| OLD | NEW |
| (Empty) |
| 1 #if defined(__x86_64__) | |
| 2 .text | |
| 3 | |
| 4 | |
| 5 | |
| 6 .globl _rsaz_512_sqr | |
| 7 .private_extern _rsaz_512_sqr | |
| 8 | |
| 9 .p2align 5 | |
| 10 _rsaz_512_sqr: | |
| 11 pushq %rbx | |
| 12 pushq %rbp | |
| 13 pushq %r12 | |
| 14 pushq %r13 | |
| 15 pushq %r14 | |
| 16 pushq %r15 | |
| 17 | |
| 18 subq $128+24,%rsp | |
| 19 L$sqr_body: | |
| 20 movq %rdx,%rbp | |
| 21 movq (%rsi),%rdx | |
| 22 movq 8(%rsi),%rax | |
| 23 movq %rcx,128(%rsp) | |
| 24 jmp L$oop_sqr | |
| 25 | |
| 26 .p2align 5 | |
| 27 L$oop_sqr: | |
| 28 movl %r8d,128+8(%rsp) | |
| 29 | |
| 30 movq %rdx,%rbx | |
| 31 mulq %rdx | |
| 32 movq %rax,%r8 | |
| 33 movq 16(%rsi),%rax | |
| 34 movq %rdx,%r9 | |
| 35 | |
| 36 mulq %rbx | |
| 37 addq %rax,%r9 | |
| 38 movq 24(%rsi),%rax | |
| 39 movq %rdx,%r10 | |
| 40 adcq $0,%r10 | |
| 41 | |
| 42 mulq %rbx | |
| 43 addq %rax,%r10 | |
| 44 movq 32(%rsi),%rax | |
| 45 movq %rdx,%r11 | |
| 46 adcq $0,%r11 | |
| 47 | |
| 48 mulq %rbx | |
| 49 addq %rax,%r11 | |
| 50 movq 40(%rsi),%rax | |
| 51 movq %rdx,%r12 | |
| 52 adcq $0,%r12 | |
| 53 | |
| 54 mulq %rbx | |
| 55 addq %rax,%r12 | |
| 56 movq 48(%rsi),%rax | |
| 57 movq %rdx,%r13 | |
| 58 adcq $0,%r13 | |
| 59 | |
| 60 mulq %rbx | |
| 61 addq %rax,%r13 | |
| 62 movq 56(%rsi),%rax | |
| 63 movq %rdx,%r14 | |
| 64 adcq $0,%r14 | |
| 65 | |
| 66 mulq %rbx | |
| 67 addq %rax,%r14 | |
| 68 movq %rbx,%rax | |
| 69 movq %rdx,%r15 | |
| 70 adcq $0,%r15 | |
| 71 | |
| 72 addq %r8,%r8 | |
| 73 movq %r9,%rcx | |
| 74 adcq %r9,%r9 | |
| 75 | |
| 76 mulq %rax | |
| 77 movq %rax,(%rsp) | |
| 78 addq %rdx,%r8 | |
| 79 adcq $0,%r9 | |
| 80 | |
| 81 movq %r8,8(%rsp) | |
| 82 shrq $63,%rcx | |
| 83 | |
| 84 | |
| 85 movq 8(%rsi),%r8 | |
| 86 movq 16(%rsi),%rax | |
| 87 mulq %r8 | |
| 88 addq %rax,%r10 | |
| 89 movq 24(%rsi),%rax | |
| 90 movq %rdx,%rbx | |
| 91 adcq $0,%rbx | |
| 92 | |
| 93 mulq %r8 | |
| 94 addq %rax,%r11 | |
| 95 movq 32(%rsi),%rax | |
| 96 adcq $0,%rdx | |
| 97 addq %rbx,%r11 | |
| 98 movq %rdx,%rbx | |
| 99 adcq $0,%rbx | |
| 100 | |
| 101 mulq %r8 | |
| 102 addq %rax,%r12 | |
| 103 movq 40(%rsi),%rax | |
| 104 adcq $0,%rdx | |
| 105 addq %rbx,%r12 | |
| 106 movq %rdx,%rbx | |
| 107 adcq $0,%rbx | |
| 108 | |
| 109 mulq %r8 | |
| 110 addq %rax,%r13 | |
| 111 movq 48(%rsi),%rax | |
| 112 adcq $0,%rdx | |
| 113 addq %rbx,%r13 | |
| 114 movq %rdx,%rbx | |
| 115 adcq $0,%rbx | |
| 116 | |
| 117 mulq %r8 | |
| 118 addq %rax,%r14 | |
| 119 movq 56(%rsi),%rax | |
| 120 adcq $0,%rdx | |
| 121 addq %rbx,%r14 | |
| 122 movq %rdx,%rbx | |
| 123 adcq $0,%rbx | |
| 124 | |
| 125 mulq %r8 | |
| 126 addq %rax,%r15 | |
| 127 movq %r8,%rax | |
| 128 adcq $0,%rdx | |
| 129 addq %rbx,%r15 | |
| 130 movq %rdx,%r8 | |
| 131 movq %r10,%rdx | |
| 132 adcq $0,%r8 | |
| 133 | |
| 134 addq %rdx,%rdx | |
| 135 leaq (%rcx,%r10,2),%r10 | |
| 136 movq %r11,%rbx | |
| 137 adcq %r11,%r11 | |
| 138 | |
| 139 mulq %rax | |
| 140 addq %rax,%r9 | |
| 141 adcq %rdx,%r10 | |
| 142 adcq $0,%r11 | |
| 143 | |
| 144 movq %r9,16(%rsp) | |
| 145 movq %r10,24(%rsp) | |
| 146 shrq $63,%rbx | |
| 147 | |
| 148 | |
| 149 movq 16(%rsi),%r9 | |
| 150 movq 24(%rsi),%rax | |
| 151 mulq %r9 | |
| 152 addq %rax,%r12 | |
| 153 movq 32(%rsi),%rax | |
| 154 movq %rdx,%rcx | |
| 155 adcq $0,%rcx | |
| 156 | |
| 157 mulq %r9 | |
| 158 addq %rax,%r13 | |
| 159 movq 40(%rsi),%rax | |
| 160 adcq $0,%rdx | |
| 161 addq %rcx,%r13 | |
| 162 movq %rdx,%rcx | |
| 163 adcq $0,%rcx | |
| 164 | |
| 165 mulq %r9 | |
| 166 addq %rax,%r14 | |
| 167 movq 48(%rsi),%rax | |
| 168 adcq $0,%rdx | |
| 169 addq %rcx,%r14 | |
| 170 movq %rdx,%rcx | |
| 171 adcq $0,%rcx | |
| 172 | |
| 173 mulq %r9 | |
| 174 movq %r12,%r10 | |
| 175 leaq (%rbx,%r12,2),%r12 | |
| 176 addq %rax,%r15 | |
| 177 movq 56(%rsi),%rax | |
| 178 adcq $0,%rdx | |
| 179 addq %rcx,%r15 | |
| 180 movq %rdx,%rcx | |
| 181 adcq $0,%rcx | |
| 182 | |
| 183 mulq %r9 | |
| 184 shrq $63,%r10 | |
| 185 addq %rax,%r8 | |
| 186 movq %r9,%rax | |
| 187 adcq $0,%rdx | |
| 188 addq %rcx,%r8 | |
| 189 movq %rdx,%r9 | |
| 190 adcq $0,%r9 | |
| 191 | |
| 192 movq %r13,%rcx | |
| 193 leaq (%r10,%r13,2),%r13 | |
| 194 | |
| 195 mulq %rax | |
| 196 addq %rax,%r11 | |
| 197 adcq %rdx,%r12 | |
| 198 adcq $0,%r13 | |
| 199 | |
| 200 movq %r11,32(%rsp) | |
| 201 movq %r12,40(%rsp) | |
| 202 shrq $63,%rcx | |
| 203 | |
| 204 | |
| 205 movq 24(%rsi),%r10 | |
| 206 movq 32(%rsi),%rax | |
| 207 mulq %r10 | |
| 208 addq %rax,%r14 | |
| 209 movq 40(%rsi),%rax | |
| 210 movq %rdx,%rbx | |
| 211 adcq $0,%rbx | |
| 212 | |
| 213 mulq %r10 | |
| 214 addq %rax,%r15 | |
| 215 movq 48(%rsi),%rax | |
| 216 adcq $0,%rdx | |
| 217 addq %rbx,%r15 | |
| 218 movq %rdx,%rbx | |
| 219 adcq $0,%rbx | |
| 220 | |
| 221 mulq %r10 | |
| 222 movq %r14,%r12 | |
| 223 leaq (%rcx,%r14,2),%r14 | |
| 224 addq %rax,%r8 | |
| 225 movq 56(%rsi),%rax | |
| 226 adcq $0,%rdx | |
| 227 addq %rbx,%r8 | |
| 228 movq %rdx,%rbx | |
| 229 adcq $0,%rbx | |
| 230 | |
| 231 mulq %r10 | |
| 232 shrq $63,%r12 | |
| 233 addq %rax,%r9 | |
| 234 movq %r10,%rax | |
| 235 adcq $0,%rdx | |
| 236 addq %rbx,%r9 | |
| 237 movq %rdx,%r10 | |
| 238 adcq $0,%r10 | |
| 239 | |
| 240 movq %r15,%rbx | |
| 241 leaq (%r12,%r15,2),%r15 | |
| 242 | |
| 243 mulq %rax | |
| 244 addq %rax,%r13 | |
| 245 adcq %rdx,%r14 | |
| 246 adcq $0,%r15 | |
| 247 | |
| 248 movq %r13,48(%rsp) | |
| 249 movq %r14,56(%rsp) | |
| 250 shrq $63,%rbx | |
| 251 | |
| 252 | |
| 253 movq 32(%rsi),%r11 | |
| 254 movq 40(%rsi),%rax | |
| 255 mulq %r11 | |
| 256 addq %rax,%r8 | |
| 257 movq 48(%rsi),%rax | |
| 258 movq %rdx,%rcx | |
| 259 adcq $0,%rcx | |
| 260 | |
| 261 mulq %r11 | |
| 262 addq %rax,%r9 | |
| 263 movq 56(%rsi),%rax | |
| 264 adcq $0,%rdx | |
| 265 movq %r8,%r12 | |
| 266 leaq (%rbx,%r8,2),%r8 | |
| 267 addq %rcx,%r9 | |
| 268 movq %rdx,%rcx | |
| 269 adcq $0,%rcx | |
| 270 | |
| 271 mulq %r11 | |
| 272 shrq $63,%r12 | |
| 273 addq %rax,%r10 | |
| 274 movq %r11,%rax | |
| 275 adcq $0,%rdx | |
| 276 addq %rcx,%r10 | |
| 277 movq %rdx,%r11 | |
| 278 adcq $0,%r11 | |
| 279 | |
| 280 movq %r9,%rcx | |
| 281 leaq (%r12,%r9,2),%r9 | |
| 282 | |
| 283 mulq %rax | |
| 284 addq %rax,%r15 | |
| 285 adcq %rdx,%r8 | |
| 286 adcq $0,%r9 | |
| 287 | |
| 288 movq %r15,64(%rsp) | |
| 289 movq %r8,72(%rsp) | |
| 290 shrq $63,%rcx | |
| 291 | |
| 292 | |
| 293 movq 40(%rsi),%r12 | |
| 294 movq 48(%rsi),%rax | |
| 295 mulq %r12 | |
| 296 addq %rax,%r10 | |
| 297 movq 56(%rsi),%rax | |
| 298 movq %rdx,%rbx | |
| 299 adcq $0,%rbx | |
| 300 | |
| 301 mulq %r12 | |
| 302 addq %rax,%r11 | |
| 303 movq %r12,%rax | |
| 304 movq %r10,%r15 | |
| 305 leaq (%rcx,%r10,2),%r10 | |
| 306 adcq $0,%rdx | |
| 307 shrq $63,%r15 | |
| 308 addq %rbx,%r11 | |
| 309 movq %rdx,%r12 | |
| 310 adcq $0,%r12 | |
| 311 | |
| 312 movq %r11,%rbx | |
| 313 leaq (%r15,%r11,2),%r11 | |
| 314 | |
| 315 mulq %rax | |
| 316 addq %rax,%r9 | |
| 317 adcq %rdx,%r10 | |
| 318 adcq $0,%r11 | |
| 319 | |
| 320 movq %r9,80(%rsp) | |
| 321 movq %r10,88(%rsp) | |
| 322 | |
| 323 | |
| 324 movq 48(%rsi),%r13 | |
| 325 movq 56(%rsi),%rax | |
| 326 mulq %r13 | |
| 327 addq %rax,%r12 | |
| 328 movq %r13,%rax | |
| 329 movq %rdx,%r13 | |
| 330 adcq $0,%r13 | |
| 331 | |
| 332 xorq %r14,%r14 | |
| 333 shlq $1,%rbx | |
| 334 adcq %r12,%r12 | |
| 335 adcq %r13,%r13 | |
| 336 adcq %r14,%r14 | |
| 337 | |
| 338 mulq %rax | |
| 339 addq %rax,%r11 | |
| 340 adcq %rdx,%r12 | |
| 341 adcq $0,%r13 | |
| 342 | |
| 343 movq %r11,96(%rsp) | |
| 344 movq %r12,104(%rsp) | |
| 345 | |
| 346 | |
| 347 movq 56(%rsi),%rax | |
| 348 mulq %rax | |
| 349 addq %rax,%r13 | |
| 350 adcq $0,%rdx | |
| 351 | |
| 352 addq %rdx,%r14 | |
| 353 | |
| 354 movq %r13,112(%rsp) | |
| 355 movq %r14,120(%rsp) | |
| 356 | |
| 357 movq (%rsp),%r8 | |
| 358 movq 8(%rsp),%r9 | |
| 359 movq 16(%rsp),%r10 | |
| 360 movq 24(%rsp),%r11 | |
| 361 movq 32(%rsp),%r12 | |
| 362 movq 40(%rsp),%r13 | |
| 363 movq 48(%rsp),%r14 | |
| 364 movq 56(%rsp),%r15 | |
| 365 | |
| 366 call __rsaz_512_reduce | |
| 367 | |
| 368 addq 64(%rsp),%r8 | |
| 369 adcq 72(%rsp),%r9 | |
| 370 adcq 80(%rsp),%r10 | |
| 371 adcq 88(%rsp),%r11 | |
| 372 adcq 96(%rsp),%r12 | |
| 373 adcq 104(%rsp),%r13 | |
| 374 adcq 112(%rsp),%r14 | |
| 375 adcq 120(%rsp),%r15 | |
| 376 sbbq %rcx,%rcx | |
| 377 | |
| 378 call __rsaz_512_subtract | |
| 379 | |
| 380 movq %r8,%rdx | |
| 381 movq %r9,%rax | |
| 382 movl 128+8(%rsp),%r8d | |
| 383 movq %rdi,%rsi | |
| 384 | |
| 385 decl %r8d | |
| 386 jnz L$oop_sqr | |
| 387 | |
| 388 leaq 128+24+48(%rsp),%rax | |
| 389 movq -48(%rax),%r15 | |
| 390 movq -40(%rax),%r14 | |
| 391 movq -32(%rax),%r13 | |
| 392 movq -24(%rax),%r12 | |
| 393 movq -16(%rax),%rbp | |
| 394 movq -8(%rax),%rbx | |
| 395 leaq (%rax),%rsp | |
| 396 L$sqr_epilogue: | |
| 397 .byte 0xf3,0xc3 | |
| 398 | |
| 399 .globl _rsaz_512_mul | |
| 400 .private_extern _rsaz_512_mul | |
| 401 | |
| 402 .p2align 5 | |
| 403 _rsaz_512_mul: | |
| 404 pushq %rbx | |
| 405 pushq %rbp | |
| 406 pushq %r12 | |
| 407 pushq %r13 | |
| 408 pushq %r14 | |
| 409 pushq %r15 | |
| 410 | |
| 411 subq $128+24,%rsp | |
| 412 L$mul_body: | |
| 413 .byte 102,72,15,110,199 | |
| 414 .byte 102,72,15,110,201 | |
| 415 movq %r8,128(%rsp) | |
| 416 movq (%rdx),%rbx | |
| 417 movq %rdx,%rbp | |
| 418 call __rsaz_512_mul | |
| 419 | |
| 420 .byte 102,72,15,126,199 | |
| 421 .byte 102,72,15,126,205 | |
| 422 | |
| 423 movq (%rsp),%r8 | |
| 424 movq 8(%rsp),%r9 | |
| 425 movq 16(%rsp),%r10 | |
| 426 movq 24(%rsp),%r11 | |
| 427 movq 32(%rsp),%r12 | |
| 428 movq 40(%rsp),%r13 | |
| 429 movq 48(%rsp),%r14 | |
| 430 movq 56(%rsp),%r15 | |
| 431 | |
| 432 call __rsaz_512_reduce | |
| 433 addq 64(%rsp),%r8 | |
| 434 adcq 72(%rsp),%r9 | |
| 435 adcq 80(%rsp),%r10 | |
| 436 adcq 88(%rsp),%r11 | |
| 437 adcq 96(%rsp),%r12 | |
| 438 adcq 104(%rsp),%r13 | |
| 439 adcq 112(%rsp),%r14 | |
| 440 adcq 120(%rsp),%r15 | |
| 441 sbbq %rcx,%rcx | |
| 442 | |
| 443 call __rsaz_512_subtract | |
| 444 | |
| 445 leaq 128+24+48(%rsp),%rax | |
| 446 movq -48(%rax),%r15 | |
| 447 movq -40(%rax),%r14 | |
| 448 movq -32(%rax),%r13 | |
| 449 movq -24(%rax),%r12 | |
| 450 movq -16(%rax),%rbp | |
| 451 movq -8(%rax),%rbx | |
| 452 leaq (%rax),%rsp | |
| 453 L$mul_epilogue: | |
| 454 .byte 0xf3,0xc3 | |
| 455 | |
| 456 .globl _rsaz_512_mul_gather4 | |
| 457 .private_extern _rsaz_512_mul_gather4 | |
| 458 | |
| 459 .p2align 5 | |
| 460 _rsaz_512_mul_gather4: | |
| 461 pushq %rbx | |
| 462 pushq %rbp | |
| 463 pushq %r12 | |
| 464 pushq %r13 | |
| 465 pushq %r14 | |
| 466 pushq %r15 | |
| 467 | |
| 468 subq $152,%rsp | |
| 469 L$mul_gather4_body: | |
| 470 movd %r9d,%xmm8 | |
| 471 movdqa L$inc+16(%rip),%xmm1 | |
| 472 movdqa L$inc(%rip),%xmm0 | |
| 473 | |
| 474 pshufd $0,%xmm8,%xmm8 | |
| 475 movdqa %xmm1,%xmm7 | |
| 476 movdqa %xmm1,%xmm2 | |
| 477 paddd %xmm0,%xmm1 | |
| 478 pcmpeqd %xmm8,%xmm0 | |
| 479 movdqa %xmm7,%xmm3 | |
| 480 paddd %xmm1,%xmm2 | |
| 481 pcmpeqd %xmm8,%xmm1 | |
| 482 movdqa %xmm7,%xmm4 | |
| 483 paddd %xmm2,%xmm3 | |
| 484 pcmpeqd %xmm8,%xmm2 | |
| 485 movdqa %xmm7,%xmm5 | |
| 486 paddd %xmm3,%xmm4 | |
| 487 pcmpeqd %xmm8,%xmm3 | |
| 488 movdqa %xmm7,%xmm6 | |
| 489 paddd %xmm4,%xmm5 | |
| 490 pcmpeqd %xmm8,%xmm4 | |
| 491 paddd %xmm5,%xmm6 | |
| 492 pcmpeqd %xmm8,%xmm5 | |
| 493 paddd %xmm6,%xmm7 | |
| 494 pcmpeqd %xmm8,%xmm6 | |
| 495 pcmpeqd %xmm8,%xmm7 | |
| 496 | |
| 497 movdqa 0(%rdx),%xmm8 | |
| 498 movdqa 16(%rdx),%xmm9 | |
| 499 movdqa 32(%rdx),%xmm10 | |
| 500 movdqa 48(%rdx),%xmm11 | |
| 501 pand %xmm0,%xmm8 | |
| 502 movdqa 64(%rdx),%xmm12 | |
| 503 pand %xmm1,%xmm9 | |
| 504 movdqa 80(%rdx),%xmm13 | |
| 505 pand %xmm2,%xmm10 | |
| 506 movdqa 96(%rdx),%xmm14 | |
| 507 pand %xmm3,%xmm11 | |
| 508 movdqa 112(%rdx),%xmm15 | |
| 509 leaq 128(%rdx),%rbp | |
| 510 pand %xmm4,%xmm12 | |
| 511 pand %xmm5,%xmm13 | |
| 512 pand %xmm6,%xmm14 | |
| 513 pand %xmm7,%xmm15 | |
| 514 por %xmm10,%xmm8 | |
| 515 por %xmm11,%xmm9 | |
| 516 por %xmm12,%xmm8 | |
| 517 por %xmm13,%xmm9 | |
| 518 por %xmm14,%xmm8 | |
| 519 por %xmm15,%xmm9 | |
| 520 | |
| 521 por %xmm9,%xmm8 | |
| 522 pshufd $0x4e,%xmm8,%xmm9 | |
| 523 por %xmm9,%xmm8 | |
| 524 .byte 102,76,15,126,195 | |
| 525 | |
| 526 movq %r8,128(%rsp) | |
| 527 movq %rdi,128+8(%rsp) | |
| 528 movq %rcx,128+16(%rsp) | |
| 529 | |
| 530 movq (%rsi),%rax | |
| 531 movq 8(%rsi),%rcx | |
| 532 mulq %rbx | |
| 533 movq %rax,(%rsp) | |
| 534 movq %rcx,%rax | |
| 535 movq %rdx,%r8 | |
| 536 | |
| 537 mulq %rbx | |
| 538 addq %rax,%r8 | |
| 539 movq 16(%rsi),%rax | |
| 540 movq %rdx,%r9 | |
| 541 adcq $0,%r9 | |
| 542 | |
| 543 mulq %rbx | |
| 544 addq %rax,%r9 | |
| 545 movq 24(%rsi),%rax | |
| 546 movq %rdx,%r10 | |
| 547 adcq $0,%r10 | |
| 548 | |
| 549 mulq %rbx | |
| 550 addq %rax,%r10 | |
| 551 movq 32(%rsi),%rax | |
| 552 movq %rdx,%r11 | |
| 553 adcq $0,%r11 | |
| 554 | |
| 555 mulq %rbx | |
| 556 addq %rax,%r11 | |
| 557 movq 40(%rsi),%rax | |
| 558 movq %rdx,%r12 | |
| 559 adcq $0,%r12 | |
| 560 | |
| 561 mulq %rbx | |
| 562 addq %rax,%r12 | |
| 563 movq 48(%rsi),%rax | |
| 564 movq %rdx,%r13 | |
| 565 adcq $0,%r13 | |
| 566 | |
| 567 mulq %rbx | |
| 568 addq %rax,%r13 | |
| 569 movq 56(%rsi),%rax | |
| 570 movq %rdx,%r14 | |
| 571 adcq $0,%r14 | |
| 572 | |
| 573 mulq %rbx | |
| 574 addq %rax,%r14 | |
| 575 movq (%rsi),%rax | |
| 576 movq %rdx,%r15 | |
| 577 adcq $0,%r15 | |
| 578 | |
| 579 leaq 8(%rsp),%rdi | |
| 580 movl $7,%ecx | |
| 581 jmp L$oop_mul_gather | |
| 582 | |
| 583 .p2align 5 | |
| 584 L$oop_mul_gather: | |
| 585 movdqa 0(%rbp),%xmm8 | |
| 586 movdqa 16(%rbp),%xmm9 | |
| 587 movdqa 32(%rbp),%xmm10 | |
| 588 movdqa 48(%rbp),%xmm11 | |
| 589 pand %xmm0,%xmm8 | |
| 590 movdqa 64(%rbp),%xmm12 | |
| 591 pand %xmm1,%xmm9 | |
| 592 movdqa 80(%rbp),%xmm13 | |
| 593 pand %xmm2,%xmm10 | |
| 594 movdqa 96(%rbp),%xmm14 | |
| 595 pand %xmm3,%xmm11 | |
| 596 movdqa 112(%rbp),%xmm15 | |
| 597 leaq 128(%rbp),%rbp | |
| 598 pand %xmm4,%xmm12 | |
| 599 pand %xmm5,%xmm13 | |
| 600 pand %xmm6,%xmm14 | |
| 601 pand %xmm7,%xmm15 | |
| 602 por %xmm10,%xmm8 | |
| 603 por %xmm11,%xmm9 | |
| 604 por %xmm12,%xmm8 | |
| 605 por %xmm13,%xmm9 | |
| 606 por %xmm14,%xmm8 | |
| 607 por %xmm15,%xmm9 | |
| 608 | |
| 609 por %xmm9,%xmm8 | |
| 610 pshufd $0x4e,%xmm8,%xmm9 | |
| 611 por %xmm9,%xmm8 | |
| 612 .byte 102,76,15,126,195 | |
| 613 | |
| 614 mulq %rbx | |
| 615 addq %rax,%r8 | |
| 616 movq 8(%rsi),%rax | |
| 617 movq %r8,(%rdi) | |
| 618 movq %rdx,%r8 | |
| 619 adcq $0,%r8 | |
| 620 | |
| 621 mulq %rbx | |
| 622 addq %rax,%r9 | |
| 623 movq 16(%rsi),%rax | |
| 624 adcq $0,%rdx | |
| 625 addq %r9,%r8 | |
| 626 movq %rdx,%r9 | |
| 627 adcq $0,%r9 | |
| 628 | |
| 629 mulq %rbx | |
| 630 addq %rax,%r10 | |
| 631 movq 24(%rsi),%rax | |
| 632 adcq $0,%rdx | |
| 633 addq %r10,%r9 | |
| 634 movq %rdx,%r10 | |
| 635 adcq $0,%r10 | |
| 636 | |
| 637 mulq %rbx | |
| 638 addq %rax,%r11 | |
| 639 movq 32(%rsi),%rax | |
| 640 adcq $0,%rdx | |
| 641 addq %r11,%r10 | |
| 642 movq %rdx,%r11 | |
| 643 adcq $0,%r11 | |
| 644 | |
| 645 mulq %rbx | |
| 646 addq %rax,%r12 | |
| 647 movq 40(%rsi),%rax | |
| 648 adcq $0,%rdx | |
| 649 addq %r12,%r11 | |
| 650 movq %rdx,%r12 | |
| 651 adcq $0,%r12 | |
| 652 | |
| 653 mulq %rbx | |
| 654 addq %rax,%r13 | |
| 655 movq 48(%rsi),%rax | |
| 656 adcq $0,%rdx | |
| 657 addq %r13,%r12 | |
| 658 movq %rdx,%r13 | |
| 659 adcq $0,%r13 | |
| 660 | |
| 661 mulq %rbx | |
| 662 addq %rax,%r14 | |
| 663 movq 56(%rsi),%rax | |
| 664 adcq $0,%rdx | |
| 665 addq %r14,%r13 | |
| 666 movq %rdx,%r14 | |
| 667 adcq $0,%r14 | |
| 668 | |
| 669 mulq %rbx | |
| 670 addq %rax,%r15 | |
| 671 movq (%rsi),%rax | |
| 672 adcq $0,%rdx | |
| 673 addq %r15,%r14 | |
| 674 movq %rdx,%r15 | |
| 675 adcq $0,%r15 | |
| 676 | |
| 677 leaq 8(%rdi),%rdi | |
| 678 | |
| 679 decl %ecx | |
| 680 jnz L$oop_mul_gather | |
| 681 | |
| 682 movq %r8,(%rdi) | |
| 683 movq %r9,8(%rdi) | |
| 684 movq %r10,16(%rdi) | |
| 685 movq %r11,24(%rdi) | |
| 686 movq %r12,32(%rdi) | |
| 687 movq %r13,40(%rdi) | |
| 688 movq %r14,48(%rdi) | |
| 689 movq %r15,56(%rdi) | |
| 690 | |
| 691 movq 128+8(%rsp),%rdi | |
| 692 movq 128+16(%rsp),%rbp | |
| 693 | |
| 694 movq (%rsp),%r8 | |
| 695 movq 8(%rsp),%r9 | |
| 696 movq 16(%rsp),%r10 | |
| 697 movq 24(%rsp),%r11 | |
| 698 movq 32(%rsp),%r12 | |
| 699 movq 40(%rsp),%r13 | |
| 700 movq 48(%rsp),%r14 | |
| 701 movq 56(%rsp),%r15 | |
| 702 | |
| 703 call __rsaz_512_reduce | |
| 704 addq 64(%rsp),%r8 | |
| 705 adcq 72(%rsp),%r9 | |
| 706 adcq 80(%rsp),%r10 | |
| 707 adcq 88(%rsp),%r11 | |
| 708 adcq 96(%rsp),%r12 | |
| 709 adcq 104(%rsp),%r13 | |
| 710 adcq 112(%rsp),%r14 | |
| 711 adcq 120(%rsp),%r15 | |
| 712 sbbq %rcx,%rcx | |
| 713 | |
| 714 call __rsaz_512_subtract | |
| 715 | |
| 716 leaq 128+24+48(%rsp),%rax | |
| 717 movq -48(%rax),%r15 | |
| 718 movq -40(%rax),%r14 | |
| 719 movq -32(%rax),%r13 | |
| 720 movq -24(%rax),%r12 | |
| 721 movq -16(%rax),%rbp | |
| 722 movq -8(%rax),%rbx | |
| 723 leaq (%rax),%rsp | |
| 724 L$mul_gather4_epilogue: | |
| 725 .byte 0xf3,0xc3 | |
| 726 | |
| 727 .globl _rsaz_512_mul_scatter4 | |
| 728 .private_extern _rsaz_512_mul_scatter4 | |
| 729 | |
| 730 .p2align 5 | |
| 731 _rsaz_512_mul_scatter4: | |
| 732 pushq %rbx | |
| 733 pushq %rbp | |
| 734 pushq %r12 | |
| 735 pushq %r13 | |
| 736 pushq %r14 | |
| 737 pushq %r15 | |
| 738 | |
| 739 movl %r9d,%r9d | |
| 740 subq $128+24,%rsp | |
| 741 L$mul_scatter4_body: | |
| 742 leaq (%r8,%r9,8),%r8 | |
| 743 .byte 102,72,15,110,199 | |
| 744 .byte 102,72,15,110,202 | |
| 745 .byte 102,73,15,110,208 | |
| 746 movq %rcx,128(%rsp) | |
| 747 | |
| 748 movq %rdi,%rbp | |
| 749 movq (%rdi),%rbx | |
| 750 call __rsaz_512_mul | |
| 751 | |
| 752 .byte 102,72,15,126,199 | |
| 753 .byte 102,72,15,126,205 | |
| 754 | |
| 755 movq (%rsp),%r8 | |
| 756 movq 8(%rsp),%r9 | |
| 757 movq 16(%rsp),%r10 | |
| 758 movq 24(%rsp),%r11 | |
| 759 movq 32(%rsp),%r12 | |
| 760 movq 40(%rsp),%r13 | |
| 761 movq 48(%rsp),%r14 | |
| 762 movq 56(%rsp),%r15 | |
| 763 | |
| 764 call __rsaz_512_reduce | |
| 765 addq 64(%rsp),%r8 | |
| 766 adcq 72(%rsp),%r9 | |
| 767 adcq 80(%rsp),%r10 | |
| 768 adcq 88(%rsp),%r11 | |
| 769 adcq 96(%rsp),%r12 | |
| 770 adcq 104(%rsp),%r13 | |
| 771 adcq 112(%rsp),%r14 | |
| 772 adcq 120(%rsp),%r15 | |
| 773 .byte 102,72,15,126,214 | |
| 774 sbbq %rcx,%rcx | |
| 775 | |
| 776 call __rsaz_512_subtract | |
| 777 | |
| 778 movq %r8,0(%rsi) | |
| 779 movq %r9,128(%rsi) | |
| 780 movq %r10,256(%rsi) | |
| 781 movq %r11,384(%rsi) | |
| 782 movq %r12,512(%rsi) | |
| 783 movq %r13,640(%rsi) | |
| 784 movq %r14,768(%rsi) | |
| 785 movq %r15,896(%rsi) | |
| 786 | |
| 787 leaq 128+24+48(%rsp),%rax | |
| 788 movq -48(%rax),%r15 | |
| 789 movq -40(%rax),%r14 | |
| 790 movq -32(%rax),%r13 | |
| 791 movq -24(%rax),%r12 | |
| 792 movq -16(%rax),%rbp | |
| 793 movq -8(%rax),%rbx | |
| 794 leaq (%rax),%rsp | |
| 795 L$mul_scatter4_epilogue: | |
| 796 .byte 0xf3,0xc3 | |
| 797 | |
| 798 .globl _rsaz_512_mul_by_one | |
| 799 .private_extern _rsaz_512_mul_by_one | |
| 800 | |
| 801 .p2align 5 | |
| 802 _rsaz_512_mul_by_one: | |
| 803 pushq %rbx | |
| 804 pushq %rbp | |
| 805 pushq %r12 | |
| 806 pushq %r13 | |
| 807 pushq %r14 | |
| 808 pushq %r15 | |
| 809 | |
| 810 subq $128+24,%rsp | |
| 811 L$mul_by_one_body: | |
| 812 movq %rdx,%rbp | |
| 813 movq %rcx,128(%rsp) | |
| 814 | |
| 815 movq (%rsi),%r8 | |
| 816 pxor %xmm0,%xmm0 | |
| 817 movq 8(%rsi),%r9 | |
| 818 movq 16(%rsi),%r10 | |
| 819 movq 24(%rsi),%r11 | |
| 820 movq 32(%rsi),%r12 | |
| 821 movq 40(%rsi),%r13 | |
| 822 movq 48(%rsi),%r14 | |
| 823 movq 56(%rsi),%r15 | |
| 824 | |
| 825 movdqa %xmm0,(%rsp) | |
| 826 movdqa %xmm0,16(%rsp) | |
| 827 movdqa %xmm0,32(%rsp) | |
| 828 movdqa %xmm0,48(%rsp) | |
| 829 movdqa %xmm0,64(%rsp) | |
| 830 movdqa %xmm0,80(%rsp) | |
| 831 movdqa %xmm0,96(%rsp) | |
| 832 call __rsaz_512_reduce | |
| 833 movq %r8,(%rdi) | |
| 834 movq %r9,8(%rdi) | |
| 835 movq %r10,16(%rdi) | |
| 836 movq %r11,24(%rdi) | |
| 837 movq %r12,32(%rdi) | |
| 838 movq %r13,40(%rdi) | |
| 839 movq %r14,48(%rdi) | |
| 840 movq %r15,56(%rdi) | |
| 841 | |
| 842 leaq 128+24+48(%rsp),%rax | |
| 843 movq -48(%rax),%r15 | |
| 844 movq -40(%rax),%r14 | |
| 845 movq -32(%rax),%r13 | |
| 846 movq -24(%rax),%r12 | |
| 847 movq -16(%rax),%rbp | |
| 848 movq -8(%rax),%rbx | |
| 849 leaq (%rax),%rsp | |
| 850 L$mul_by_one_epilogue: | |
| 851 .byte 0xf3,0xc3 | |
| 852 | |
| 853 | |
| 854 .p2align 5 | |
| 855 __rsaz_512_reduce: | |
| 856 movq %r8,%rbx | |
| 857 imulq 128+8(%rsp),%rbx | |
| 858 movq 0(%rbp),%rax | |
| 859 movl $8,%ecx | |
| 860 jmp L$reduction_loop | |
| 861 | |
| 862 .p2align 5 | |
| 863 L$reduction_loop: | |
| 864 mulq %rbx | |
| 865 movq 8(%rbp),%rax | |
| 866 negq %r8 | |
| 867 movq %rdx,%r8 | |
| 868 adcq $0,%r8 | |
| 869 | |
| 870 mulq %rbx | |
| 871 addq %rax,%r9 | |
| 872 movq 16(%rbp),%rax | |
| 873 adcq $0,%rdx | |
| 874 addq %r9,%r8 | |
| 875 movq %rdx,%r9 | |
| 876 adcq $0,%r9 | |
| 877 | |
| 878 mulq %rbx | |
| 879 addq %rax,%r10 | |
| 880 movq 24(%rbp),%rax | |
| 881 adcq $0,%rdx | |
| 882 addq %r10,%r9 | |
| 883 movq %rdx,%r10 | |
| 884 adcq $0,%r10 | |
| 885 | |
| 886 mulq %rbx | |
| 887 addq %rax,%r11 | |
| 888 movq 32(%rbp),%rax | |
| 889 adcq $0,%rdx | |
| 890 addq %r11,%r10 | |
| 891 movq 128+8(%rsp),%rsi | |
| 892 | |
| 893 | |
| 894 adcq $0,%rdx | |
| 895 movq %rdx,%r11 | |
| 896 | |
| 897 mulq %rbx | |
| 898 addq %rax,%r12 | |
| 899 movq 40(%rbp),%rax | |
| 900 adcq $0,%rdx | |
| 901 imulq %r8,%rsi | |
| 902 addq %r12,%r11 | |
| 903 movq %rdx,%r12 | |
| 904 adcq $0,%r12 | |
| 905 | |
| 906 mulq %rbx | |
| 907 addq %rax,%r13 | |
| 908 movq 48(%rbp),%rax | |
| 909 adcq $0,%rdx | |
| 910 addq %r13,%r12 | |
| 911 movq %rdx,%r13 | |
| 912 adcq $0,%r13 | |
| 913 | |
| 914 mulq %rbx | |
| 915 addq %rax,%r14 | |
| 916 movq 56(%rbp),%rax | |
| 917 adcq $0,%rdx | |
| 918 addq %r14,%r13 | |
| 919 movq %rdx,%r14 | |
| 920 adcq $0,%r14 | |
| 921 | |
| 922 mulq %rbx | |
| 923 movq %rsi,%rbx | |
| 924 addq %rax,%r15 | |
| 925 movq 0(%rbp),%rax | |
| 926 adcq $0,%rdx | |
| 927 addq %r15,%r14 | |
| 928 movq %rdx,%r15 | |
| 929 adcq $0,%r15 | |
| 930 | |
| 931 decl %ecx | |
| 932 jne L$reduction_loop | |
| 933 | |
| 934 .byte 0xf3,0xc3 | |
| 935 | |
| 936 | |
| 937 .p2align 5 | |
| 938 __rsaz_512_subtract: | |
| 939 movq %r8,(%rdi) | |
| 940 movq %r9,8(%rdi) | |
| 941 movq %r10,16(%rdi) | |
| 942 movq %r11,24(%rdi) | |
| 943 movq %r12,32(%rdi) | |
| 944 movq %r13,40(%rdi) | |
| 945 movq %r14,48(%rdi) | |
| 946 movq %r15,56(%rdi) | |
| 947 | |
| 948 movq 0(%rbp),%r8 | |
| 949 movq 8(%rbp),%r9 | |
| 950 negq %r8 | |
| 951 notq %r9 | |
| 952 andq %rcx,%r8 | |
| 953 movq 16(%rbp),%r10 | |
| 954 andq %rcx,%r9 | |
| 955 notq %r10 | |
| 956 movq 24(%rbp),%r11 | |
| 957 andq %rcx,%r10 | |
| 958 notq %r11 | |
| 959 movq 32(%rbp),%r12 | |
| 960 andq %rcx,%r11 | |
| 961 notq %r12 | |
| 962 movq 40(%rbp),%r13 | |
| 963 andq %rcx,%r12 | |
| 964 notq %r13 | |
| 965 movq 48(%rbp),%r14 | |
| 966 andq %rcx,%r13 | |
| 967 notq %r14 | |
| 968 movq 56(%rbp),%r15 | |
| 969 andq %rcx,%r14 | |
| 970 notq %r15 | |
| 971 andq %rcx,%r15 | |
| 972 | |
| 973 addq (%rdi),%r8 | |
| 974 adcq 8(%rdi),%r9 | |
| 975 adcq 16(%rdi),%r10 | |
| 976 adcq 24(%rdi),%r11 | |
| 977 adcq 32(%rdi),%r12 | |
| 978 adcq 40(%rdi),%r13 | |
| 979 adcq 48(%rdi),%r14 | |
| 980 adcq 56(%rdi),%r15 | |
| 981 | |
| 982 movq %r8,(%rdi) | |
| 983 movq %r9,8(%rdi) | |
| 984 movq %r10,16(%rdi) | |
| 985 movq %r11,24(%rdi) | |
| 986 movq %r12,32(%rdi) | |
| 987 movq %r13,40(%rdi) | |
| 988 movq %r14,48(%rdi) | |
| 989 movq %r15,56(%rdi) | |
| 990 | |
| 991 .byte 0xf3,0xc3 | |
| 992 | |
| 993 | |
| 994 .p2align 5 | |
| 995 __rsaz_512_mul: | |
| 996 leaq 8(%rsp),%rdi | |
| 997 | |
| 998 movq (%rsi),%rax | |
| 999 mulq %rbx | |
| 1000 movq %rax,(%rdi) | |
| 1001 movq 8(%rsi),%rax | |
| 1002 movq %rdx,%r8 | |
| 1003 | |
| 1004 mulq %rbx | |
| 1005 addq %rax,%r8 | |
| 1006 movq 16(%rsi),%rax | |
| 1007 movq %rdx,%r9 | |
| 1008 adcq $0,%r9 | |
| 1009 | |
| 1010 mulq %rbx | |
| 1011 addq %rax,%r9 | |
| 1012 movq 24(%rsi),%rax | |
| 1013 movq %rdx,%r10 | |
| 1014 adcq $0,%r10 | |
| 1015 | |
| 1016 mulq %rbx | |
| 1017 addq %rax,%r10 | |
| 1018 movq 32(%rsi),%rax | |
| 1019 movq %rdx,%r11 | |
| 1020 adcq $0,%r11 | |
| 1021 | |
| 1022 mulq %rbx | |
| 1023 addq %rax,%r11 | |
| 1024 movq 40(%rsi),%rax | |
| 1025 movq %rdx,%r12 | |
| 1026 adcq $0,%r12 | |
| 1027 | |
| 1028 mulq %rbx | |
| 1029 addq %rax,%r12 | |
| 1030 movq 48(%rsi),%rax | |
| 1031 movq %rdx,%r13 | |
| 1032 adcq $0,%r13 | |
| 1033 | |
| 1034 mulq %rbx | |
| 1035 addq %rax,%r13 | |
| 1036 movq 56(%rsi),%rax | |
| 1037 movq %rdx,%r14 | |
| 1038 adcq $0,%r14 | |
| 1039 | |
| 1040 mulq %rbx | |
| 1041 addq %rax,%r14 | |
| 1042 movq (%rsi),%rax | |
| 1043 movq %rdx,%r15 | |
| 1044 adcq $0,%r15 | |
| 1045 | |
| 1046 leaq 8(%rbp),%rbp | |
| 1047 leaq 8(%rdi),%rdi | |
| 1048 | |
| 1049 movl $7,%ecx | |
| 1050 jmp L$oop_mul | |
| 1051 | |
| 1052 .p2align 5 | |
| 1053 L$oop_mul: | |
| 1054 movq (%rbp),%rbx | |
| 1055 mulq %rbx | |
| 1056 addq %rax,%r8 | |
| 1057 movq 8(%rsi),%rax | |
| 1058 movq %r8,(%rdi) | |
| 1059 movq %rdx,%r8 | |
| 1060 adcq $0,%r8 | |
| 1061 | |
| 1062 mulq %rbx | |
| 1063 addq %rax,%r9 | |
| 1064 movq 16(%rsi),%rax | |
| 1065 adcq $0,%rdx | |
| 1066 addq %r9,%r8 | |
| 1067 movq %rdx,%r9 | |
| 1068 adcq $0,%r9 | |
| 1069 | |
| 1070 mulq %rbx | |
| 1071 addq %rax,%r10 | |
| 1072 movq 24(%rsi),%rax | |
| 1073 adcq $0,%rdx | |
| 1074 addq %r10,%r9 | |
| 1075 movq %rdx,%r10 | |
| 1076 adcq $0,%r10 | |
| 1077 | |
| 1078 mulq %rbx | |
| 1079 addq %rax,%r11 | |
| 1080 movq 32(%rsi),%rax | |
| 1081 adcq $0,%rdx | |
| 1082 addq %r11,%r10 | |
| 1083 movq %rdx,%r11 | |
| 1084 adcq $0,%r11 | |
| 1085 | |
| 1086 mulq %rbx | |
| 1087 addq %rax,%r12 | |
| 1088 movq 40(%rsi),%rax | |
| 1089 adcq $0,%rdx | |
| 1090 addq %r12,%r11 | |
| 1091 movq %rdx,%r12 | |
| 1092 adcq $0,%r12 | |
| 1093 | |
| 1094 mulq %rbx | |
| 1095 addq %rax,%r13 | |
| 1096 movq 48(%rsi),%rax | |
| 1097 adcq $0,%rdx | |
| 1098 addq %r13,%r12 | |
| 1099 movq %rdx,%r13 | |
| 1100 adcq $0,%r13 | |
| 1101 | |
| 1102 mulq %rbx | |
| 1103 addq %rax,%r14 | |
| 1104 movq 56(%rsi),%rax | |
| 1105 adcq $0,%rdx | |
| 1106 addq %r14,%r13 | |
| 1107 movq %rdx,%r14 | |
| 1108 leaq 8(%rbp),%rbp | |
| 1109 adcq $0,%r14 | |
| 1110 | |
| 1111 mulq %rbx | |
| 1112 addq %rax,%r15 | |
| 1113 movq (%rsi),%rax | |
| 1114 adcq $0,%rdx | |
| 1115 addq %r15,%r14 | |
| 1116 movq %rdx,%r15 | |
| 1117 adcq $0,%r15 | |
| 1118 | |
| 1119 leaq 8(%rdi),%rdi | |
| 1120 | |
| 1121 decl %ecx | |
| 1122 jnz L$oop_mul | |
| 1123 | |
| 1124 movq %r8,(%rdi) | |
| 1125 movq %r9,8(%rdi) | |
| 1126 movq %r10,16(%rdi) | |
| 1127 movq %r11,24(%rdi) | |
| 1128 movq %r12,32(%rdi) | |
| 1129 movq %r13,40(%rdi) | |
| 1130 movq %r14,48(%rdi) | |
| 1131 movq %r15,56(%rdi) | |
| 1132 | |
| 1133 .byte 0xf3,0xc3 | |
| 1134 | |
| 1135 .globl _rsaz_512_scatter4 | |
| 1136 .private_extern _rsaz_512_scatter4 | |
| 1137 | |
| 1138 .p2align 4 | |
| 1139 _rsaz_512_scatter4: | |
| 1140 leaq (%rdi,%rdx,8),%rdi | |
| 1141 movl $8,%r9d | |
| 1142 jmp L$oop_scatter | |
| 1143 .p2align 4 | |
| 1144 L$oop_scatter: | |
| 1145 movq (%rsi),%rax | |
| 1146 leaq 8(%rsi),%rsi | |
| 1147 movq %rax,(%rdi) | |
| 1148 leaq 128(%rdi),%rdi | |
| 1149 decl %r9d | |
| 1150 jnz L$oop_scatter | |
| 1151 .byte 0xf3,0xc3 | |
| 1152 | |
| 1153 | |
| 1154 .globl _rsaz_512_gather4 | |
| 1155 .private_extern _rsaz_512_gather4 | |
| 1156 | |
| 1157 .p2align 4 | |
| 1158 _rsaz_512_gather4: | |
| 1159 movd %edx,%xmm8 | |
| 1160 movdqa L$inc+16(%rip),%xmm1 | |
| 1161 movdqa L$inc(%rip),%xmm0 | |
| 1162 | |
| 1163 pshufd $0,%xmm8,%xmm8 | |
| 1164 movdqa %xmm1,%xmm7 | |
| 1165 movdqa %xmm1,%xmm2 | |
| 1166 paddd %xmm0,%xmm1 | |
| 1167 pcmpeqd %xmm8,%xmm0 | |
| 1168 movdqa %xmm7,%xmm3 | |
| 1169 paddd %xmm1,%xmm2 | |
| 1170 pcmpeqd %xmm8,%xmm1 | |
| 1171 movdqa %xmm7,%xmm4 | |
| 1172 paddd %xmm2,%xmm3 | |
| 1173 pcmpeqd %xmm8,%xmm2 | |
| 1174 movdqa %xmm7,%xmm5 | |
| 1175 paddd %xmm3,%xmm4 | |
| 1176 pcmpeqd %xmm8,%xmm3 | |
| 1177 movdqa %xmm7,%xmm6 | |
| 1178 paddd %xmm4,%xmm5 | |
| 1179 pcmpeqd %xmm8,%xmm4 | |
| 1180 paddd %xmm5,%xmm6 | |
| 1181 pcmpeqd %xmm8,%xmm5 | |
| 1182 paddd %xmm6,%xmm7 | |
| 1183 pcmpeqd %xmm8,%xmm6 | |
| 1184 pcmpeqd %xmm8,%xmm7 | |
| 1185 movl $8,%r9d | |
| 1186 jmp L$oop_gather | |
| 1187 .p2align 4 | |
| 1188 L$oop_gather: | |
| 1189 movdqa 0(%rsi),%xmm8 | |
| 1190 movdqa 16(%rsi),%xmm9 | |
| 1191 movdqa 32(%rsi),%xmm10 | |
| 1192 movdqa 48(%rsi),%xmm11 | |
| 1193 pand %xmm0,%xmm8 | |
| 1194 movdqa 64(%rsi),%xmm12 | |
| 1195 pand %xmm1,%xmm9 | |
| 1196 movdqa 80(%rsi),%xmm13 | |
| 1197 pand %xmm2,%xmm10 | |
| 1198 movdqa 96(%rsi),%xmm14 | |
| 1199 pand %xmm3,%xmm11 | |
| 1200 movdqa 112(%rsi),%xmm15 | |
| 1201 leaq 128(%rsi),%rsi | |
| 1202 pand %xmm4,%xmm12 | |
| 1203 pand %xmm5,%xmm13 | |
| 1204 pand %xmm6,%xmm14 | |
| 1205 pand %xmm7,%xmm15 | |
| 1206 por %xmm10,%xmm8 | |
| 1207 por %xmm11,%xmm9 | |
| 1208 por %xmm12,%xmm8 | |
| 1209 por %xmm13,%xmm9 | |
| 1210 por %xmm14,%xmm8 | |
| 1211 por %xmm15,%xmm9 | |
| 1212 | |
| 1213 por %xmm9,%xmm8 | |
| 1214 pshufd $0x4e,%xmm8,%xmm9 | |
| 1215 por %xmm9,%xmm8 | |
| 1216 movq %xmm8,(%rdi) | |
| 1217 leaq 8(%rdi),%rdi | |
| 1218 decl %r9d | |
| 1219 jnz L$oop_gather | |
| 1220 .byte 0xf3,0xc3 | |
| 1221 L$SEH_end_rsaz_512_gather4: | |
| 1222 | |
| 1223 | |
| 1224 .p2align 6 | |
| 1225 L$inc: | |
| 1226 .long 0,0, 1,1 | |
| 1227 .long 2,2, 2,2 | |
| 1228 #endif | |
| OLD | NEW |