%xmm8", 1298 "vpermilps $0x92, (%%rax), %%xmm8") 1299 1300 GEN_test_RandM(VBLENDPS_128_1of3, 1301 "vblendps $0, %%xmm6, %%xmm8, %%xmm7", 1302 "vblendps $3, (%%rax), %%xmm8, %%xmm7") 1303 GEN_test_RandM(VBLENDPS_128_2of3, 1304 "vblendps $6, %%xmm6, %%xmm8, %%xmm7", 1305 "vblendps $9, (%%rax), %%xmm8, %%xmm7") 1306 GEN_test_RandM(VBLENDPS_128_3of3, 1307 "vblendps $12, %%xmm6, %%xmm8, %%xmm7", 1308 "vblendps $15, (%%rax), %%xmm8, %%xmm7") 1309 1310 GEN_test_RandM(VBLENDPD_128_1of2, 1311 "vblendpd $0, %%xmm6, %%xmm8, %%xmm7", 1312 "vblendpd $1, (%%rax), %%xmm8, %%xmm7") 1313 GEN_test_RandM(VBLENDPD_128_2of2, 1314 "vblendpd $2, %%xmm6, %%xmm8, %%xmm7", 1315 "vblendpd $3, (%%rax), %%xmm8, %%xmm7") 1316 1317 GEN_test_RandM(VBLENDPD_256_1of3, 1318 "vblendpd $0, %%ymm6, %%ymm8, %%ymm7", 1319 "vblendpd $3, (%%rax), %%ymm8, %%ymm7") 1320 GEN_test_RandM(VBLENDPD_256_2of3, 1321 "vblendpd $6, %%ymm6, %%ymm8, %%ymm7", 1322 "vblendpd $9, (%%rax), %%ymm8, %%ymm7") 1323 GEN_test_RandM(VBLENDPD_256_3of3, 1324 "vblendpd $12, %%ymm6, %%ymm8, %%ymm7", 1325 "vblendpd $15, (%%rax), %%ymm8, %%ymm7") 1326 1327 GEN_test_RandM(VPBLENDW_128_0x00, 1328 "vpblendw $0x00, %%xmm6, %%xmm8, %%xmm7", 1329 "vpblendw $0x01, (%%rax), %%xmm8, %%xmm7") 1330 GEN_test_RandM(VPBLENDW_128_0xFE, 1331 "vpblendw $0xFE, %%xmm6, %%xmm8, %%xmm7", 1332 "vpblendw $0xFF, (%%rax), %%xmm8, %%xmm7") 1333 GEN_test_RandM(VPBLENDW_128_0x30, 1334 "vpblendw $0x30, %%xmm6, %%xmm8, %%xmm7", 1335 "vpblendw $0x03, (%%rax), %%xmm8, %%xmm7") 1336 GEN_test_RandM(VPBLENDW_128_0x21, 1337 "vpblendw $0x21, %%xmm6, %%xmm8, %%xmm7", 1338 "vpblendw $0x12, (%%rax), %%xmm8, %%xmm7") 1339 GEN_test_RandM(VPBLENDW_128_0xD7, 1340 "vpblendw $0xD7, %%xmm6, %%xmm8, %%xmm7", 1341 "vpblendw $0x6C, (%%rax), %%xmm8, %%xmm7") 1342 GEN_test_RandM(VPBLENDW_128_0xB5, 1343 "vpblendw $0xB5, %%xmm6, %%xmm8, %%xmm7", 1344 "vpblendw $0x4A, (%%rax), %%xmm8, %%xmm7") 1345 GEN_test_RandM(VPBLENDW_128_0x85, 1346 "vpblendw $0x85, %%xmm6, %%xmm8, %%xmm7", 1347 "vpblendw $0xDC, (%%rax), %%xmm8, %%xmm7") 1348 GEN_test_RandM(VPBLENDW_128_0x29, 1349 "vpblendw $0x29, %%xmm6, %%xmm8, %%xmm7", 1350 "vpblendw $0x92, (%%rax), %%xmm8, %%xmm7") 1351 1352 GEN_test_RandM(VMOVUPS_EtoG_256, 1353 "vmovups %%ymm6, %%ymm9", 1354 "vmovups (%%rax), %%ymm7") 1355 1356 GEN_test_RandM(VSQRTSS_128, 1357 "vsqrtss %%xmm6, %%xmm8, %%xmm7", 1358 "vsqrtss (%%rax), %%xmm8, %%xmm7") 1359 1360 GEN_test_RandM(VSQRTPS_128, 1361 "vsqrtps %%xmm6, %%xmm8", 1362 "vsqrtps (%%rax), %%xmm8") 1363 1364 GEN_test_RandM(VSQRTPS_256, 1365 "vsqrtps %%ymm6, %%ymm8", 1366 "vsqrtps (%%rax), %%ymm8") 1367 1368 GEN_test_RandM(VSQRTPD_128, 1369 "vsqrtpd %%xmm6, %%xmm8", 1370 "vsqrtpd (%%rax), %%xmm8") 1371 1372 GEN_test_RandM(VSQRTPD_256, 1373 "vsqrtpd %%ymm6, %%ymm8", 1374 "vsqrtpd (%%rax), %%ymm8") 1375 1376 GEN_test_RandM(VRSQRTSS_128, 1377 "vrsqrtss %%xmm6, %%xmm8, %%xmm7", 1378 "vrsqrtss (%%rax), %%xmm8, %%xmm7") 1379 1380 GEN_test_RandM(VRSQRTPS_128, 1381 "vrsqrtps %%xmm6, %%xmm8", 1382 "vrsqrtps (%%rax), %%xmm8") 1383 1384 GEN_test_RandM(VRSQRTPS_256, 1385 "vrsqrtps %%ymm6, %%ymm8", 1386 "vrsqrtps (%%rax), %%ymm8") 1387 1388 GEN_test_RandM(VMOVDQU_GtoE_256, 1389 "vmovdqu %%ymm9, %%ymm6", 1390 "vmovdqu %%ymm7, (%%rax)") 1391 1392 GEN_test_RandM(VCVTPS2PD_256, 1393 "vcvtps2pd %%xmm9, %%ymm6", 1394 "vcvtps2pd (%%rax), %%ymm7") 1395 1396 GEN_test_RandM(VCVTTPS2DQ_128, 1397 "vcvttps2dq %%xmm9, %%xmm6", 1398 "vcvttps2dq (%%rax), %%xmm7") 1399 1400 GEN_test_RandM(VCVTTPS2DQ_256, 1401 "vcvttps2dq %%ymm9, %%ymm6", 1402 "vcvttps2dq (%%rax), %%ymm7") 1403 1404 GEN_test_RandM(VCVTDQ2PS_128, 1405 "vcvtdq2ps %%xmm9, %%xmm6", 1406 "vcvtdq2ps (%%rax), %%xmm7") 1407 1408 GEN_test_RandM(VCVTDQ2PS_256, 1409 "vcvtdq2ps %%ymm9, %%ymm6", 1410 "vcvtdq2ps (%%rax), %%ymm7") 1411 1412 GEN_test_RandM(VCVTTPD2DQ_128, 1413 "vcvttpd2dqx %%xmm9, %%xmm6", 1414 "vcvttpd2dqx (%%rax), %%xmm7") 1415 1416 GEN_test_RandM(VCVTTPD2DQ_256, 1417 "vcvttpd2dqy %%ymm9, %%xmm6", 1418 "vcvttpd2dqy (%%rax), %%xmm7") 1419 1420 GEN_test_RandM(VCVTPD2DQ_128, 1421 "vcvtpd2dqx %%xmm9, %%xmm6", 1422 "vcvtpd2dqx (%%rax), %%xmm7") 1423 1424 GEN_test_RandM(VCVTPD2DQ_256, 1425 "vcvtpd2dqy %%ymm9, %%xmm6", 1426 "vcvtpd2dqy (%%rax), %%xmm7") 1427 1428 GEN_test_RandM(VMOVSLDUP_128, 1429 "vmovsldup %%xmm9, %%xmm6", 1430 "vmovsldup (%%rax), %%xmm7") 1431 1432 GEN_test_RandM(VMOVSLDUP_256, 1433 "vmovsldup %%ymm9, %%ymm6", 1434 "vmovsldup (%%rax), %%ymm7") 1435 1436 GEN_test_RandM(VMOVSHDUP_128, 1437 "vmovshdup %%xmm9, %%xmm6", 1438 "vmovshdup (%%rax), %%xmm7") 1439 1440 GEN_test_RandM(VMOVSHDUP_256, 1441 "vmovshdup %%ymm9, %%ymm6", 1442 "vmovshdup (%%rax), %%ymm7") 1443 1444 GEN_test_RandM(VPERMILPS_VAR_128, 1445 "vpermilps %%xmm6, %%xmm8, %%xmm7", 1446 "vpermilps (%%rax), %%xmm8, %%xmm7") 1447 1448 GEN_test_RandM(VPERMILPD_VAR_128, 1449 "vpermilpd %%xmm6, %%xmm8, %%xmm7", 1450 "vpermilpd (%%rax), %%xmm8, %%xmm7") 1451 1452 GEN_test_RandM(VPERMILPS_VAR_256, 1453 "vpermilps %%ymm6, %%ymm8, %%ymm7", 1454 "vpermilps (%%rax), %%ymm8, %%ymm7") 1455 1456 GEN_test_RandM(VPERMILPD_VAR_256, 1457 "vpermilpd %%ymm6, %%ymm8, %%ymm7", 1458 "vpermilpd (%%rax), %%ymm8, %%ymm7") 1459 1460 GEN_test_RandM(VPSLLW_128, 1461 "andl $15, %%r14d;" 1462 "vmovd %%r14d, %%xmm6;" 1463 "vpsllw %%xmm6, %%xmm8, %%xmm9", 1464 "andq $15, 128(%%rax);" 1465 "vpsllw 128(%%rax), %%xmm8, %%xmm9") 1466 1467 GEN_test_RandM(VPSRLW_128, 1468 "andl $15, %%r14d;" 1469 "vmovd %%r14d, %%xmm6;" 1470 "vpsrlw %%xmm6, %%xmm8, %%xmm9", 1471 "andq $15, 128(%%rax);" 1472 "vpsrlw 128(%%rax), %%xmm8, %%xmm9") 1473 1474 GEN_test_RandM(VPSRAW_128, 1475 "andl $31, %%r14d;" 1476 "vmovd %%r14d, %%xmm6;" 1477 "vpsraw %%xmm6, %%xmm8, %%xmm9", 1478 "andq $15, 128(%%rax);" 1479 "vpsraw 128(%%rax), %%xmm8, %%xmm9") 1480 1481 GEN_test_RandM(VPSLLD_128, 1482 "andl $31, %%r14d;" 1483 "vmovd %%r14d, %%xmm6;" 1484 "vpslld %%xmm6, %%xmm8, %%xmm9", 1485 "andq $31, 128(%%rax);" 1486 "vpslld 128(%%rax), %%xmm8, %%xmm9") 1487 1488 GEN_test_RandM(VPSRLD_128, 1489 "andl $31, %%r14d;" 1490 "vmovd %%r14d, %%xmm6;" 1491 "vpsrld %%xmm6, %%xmm8, %%xmm9", 1492 "andq $31, 128(%%rax);" 1493 "vpsrld 128(%%rax), %%xmm8, %%xmm9") 1494 1495 GEN_test_RandM(VPSRAD_128, 1496 "andl $31, %%r14d;" 1497 "vmovd %%r14d, %%xmm6;" 1498 "vpsrad %%xmm6, %%xmm8, %%xmm9", 1499 "andq $31, 128(%%rax);" 1500 "vpsrad 128(%%rax), %%xmm8, %%xmm9") 1501 1502 GEN_test_RandM(VPSLLQ_128, 1503 "andl $63, %%r14d;" 1504 "vmovd %%r14d, %%xmm6;" 1505 "vpsllq %%xmm6, %%xmm8, %%xmm9", 1506 "andq $63, 128(%%rax);" 1507 "vpsllq 128(%%rax), %%xmm8, %%xmm9") 1508 1509 GEN_test_RandM(VPSRLQ_128, 1510 "andl $63, %%r14d;" 1511 "vmovd %%r14d, %%xmm6;" 1512 "vpsrlq %%xmm6, %%xmm8, %%xmm9", 1513 "andq $63, 128(%%rax);" 1514 "vpsrlq 128(%%rax), %%xmm8, %%xmm9") 1515 1516 GEN_test_RandM(VROUNDPS_128_0x0, 1517 "vroundps $0x0, %%xmm8, %%xmm9", 1518 "vroundps $0x0, (%%rax), %%xmm9") 1519 GEN_test_RandM(VROUNDPS_128_0x1, 1520 "vroundps $0x1, %%xmm8, %%xmm9", 1521 "vroundps $0x1, (%%rax), %%xmm9") 1522 GEN_test_RandM(VROUNDPS_128_0x2, 1523 "vroundps $0x2, %%xmm8, %%xmm9", 1524 "vroundps $0x2, (%%rax), %%xmm9") 1525 GEN_test_RandM(VROUNDPS_128_0x3, 1526 "vroundps $0x3, %%xmm8, %%xmm9", 1527 "vroundps $0x3, (%%rax), %%xmm9") 1528 GEN_test_RandM(VROUNDPS_128_0x4, 1529 "vroundps $0x4, %%xmm8, %%xmm9", 1530 "vroundps $0x4, (%%rax), %%xmm9") 1531 1532 GEN_test_RandM(VROUNDPS_256_0x0, 1533 "vroundps $0x0, %%ymm8, %%ymm9", 1534 "vroundps $0x0, (%%rax), %%ymm9") 1535 GEN_test_RandM(VROUNDPS_256_0x1, 1536 "vroundps $0x1, %%ymm8, %%ymm9", 1537 "vroundps $0x1, (%%rax), %%ymm9") 1538 GEN_test_RandM(VROUNDPS_256_0x2, 1539 "vroundps $0x2, %%ymm8, %%ymm9", 1540 "vroundps $0x2, (%%rax), %%ymm9") 1541 GEN_test_RandM(VROUNDPS_256_0x3, 1542 "vroundps $0x3, %%ymm8, %%ymm9", 1543 "vroundps $0x3, (%%rax), %%ymm9") 1544 GEN_test_RandM(VROUNDPS_256_0x4, 1545 "vroundps $0x4, %%ymm8, %%ymm9", 1546 "vroundps $0x4, (%%rax), %%ymm9") 1547 1548 GEN_test_RandM(VROUNDPD_128_0x0, 1549 "vroundpd $0x0, %%xmm8, %%xmm9", 1550 "vroundpd $0x0, (%%rax), %%xmm9") 1551 GEN_test_RandM(VROUNDPD_128_0x1, 1552 "vroundpd $0x1, %%xmm8, %%xmm9", 1553 "vroundpd $0x1, (%%rax), %%xmm9") 1554 GEN_test_RandM(VROUNDPD_128_0x2, 1555 "vroundpd $0x2, %%xmm8, %%xmm9", 1556 "vroundpd $0x2, (%%rax), %%xmm9") 1557 GEN_test_RandM(VROUNDPD_128_0x3, 1558 "vroundpd $0x3, %%xmm8, %%xmm9", 1559 "vroundpd $0x3, (%%rax), %%xmm9") 1560 GEN_test_RandM(VROUNDPD_128_0x4, 1561 "vroundpd $0x4, %%xmm8, %%xmm9", 1562 "vroundpd $0x4, (%%rax), %%xmm9") 1563 1564 GEN_test_RandM(VROUNDPD_256_0x0, 1565 "vroundpd $0x0, %%ymm8, %%ymm9", 1566 "vroundpd $0x0, (%%rax), %%ymm9") 1567 GEN_test_RandM(VROUNDPD_256_0x1, 1568 "vroundpd $0x1, %%ymm8, %%ymm9", 1569 "vroundpd $0x1, (%%rax), %%ymm9") 1570 GEN_test_RandM(VROUNDPD_256_0x2, 1571 "vroundpd $0x2, %%ymm8, %%ymm9", 1572 "vroundpd $0x2, (%%rax), %%ymm9") 1573 GEN_test_RandM(VROUNDPD_256_0x3, 1574 "vroundpd $0x3, %%ymm8, %%ymm9", 1575 "vroundpd $0x3, (%%rax), %%ymm9") 1576 GEN_test_RandM(VROUNDPD_256_0x4, 1577 "vroundpd $0x4, %%ymm8, %%ymm9", 1578 "vroundpd $0x4, (%%rax), %%ymm9") 1579 1580 GEN_test_RandM(VPMADDWD_128, 1581 "vpmaddwd %%xmm6, %%xmm8, %%xmm7", 1582 "vpmaddwd (%%rax), %%xmm8, %%xmm7") 1583 1584 GEN_test_RandM(VADDSUBPS_128, 1585 "vaddsubps %%xmm6, %%xmm8, %%xmm7", 1586 "vaddsubps (%%rax), %%xmm8, %%xmm7") 1587 1588 GEN_test_RandM(VADDSUBPS_256, 1589 "vaddsubps %%ymm6, %%ymm8, %%ymm7", 1590 "vaddsubps (%%rax), %%ymm8, %%ymm7") 1591 1592 GEN_test_RandM(VADDSUBPD_128, 1593 "vaddsubpd %%xmm6, %%xmm8, %%xmm7", 1594 "vaddsubpd (%%rax), %%xmm8, %%xmm7") 1595 1596 GEN_test_RandM(VADDSUBPD_256, 1597 "vaddsubpd %%ymm6, %%ymm8, %%ymm7", 1598 "vaddsubpd (%%rax), %%ymm8, %%ymm7") 1599 1600 GEN_test_RandM(VROUNDSS_0x0, 1601 "vroundss $0x0, %%xmm8, %%xmm6, %%xmm9", 1602 "vroundss $0x0, (%%rax), %%xmm6, %%xmm9") 1603 GEN_test_RandM(VROUNDSS_0x1, 1604 "vroundss $0x1, %%xmm8, %%xmm6, %%xmm9", 1605 "vroundss $0x1, (%%rax), %%xmm6, %%xmm9") 1606 GEN_test_RandM(VROUNDSS_0x2, 1607 "vroundss $0x2, %%xmm8, %%xmm6, %%xmm9", 1608 "vroundss $0x2, (%%rax), %%xmm6, %%xmm9") 1609 GEN_test_RandM(VROUNDSS_0x3, 1610 "vroundss $0x3, %%xmm8, %%xmm6, %%xmm9", 1611 "vroundss $0x3, (%%rax), %%xmm6, %%xmm9") 1612 GEN_test_RandM(VROUNDSS_0x4, 1613 "vroundss $0x4, %%xmm8, %%xmm6, %%xmm9", 1614 "vroundss $0x4, (%%rax), %%xmm6, %%xmm9") 1615 GEN_test_RandM(VROUNDSS_0x5, 1616 "vroundss $0x5, %%xmm8, %%xmm6, %%xmm9", 1617 "vroundss $0x5, (%%rax), %%xmm6, %%xmm9") 1618 1619 GEN_test_RandM(VROUNDSD_0x0, 1620 "vroundsd $0x0, %%xmm8, %%xmm6, %%xmm9", 1621 "vroundsd $0x0, (%%rax), %%xmm6, %%xmm9") 1622 GEN_test_RandM(VROUNDSD_0x1, 1623 "vroundsd $0x1, %%xmm8, %%xmm6, %%xmm9", 1624 "vroundsd $0x1, (%%rax), %%xmm6, %%xmm9") 1625 GEN_test_RandM(VROUNDSD_0x2, 1626 "vroundsd $0x2, %%xmm8, %%xmm6, %%xmm9", 1627 "vroundsd $0x2, (%%rax), %%xmm6, %%xmm9") 1628 GEN_test_RandM(VROUNDSD_0x3, 1629 "vroundsd $0x3, %%xmm8, %%xmm6, %%xmm9", 1630 "vroundsd $0x3, (%%rax), %%xmm6, %%xmm9") 1631 GEN_test_RandM(VROUNDSD_0x4, 1632 "vroundsd $0x4, %%xmm8, %%xmm6, %%xmm9", 1633 "vroundsd $0x4, (%%rax), %%xmm6, %%xmm9") 1634 GEN_test_RandM(VROUNDSD_0x5, 1635 "vroundsd $0x5, %%xmm8, %%xmm6, %%xmm9", 1636 "vroundsd $0x5, (%%rax), %%xmm6, %%xmm9") 1637 1638 GEN_test_RandM(VPTEST_128_1, 1639 "vptest %%xmm6, %%xmm8; " 1640 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1641 "vptest (%%rax), %%xmm8; " 1642 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1643 1644 /* Here we ignore the boilerplate-supplied data and try to do 1645 x AND x and x AND NOT x. Not a great test but better 1646 than nothing. */ 1647 GEN_test_RandM(VPTEST_128_2, 1648 "vmovups %%xmm6, %%xmm8;" 1649 "vptest %%xmm6, %%xmm8; " 1650 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1651 "vmovups (%%rax), %%xmm8;" 1652 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1653 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1654 "vptest (%%rax), %%xmm8; " 1655 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1656 1657 GEN_test_RandM(VPTEST_256_1, 1658 "vptest %%ymm6, %%ymm8; " 1659 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1660 "vptest (%%rax), %%ymm8; " 1661 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1662 1663 /* Here we ignore the boilerplate-supplied data and try to do 1664 x AND x and x AND NOT x. Not a great test but better 1665 than nothing. */ 1666 GEN_test_RandM(VPTEST_256_2, 1667 "vmovups %%ymm6, %%ymm8;" 1668 "vptest %%ymm6, %%ymm8; " 1669 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1670 "vmovups (%%rax), %%ymm8;" 1671 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1672 "subq $1024, %%rsp;" 1673 "vmovups %%xmm7,512(%%rsp);" 1674 "vmovups %%xmm7,528(%%rsp);" 1675 "vmovups 512(%%rsp), %%ymm7;" 1676 "addq $1024, %%rsp;" 1677 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1678 "vptest (%%rax), %%ymm8; " 1679 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1680 1681 1682 /* VTESTPS/VTESTPD: test once with all-0 operands, once with 1683 one all-0s and one all 1s, and once with random data. */ 1684 1685 GEN_test_RandM(VTESTPS_128_1, 1686 "vtestps %%xmm6, %%xmm8; " 1687 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1688 "vtestps (%%rax), %%xmm8; " 1689 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1690 1691 /* Here we ignore the boilerplate-supplied data and try to do 1692 x AND x and x AND NOT x. Not a great test but better 1693 than nothing. */ 1694 GEN_test_RandM(VTESTPS_128_2, 1695 "vmovups %%xmm6, %%xmm8;" 1696 "vtestps %%xmm6, %%xmm8; " 1697 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1698 "vmovups (%%rax), %%xmm8;" 1699 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1700 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1701 "vtestps (%%rax), %%xmm8; " 1702 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1703 1704 GEN_test_RandM(VTESTPS_128_3, 1705 "vtestps %%xmm8, %%xmm9; " 1706 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1707 "vtestps (%%rax), %%xmm9; " 1708 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1709 1710 1711 1712 1713 GEN_test_RandM(VTESTPS_256_1, 1714 "vtestps %%ymm6, %%ymm8; " 1715 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1716 "vtestps (%%rax), %%ymm8; " 1717 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1718 1719 /* Here we ignore the boilerplate-supplied data and try to do 1720 x AND x and x AND NOT x. Not a great test but better 1721 than nothing. */ 1722 GEN_test_RandM(VTESTPS_256_2, 1723 "vmovups %%ymm6, %%ymm8;" 1724 "vtestps %%ymm6, %%ymm8; " 1725 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1726 "vmovups (%%rax), %%ymm8;" 1727 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1728 "subq $1024, %%rsp;" 1729 "vmovups %%xmm7,512(%%rsp);" 1730 "vmovups %%xmm7,528(%%rsp);" 1731 "vmovups 512(%%rsp), %%ymm7;" 1732 "addq $1024, %%rsp;" 1733 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1734 "vtestps (%%rax), %%ymm8; " 1735 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1736 1737 GEN_test_RandM(VTESTPS_256_3, 1738 "vtestps %%ymm8, %%ymm9; " 1739 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1740 "vtestps (%%rax), %%ymm9; " 1741 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1742 1743 1744 1745 GEN_test_RandM(VTESTPD_128_1, 1746 "vtestpd %%xmm6, %%xmm8; " 1747 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1748 "vtestpd (%%rax), %%xmm8; " 1749 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1750 1751 /* Here we ignore the boilerplate-supplied data and try to do 1752 x AND x and x AND NOT x. Not a great test but better 1753 than nothing. */ 1754 GEN_test_RandM(VTESTPD_128_2, 1755 "vmovups %%xmm6, %%xmm8;" 1756 "vtestpd %%xmm6, %%xmm8; " 1757 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1758 "vmovups (%%rax), %%xmm8;" 1759 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1760 "vxorpd %%xmm8,%%xmm7,%%xmm8;" 1761 "vtestpd (%%rax), %%xmm8; " 1762 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1763 1764 GEN_test_RandM(VTESTPD_128_3, 1765 "vtestpd %%xmm8, %%xmm9; " 1766 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1767 "vtestpd (%%rax), %%xmm9; " 1768 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1769 1770 1771 1772 1773 GEN_test_RandM(VTESTPD_256_1, 1774 "vtestpd %%ymm6, %%ymm8; " 1775 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1776 "vtestpd (%%rax), %%ymm8; " 1777 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1778 1779 /* Here we ignore the boilerplate-supplied data and try to do 1780 x AND x and x AND NOT x. Not a great test but better 1781 than nothing. */ 1782 GEN_test_RandM(VTESTPD_256_2, 1783 "vmovups %%ymm6, %%ymm8;" 1784 "vtestpd %%ymm6, %%ymm8; " 1785 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1786 "vmovups (%%rax), %%ymm8;" 1787 "vcmpeqpd %%xmm8,%%xmm8,%%xmm7;" 1788 "subq $1024, %%rsp;" 1789 "vmovups %%xmm7,512(%%rsp);" 1790 "vmovups %%xmm7,528(%%rsp);" 1791 "vmovups 512(%%rsp), %%ymm7;" 1792 "addq $1024, %%rsp;" 1793 "vxorpd %%ymm8,%%ymm7,%%ymm8;" 1794 "vtestpd (%%rax), %%ymm8; " 1795 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1796 1797 GEN_test_RandM(VTESTPD_256_3, 1798 "vtestpd %%ymm8, %%ymm9; " 1799 "pushfq; popq %%r14; andq $0x8D5, %%r14", 1800 "vtestpd (%%rax), %%ymm9; " 1801 "pushfq; popq %%r14; andq $0x8D5, %%r14") 1802 1803 GEN_test_RandM(VBLENDVPS_128, 1804 "vblendvps %%xmm9, %%xmm6, %%xmm8, %%xmm7", 1805 "vblendvps %%xmm9, (%%rax), %%xmm8, %%xmm7") 1806 1807 GEN_test_RandM(VBLENDVPS_256, 1808 "vblendvps %%ymm9, %%ymm6, %%ymm8, %%ymm7", 1809 "vblendvps %%ymm9, (%%rax), %%ymm8, %%ymm7") 1810 1811 GEN_test_RandM(VBLENDVPD_128, 1812 "vblendvpd %%xmm9, %%xmm6, %%xmm8, %%xmm7", 1813 "vblendvpd %%xmm9, (%%rax), %%xmm8, %%xmm7") 1814 1815 GEN_test_RandM(VBLENDVPD_256, 1816 "vblendvpd %%ymm9, %%ymm6, %%ymm8, %%ymm7", 1817 "vblendvpd %%ymm9, (%%rax), %%ymm8, %%ymm7") 1818 1819 1820 GEN_test_RandM(VHADDPS_128, 1821 "vhaddps %%xmm6, %%xmm8, %%xmm7", 1822 "vhaddps (%%rax), %%xmm8, %%xmm7") 1823 1824 GEN_test_RandM(VHADDPS_256, 1825 "vhaddps %%ymm6, %%ymm8, %%ymm7", 1826 "vhaddps (%%rax), %%ymm8, %%ymm7") 1827 1828 GEN_test_RandM(VHADDPD_128, 1829 "vhaddpd %%xmm6, %%xmm8, %%xmm7", 1830 "vhaddpd (%%rax), %%xmm8, %%xmm7") 1831 1832 GEN_test_RandM(VHADDPD_256, 1833 "vhaddpd %%ymm6, %%ymm8, %%ymm7", 1834 "vhaddpd (%%rax), %%ymm8, %%ymm7") 1835 1836 GEN_test_RandM(VHSUBPS_128, 1837 "vhsubps %%xmm6, %%xmm8, %%xmm7", 1838 "vhsubps (%%rax), %%xmm8, %%xmm7") 1839 1840 GEN_test_RandM(VHSUBPS_256, 1841 "vhsubps %%ymm6, %%ymm8, %%ymm7", 1842 "vhsubps (%%rax), %%ymm8, %%ymm7") 1843 1844 GEN_test_RandM(VHSUBPD_128, 1845 "vhsubpd %%xmm6, %%xmm8, %%xmm7", 1846 "vhsubpd (%%rax), %%xmm8, %%xmm7") 1847 1848 GEN_test_RandM(VHSUBPD_256, 1849 "vhsubpd %%ymm6, %%ymm8, %%ymm7", 1850 "vhsubpd (%%rax), %%ymm8, %%ymm7") 1851 1852 GEN_test_RandM(VEXTRACTPS_0x0, 1853 "vextractps $0, %%xmm8, %%r14d", 1854 "vextractps $0, %%xmm8, (%%rax)") 1855 1856 GEN_test_RandM(VEXTRACTPS_0x1, 1857 "vextractps $1, %%xmm8, %%r14d", 1858 "vextractps $1, %%xmm8, (%%rax)") 1859 1860 GEN_test_RandM(VEXTRACTPS_0x2, 1861 "vextractps $2, %%xmm8, %%r14d", 1862 "vextractps $2, %%xmm8, (%%rax)") 1863 1864 GEN_test_RandM(VEXTRACTPS_0x3, 1865 "vextractps $3, %%xmm8, %%r14d", 1866 "vextractps $3, %%xmm8, (%%rax)") 1867 1868 GEN_test_Monly(VLDDQU_128, 1869 "vlddqu 1(%%rax), %%xmm8") 1870 1871 GEN_test_Monly(VLDDQU_256, 1872 "vlddqu 1(%%rax), %%ymm8") 1873 1874 GEN_test_Monly(VMOVNTDQA_128, 1875 "vmovntdqa (%%rax), %%xmm9") 1876 1877 GEN_test_Monly(VMASKMOVDQU_128, 1878 "xchgq %%rax, %%rdi;" 1879 "vmaskmovdqu %%xmm8, %%xmm9;" 1880 "xchgq %%rax, %%rdi") 1881 1882 GEN_test_Ronly(VMOVMSKPD_128, 1883 "vmovmskpd %%xmm9, %%r14d") 1884 1885 GEN_test_Ronly(VMOVMSKPD_256, 1886 "vmovmskpd %%ymm9, %%r14d") 1887 1888 GEN_test_Ronly(VMOVMSKPS_128, 1889 "vmovmskps %%xmm9, %%r14d") 1890 1891 GEN_test_Ronly(VMOVMSKPS_256, 1892 "vmovmskps %%ymm9, %%r14d") 1893 1894 GEN_test_Monly(VMOVNTPD_128, 1895 "vmovntpd %%xmm9, (%%rax)") 1896 1897 GEN_test_Monly(VMOVNTPD_256, 1898 "vmovntpd %%ymm9, (%%rax)") 1899 1900 GEN_test_Monly(VMOVNTPS_128, 1901 "vmovntps %%xmm9, (%%rax)") 1902 1903 GEN_test_Monly(VMOVNTPS_256, 1904 "vmovntps %%ymm9, (%%rax)") 1905 1906 GEN_test_RandM(VPACKSSWB_128, 1907 "vpacksswb %%xmm6, %%xmm8, %%xmm7", 1908 "vpacksswb (%%rax), %%xmm8, %%xmm7") 1909 1910 GEN_test_RandM(VPAVGB_128, 1911 "vpavgb %%xmm6, %%xmm8, %%xmm7", 1912 "vpavgb (%%rax), %%xmm8, %%xmm7") 1913 1914 GEN_test_RandM(VPAVGW_128, 1915 "vpavgw %%xmm6, %%xmm8, %%xmm7", 1916 "vpavgw (%%rax), %%xmm8, %%xmm7") 1917 1918 GEN_test_RandM(VPADDSB_128, 1919 "vpaddsb %%xmm6, %%xmm8, %%xmm7", 1920 "vpaddsb (%%rax), %%xmm8, %%xmm7") 1921 1922 GEN_test_RandM(VPADDSW_128, 1923 "vpaddsw %%xmm6, %%xmm8, %%xmm7", 1924 "vpaddsw (%%rax), %%xmm8, %%xmm7") 1925 1926 GEN_test_RandM(VPHADDW_128, 1927 "vphaddw %%xmm6, %%xmm8, %%xmm7", 1928 "vphaddw (%%rax), %%xmm8, %%xmm7") 1929 1930 GEN_test_RandM(VPHADDD_128, 1931 "vphaddd %%xmm6, %%xmm8, %%xmm7", 1932 "vphaddd (%%rax), %%xmm8, %%xmm7") 1933 1934 GEN_test_RandM(VPHADDSW_128, 1935 "vphaddsw %%xmm6, %%xmm8, %%xmm7", 1936 "vphaddsw (%%rax), %%xmm8, %%xmm7") 1937 1938 GEN_test_RandM(VPMADDUBSW_128, 1939 "vpmaddubsw %%xmm6, %%xmm8, %%xmm7", 1940 "vpmaddubsw (%%rax), %%xmm8, %%xmm7") 1941 1942 GEN_test_RandM(VPHSUBW_128, 1943 "vphsubw %%xmm6, %%xmm8, %%xmm7", 1944 "vphsubw (%%rax), %%xmm8, %%xmm7") 1945 1946 GEN_test_RandM(VPHSUBD_128, 1947 "vphsubd %%xmm6, %%xmm8, %%xmm7", 1948 "vphsubd (%%rax), %%xmm8, %%xmm7") 1949 1950 GEN_test_RandM(VPHSUBSW_128, 1951 "vphsubsw %%xmm6, %%xmm8, %%xmm7", 1952 "vphsubsw (%%rax), %%xmm8, %%xmm7") 1953 1954 GEN_test_RandM(VPABSB_128, 1955 "vpabsb %%xmm6, %%xmm7", 1956 "vpabsb (%%rax), %%xmm7") 1957 1958 GEN_test_RandM(VPABSW_128, 1959 "vpabsw %%xmm6, %%xmm7", 1960 "vpabsw (%%rax), %%xmm7") 1961 1962 GEN_test_RandM(VPMOVSXBQ_128, 1963 "vpmovsxbq %%xmm6, %%xmm8", 1964 "vpmovsxbq (%%rax), %%xmm8") 1965 1966 GEN_test_RandM(VPMOVSXWQ_128, 1967 "vpmovsxwq %%xmm6, %%xmm8", 1968 "vpmovsxwq (%%rax), %%xmm8") 1969 1970 GEN_test_RandM(VPACKUSDW_128, 1971 "vpackusdw %%xmm6, %%xmm8, %%xmm7", 1972 "vpackusdw (%%rax), %%xmm8, %%xmm7") 1973 1974 GEN_test_RandM(VPMOVZXBQ_128, 1975 "vpmovzxbq %%xmm6, %%xmm8", 1976 "vpmovzxbq (%%rax), %%xmm8") 1977 1978 GEN_test_RandM(VPMOVZXWQ_128, 1979 "vpmovzxwq %%xmm6, %%xmm8", 1980 "vpmovzxwq (%%rax), %%xmm8") 1981 1982 GEN_test_RandM(VPMOVZXDQ_128, 1983 "vpmovzxdq %%xmm6, %%xmm8", 1984 "vpmovzxdq (%%rax), %%xmm8") 1985 1986 GEN_test_RandM(VMPSADBW_128_0x0, 1987 "vmpsadbw $0, %%xmm6, %%xmm8, %%xmm7", 1988 "vmpsadbw $0, (%%rax), %%xmm8, %%xmm7") 1989 GEN_test_RandM(VMPSADBW_128_0x1, 1990 "vmpsadbw $1, %%xmm6, %%xmm8, %%xmm7", 1991 "vmpsadbw $1, (%%rax), %%xmm8, %%xmm7") 1992 GEN_test_RandM(VMPSADBW_128_0x2, 1993 "vmpsadbw $2, %%xmm6, %%xmm8, %%xmm7", 1994 "vmpsadbw $2, (%%rax), %%xmm8, %%xmm7") 1995 GEN_test_RandM(VMPSADBW_128_0x3, 1996 "vmpsadbw $3, %%xmm6, %%xmm8, %%xmm7", 1997 "vmpsadbw $3, (%%rax), %%xmm8, %%xmm7") 1998 GEN_test_RandM(VMPSADBW_128_0x4, 1999 "vmpsadbw $4, %%xmm6, %%xmm8, %%xmm7", 2000 "vmpsadbw $4, (%%rax), %%xmm8, %%xmm7") 2001 GEN_test_RandM(VMPSADBW_128_0x5, 2002 "vmpsadbw $5, %%xmm6, %%xmm8, %%xmm7", 2003 "vmpsadbw $5, (%%rax), %%xmm8, %%xmm7") 2004 GEN_test_RandM(VMPSADBW_128_0x6, 2005 "vmpsadbw $6, %%xmm6, %%xmm8, %%xmm7", 2006 "vmpsadbw $6, (%%rax), %%xmm8, %%xmm7") 2007 GEN_test_RandM(VMPSADBW_128_0x7, 2008 "vmpsadbw $7, %%xmm6, %%xmm8, %%xmm7", 2009 "vmpsadbw $7, (%%rax), %%xmm8, %%xmm7") 2010 2011 GEN_test_RandM(VMOVDDUP_YMMorMEM256_to_YMM, 2012 "vmovddup %%ymm8, %%ymm7", 2013 "vmovddup (%%rax), %%ymm9") 2014 2015 GEN_test_Monly(VMOVLPS_128_M64_XMM_XMM, "vmovlps (%%rax), %%xmm8, %%xmm7") 2016 2017 GEN_test_Monly(VMOVLPS_128_XMM_M64, "vmovlps %%xmm7, (%%rax)") 2018 2019 GEN_test_RandM(VRCPSS_128, 2020 "vrcpss %%xmm6, %%xmm8, %%xmm7", 2021 "vrcpss (%%rax), %%xmm8, %%xmm7") 2022 2023 GEN_test_RandM(VRCPPS_128, 2024 "vrcpps %%xmm6, %%xmm8", 2025 "vrcpps (%%rax), %%xmm8") 2026 2027 GEN_test_RandM(VRCPPS_256, 2028 "vrcpps %%ymm6, %%ymm8", 2029 "vrcpps (%%rax), %%ymm8") 2030 2031 GEN_test_RandM(VPSADBW_128, 2032 "vpsadbw %%xmm6, %%xmm8, %%xmm7", 2033 "vpsadbw (%%rax), %%xmm8, %%xmm7") 2034 2035 GEN_test_RandM(VPSIGNB_128, 2036 "vpsignb %%xmm6, %%xmm8, %%xmm7", 2037 "vpsignb (%%rax), %%xmm8, %%xmm7") 2038 2039 GEN_test_RandM(VPSIGNW_128, 2040 "vpsignw %%xmm6, %%xmm8, %%xmm7", 2041 "vpsignw (%%rax), %%xmm8, %%xmm7") 2042 2043 GEN_test_RandM(VPSIGND_128, 2044 "vpsignd %%xmm6, %%xmm8, %%xmm7", 2045 "vpsignd (%%rax), %%xmm8, %%xmm7") 2046 2047 GEN_test_RandM(VPMULHRSW_128, 2048 "vpmulhrsw %%xmm6, %%xmm8, %%xmm7", 2049 "vpmulhrsw (%%rax), %%xmm8, %%xmm7") 2050 2051 GEN_test_Monly(VBROADCASTF128, 2052 "vbroadcastf128 (%%rax), %%ymm9") 2053 2054 GEN_test_RandM(VPEXTRW_128_0x0, 2055 "vpextrw $0x0, %%xmm7, %%r14d", 2056 "vpextrw $0x0, %%xmm7, (%%rax)") 2057 GEN_test_RandM(VPEXTRW_128_0x1, 2058 "vpextrw $0x1, %%xmm7, %%r14d", 2059 "vpextrw $0x1, %%xmm7, (%%rax)") 2060 GEN_test_RandM(VPEXTRW_128_0x2, 2061 "vpextrw $0x2, %%xmm7, %%r14d", 2062 "vpextrw $0x2, %%xmm7, (%%rax)") 2063 GEN_test_RandM(VPEXTRW_128_0x3, 2064 "vpextrw $0x3, %%xmm7, %%r14d", 2065 "vpextrw $0x3, %%xmm7, (%%rax)") 2066 GEN_test_RandM(VPEXTRW_128_0x4, 2067 "vpextrw $0x4, %%xmm7, %%r14d", 2068 "vpextrw $0x4, %%xmm7, (%%rax)") 2069 GEN_test_RandM(VPEXTRW_128_0x5, 2070 "vpextrw $0x5, %%xmm7, %%r14d", 2071 "vpextrw $0x5, %%xmm7, (%%rax)") 2072 GEN_test_RandM(VPEXTRW_128_0x6, 2073 "vpextrw $0x6, %%xmm7, %%r14d", 2074 "vpextrw $0x6, %%xmm7, (%%rax)") 2075 GEN_test_RandM(VPEXTRW_128_0x7, 2076 "vpextrw $0x7, %%xmm7, %%r14d", 2077 "vpextrw $0x7, %%xmm7, (%%rax)") 2078 2079 GEN_test_RandM(VAESENC, 2080 "vaesenc %%xmm6, %%xmm8, %%xmm7", 2081 "vaesenc (%%rax), %%xmm8, %%xmm7") 2082 2083 GEN_test_RandM(VAESENCLAST, 2084 "vaesenclast %%xmm6, %%xmm8, %%xmm7", 2085 "vaesenclast (%%rax), %%xmm8, %%xmm7") 2086 2087 GEN_test_RandM(VAESDEC, 2088 "vaesdec %%xmm6, %%xmm8, %%xmm7", 2089 "vaesdec (%%rax), %%xmm8, %%xmm7") 2090 2091 GEN_test_RandM(VAESDECLAST, 2092 "vaesdeclast %%xmm6, %%xmm8, %%xmm7", 2093 "vaesdeclast (%%rax), %%xmm8, %%xmm7") 2094 2095 GEN_test_RandM(VAESIMC, 2096 "vaesimc %%xmm6, %%xmm7", 2097 "vaesimc (%%rax), %%xmm7") 2098 2099 GEN_test_RandM(VAESKEYGENASSIST_0x00, 2100 "vaeskeygenassist $0x00, %%xmm6, %%xmm7", 2101 "vaeskeygenassist $0x00, (%%rax), %%xmm7") 2102 GEN_test_RandM(VAESKEYGENASSIST_0x31, 2103 "vaeskeygenassist $0x31, %%xmm6, %%xmm7", 2104 "vaeskeygenassist $0x31, (%%rax), %%xmm7") 2105 GEN_test_RandM(VAESKEYGENASSIST_0xB2, 2106 "vaeskeygenassist $0xb2, %%xmm6, %%xmm7", 2107 "vaeskeygenassist $0xb2, (%%rax), %%xmm7") 2108 GEN_test_RandM(VAESKEYGENASSIST_0xFF, 2109 "vaeskeygenassist $0xFF, %%xmm6, %%xmm7", 2110 "vaeskeygenassist $0xFF, (%%rax), %%xmm7") 2111 2112 GEN_test_RandM(VPCLMULQDQ_0x00, 2113 "vpclmulqdq $0x00, %%xmm6, %%xmm8, %%xmm7", 2114 "vpclmulqdq $0x00, (%%rax), %%xmm8, %%xmm7") 2115 GEN_test_RandM(VPCLMULQDQ_0x01, 2116 "vpclmulqdq $0x01, %%xmm6, %%xmm8, %%xmm7", 2117 "vpclmulqdq $0x01, (%%rax), %%xmm8, %%xmm7") 2118 GEN_test_RandM(VPCLMULQDQ_0x10, 2119 "vpclmulqdq $0x10, %%xmm6, %%xmm8, %%xmm7", 2120 "vpclmulqdq $0x10, (%%rax), %%xmm8, %%xmm7") 2121 GEN_test_RandM(VPCLMULQDQ_0x11, 2122 "vpclmulqdq $0x11, %%xmm6, %%xmm8, %%xmm7", 2123 "vpclmulqdq $0x11, (%%rax), %%xmm8, %%xmm7") 2124 GEN_test_RandM(VPCLMULQDQ_0xFF, 2125 "vpclmulqdq $0xFF, %%xmm6, %%xmm8, %%xmm7", 2126 "vpclmulqdq $0xFF, (%%rax), %%xmm8, %%xmm7") 2127 2128 GEN_test_RandM(VCMPSS_128_0x9, 2129 "vcmpss $0x9, %%xmm6, %%xmm8, %%xmm7", 2130 "vcmpss $0x9, (%%rax), %%xmm8, %%xmm7") 2131 2132 /* Comment duplicated above, for convenient reference: 2133 Allowed operands in test insns: 2134 Reg form: %ymm6, %ymm7, %ymm8, %ymm9 and %r14. 2135 Mem form: (%rax), %ymm7, %ymm8, %ymm9 and %r14. 2136 Imm8 etc fields are also allowed, where they make sense. 2137 */ 2138 2139 #define N_DEFAULT_ITERS 3 2140 2141 // Do the specified test some number of times 2142 #define DO_N(_iters, _testfn) \ 2143 do { int i; for (i = 0; i < (_iters); i++) { test_##_testfn(); } } while (0) 2144 2145 // Do the specified test the default number of times 2146 #define DO_D(_testfn) DO_N(N_DEFAULT_ITERS, _testfn) 2147 2148 2149 int main ( void ) 2150 { 2151 DO_D( VMOVUPD_EtoG_256 ); 2152 DO_D( VMOVUPD_GtoE_256 ); 2153 DO_D( VPSUBW_128 ); 2154 DO_D( VPSUBQ_128 ); 2155 DO_D( VPADDQ_128 ); 2156 DO_D( VPINSRQ_128 ); 2157 DO_D( VUCOMISS_128 ); 2158 DO_D( VUCOMISD_128 ); 2159 DO_D( VCVTPS2PD_128 ); 2160 DO_D( VANDNPD_128 ); 2161 DO_D( VORPD_128 ); 2162 DO_D( VXORPD_128 ); 2163 DO_D( VXORPS_128 ); 2164 DO_D( VMULSD_128 ); 2165 DO_D( VADDSD_128 ); 2166 DO_D( VMINSD_128 ); 2167 DO_D( VSUBSD_128 ); 2168 DO_D( VDIVSD_128 ); 2169 DO_D( VMAXSD_128 ); 2170 DO_D( VPSHUFD_0x39_128 ); 2171 DO_D( VPCMPEQD_128 ); 2172 DO_D( VPEXTRD_128_0x3 ); 2173 DO_D( VPEXTRD_128_0x0 ); 2174 DO_D( VINSERTF128_0x0 ); 2175 DO_D( VINSERTF128_0x1 ); 2176 DO_D( VEXTRACTF128_0x0 ); 2177 DO_D( VEXTRACTF128_0x1 ); 2178 DO_D( VCVTPD2PS_128 ); 2179 /* Test all CMPSS variants; this code is tricky. */ 2180 DO_D( VCMPSS_128_0x0 ); 2181 DO_D( VCMPSS_128_0x1 ); 2182 DO_D( VCMPSS_128_0x2 ); 2183 DO_D( VCMPSS_128_0x3 ); 2184 DO_D( VCMPSS_128_0x4 ); 2185 DO_D( VCMPSS_128_0x5 ); 2186 DO_D( VCMPSS_128_0x6 ); 2187 DO_D( VCMPSS_128_0x7 ); 2188 DO_D( VCMPSS_128_0xA ); 2189 DO_D( VCMPSS_128_0xC ); 2190 DO_D( VCMPSS_128_0xC ); 2191 DO_D( VCMPSS_128_0xD ); 2192 DO_D( VCMPSS_128_0xE ); 2193 DO_D( VCMPSS_128_0x11 ); 2194 DO_D( VCMPSS_128_0x12); 2195 DO_D( VCMPSS_128_0x16 ); 2196 DO_D( VCMPSS_128_0x1E ); 2197 DO_D( VMOVDDUP_XMMorMEM64_to_XMM ); 2198 DO_D( VMOVD_IREGorMEM32_to_XMM ); 2199 DO_D( VMOVQ_XMM_MEM64 ); 2200 DO_D( VMOVDQA_GtoE_256 ); 2201 DO_D( VMOVDQA_GtoE_128 ); 2202 DO_D( VMOVDQU_GtoE_128 ); 2203 DO_D( VMOVDQA_EtoG_256 ); 2204 DO_D( VMOVDQA_EtoG_128 ); 2205 DO_D( VMOVDQU_EtoG_128 ); 2206 DO_D( VMOVAPD_GtoE_128 ); 2207 DO_D( VMOVAPD_GtoE_256 ); 2208 DO_D( VMOVAPS_GtoE_128 ); 2209 DO_D( VMOVAPS_GtoE_256 ); 2210 DO_D( VMOVAPS_EtoG_128 ); 2211 DO_D( VMOVAPD_EtoG_256 ); 2212 DO_D( VMOVAPD_EtoG_128 ); 2213 DO_D( VMOVUPD_GtoE_128 ); 2214 DO_D( VMOVSS_XMM_M32 ); 2215 DO_D( VMOVSD_XMM_M64 ); 2216 DO_D( VMOVSS_M64_XMM ); 2217 DO_D( VMOVSD_M64_XMM ); 2218 DO_D( VINSERTPS_0x39_128 ); 2219 DO_D( VPUNPCKLDQ_128 ); 2220 DO_D( VPACKSSDW_128 ); 2221 DO_D( VPADDW_128 ); 2222 DO_D( VPSRLW_0x05_128 ); 2223 DO_D( VPSLLW_0x05_128 ); 2224 DO_D( VPUNPCKLQDQ_128 ); 2225 DO_D( VPINSRD_128 ); 2226 DO_D( VMOVD_XMM_to_MEM32 ); 2227 DO_D( VPANDN_128 ); 2228 DO_D( VPSLLDQ_0x05_128 ); 2229 DO_D( VPSRLDQ_0x05_128 ); 2230 DO_D( VPSUBUSB_128 ); 2231 DO_D( VPSUBSB_128 ); 2232 DO_D( VPSLLD_0x05_128 ); 2233 DO_D( VPSRLD_0x05_128 ); 2234 DO_D( VPSRAD_0x05_128 ); 2235 DO_D( VPUNPCKLWD_128 ); 2236 DO_D( VPUNPCKHWD_128 ); 2237 DO_D( VPADDUSB_128 ); 2238 DO_D( VPMULHUW_128 ); 2239 DO_D( VPADDUSW_128 ); 2240 DO_D( VPMULLW_128 ); 2241 DO_D( VPSHUFHW_0x39_128 ); 2242 DO_D( VPSHUFLW_0x39_128 ); 2243 DO_D( VCVTPS2DQ_128 ); 2244 DO_D( VSUBPS_128 ); 2245 DO_D( VADDPS_128 ); 2246 DO_D( VMULPS_128 ); 2247 DO_D( VMAXPS_128 ); 2248 DO_D( VMINPS_128 ); 2249 DO_D( VSHUFPS_0x39_128 ); 2250 DO_D( VPCMPEQB_128 ); 2251 DO_D( VMOVHPD_128_StoreForm ); 2252 DO_D( VPAND_128 ); 2253 DO_D( VPMOVMSKB_128 ); 2254 DO_D( VCVTTSS2SI_64 ); 2255 DO_D( VPACKUSWB_128 ); 2256 DO_D( VCVTSS2SD_128 ); 2257 DO_D( VCVTSD2SS_128 ); 2258 DO_D( VMOVD_XMM_to_IREG32 ); 2259 DO_D( VPCMPESTRM_0x45_128 ); 2260 DO_D( VMOVQ_IREGorMEM64_to_XMM ); 2261 DO_D( VMOVUPS_XMM_to_XMMorMEM ); 2262 DO_D( VMOVNTDQ_128 ); 2263 DO_D( VMOVLHPS_128 ); 2264 DO_D( VPABSD_128 ); 2265 DO_D( VMOVHLPS_128 ); 2266 DO_D( VMOVQ_XMM_to_IREG64 ); 2267 DO_D( VMOVQ_XMMorMEM64_to_XMM ); 2268 DO_D( VCVTTSS2SI_32 ); 2269 DO_D( VPUNPCKLBW_128 ); 2270 DO_D( VPUNPCKHBW_128 ); 2271 DO_D( VMULSS_128 ); 2272 DO_D( VSUBSS_128 ); 2273 DO_D( VADDSS_128 ); 2274 DO_D( VDIVSS_128 ); 2275 DO_D( VUNPCKLPS_128 ); 2276 DO_D( VCVTSI2SS_128 ); 2277 DO_D( VANDPS_128 ); 2278 DO_D( VMINSS_128 ); 2279 DO_D( VMAXSS_128 ); 2280 DO_D( VANDNPS_128 ); 2281 DO_D( VORPS_128 ); 2282 DO_D( VSQRTSD_128 ); 2283 /* Test all CMPSS variants; this code is tricky. */ 2284 DO_D( VCMPSD_128_0x0 ); 2285 DO_D( VCMPSD_128_0x1 ); 2286 DO_D( VCMPSD_128_0x2 ); 2287 DO_D( VCMPSD_128_0x3 ); 2288 DO_D( VCMPSD_128_0x4 ); 2289 DO_D( VCMPSD_128_0x5 ); 2290 DO_D( VCMPSD_128_0x6 ); 2291 DO_D( VCMPSD_128_0x7 ); 2292 DO_D( VCMPSD_128_0xA ); 2293 DO_D( VCMPSD_128_0xC ); 2294 DO_D( VCMPSD_128_0xD ); 2295 DO_D( VCMPSD_128_0xE ); 2296 DO_D( VCMPSD_128_0x11 ); 2297 DO_D( VCMPSD_128_0x12 ); 2298 DO_D( VCMPSD_128_0x16 ); 2299 DO_D( VCMPSD_128_0x1E ); 2300 DO_D( VPSHUFB_128 ); 2301 DO_D( VCVTTSD2SI_32 ); 2302 DO_D( VCVTTSD2SI_64 ); 2303 DO_D( VCVTSI2SS_64 ); 2304 DO_D( VCVTSI2SD_64 ); 2305 DO_D( VCVTSI2SD_32 ); 2306 DO_D( VPOR_128 ); 2307 DO_D( VPXOR_128 ); 2308 DO_D( VPSUBB_128 ); 2309 DO_D( VPSUBD_128 ); 2310 DO_D( VPADDD_128 ); 2311 DO_D( VPMOVZXBW_128 ); 2312 DO_D( VPMOVZXWD_128 ); 2313 DO_D( VPBLENDVB_128 ); 2314 DO_D( VPMINSD_128 ); 2315 DO_D( VPMAXSD_128 ); 2316 DO_D( VANDPD_128 ); 2317 DO_D( VMULPD_256 ); 2318 DO_D( VMOVUPD_EtoG_128 ); 2319 DO_D( VADDPD_256 ); 2320 DO_D( VSUBPD_256 ); 2321 DO_D( VDIVPD_256 ); 2322 DO_D( VPCMPEQQ_128 ); 2323 DO_D( VSUBPD_128 ); 2324 DO_D( VADDPD_128 ); 2325 DO_D( VUNPCKLPD_128 ); 2326 DO_D( VUNPCKHPD_128 ); 2327 DO_D( VUNPCKHPS_128 ); 2328 DO_D( VMOVUPS_EtoG_128 ); 2329 DO_D( VADDPS_256 ); 2330 DO_D( VSUBPS_256 ); 2331 DO_D( VMULPS_256 ); 2332 DO_D( VDIVPS_256 ); 2333 DO_D( VPCMPGTQ_128 ); 2334 DO_D( VPEXTRQ_128_0x0 ); 2335 DO_D( VPEXTRQ_128_0x1 ); 2336 DO_D( VPSRLQ_0x05_128 ); 2337 DO_D( VPMULUDQ_128 ); 2338 DO_D( VPSLLQ_0x05_128 ); 2339 DO_D( VPMAXUD_128 ); 2340 DO_D( VPMINUD_128 ); 2341 DO_D( VPMULLD_128 ); 2342 DO_D( VPMAXUW_128 ); 2343 DO_D( VPEXTRW_128_EregOnly_toG_0x0 ); 2344 DO_D( VPEXTRW_128_EregOnly_toG_0x7 ); 2345 DO_D( VPMINUW_128 ); 2346 DO_D( VPHMINPOSUW_128 ); 2347 DO_D( VPMAXSW_128 ); 2348 DO_D( VPMINSW_128 ); 2349 DO_D( VPMAXUB_128 ); 2350 DO_D( VPEXTRB_GtoE_128_0x0 ); 2351 DO_D( VPEXTRB_GtoE_128_0x1 ); 2352 DO_D( VPEXTRB_GtoE_128_0x2 ); 2353 DO_D( VPEXTRB_GtoE_128_0x3 ); 2354 DO_D( VPEXTRB_GtoE_128_0x4 ); 2355 DO_D( VPEXTRB_GtoE_128_0x9 ); 2356 DO_D( VPEXTRB_GtoE_128_0xE ); 2357 DO_D( VPEXTRB_GtoE_128_0xF ); 2358 DO_D( VPMINUB_128 ); 2359 DO_D( VPMAXSB_128 ); 2360 DO_D( VPMINSB_128 ); 2361 DO_D( VPERM2F128_0x00 ); 2362 DO_D( VPERM2F128_0xFF ); 2363 DO_D( VPERM2F128_0x30 ); 2364 DO_D( VPERM2F128_0x21 ); 2365 DO_D( VPERM2F128_0x12 ); 2366 DO_D( VPERM2F128_0x03 ); 2367 DO_D( VPERM2F128_0x85 ); 2368 DO_D( VPERM2F128_0x5A ); 2369 DO_D( VPERMILPD_256_0x0 ); 2370 DO_D( VPERMILPD_256_0xF ); 2371 DO_D( VPERMILPD_256_0xA ); 2372 DO_D( VPERMILPD_256_0x5 ); 2373 DO_D( VPERMILPD_128_0x0 ); 2374 DO_D( VPERMILPD_128_0x3 ); 2375 DO_D( VUNPCKLPD_256 ); 2376 DO_D( VUNPCKHPD_256 ); 2377 DO_D( VSHUFPS_0x39_256 ); 2378 DO_D( VUNPCKLPS_256 ); 2379 DO_D( VUNPCKHPS_256 ); 2380 DO_D( VXORPD_256 ); 2381 DO_D( VBROADCASTSD_256 ); 2382 DO_D( VCMPPD_128_0x4 ); 2383 DO_D( VCVTDQ2PD_128 ); 2384 DO_D( VDIVPD_128 ); 2385 DO_D( VANDPD_256 ); 2386 DO_D( VPMOVSXBW_128 ); 2387 DO_D( VPSUBUSW_128 ); 2388 DO_D( VPSUBSW_128 ); 2389 DO_D( VPCMPEQW_128 ); 2390 DO_D( VPADDB_128 ); 2391 DO_D( VMOVAPS_EtoG_256 ); 2392 DO_D( VCVTDQ2PD_256 ); 2393 DO_D( VMOVHPD_128_LoadForm ); 2394 DO_D( VCVTPD2PS_256 ); 2395 DO_D( VPUNPCKHDQ_128 ); 2396 DO_D( VBROADCASTSS_128 ); 2397 DO_D( VPMOVSXDQ_128 ); 2398 DO_D( VPMOVSXWD_128 ); 2399 DO_D( VDIVPS_128 ); 2400 DO_D( VANDPS_256 ); 2401 DO_D( VXORPS_256 ); 2402 DO_D( VORPS_256 ); 2403 DO_D( VANDNPD_256 ); 2404 DO_D( VANDNPS_256 ); 2405 DO_D( VORPD_256 ); 2406 DO_D( VPERMILPS_256_0x0F ); 2407 DO_D( VPERMILPS_256_0xFA ); 2408 DO_D( VPERMILPS_256_0xA3 ); 2409 DO_D( VPERMILPS_256_0x5A ); 2410 DO_D( VPMULHW_128 ); 2411 DO_D( VPUNPCKHQDQ_128 ); 2412 DO_D( VPSRAW_0x05_128 ); 2413 DO_D( VPCMPGTD_128 ); 2414 DO_D( VPMOVZXBD_128 ); 2415 DO_D( VPMOVSXBD_128 ); 2416 DO_D( VPINSRB_128_1of3 ); 2417 DO_D( VPINSRB_128_2of3 ); 2418 DO_D( VPINSRB_128_3of3 ); 2419 DO_D( VCOMISD_128 ); 2420 DO_D( VCOMISS_128 ); 2421 DO_D( VMOVUPS_YMM_to_YMMorMEM ); 2422 DO_D( VDPPD_128_1of4 ); 2423 DO_D( VDPPD_128_2of4 ); 2424 DO_D( VDPPD_128_3of4 ); 2425 DO_D( VDPPD_128_4of4 ); 2426 DO_D( VPINSRW_128_1of4 ); 2427 DO_D( VPINSRW_128_2of4 ); 2428 DO_D( VPINSRW_128_3of4 ); 2429 DO_D( VPINSRW_128_4of4 ); 2430 DO_D( VBROADCASTSS_256 ); 2431 DO_D( VPALIGNR_128_1of3 ); 2432 DO_D( VPALIGNR_128_2of3 ); 2433 DO_D( VPALIGNR_128_3of3 ); 2434 DO_D( VMOVSD_REG_XMM ); 2435 DO_D( VMOVSS_REG_XMM ); 2436 DO_D( VMOVLPD_128_M64_XMM_XMM ); 2437 DO_D( VMOVLPD_128_XMM_M64 ); 2438 DO_D( VSHUFPD_128_1of2 ); 2439 DO_D( VSHUFPD_128_2of2 ); 2440 DO_D( VSHUFPD_256_1of2 ); 2441 DO_D( VSHUFPD_256_2of2 ); 2442 DO_D( VPERMILPS_128_0x00 ); 2443 DO_D( VPERMILPS_128_0xFE ); 2444 DO_D( VPERMILPS_128_0x30 ); 2445 DO_D( VPERMILPS_128_0x21 ); 2446 DO_D( VPERMILPS_128_0xD7 ); 2447 DO_D( VPERMILPS_128_0xB5 ); 2448 DO_D( VPERMILPS_128_0x85 ); 2449 DO_D( VPERMILPS_128_0x29 ); 2450 DO_D( VBLENDPS_128_1of3 ); 2451 DO_D( VBLENDPS_128_2of3 ); 2452 DO_D( VBLENDPS_128_3of3 ); 2453 DO_D( VBLENDPD_128_1of2 ); 2454 DO_D( VBLENDPD_128_2of2 ); 2455 DO_D( VBLENDPD_256_1of3 ); 2456 DO_D( VBLENDPD_256_2of3 ); 2457 DO_D( VBLENDPD_256_3of3 ); 2458 DO_D( VPBLENDW_128_0x00 ); 2459 DO_D( VPBLENDW_128_0xFE ); 2460 DO_D( VPBLENDW_128_0x30 ); 2461 DO_D( VPBLENDW_128_0x21 ); 2462 DO_D( VPBLENDW_128_0xD7 ); 2463 DO_D( VPBLENDW_128_0xB5 ); 2464 DO_D( VPBLENDW_128_0x85 ); 2465 DO_D( VPBLENDW_128_0x29 ); 2466 DO_D( VMOVUPS_EtoG_256 ); 2467 DO_D( VSQRTSS_128 ); 2468 DO_D( VSQRTPS_128 ); 2469 DO_D( VSQRTPS_256 ); 2470 DO_D( VSQRTPD_128 ); 2471 DO_D( VSQRTPD_256 ); 2472 DO_D( VRSQRTSS_128 ); 2473 DO_D( VRSQRTPS_128 ); 2474 DO_D( VRSQRTPS_256 ); 2475 DO_D( VMOVDQU_GtoE_256 ); 2476 DO_D( VCVTPS2PD_256 ); 2477 DO_D( VCVTTPS2DQ_128 ); 2478 DO_D( VCVTTPS2DQ_256 ); 2479 DO_D( VCVTDQ2PS_128 ); 2480 DO_D( VCVTDQ2PS_256 ); 2481 DO_D( VCVTTPD2DQ_128 ); 2482 DO_D( VCVTTPD2DQ_256 ); 2483 DO_D( VCVTPD2DQ_128 ); 2484 DO_D( VCVTPD2DQ_256 ); 2485 DO_D( VMOVSLDUP_128 ); 2486 DO_D( VMOVSLDUP_256 ); 2487 DO_D( VMOVSHDUP_128 ); 2488 DO_D( VMOVSHDUP_256 ); 2489 DO_D( VPERMILPS_VAR_128 ); 2490 DO_D( VPERMILPD_VAR_128 ); 2491 DO_D( VPERMILPS_VAR_256 ); 2492 DO_D( VPERMILPD_VAR_256 ); 2493 DO_D( VPSLLW_128 ); 2494 DO_D( VPSRLW_128 ); 2495 DO_D( VPSRAW_128 ); 2496 DO_D( VPSLLD_128 ); 2497 DO_D( VPSRLD_128 ); 2498 DO_D( VPSRAD_128 ); 2499 DO_D( VPSLLQ_128 ); 2500 DO_D( VPSRLQ_128 ); 2501 DO_D( VROUNDPS_128_0x0 ); 2502 DO_D( VROUNDPS_128_0x1 ); 2503 DO_D( VROUNDPS_128_0x2 ); 2504 DO_D( VROUNDPS_128_0x3 ); 2505 DO_D( VROUNDPS_128_0x4 ); 2506 DO_D( VROUNDPS_256_0x0 ); 2507 DO_D( VROUNDPS_256_0x1 ); 2508 DO_D( VROUNDPS_256_0x2 ); 2509 DO_D( VROUNDPS_256_0x3 ); 2510 DO_D( VROUNDPS_256_0x4 ); 2511 DO_D( VROUNDPD_128_0x0 ); 2512 DO_D( VROUNDPD_128_0x1 ); 2513 DO_D( VROUNDPD_128_0x2 ); 2514 DO_D( VROUNDPD_128_0x3 ); 2515 DO_D( VROUNDPD_128_0x4 ); 2516 DO_D( VROUNDPD_256_0x0 ); 2517 DO_D( VROUNDPD_256_0x1 ); 2518 DO_D( VROUNDPD_256_0x2 ); 2519 DO_D( VROUNDPD_256_0x3 ); 2520 DO_D( VROUNDPD_256_0x4 ); 2521 DO_D( VROUNDSS_0x0 ); 2522 DO_D( VROUNDSS_0x1 ); 2523 DO_D( VROUNDSS_0x2 ); 2524 DO_D( VROUNDSS_0x3 ); 2525 DO_D( VROUNDSS_0x4 ); 2526 DO_D( VROUNDSS_0x5 ); 2527 DO_D( VROUNDSD_0x0 ); 2528 DO_D( VROUNDSD_0x1 ); 2529 DO_D( VROUNDSD_0x2 ); 2530 DO_D( VROUNDSD_0x3 ); 2531 DO_D( VROUNDSD_0x4 ); 2532 DO_D( VROUNDSD_0x5 ); 2533 DO_D( VPTEST_128_1 ); 2534 DO_D( VPTEST_128_2 ); 2535 DO_D( VPTEST_256_1 ); 2536 DO_D( VPTEST_256_2 ); 2537 DO_D( VTESTPS_128_1 ); 2538 DO_D( VTESTPS_128_2 ); 2539 DO_N( 10, VTESTPS_128_3 ); 2540 DO_D( VTESTPS_256_1 ); 2541 DO_D( VTESTPS_256_2 ); 2542 DO_N( 10, VTESTPS_256_3 ); 2543 DO_D( VTESTPD_128_1 ); 2544 DO_D( VTESTPD_128_2 ); 2545 DO_N( 10, VTESTPD_128_3 ); 2546 DO_D( VTESTPD_256_1 ); 2547 DO_D( VTESTPD_256_2 ); 2548 DO_N( 10, VTESTPD_256_3 ); 2549 DO_D( VBLENDVPS_128 ); 2550 DO_D( VBLENDVPS_256 ); 2551 DO_D( VBLENDVPD_128 ); 2552 DO_D( VBLENDVPD_256 ); 2553 DO_D( VPMULDQ_128 ); 2554 DO_D( VCMPPD_256_0x4 ); 2555 DO_D( VCMPPS_128_0x4 ); 2556 DO_D( VCMPPS_256_0x4 ); 2557 DO_D( VPCMPGTB_128 ); 2558 DO_D( VPCMPGTW_128 ); 2559 DO_D( VPMADDWD_128 ); 2560 DO_D( VADDSUBPS_128 ); 2561 DO_D( VADDSUBPS_256 ); 2562 DO_D( VADDSUBPD_128 ); 2563 DO_D( VADDSUBPD_256 ); 2564 DO_D( VCVTSS2SI_64 ); 2565 DO_D( VCVTSS2SI_32 ); 2566 DO_D( VCVTSD2SI_32 ); 2567 DO_D( VCVTSD2SI_64 ); 2568 DO_D( VDPPS_128_1of4 ); 2569 DO_D( VDPPS_128_2of4 ); 2570 DO_D( VDPPS_128_3of4 ); 2571 DO_D( VDPPS_128_4of4 ); 2572 DO_D( VDPPS_256_1of4 ); 2573 DO_D( VDPPS_256_2of4 ); 2574 DO_D( VDPPS_256_3of4 ); 2575 DO_D( VDPPS_256_4of4 ); 2576 DO_D( VHADDPS_128 ); 2577 DO_D( VHADDPS_256 ); 2578 DO_D( VHADDPD_128 ); 2579 DO_D( VHADDPD_256 ); 2580 DO_D( VHSUBPS_128 ); 2581 DO_D( VHSUBPS_256 ); 2582 DO_D( VHSUBPD_128 ); 2583 DO_D( VHSUBPD_256 ); 2584 DO_D( VEXTRACTPS_0x0 ); 2585 DO_D( VEXTRACTPS_0x1 ); 2586 DO_D( VEXTRACTPS_0x2 ); 2587 DO_D( VEXTRACTPS_0x3 ); 2588 DO_D( VLDDQU_128 ); 2589 DO_D( VLDDQU_256 ); 2590 DO_D( VMAXPS_256 ); 2591 DO_D( VMAXPD_128 ); 2592 DO_D( VMAXPD_256 ); 2593 DO_D( VMINPS_256 ); 2594 DO_D( VMINPD_128 ); 2595 DO_D( VMINPD_256 ); 2596 DO_D( VMOVHPS_128_StoreForm ); 2597 DO_D( VMOVNTDQ_256 ); 2598 DO_D( VMOVHPS_128_LoadForm ); 2599 DO_D( VMOVNTDQA_128 ); 2600 DO_D( VMASKMOVDQU_128 ); 2601 DO_D( VMOVMSKPD_128 ); 2602 DO_D( VMOVMSKPD_256 ); 2603 DO_D( VMOVMSKPS_128 ); 2604 DO_D( VMOVMSKPS_256 ); 2605 DO_D( VMOVNTPD_128 ); 2606 DO_D( VMOVNTPD_256 ); 2607 DO_D( VMOVNTPS_128 ); 2608 DO_D( VMOVNTPS_256 ); 2609 DO_D( VPACKSSWB_128 ); 2610 DO_D( VPAVGB_128 ); 2611 DO_D( VPAVGW_128 ); 2612 DO_D( VPADDSB_128 ); 2613 DO_D( VPADDSW_128 ); 2614 DO_D( VPHADDW_128 ); 2615 DO_D( VPHADDD_128 ); 2616 DO_D( VPHADDSW_128 ); 2617 DO_D( VPMADDUBSW_128 ); 2618 DO_D( VPHSUBW_128 ); 2619 DO_D( VPHSUBD_128 ); 2620 DO_D( VPHSUBSW_128 ); 2621 DO_D( VPABSB_128 ); 2622 DO_D( VPABSW_128 ); 2623 DO_D( VPMOVSXBQ_128 ); 2624 DO_D( VPMOVSXWQ_128 ); 2625 DO_D( VPACKUSDW_128 ); 2626 DO_D( VPMOVZXBQ_128 ); 2627 DO_D( VPMOVZXWQ_128 ); 2628 DO_D( VPMOVZXDQ_128 ); 2629 DO_D( VMPSADBW_128_0x0 ); 2630 DO_D( VMPSADBW_128_0x1 ); 2631 DO_D( VMPSADBW_128_0x2 ); 2632 DO_D( VMPSADBW_128_0x3 ); 2633 DO_D( VMPSADBW_128_0x4 ); 2634 DO_D( VMPSADBW_128_0x5 ); 2635 DO_D( VMPSADBW_128_0x6 ); 2636 DO_D( VMPSADBW_128_0x7 ); 2637 DO_D( VMOVDDUP_YMMorMEM256_to_YMM ); 2638 DO_D( VMOVLPS_128_M64_XMM_XMM ); 2639 DO_D( VMOVLPS_128_XMM_M64 ); 2640 DO_D( VRCPSS_128 ); 2641 DO_D( VRCPPS_128 ); 2642 DO_D( VRCPPS_256 ); 2643 DO_D( VPSADBW_128 ); 2644 DO_D( VPSIGNB_128 ); 2645 DO_D( VPSIGNW_128 ); 2646 DO_D( VPSIGND_128 ); 2647 DO_D( VPMULHRSW_128 ); 2648 DO_D( VBROADCASTF128 ); 2649 DO_D( VPEXTRW_128_0x0 ); 2650 DO_D( VPEXTRW_128_0x1 ); 2651 DO_D( VPEXTRW_128_0x2 ); 2652 DO_D( VPEXTRW_128_0x3 ); 2653 DO_D( VPEXTRW_128_0x4 ); 2654 DO_D( VPEXTRW_128_0x5 ); 2655 DO_D( VPEXTRW_128_0x6 ); 2656 DO_D( VPEXTRW_128_0x7 ); 2657 DO_D( VAESENC ); 2658 DO_D( VAESENCLAST ); 2659 DO_D( VAESDEC ); 2660 DO_D( VAESDECLAST ); 2661 DO_D( VAESIMC ); 2662 DO_D( VAESKEYGENASSIST_0x00 ); 2663 DO_D( VAESKEYGENASSIST_0x31 ); 2664 DO_D( VAESKEYGENASSIST_0xB2 ); 2665 DO_D( VAESKEYGENASSIST_0xFF ); 2666 DO_D( VPCLMULQDQ_0x00 ); 2667 DO_D( VPCLMULQDQ_0x01 ); 2668 DO_D( VPCLMULQDQ_0x10 ); 2669 DO_D( VPCLMULQDQ_0x11 ); 2670 DO_D( VPCLMULQDQ_0xFF ); 2671 DO_D( VCMPSS_128_0x9 ); 2672 return 0; 2673 } 2674 2675
