| OLD | NEW |
| (Empty) |
| 1 #if defined(__x86_64__) && !defined(OPENSSL_NO_ASM) | |
| 2 .text | |
| 3 | |
| 4 | |
| 5 | |
| 6 .p2align 6 | |
| 7 L$poly: | |
| 8 .quad 0xffffffffffffffff, 0x00000000ffffffff, 0x0000000000000000, 0xffffffff00
000001 | |
| 9 | |
| 10 L$One: | |
| 11 .long 1,1,1,1,1,1,1,1 | |
| 12 L$Two: | |
| 13 .long 2,2,2,2,2,2,2,2 | |
| 14 L$Three: | |
| 15 .long 3,3,3,3,3,3,3,3 | |
| 16 L$ONE_mont: | |
| 17 .quad 0x0000000000000001, 0xffffffff00000000, 0xffffffffffffffff, 0x00000000ff
fffffe | |
| 18 | |
| 19 | |
| 20 | |
| 21 .globl _ecp_nistz256_neg | |
| 22 .private_extern _ecp_nistz256_neg | |
| 23 | |
| 24 .p2align 5 | |
| 25 _ecp_nistz256_neg: | |
| 26 pushq %r12 | |
| 27 pushq %r13 | |
| 28 | |
| 29 xorq %r8,%r8 | |
| 30 xorq %r9,%r9 | |
| 31 xorq %r10,%r10 | |
| 32 xorq %r11,%r11 | |
| 33 xorq %r13,%r13 | |
| 34 | |
| 35 subq 0(%rsi),%r8 | |
| 36 sbbq 8(%rsi),%r9 | |
| 37 sbbq 16(%rsi),%r10 | |
| 38 movq %r8,%rax | |
| 39 sbbq 24(%rsi),%r11 | |
| 40 leaq L$poly(%rip),%rsi | |
| 41 movq %r9,%rdx | |
| 42 sbbq $0,%r13 | |
| 43 | |
| 44 addq 0(%rsi),%r8 | |
| 45 movq %r10,%rcx | |
| 46 adcq 8(%rsi),%r9 | |
| 47 adcq 16(%rsi),%r10 | |
| 48 movq %r11,%r12 | |
| 49 adcq 24(%rsi),%r11 | |
| 50 testq %r13,%r13 | |
| 51 | |
| 52 cmovzq %rax,%r8 | |
| 53 cmovzq %rdx,%r9 | |
| 54 movq %r8,0(%rdi) | |
| 55 cmovzq %rcx,%r10 | |
| 56 movq %r9,8(%rdi) | |
| 57 cmovzq %r12,%r11 | |
| 58 movq %r10,16(%rdi) | |
| 59 movq %r11,24(%rdi) | |
| 60 | |
| 61 popq %r13 | |
| 62 popq %r12 | |
| 63 .byte 0xf3,0xc3 | |
| 64 | |
| 65 | |
| 66 | |
| 67 | |
| 68 | |
| 69 | |
| 70 | |
| 71 .globl _ecp_nistz256_mul_mont | |
| 72 .private_extern _ecp_nistz256_mul_mont | |
| 73 | |
| 74 .p2align 5 | |
| 75 _ecp_nistz256_mul_mont: | |
| 76 L$mul_mont: | |
| 77 pushq %rbp | |
| 78 pushq %rbx | |
| 79 pushq %r12 | |
| 80 pushq %r13 | |
| 81 pushq %r14 | |
| 82 pushq %r15 | |
| 83 movq %rdx,%rbx | |
| 84 movq 0(%rdx),%rax | |
| 85 movq 0(%rsi),%r9 | |
| 86 movq 8(%rsi),%r10 | |
| 87 movq 16(%rsi),%r11 | |
| 88 movq 24(%rsi),%r12 | |
| 89 | |
| 90 call __ecp_nistz256_mul_montq | |
| 91 L$mul_mont_done: | |
| 92 popq %r15 | |
| 93 popq %r14 | |
| 94 popq %r13 | |
| 95 popq %r12 | |
| 96 popq %rbx | |
| 97 popq %rbp | |
| 98 .byte 0xf3,0xc3 | |
| 99 | |
| 100 | |
| 101 | |
| 102 .p2align 5 | |
| 103 __ecp_nistz256_mul_montq: | |
| 104 | |
| 105 | |
| 106 movq %rax,%rbp | |
| 107 mulq %r9 | |
| 108 movq L$poly+8(%rip),%r14 | |
| 109 movq %rax,%r8 | |
| 110 movq %rbp,%rax | |
| 111 movq %rdx,%r9 | |
| 112 | |
| 113 mulq %r10 | |
| 114 movq L$poly+24(%rip),%r15 | |
| 115 addq %rax,%r9 | |
| 116 movq %rbp,%rax | |
| 117 adcq $0,%rdx | |
| 118 movq %rdx,%r10 | |
| 119 | |
| 120 mulq %r11 | |
| 121 addq %rax,%r10 | |
| 122 movq %rbp,%rax | |
| 123 adcq $0,%rdx | |
| 124 movq %rdx,%r11 | |
| 125 | |
| 126 mulq %r12 | |
| 127 addq %rax,%r11 | |
| 128 movq %r8,%rax | |
| 129 adcq $0,%rdx | |
| 130 xorq %r13,%r13 | |
| 131 movq %rdx,%r12 | |
| 132 | |
| 133 | |
| 134 | |
| 135 | |
| 136 | |
| 137 | |
| 138 | |
| 139 | |
| 140 | |
| 141 | |
| 142 movq %r8,%rbp | |
| 143 shlq $32,%r8 | |
| 144 mulq %r15 | |
| 145 shrq $32,%rbp | |
| 146 addq %r8,%r9 | |
| 147 adcq %rbp,%r10 | |
| 148 adcq %rax,%r11 | |
| 149 movq 8(%rbx),%rax | |
| 150 adcq %rdx,%r12 | |
| 151 adcq $0,%r13 | |
| 152 xorq %r8,%r8 | |
| 153 | |
| 154 | |
| 155 | |
| 156 movq %rax,%rbp | |
| 157 mulq 0(%rsi) | |
| 158 addq %rax,%r9 | |
| 159 movq %rbp,%rax | |
| 160 adcq $0,%rdx | |
| 161 movq %rdx,%rcx | |
| 162 | |
| 163 mulq 8(%rsi) | |
| 164 addq %rcx,%r10 | |
| 165 adcq $0,%rdx | |
| 166 addq %rax,%r10 | |
| 167 movq %rbp,%rax | |
| 168 adcq $0,%rdx | |
| 169 movq %rdx,%rcx | |
| 170 | |
| 171 mulq 16(%rsi) | |
| 172 addq %rcx,%r11 | |
| 173 adcq $0,%rdx | |
| 174 addq %rax,%r11 | |
| 175 movq %rbp,%rax | |
| 176 adcq $0,%rdx | |
| 177 movq %rdx,%rcx | |
| 178 | |
| 179 mulq 24(%rsi) | |
| 180 addq %rcx,%r12 | |
| 181 adcq $0,%rdx | |
| 182 addq %rax,%r12 | |
| 183 movq %r9,%rax | |
| 184 adcq %rdx,%r13 | |
| 185 adcq $0,%r8 | |
| 186 | |
| 187 | |
| 188 | |
| 189 movq %r9,%rbp | |
| 190 shlq $32,%r9 | |
| 191 mulq %r15 | |
| 192 shrq $32,%rbp | |
| 193 addq %r9,%r10 | |
| 194 adcq %rbp,%r11 | |
| 195 adcq %rax,%r12 | |
| 196 movq 16(%rbx),%rax | |
| 197 adcq %rdx,%r13 | |
| 198 adcq $0,%r8 | |
| 199 xorq %r9,%r9 | |
| 200 | |
| 201 | |
| 202 | |
| 203 movq %rax,%rbp | |
| 204 mulq 0(%rsi) | |
| 205 addq %rax,%r10 | |
| 206 movq %rbp,%rax | |
| 207 adcq $0,%rdx | |
| 208 movq %rdx,%rcx | |
| 209 | |
| 210 mulq 8(%rsi) | |
| 211 addq %rcx,%r11 | |
| 212 adcq $0,%rdx | |
| 213 addq %rax,%r11 | |
| 214 movq %rbp,%rax | |
| 215 adcq $0,%rdx | |
| 216 movq %rdx,%rcx | |
| 217 | |
| 218 mulq 16(%rsi) | |
| 219 addq %rcx,%r12 | |
| 220 adcq $0,%rdx | |
| 221 addq %rax,%r12 | |
| 222 movq %rbp,%rax | |
| 223 adcq $0,%rdx | |
| 224 movq %rdx,%rcx | |
| 225 | |
| 226 mulq 24(%rsi) | |
| 227 addq %rcx,%r13 | |
| 228 adcq $0,%rdx | |
| 229 addq %rax,%r13 | |
| 230 movq %r10,%rax | |
| 231 adcq %rdx,%r8 | |
| 232 adcq $0,%r9 | |
| 233 | |
| 234 | |
| 235 | |
| 236 movq %r10,%rbp | |
| 237 shlq $32,%r10 | |
| 238 mulq %r15 | |
| 239 shrq $32,%rbp | |
| 240 addq %r10,%r11 | |
| 241 adcq %rbp,%r12 | |
| 242 adcq %rax,%r13 | |
| 243 movq 24(%rbx),%rax | |
| 244 adcq %rdx,%r8 | |
| 245 adcq $0,%r9 | |
| 246 xorq %r10,%r10 | |
| 247 | |
| 248 | |
| 249 | |
| 250 movq %rax,%rbp | |
| 251 mulq 0(%rsi) | |
| 252 addq %rax,%r11 | |
| 253 movq %rbp,%rax | |
| 254 adcq $0,%rdx | |
| 255 movq %rdx,%rcx | |
| 256 | |
| 257 mulq 8(%rsi) | |
| 258 addq %rcx,%r12 | |
| 259 adcq $0,%rdx | |
| 260 addq %rax,%r12 | |
| 261 movq %rbp,%rax | |
| 262 adcq $0,%rdx | |
| 263 movq %rdx,%rcx | |
| 264 | |
| 265 mulq 16(%rsi) | |
| 266 addq %rcx,%r13 | |
| 267 adcq $0,%rdx | |
| 268 addq %rax,%r13 | |
| 269 movq %rbp,%rax | |
| 270 adcq $0,%rdx | |
| 271 movq %rdx,%rcx | |
| 272 | |
| 273 mulq 24(%rsi) | |
| 274 addq %rcx,%r8 | |
| 275 adcq $0,%rdx | |
| 276 addq %rax,%r8 | |
| 277 movq %r11,%rax | |
| 278 adcq %rdx,%r9 | |
| 279 adcq $0,%r10 | |
| 280 | |
| 281 | |
| 282 | |
| 283 movq %r11,%rbp | |
| 284 shlq $32,%r11 | |
| 285 mulq %r15 | |
| 286 shrq $32,%rbp | |
| 287 addq %r11,%r12 | |
| 288 adcq %rbp,%r13 | |
| 289 movq %r12,%rcx | |
| 290 adcq %rax,%r8 | |
| 291 adcq %rdx,%r9 | |
| 292 movq %r13,%rbp | |
| 293 adcq $0,%r10 | |
| 294 | |
| 295 | |
| 296 | |
| 297 subq $-1,%r12 | |
| 298 movq %r8,%rbx | |
| 299 sbbq %r14,%r13 | |
| 300 sbbq $0,%r8 | |
| 301 movq %r9,%rdx | |
| 302 sbbq %r15,%r9 | |
| 303 sbbq $0,%r10 | |
| 304 | |
| 305 cmovcq %rcx,%r12 | |
| 306 cmovcq %rbp,%r13 | |
| 307 movq %r12,0(%rdi) | |
| 308 cmovcq %rbx,%r8 | |
| 309 movq %r13,8(%rdi) | |
| 310 cmovcq %rdx,%r9 | |
| 311 movq %r8,16(%rdi) | |
| 312 movq %r9,24(%rdi) | |
| 313 | |
| 314 .byte 0xf3,0xc3 | |
| 315 | |
| 316 | |
| 317 | |
| 318 | |
| 319 | |
| 320 | |
| 321 | |
| 322 | |
| 323 | |
| 324 .globl _ecp_nistz256_sqr_mont | |
| 325 .private_extern _ecp_nistz256_sqr_mont | |
| 326 | |
| 327 .p2align 5 | |
| 328 _ecp_nistz256_sqr_mont: | |
| 329 pushq %rbp | |
| 330 pushq %rbx | |
| 331 pushq %r12 | |
| 332 pushq %r13 | |
| 333 pushq %r14 | |
| 334 pushq %r15 | |
| 335 movq 0(%rsi),%rax | |
| 336 movq 8(%rsi),%r14 | |
| 337 movq 16(%rsi),%r15 | |
| 338 movq 24(%rsi),%r8 | |
| 339 | |
| 340 call __ecp_nistz256_sqr_montq | |
| 341 L$sqr_mont_done: | |
| 342 popq %r15 | |
| 343 popq %r14 | |
| 344 popq %r13 | |
| 345 popq %r12 | |
| 346 popq %rbx | |
| 347 popq %rbp | |
| 348 .byte 0xf3,0xc3 | |
| 349 | |
| 350 | |
| 351 | |
| 352 .p2align 5 | |
| 353 __ecp_nistz256_sqr_montq: | |
| 354 movq %rax,%r13 | |
| 355 mulq %r14 | |
| 356 movq %rax,%r9 | |
| 357 movq %r15,%rax | |
| 358 movq %rdx,%r10 | |
| 359 | |
| 360 mulq %r13 | |
| 361 addq %rax,%r10 | |
| 362 movq %r8,%rax | |
| 363 adcq $0,%rdx | |
| 364 movq %rdx,%r11 | |
| 365 | |
| 366 mulq %r13 | |
| 367 addq %rax,%r11 | |
| 368 movq %r15,%rax | |
| 369 adcq $0,%rdx | |
| 370 movq %rdx,%r12 | |
| 371 | |
| 372 | |
| 373 mulq %r14 | |
| 374 addq %rax,%r11 | |
| 375 movq %r8,%rax | |
| 376 adcq $0,%rdx | |
| 377 movq %rdx,%rbp | |
| 378 | |
| 379 mulq %r14 | |
| 380 addq %rax,%r12 | |
| 381 movq %r8,%rax | |
| 382 adcq $0,%rdx | |
| 383 addq %rbp,%r12 | |
| 384 movq %rdx,%r13 | |
| 385 adcq $0,%r13 | |
| 386 | |
| 387 | |
| 388 mulq %r15 | |
| 389 xorq %r15,%r15 | |
| 390 addq %rax,%r13 | |
| 391 movq 0(%rsi),%rax | |
| 392 movq %rdx,%r14 | |
| 393 adcq $0,%r14 | |
| 394 | |
| 395 addq %r9,%r9 | |
| 396 adcq %r10,%r10 | |
| 397 adcq %r11,%r11 | |
| 398 adcq %r12,%r12 | |
| 399 adcq %r13,%r13 | |
| 400 adcq %r14,%r14 | |
| 401 adcq $0,%r15 | |
| 402 | |
| 403 mulq %rax | |
| 404 movq %rax,%r8 | |
| 405 movq 8(%rsi),%rax | |
| 406 movq %rdx,%rcx | |
| 407 | |
| 408 mulq %rax | |
| 409 addq %rcx,%r9 | |
| 410 adcq %rax,%r10 | |
| 411 movq 16(%rsi),%rax | |
| 412 adcq $0,%rdx | |
| 413 movq %rdx,%rcx | |
| 414 | |
| 415 mulq %rax | |
| 416 addq %rcx,%r11 | |
| 417 adcq %rax,%r12 | |
| 418 movq 24(%rsi),%rax | |
| 419 adcq $0,%rdx | |
| 420 movq %rdx,%rcx | |
| 421 | |
| 422 mulq %rax | |
| 423 addq %rcx,%r13 | |
| 424 adcq %rax,%r14 | |
| 425 movq %r8,%rax | |
| 426 adcq %rdx,%r15 | |
| 427 | |
| 428 movq L$poly+8(%rip),%rsi | |
| 429 movq L$poly+24(%rip),%rbp | |
| 430 | |
| 431 | |
| 432 | |
| 433 | |
| 434 movq %r8,%rcx | |
| 435 shlq $32,%r8 | |
| 436 mulq %rbp | |
| 437 shrq $32,%rcx | |
| 438 addq %r8,%r9 | |
| 439 adcq %rcx,%r10 | |
| 440 adcq %rax,%r11 | |
| 441 movq %r9,%rax | |
| 442 adcq $0,%rdx | |
| 443 | |
| 444 | |
| 445 | |
| 446 movq %r9,%rcx | |
| 447 shlq $32,%r9 | |
| 448 movq %rdx,%r8 | |
| 449 mulq %rbp | |
| 450 shrq $32,%rcx | |
| 451 addq %r9,%r10 | |
| 452 adcq %rcx,%r11 | |
| 453 adcq %rax,%r8 | |
| 454 movq %r10,%rax | |
| 455 adcq $0,%rdx | |
| 456 | |
| 457 | |
| 458 | |
| 459 movq %r10,%rcx | |
| 460 shlq $32,%r10 | |
| 461 movq %rdx,%r9 | |
| 462 mulq %rbp | |
| 463 shrq $32,%rcx | |
| 464 addq %r10,%r11 | |
| 465 adcq %rcx,%r8 | |
| 466 adcq %rax,%r9 | |
| 467 movq %r11,%rax | |
| 468 adcq $0,%rdx | |
| 469 | |
| 470 | |
| 471 | |
| 472 movq %r11,%rcx | |
| 473 shlq $32,%r11 | |
| 474 movq %rdx,%r10 | |
| 475 mulq %rbp | |
| 476 shrq $32,%rcx | |
| 477 addq %r11,%r8 | |
| 478 adcq %rcx,%r9 | |
| 479 adcq %rax,%r10 | |
| 480 adcq $0,%rdx | |
| 481 xorq %r11,%r11 | |
| 482 | |
| 483 | |
| 484 | |
| 485 addq %r8,%r12 | |
| 486 adcq %r9,%r13 | |
| 487 movq %r12,%r8 | |
| 488 adcq %r10,%r14 | |
| 489 adcq %rdx,%r15 | |
| 490 movq %r13,%r9 | |
| 491 adcq $0,%r11 | |
| 492 | |
| 493 subq $-1,%r12 | |
| 494 movq %r14,%r10 | |
| 495 sbbq %rsi,%r13 | |
| 496 sbbq $0,%r14 | |
| 497 movq %r15,%rcx | |
| 498 sbbq %rbp,%r15 | |
| 499 sbbq $0,%r11 | |
| 500 | |
| 501 cmovcq %r8,%r12 | |
| 502 cmovcq %r9,%r13 | |
| 503 movq %r12,0(%rdi) | |
| 504 cmovcq %r10,%r14 | |
| 505 movq %r13,8(%rdi) | |
| 506 cmovcq %rcx,%r15 | |
| 507 movq %r14,16(%rdi) | |
| 508 movq %r15,24(%rdi) | |
| 509 | |
| 510 .byte 0xf3,0xc3 | |
| 511 | |
| 512 | |
| 513 | |
| 514 .globl _ecp_nistz256_select_w5 | |
| 515 .private_extern _ecp_nistz256_select_w5 | |
| 516 | |
| 517 .p2align 5 | |
| 518 _ecp_nistz256_select_w5: | |
| 519 movdqa L$One(%rip),%xmm0 | |
| 520 movd %edx,%xmm1 | |
| 521 | |
| 522 pxor %xmm2,%xmm2 | |
| 523 pxor %xmm3,%xmm3 | |
| 524 pxor %xmm4,%xmm4 | |
| 525 pxor %xmm5,%xmm5 | |
| 526 pxor %xmm6,%xmm6 | |
| 527 pxor %xmm7,%xmm7 | |
| 528 | |
| 529 movdqa %xmm0,%xmm8 | |
| 530 pshufd $0,%xmm1,%xmm1 | |
| 531 | |
| 532 movq $16,%rax | |
| 533 L$select_loop_sse_w5: | |
| 534 | |
| 535 movdqa %xmm8,%xmm15 | |
| 536 paddd %xmm0,%xmm8 | |
| 537 pcmpeqd %xmm1,%xmm15 | |
| 538 | |
| 539 movdqa 0(%rsi),%xmm9 | |
| 540 movdqa 16(%rsi),%xmm10 | |
| 541 movdqa 32(%rsi),%xmm11 | |
| 542 movdqa 48(%rsi),%xmm12 | |
| 543 movdqa 64(%rsi),%xmm13 | |
| 544 movdqa 80(%rsi),%xmm14 | |
| 545 leaq 96(%rsi),%rsi | |
| 546 | |
| 547 pand %xmm15,%xmm9 | |
| 548 pand %xmm15,%xmm10 | |
| 549 por %xmm9,%xmm2 | |
| 550 pand %xmm15,%xmm11 | |
| 551 por %xmm10,%xmm3 | |
| 552 pand %xmm15,%xmm12 | |
| 553 por %xmm11,%xmm4 | |
| 554 pand %xmm15,%xmm13 | |
| 555 por %xmm12,%xmm5 | |
| 556 pand %xmm15,%xmm14 | |
| 557 por %xmm13,%xmm6 | |
| 558 por %xmm14,%xmm7 | |
| 559 | |
| 560 decq %rax | |
| 561 jnz L$select_loop_sse_w5 | |
| 562 | |
| 563 movdqu %xmm2,0(%rdi) | |
| 564 movdqu %xmm3,16(%rdi) | |
| 565 movdqu %xmm4,32(%rdi) | |
| 566 movdqu %xmm5,48(%rdi) | |
| 567 movdqu %xmm6,64(%rdi) | |
| 568 movdqu %xmm7,80(%rdi) | |
| 569 .byte 0xf3,0xc3 | |
| 570 | |
| 571 | |
| 572 | |
| 573 | |
| 574 .globl _ecp_nistz256_select_w7 | |
| 575 .private_extern _ecp_nistz256_select_w7 | |
| 576 | |
| 577 .p2align 5 | |
| 578 _ecp_nistz256_select_w7: | |
| 579 movdqa L$One(%rip),%xmm8 | |
| 580 movd %edx,%xmm1 | |
| 581 | |
| 582 pxor %xmm2,%xmm2 | |
| 583 pxor %xmm3,%xmm3 | |
| 584 pxor %xmm4,%xmm4 | |
| 585 pxor %xmm5,%xmm5 | |
| 586 | |
| 587 movdqa %xmm8,%xmm0 | |
| 588 pshufd $0,%xmm1,%xmm1 | |
| 589 movq $64,%rax | |
| 590 | |
| 591 L$select_loop_sse_w7: | |
| 592 movdqa %xmm8,%xmm15 | |
| 593 paddd %xmm0,%xmm8 | |
| 594 movdqa 0(%rsi),%xmm9 | |
| 595 movdqa 16(%rsi),%xmm10 | |
| 596 pcmpeqd %xmm1,%xmm15 | |
| 597 movdqa 32(%rsi),%xmm11 | |
| 598 movdqa 48(%rsi),%xmm12 | |
| 599 leaq 64(%rsi),%rsi | |
| 600 | |
| 601 pand %xmm15,%xmm9 | |
| 602 pand %xmm15,%xmm10 | |
| 603 por %xmm9,%xmm2 | |
| 604 pand %xmm15,%xmm11 | |
| 605 por %xmm10,%xmm3 | |
| 606 pand %xmm15,%xmm12 | |
| 607 por %xmm11,%xmm4 | |
| 608 prefetcht0 255(%rsi) | |
| 609 por %xmm12,%xmm5 | |
| 610 | |
| 611 decq %rax | |
| 612 jnz L$select_loop_sse_w7 | |
| 613 | |
| 614 movdqu %xmm2,0(%rdi) | |
| 615 movdqu %xmm3,16(%rdi) | |
| 616 movdqu %xmm4,32(%rdi) | |
| 617 movdqu %xmm5,48(%rdi) | |
| 618 .byte 0xf3,0xc3 | |
| 619 | |
| 620 .globl _ecp_nistz256_avx2_select_w7 | |
| 621 .private_extern _ecp_nistz256_avx2_select_w7 | |
| 622 | |
| 623 .p2align 5 | |
| 624 _ecp_nistz256_avx2_select_w7: | |
| 625 .byte 0x0f,0x0b | |
| 626 .byte 0xf3,0xc3 | |
| 627 | |
| 628 | |
| 629 .p2align 5 | |
| 630 __ecp_nistz256_add_toq: | |
| 631 xorq %r11,%r11 | |
| 632 addq 0(%rbx),%r12 | |
| 633 adcq 8(%rbx),%r13 | |
| 634 movq %r12,%rax | |
| 635 adcq 16(%rbx),%r8 | |
| 636 adcq 24(%rbx),%r9 | |
| 637 movq %r13,%rbp | |
| 638 adcq $0,%r11 | |
| 639 | |
| 640 subq $-1,%r12 | |
| 641 movq %r8,%rcx | |
| 642 sbbq %r14,%r13 | |
| 643 sbbq $0,%r8 | |
| 644 movq %r9,%r10 | |
| 645 sbbq %r15,%r9 | |
| 646 sbbq $0,%r11 | |
| 647 | |
| 648 cmovcq %rax,%r12 | |
| 649 cmovcq %rbp,%r13 | |
| 650 movq %r12,0(%rdi) | |
| 651 cmovcq %rcx,%r8 | |
| 652 movq %r13,8(%rdi) | |
| 653 cmovcq %r10,%r9 | |
| 654 movq %r8,16(%rdi) | |
| 655 movq %r9,24(%rdi) | |
| 656 | |
| 657 .byte 0xf3,0xc3 | |
| 658 | |
| 659 | |
| 660 | |
| 661 .p2align 5 | |
| 662 __ecp_nistz256_sub_fromq: | |
| 663 subq 0(%rbx),%r12 | |
| 664 sbbq 8(%rbx),%r13 | |
| 665 movq %r12,%rax | |
| 666 sbbq 16(%rbx),%r8 | |
| 667 sbbq 24(%rbx),%r9 | |
| 668 movq %r13,%rbp | |
| 669 sbbq %r11,%r11 | |
| 670 | |
| 671 addq $-1,%r12 | |
| 672 movq %r8,%rcx | |
| 673 adcq %r14,%r13 | |
| 674 adcq $0,%r8 | |
| 675 movq %r9,%r10 | |
| 676 adcq %r15,%r9 | |
| 677 testq %r11,%r11 | |
| 678 | |
| 679 cmovzq %rax,%r12 | |
| 680 cmovzq %rbp,%r13 | |
| 681 movq %r12,0(%rdi) | |
| 682 cmovzq %rcx,%r8 | |
| 683 movq %r13,8(%rdi) | |
| 684 cmovzq %r10,%r9 | |
| 685 movq %r8,16(%rdi) | |
| 686 movq %r9,24(%rdi) | |
| 687 | |
| 688 .byte 0xf3,0xc3 | |
| 689 | |
| 690 | |
| 691 | |
| 692 .p2align 5 | |
| 693 __ecp_nistz256_subq: | |
| 694 subq %r12,%rax | |
| 695 sbbq %r13,%rbp | |
| 696 movq %rax,%r12 | |
| 697 sbbq %r8,%rcx | |
| 698 sbbq %r9,%r10 | |
| 699 movq %rbp,%r13 | |
| 700 sbbq %r11,%r11 | |
| 701 | |
| 702 addq $-1,%rax | |
| 703 movq %rcx,%r8 | |
| 704 adcq %r14,%rbp | |
| 705 adcq $0,%rcx | |
| 706 movq %r10,%r9 | |
| 707 adcq %r15,%r10 | |
| 708 testq %r11,%r11 | |
| 709 | |
| 710 cmovnzq %rax,%r12 | |
| 711 cmovnzq %rbp,%r13 | |
| 712 cmovnzq %rcx,%r8 | |
| 713 cmovnzq %r10,%r9 | |
| 714 | |
| 715 .byte 0xf3,0xc3 | |
| 716 | |
| 717 | |
| 718 | |
| 719 .p2align 5 | |
| 720 __ecp_nistz256_mul_by_2q: | |
| 721 xorq %r11,%r11 | |
| 722 addq %r12,%r12 | |
| 723 adcq %r13,%r13 | |
| 724 movq %r12,%rax | |
| 725 adcq %r8,%r8 | |
| 726 adcq %r9,%r9 | |
| 727 movq %r13,%rbp | |
| 728 adcq $0,%r11 | |
| 729 | |
| 730 subq $-1,%r12 | |
| 731 movq %r8,%rcx | |
| 732 sbbq %r14,%r13 | |
| 733 sbbq $0,%r8 | |
| 734 movq %r9,%r10 | |
| 735 sbbq %r15,%r9 | |
| 736 sbbq $0,%r11 | |
| 737 | |
| 738 cmovcq %rax,%r12 | |
| 739 cmovcq %rbp,%r13 | |
| 740 movq %r12,0(%rdi) | |
| 741 cmovcq %rcx,%r8 | |
| 742 movq %r13,8(%rdi) | |
| 743 cmovcq %r10,%r9 | |
| 744 movq %r8,16(%rdi) | |
| 745 movq %r9,24(%rdi) | |
| 746 | |
| 747 .byte 0xf3,0xc3 | |
| 748 | |
| 749 .globl _ecp_nistz256_point_double | |
| 750 .private_extern _ecp_nistz256_point_double | |
| 751 | |
| 752 .p2align 5 | |
| 753 _ecp_nistz256_point_double: | |
| 754 pushq %rbp | |
| 755 pushq %rbx | |
| 756 pushq %r12 | |
| 757 pushq %r13 | |
| 758 pushq %r14 | |
| 759 pushq %r15 | |
| 760 subq $160+8,%rsp | |
| 761 | |
| 762 L$point_double_shortcutq: | |
| 763 movdqu 0(%rsi),%xmm0 | |
| 764 movq %rsi,%rbx | |
| 765 movdqu 16(%rsi),%xmm1 | |
| 766 movq 32+0(%rsi),%r12 | |
| 767 movq 32+8(%rsi),%r13 | |
| 768 movq 32+16(%rsi),%r8 | |
| 769 movq 32+24(%rsi),%r9 | |
| 770 movq L$poly+8(%rip),%r14 | |
| 771 movq L$poly+24(%rip),%r15 | |
| 772 movdqa %xmm0,96(%rsp) | |
| 773 movdqa %xmm1,96+16(%rsp) | |
| 774 leaq 32(%rdi),%r10 | |
| 775 leaq 64(%rdi),%r11 | |
| 776 .byte 102,72,15,110,199 | |
| 777 .byte 102,73,15,110,202 | |
| 778 .byte 102,73,15,110,211 | |
| 779 | |
| 780 leaq 0(%rsp),%rdi | |
| 781 call __ecp_nistz256_mul_by_2q | |
| 782 | |
| 783 movq 64+0(%rsi),%rax | |
| 784 movq 64+8(%rsi),%r14 | |
| 785 movq 64+16(%rsi),%r15 | |
| 786 movq 64+24(%rsi),%r8 | |
| 787 leaq 64-0(%rsi),%rsi | |
| 788 leaq 64(%rsp),%rdi | |
| 789 call __ecp_nistz256_sqr_montq | |
| 790 | |
| 791 movq 0+0(%rsp),%rax | |
| 792 movq 8+0(%rsp),%r14 | |
| 793 leaq 0+0(%rsp),%rsi | |
| 794 movq 16+0(%rsp),%r15 | |
| 795 movq 24+0(%rsp),%r8 | |
| 796 leaq 0(%rsp),%rdi | |
| 797 call __ecp_nistz256_sqr_montq | |
| 798 | |
| 799 movq 32(%rbx),%rax | |
| 800 movq 64+0(%rbx),%r9 | |
| 801 movq 64+8(%rbx),%r10 | |
| 802 movq 64+16(%rbx),%r11 | |
| 803 movq 64+24(%rbx),%r12 | |
| 804 leaq 64-0(%rbx),%rsi | |
| 805 leaq 32(%rbx),%rbx | |
| 806 .byte 102,72,15,126,215 | |
| 807 call __ecp_nistz256_mul_montq | |
| 808 call __ecp_nistz256_mul_by_2q | |
| 809 | |
| 810 movq 96+0(%rsp),%r12 | |
| 811 movq 96+8(%rsp),%r13 | |
| 812 leaq 64(%rsp),%rbx | |
| 813 movq 96+16(%rsp),%r8 | |
| 814 movq 96+24(%rsp),%r9 | |
| 815 leaq 32(%rsp),%rdi | |
| 816 call __ecp_nistz256_add_toq | |
| 817 | |
| 818 movq 96+0(%rsp),%r12 | |
| 819 movq 96+8(%rsp),%r13 | |
| 820 leaq 64(%rsp),%rbx | |
| 821 movq 96+16(%rsp),%r8 | |
| 822 movq 96+24(%rsp),%r9 | |
| 823 leaq 64(%rsp),%rdi | |
| 824 call __ecp_nistz256_sub_fromq | |
| 825 | |
| 826 movq 0+0(%rsp),%rax | |
| 827 movq 8+0(%rsp),%r14 | |
| 828 leaq 0+0(%rsp),%rsi | |
| 829 movq 16+0(%rsp),%r15 | |
| 830 movq 24+0(%rsp),%r8 | |
| 831 .byte 102,72,15,126,207 | |
| 832 call __ecp_nistz256_sqr_montq | |
| 833 xorq %r9,%r9 | |
| 834 movq %r12,%rax | |
| 835 addq $-1,%r12 | |
| 836 movq %r13,%r10 | |
| 837 adcq %rsi,%r13 | |
| 838 movq %r14,%rcx | |
| 839 adcq $0,%r14 | |
| 840 movq %r15,%r8 | |
| 841 adcq %rbp,%r15 | |
| 842 adcq $0,%r9 | |
| 843 xorq %rsi,%rsi | |
| 844 testq $1,%rax | |
| 845 | |
| 846 cmovzq %rax,%r12 | |
| 847 cmovzq %r10,%r13 | |
| 848 cmovzq %rcx,%r14 | |
| 849 cmovzq %r8,%r15 | |
| 850 cmovzq %rsi,%r9 | |
| 851 | |
| 852 movq %r13,%rax | |
| 853 shrq $1,%r12 | |
| 854 shlq $63,%rax | |
| 855 movq %r14,%r10 | |
| 856 shrq $1,%r13 | |
| 857 orq %rax,%r12 | |
| 858 shlq $63,%r10 | |
| 859 movq %r15,%rcx | |
| 860 shrq $1,%r14 | |
| 861 orq %r10,%r13 | |
| 862 shlq $63,%rcx | |
| 863 movq %r12,0(%rdi) | |
| 864 shrq $1,%r15 | |
| 865 movq %r13,8(%rdi) | |
| 866 shlq $63,%r9 | |
| 867 orq %rcx,%r14 | |
| 868 orq %r9,%r15 | |
| 869 movq %r14,16(%rdi) | |
| 870 movq %r15,24(%rdi) | |
| 871 movq 64(%rsp),%rax | |
| 872 leaq 64(%rsp),%rbx | |
| 873 movq 0+32(%rsp),%r9 | |
| 874 movq 8+32(%rsp),%r10 | |
| 875 leaq 0+32(%rsp),%rsi | |
| 876 movq 16+32(%rsp),%r11 | |
| 877 movq 24+32(%rsp),%r12 | |
| 878 leaq 32(%rsp),%rdi | |
| 879 call __ecp_nistz256_mul_montq | |
| 880 | |
| 881 leaq 128(%rsp),%rdi | |
| 882 call __ecp_nistz256_mul_by_2q | |
| 883 | |
| 884 leaq 32(%rsp),%rbx | |
| 885 leaq 32(%rsp),%rdi | |
| 886 call __ecp_nistz256_add_toq | |
| 887 | |
| 888 movq 96(%rsp),%rax | |
| 889 leaq 96(%rsp),%rbx | |
| 890 movq 0+0(%rsp),%r9 | |
| 891 movq 8+0(%rsp),%r10 | |
| 892 leaq 0+0(%rsp),%rsi | |
| 893 movq 16+0(%rsp),%r11 | |
| 894 movq 24+0(%rsp),%r12 | |
| 895 leaq 0(%rsp),%rdi | |
| 896 call __ecp_nistz256_mul_montq | |
| 897 | |
| 898 leaq 128(%rsp),%rdi | |
| 899 call __ecp_nistz256_mul_by_2q | |
| 900 | |
| 901 movq 0+32(%rsp),%rax | |
| 902 movq 8+32(%rsp),%r14 | |
| 903 leaq 0+32(%rsp),%rsi | |
| 904 movq 16+32(%rsp),%r15 | |
| 905 movq 24+32(%rsp),%r8 | |
| 906 .byte 102,72,15,126,199 | |
| 907 call __ecp_nistz256_sqr_montq | |
| 908 | |
| 909 leaq 128(%rsp),%rbx | |
| 910 movq %r14,%r8 | |
| 911 movq %r15,%r9 | |
| 912 movq %rsi,%r14 | |
| 913 movq %rbp,%r15 | |
| 914 call __ecp_nistz256_sub_fromq | |
| 915 | |
| 916 movq 0+0(%rsp),%rax | |
| 917 movq 0+8(%rsp),%rbp | |
| 918 movq 0+16(%rsp),%rcx | |
| 919 movq 0+24(%rsp),%r10 | |
| 920 leaq 0(%rsp),%rdi | |
| 921 call __ecp_nistz256_subq | |
| 922 | |
| 923 movq 32(%rsp),%rax | |
| 924 leaq 32(%rsp),%rbx | |
| 925 movq %r12,%r14 | |
| 926 xorl %ecx,%ecx | |
| 927 movq %r12,0+0(%rsp) | |
| 928 movq %r13,%r10 | |
| 929 movq %r13,0+8(%rsp) | |
| 930 cmovzq %r8,%r11 | |
| 931 movq %r8,0+16(%rsp) | |
| 932 leaq 0-0(%rsp),%rsi | |
| 933 cmovzq %r9,%r12 | |
| 934 movq %r9,0+24(%rsp) | |
| 935 movq %r14,%r9 | |
| 936 leaq 0(%rsp),%rdi | |
| 937 call __ecp_nistz256_mul_montq | |
| 938 | |
| 939 .byte 102,72,15,126,203 | |
| 940 .byte 102,72,15,126,207 | |
| 941 call __ecp_nistz256_sub_fromq | |
| 942 | |
| 943 addq $160+8,%rsp | |
| 944 popq %r15 | |
| 945 popq %r14 | |
| 946 popq %r13 | |
| 947 popq %r12 | |
| 948 popq %rbx | |
| 949 popq %rbp | |
| 950 .byte 0xf3,0xc3 | |
| 951 | |
| 952 .globl _ecp_nistz256_point_add | |
| 953 .private_extern _ecp_nistz256_point_add | |
| 954 | |
| 955 .p2align 5 | |
| 956 _ecp_nistz256_point_add: | |
| 957 pushq %rbp | |
| 958 pushq %rbx | |
| 959 pushq %r12 | |
| 960 pushq %r13 | |
| 961 pushq %r14 | |
| 962 pushq %r15 | |
| 963 subq $576+8,%rsp | |
| 964 | |
| 965 movdqu 0(%rsi),%xmm0 | |
| 966 movdqu 16(%rsi),%xmm1 | |
| 967 movdqu 32(%rsi),%xmm2 | |
| 968 movdqu 48(%rsi),%xmm3 | |
| 969 movdqu 64(%rsi),%xmm4 | |
| 970 movdqu 80(%rsi),%xmm5 | |
| 971 movq %rsi,%rbx | |
| 972 movq %rdx,%rsi | |
| 973 movdqa %xmm0,384(%rsp) | |
| 974 movdqa %xmm1,384+16(%rsp) | |
| 975 movdqa %xmm2,416(%rsp) | |
| 976 movdqa %xmm3,416+16(%rsp) | |
| 977 movdqa %xmm4,448(%rsp) | |
| 978 movdqa %xmm5,448+16(%rsp) | |
| 979 por %xmm4,%xmm5 | |
| 980 | |
| 981 movdqu 0(%rsi),%xmm0 | |
| 982 pshufd $0xb1,%xmm5,%xmm3 | |
| 983 movdqu 16(%rsi),%xmm1 | |
| 984 movdqu 32(%rsi),%xmm2 | |
| 985 por %xmm3,%xmm5 | |
| 986 movdqu 48(%rsi),%xmm3 | |
| 987 movq 64+0(%rsi),%rax | |
| 988 movq 64+8(%rsi),%r14 | |
| 989 movq 64+16(%rsi),%r15 | |
| 990 movq 64+24(%rsi),%r8 | |
| 991 movdqa %xmm0,480(%rsp) | |
| 992 pshufd $0x1e,%xmm5,%xmm4 | |
| 993 movdqa %xmm1,480+16(%rsp) | |
| 994 movdqu 64(%rsi),%xmm0 | |
| 995 movdqu 80(%rsi),%xmm1 | |
| 996 movdqa %xmm2,512(%rsp) | |
| 997 movdqa %xmm3,512+16(%rsp) | |
| 998 por %xmm4,%xmm5 | |
| 999 pxor %xmm4,%xmm4 | |
| 1000 por %xmm0,%xmm1 | |
| 1001 .byte 102,72,15,110,199 | |
| 1002 | |
| 1003 leaq 64-0(%rsi),%rsi | |
| 1004 movq %rax,544+0(%rsp) | |
| 1005 movq %r14,544+8(%rsp) | |
| 1006 movq %r15,544+16(%rsp) | |
| 1007 movq %r8,544+24(%rsp) | |
| 1008 leaq 96(%rsp),%rdi | |
| 1009 call __ecp_nistz256_sqr_montq | |
| 1010 | |
| 1011 pcmpeqd %xmm4,%xmm5 | |
| 1012 pshufd $0xb1,%xmm1,%xmm4 | |
| 1013 por %xmm1,%xmm4 | |
| 1014 pshufd $0,%xmm5,%xmm5 | |
| 1015 pshufd $0x1e,%xmm4,%xmm3 | |
| 1016 por %xmm3,%xmm4 | |
| 1017 pxor %xmm3,%xmm3 | |
| 1018 pcmpeqd %xmm3,%xmm4 | |
| 1019 pshufd $0,%xmm4,%xmm4 | |
| 1020 movq 64+0(%rbx),%rax | |
| 1021 movq 64+8(%rbx),%r14 | |
| 1022 movq 64+16(%rbx),%r15 | |
| 1023 movq 64+24(%rbx),%r8 | |
| 1024 .byte 102,72,15,110,203 | |
| 1025 | |
| 1026 leaq 64-0(%rbx),%rsi | |
| 1027 leaq 32(%rsp),%rdi | |
| 1028 call __ecp_nistz256_sqr_montq | |
| 1029 | |
| 1030 movq 544(%rsp),%rax | |
| 1031 leaq 544(%rsp),%rbx | |
| 1032 movq 0+96(%rsp),%r9 | |
| 1033 movq 8+96(%rsp),%r10 | |
| 1034 leaq 0+96(%rsp),%rsi | |
| 1035 movq 16+96(%rsp),%r11 | |
| 1036 movq 24+96(%rsp),%r12 | |
| 1037 leaq 224(%rsp),%rdi | |
| 1038 call __ecp_nistz256_mul_montq | |
| 1039 | |
| 1040 movq 448(%rsp),%rax | |
| 1041 leaq 448(%rsp),%rbx | |
| 1042 movq 0+32(%rsp),%r9 | |
| 1043 movq 8+32(%rsp),%r10 | |
| 1044 leaq 0+32(%rsp),%rsi | |
| 1045 movq 16+32(%rsp),%r11 | |
| 1046 movq 24+32(%rsp),%r12 | |
| 1047 leaq 256(%rsp),%rdi | |
| 1048 call __ecp_nistz256_mul_montq | |
| 1049 | |
| 1050 movq 416(%rsp),%rax | |
| 1051 leaq 416(%rsp),%rbx | |
| 1052 movq 0+224(%rsp),%r9 | |
| 1053 movq 8+224(%rsp),%r10 | |
| 1054 leaq 0+224(%rsp),%rsi | |
| 1055 movq 16+224(%rsp),%r11 | |
| 1056 movq 24+224(%rsp),%r12 | |
| 1057 leaq 224(%rsp),%rdi | |
| 1058 call __ecp_nistz256_mul_montq | |
| 1059 | |
| 1060 movq 512(%rsp),%rax | |
| 1061 leaq 512(%rsp),%rbx | |
| 1062 movq 0+256(%rsp),%r9 | |
| 1063 movq 8+256(%rsp),%r10 | |
| 1064 leaq 0+256(%rsp),%rsi | |
| 1065 movq 16+256(%rsp),%r11 | |
| 1066 movq 24+256(%rsp),%r12 | |
| 1067 leaq 256(%rsp),%rdi | |
| 1068 call __ecp_nistz256_mul_montq | |
| 1069 | |
| 1070 leaq 224(%rsp),%rbx | |
| 1071 leaq 64(%rsp),%rdi | |
| 1072 call __ecp_nistz256_sub_fromq | |
| 1073 | |
| 1074 orq %r13,%r12 | |
| 1075 movdqa %xmm4,%xmm2 | |
| 1076 orq %r8,%r12 | |
| 1077 orq %r9,%r12 | |
| 1078 por %xmm5,%xmm2 | |
| 1079 .byte 102,73,15,110,220 | |
| 1080 | |
| 1081 movq 384(%rsp),%rax | |
| 1082 leaq 384(%rsp),%rbx | |
| 1083 movq 0+96(%rsp),%r9 | |
| 1084 movq 8+96(%rsp),%r10 | |
| 1085 leaq 0+96(%rsp),%rsi | |
| 1086 movq 16+96(%rsp),%r11 | |
| 1087 movq 24+96(%rsp),%r12 | |
| 1088 leaq 160(%rsp),%rdi | |
| 1089 call __ecp_nistz256_mul_montq | |
| 1090 | |
| 1091 movq 480(%rsp),%rax | |
| 1092 leaq 480(%rsp),%rbx | |
| 1093 movq 0+32(%rsp),%r9 | |
| 1094 movq 8+32(%rsp),%r10 | |
| 1095 leaq 0+32(%rsp),%rsi | |
| 1096 movq 16+32(%rsp),%r11 | |
| 1097 movq 24+32(%rsp),%r12 | |
| 1098 leaq 192(%rsp),%rdi | |
| 1099 call __ecp_nistz256_mul_montq | |
| 1100 | |
| 1101 leaq 160(%rsp),%rbx | |
| 1102 leaq 0(%rsp),%rdi | |
| 1103 call __ecp_nistz256_sub_fromq | |
| 1104 | |
| 1105 orq %r13,%r12 | |
| 1106 orq %r8,%r12 | |
| 1107 orq %r9,%r12 | |
| 1108 | |
| 1109 .byte 0x3e | |
| 1110 jnz L$add_proceedq | |
| 1111 .byte 102,73,15,126,208 | |
| 1112 .byte 102,73,15,126,217 | |
| 1113 testq %r8,%r8 | |
| 1114 jnz L$add_proceedq | |
| 1115 testq %r9,%r9 | |
| 1116 jz L$add_doubleq | |
| 1117 | |
| 1118 .byte 102,72,15,126,199 | |
| 1119 pxor %xmm0,%xmm0 | |
| 1120 movdqu %xmm0,0(%rdi) | |
| 1121 movdqu %xmm0,16(%rdi) | |
| 1122 movdqu %xmm0,32(%rdi) | |
| 1123 movdqu %xmm0,48(%rdi) | |
| 1124 movdqu %xmm0,64(%rdi) | |
| 1125 movdqu %xmm0,80(%rdi) | |
| 1126 jmp L$add_doneq | |
| 1127 | |
| 1128 .p2align 5 | |
| 1129 L$add_doubleq: | |
| 1130 .byte 102,72,15,126,206 | |
| 1131 .byte 102,72,15,126,199 | |
| 1132 addq $416,%rsp | |
| 1133 jmp L$point_double_shortcutq | |
| 1134 | |
| 1135 .p2align 5 | |
| 1136 L$add_proceedq: | |
| 1137 movq 0+64(%rsp),%rax | |
| 1138 movq 8+64(%rsp),%r14 | |
| 1139 leaq 0+64(%rsp),%rsi | |
| 1140 movq 16+64(%rsp),%r15 | |
| 1141 movq 24+64(%rsp),%r8 | |
| 1142 leaq 96(%rsp),%rdi | |
| 1143 call __ecp_nistz256_sqr_montq | |
| 1144 | |
| 1145 movq 448(%rsp),%rax | |
| 1146 leaq 448(%rsp),%rbx | |
| 1147 movq 0+0(%rsp),%r9 | |
| 1148 movq 8+0(%rsp),%r10 | |
| 1149 leaq 0+0(%rsp),%rsi | |
| 1150 movq 16+0(%rsp),%r11 | |
| 1151 movq 24+0(%rsp),%r12 | |
| 1152 leaq 352(%rsp),%rdi | |
| 1153 call __ecp_nistz256_mul_montq | |
| 1154 | |
| 1155 movq 0+0(%rsp),%rax | |
| 1156 movq 8+0(%rsp),%r14 | |
| 1157 leaq 0+0(%rsp),%rsi | |
| 1158 movq 16+0(%rsp),%r15 | |
| 1159 movq 24+0(%rsp),%r8 | |
| 1160 leaq 32(%rsp),%rdi | |
| 1161 call __ecp_nistz256_sqr_montq | |
| 1162 | |
| 1163 movq 544(%rsp),%rax | |
| 1164 leaq 544(%rsp),%rbx | |
| 1165 movq 0+352(%rsp),%r9 | |
| 1166 movq 8+352(%rsp),%r10 | |
| 1167 leaq 0+352(%rsp),%rsi | |
| 1168 movq 16+352(%rsp),%r11 | |
| 1169 movq 24+352(%rsp),%r12 | |
| 1170 leaq 352(%rsp),%rdi | |
| 1171 call __ecp_nistz256_mul_montq | |
| 1172 | |
| 1173 movq 0(%rsp),%rax | |
| 1174 leaq 0(%rsp),%rbx | |
| 1175 movq 0+32(%rsp),%r9 | |
| 1176 movq 8+32(%rsp),%r10 | |
| 1177 leaq 0+32(%rsp),%rsi | |
| 1178 movq 16+32(%rsp),%r11 | |
| 1179 movq 24+32(%rsp),%r12 | |
| 1180 leaq 128(%rsp),%rdi | |
| 1181 call __ecp_nistz256_mul_montq | |
| 1182 | |
| 1183 movq 160(%rsp),%rax | |
| 1184 leaq 160(%rsp),%rbx | |
| 1185 movq 0+32(%rsp),%r9 | |
| 1186 movq 8+32(%rsp),%r10 | |
| 1187 leaq 0+32(%rsp),%rsi | |
| 1188 movq 16+32(%rsp),%r11 | |
| 1189 movq 24+32(%rsp),%r12 | |
| 1190 leaq 192(%rsp),%rdi | |
| 1191 call __ecp_nistz256_mul_montq | |
| 1192 | |
| 1193 | |
| 1194 | |
| 1195 | |
| 1196 xorq %r11,%r11 | |
| 1197 addq %r12,%r12 | |
| 1198 leaq 96(%rsp),%rsi | |
| 1199 adcq %r13,%r13 | |
| 1200 movq %r12,%rax | |
| 1201 adcq %r8,%r8 | |
| 1202 adcq %r9,%r9 | |
| 1203 movq %r13,%rbp | |
| 1204 adcq $0,%r11 | |
| 1205 | |
| 1206 subq $-1,%r12 | |
| 1207 movq %r8,%rcx | |
| 1208 sbbq %r14,%r13 | |
| 1209 sbbq $0,%r8 | |
| 1210 movq %r9,%r10 | |
| 1211 sbbq %r15,%r9 | |
| 1212 sbbq $0,%r11 | |
| 1213 | |
| 1214 cmovcq %rax,%r12 | |
| 1215 movq 0(%rsi),%rax | |
| 1216 cmovcq %rbp,%r13 | |
| 1217 movq 8(%rsi),%rbp | |
| 1218 cmovcq %rcx,%r8 | |
| 1219 movq 16(%rsi),%rcx | |
| 1220 cmovcq %r10,%r9 | |
| 1221 movq 24(%rsi),%r10 | |
| 1222 | |
| 1223 call __ecp_nistz256_subq | |
| 1224 | |
| 1225 leaq 128(%rsp),%rbx | |
| 1226 leaq 288(%rsp),%rdi | |
| 1227 call __ecp_nistz256_sub_fromq | |
| 1228 | |
| 1229 movq 192+0(%rsp),%rax | |
| 1230 movq 192+8(%rsp),%rbp | |
| 1231 movq 192+16(%rsp),%rcx | |
| 1232 movq 192+24(%rsp),%r10 | |
| 1233 leaq 320(%rsp),%rdi | |
| 1234 | |
| 1235 call __ecp_nistz256_subq | |
| 1236 | |
| 1237 movq %r12,0(%rdi) | |
| 1238 movq %r13,8(%rdi) | |
| 1239 movq %r8,16(%rdi) | |
| 1240 movq %r9,24(%rdi) | |
| 1241 movq 128(%rsp),%rax | |
| 1242 leaq 128(%rsp),%rbx | |
| 1243 movq 0+224(%rsp),%r9 | |
| 1244 movq 8+224(%rsp),%r10 | |
| 1245 leaq 0+224(%rsp),%rsi | |
| 1246 movq 16+224(%rsp),%r11 | |
| 1247 movq 24+224(%rsp),%r12 | |
| 1248 leaq 256(%rsp),%rdi | |
| 1249 call __ecp_nistz256_mul_montq | |
| 1250 | |
| 1251 movq 320(%rsp),%rax | |
| 1252 leaq 320(%rsp),%rbx | |
| 1253 movq 0+64(%rsp),%r9 | |
| 1254 movq 8+64(%rsp),%r10 | |
| 1255 leaq 0+64(%rsp),%rsi | |
| 1256 movq 16+64(%rsp),%r11 | |
| 1257 movq 24+64(%rsp),%r12 | |
| 1258 leaq 320(%rsp),%rdi | |
| 1259 call __ecp_nistz256_mul_montq | |
| 1260 | |
| 1261 leaq 256(%rsp),%rbx | |
| 1262 leaq 320(%rsp),%rdi | |
| 1263 call __ecp_nistz256_sub_fromq | |
| 1264 | |
| 1265 .byte 102,72,15,126,199 | |
| 1266 | |
| 1267 movdqa %xmm5,%xmm0 | |
| 1268 movdqa %xmm5,%xmm1 | |
| 1269 pandn 352(%rsp),%xmm0 | |
| 1270 movdqa %xmm5,%xmm2 | |
| 1271 pandn 352+16(%rsp),%xmm1 | |
| 1272 movdqa %xmm5,%xmm3 | |
| 1273 pand 544(%rsp),%xmm2 | |
| 1274 pand 544+16(%rsp),%xmm3 | |
| 1275 por %xmm0,%xmm2 | |
| 1276 por %xmm1,%xmm3 | |
| 1277 | |
| 1278 movdqa %xmm4,%xmm0 | |
| 1279 movdqa %xmm4,%xmm1 | |
| 1280 pandn %xmm2,%xmm0 | |
| 1281 movdqa %xmm4,%xmm2 | |
| 1282 pandn %xmm3,%xmm1 | |
| 1283 movdqa %xmm4,%xmm3 | |
| 1284 pand 448(%rsp),%xmm2 | |
| 1285 pand 448+16(%rsp),%xmm3 | |
| 1286 por %xmm0,%xmm2 | |
| 1287 por %xmm1,%xmm3 | |
| 1288 movdqu %xmm2,64(%rdi) | |
| 1289 movdqu %xmm3,80(%rdi) | |
| 1290 | |
| 1291 movdqa %xmm5,%xmm0 | |
| 1292 movdqa %xmm5,%xmm1 | |
| 1293 pandn 288(%rsp),%xmm0 | |
| 1294 movdqa %xmm5,%xmm2 | |
| 1295 pandn 288+16(%rsp),%xmm1 | |
| 1296 movdqa %xmm5,%xmm3 | |
| 1297 pand 480(%rsp),%xmm2 | |
| 1298 pand 480+16(%rsp),%xmm3 | |
| 1299 por %xmm0,%xmm2 | |
| 1300 por %xmm1,%xmm3 | |
| 1301 | |
| 1302 movdqa %xmm4,%xmm0 | |
| 1303 movdqa %xmm4,%xmm1 | |
| 1304 pandn %xmm2,%xmm0 | |
| 1305 movdqa %xmm4,%xmm2 | |
| 1306 pandn %xmm3,%xmm1 | |
| 1307 movdqa %xmm4,%xmm3 | |
| 1308 pand 384(%rsp),%xmm2 | |
| 1309 pand 384+16(%rsp),%xmm3 | |
| 1310 por %xmm0,%xmm2 | |
| 1311 por %xmm1,%xmm3 | |
| 1312 movdqu %xmm2,0(%rdi) | |
| 1313 movdqu %xmm3,16(%rdi) | |
| 1314 | |
| 1315 movdqa %xmm5,%xmm0 | |
| 1316 movdqa %xmm5,%xmm1 | |
| 1317 pandn 320(%rsp),%xmm0 | |
| 1318 movdqa %xmm5,%xmm2 | |
| 1319 pandn 320+16(%rsp),%xmm1 | |
| 1320 movdqa %xmm5,%xmm3 | |
| 1321 pand 512(%rsp),%xmm2 | |
| 1322 pand 512+16(%rsp),%xmm3 | |
| 1323 por %xmm0,%xmm2 | |
| 1324 por %xmm1,%xmm3 | |
| 1325 | |
| 1326 movdqa %xmm4,%xmm0 | |
| 1327 movdqa %xmm4,%xmm1 | |
| 1328 pandn %xmm2,%xmm0 | |
| 1329 movdqa %xmm4,%xmm2 | |
| 1330 pandn %xmm3,%xmm1 | |
| 1331 movdqa %xmm4,%xmm3 | |
| 1332 pand 416(%rsp),%xmm2 | |
| 1333 pand 416+16(%rsp),%xmm3 | |
| 1334 por %xmm0,%xmm2 | |
| 1335 por %xmm1,%xmm3 | |
| 1336 movdqu %xmm2,32(%rdi) | |
| 1337 movdqu %xmm3,48(%rdi) | |
| 1338 | |
| 1339 L$add_doneq: | |
| 1340 addq $576+8,%rsp | |
| 1341 popq %r15 | |
| 1342 popq %r14 | |
| 1343 popq %r13 | |
| 1344 popq %r12 | |
| 1345 popq %rbx | |
| 1346 popq %rbp | |
| 1347 .byte 0xf3,0xc3 | |
| 1348 | |
| 1349 .globl _ecp_nistz256_point_add_affine | |
| 1350 .private_extern _ecp_nistz256_point_add_affine | |
| 1351 | |
| 1352 .p2align 5 | |
| 1353 _ecp_nistz256_point_add_affine: | |
| 1354 pushq %rbp | |
| 1355 pushq %rbx | |
| 1356 pushq %r12 | |
| 1357 pushq %r13 | |
| 1358 pushq %r14 | |
| 1359 pushq %r15 | |
| 1360 subq $480+8,%rsp | |
| 1361 | |
| 1362 movdqu 0(%rsi),%xmm0 | |
| 1363 movq %rdx,%rbx | |
| 1364 movdqu 16(%rsi),%xmm1 | |
| 1365 movdqu 32(%rsi),%xmm2 | |
| 1366 movdqu 48(%rsi),%xmm3 | |
| 1367 movdqu 64(%rsi),%xmm4 | |
| 1368 movdqu 80(%rsi),%xmm5 | |
| 1369 movq 64+0(%rsi),%rax | |
| 1370 movq 64+8(%rsi),%r14 | |
| 1371 movq 64+16(%rsi),%r15 | |
| 1372 movq 64+24(%rsi),%r8 | |
| 1373 movdqa %xmm0,320(%rsp) | |
| 1374 movdqa %xmm1,320+16(%rsp) | |
| 1375 movdqa %xmm2,352(%rsp) | |
| 1376 movdqa %xmm3,352+16(%rsp) | |
| 1377 movdqa %xmm4,384(%rsp) | |
| 1378 movdqa %xmm5,384+16(%rsp) | |
| 1379 por %xmm4,%xmm5 | |
| 1380 | |
| 1381 movdqu 0(%rbx),%xmm0 | |
| 1382 pshufd $0xb1,%xmm5,%xmm3 | |
| 1383 movdqu 16(%rbx),%xmm1 | |
| 1384 movdqu 32(%rbx),%xmm2 | |
| 1385 por %xmm3,%xmm5 | |
| 1386 movdqu 48(%rbx),%xmm3 | |
| 1387 movdqa %xmm0,416(%rsp) | |
| 1388 pshufd $0x1e,%xmm5,%xmm4 | |
| 1389 movdqa %xmm1,416+16(%rsp) | |
| 1390 por %xmm0,%xmm1 | |
| 1391 .byte 102,72,15,110,199 | |
| 1392 movdqa %xmm2,448(%rsp) | |
| 1393 movdqa %xmm3,448+16(%rsp) | |
| 1394 por %xmm2,%xmm3 | |
| 1395 por %xmm4,%xmm5 | |
| 1396 pxor %xmm4,%xmm4 | |
| 1397 por %xmm1,%xmm3 | |
| 1398 | |
| 1399 leaq 64-0(%rsi),%rsi | |
| 1400 leaq 32(%rsp),%rdi | |
| 1401 call __ecp_nistz256_sqr_montq | |
| 1402 | |
| 1403 pcmpeqd %xmm4,%xmm5 | |
| 1404 pshufd $0xb1,%xmm3,%xmm4 | |
| 1405 movq 0(%rbx),%rax | |
| 1406 | |
| 1407 movq %r12,%r9 | |
| 1408 por %xmm3,%xmm4 | |
| 1409 pshufd $0,%xmm5,%xmm5 | |
| 1410 pshufd $0x1e,%xmm4,%xmm3 | |
| 1411 movq %r13,%r10 | |
| 1412 por %xmm3,%xmm4 | |
| 1413 pxor %xmm3,%xmm3 | |
| 1414 movq %r14,%r11 | |
| 1415 pcmpeqd %xmm3,%xmm4 | |
| 1416 pshufd $0,%xmm4,%xmm4 | |
| 1417 | |
| 1418 leaq 32-0(%rsp),%rsi | |
| 1419 movq %r15,%r12 | |
| 1420 leaq 0(%rsp),%rdi | |
| 1421 call __ecp_nistz256_mul_montq | |
| 1422 | |
| 1423 leaq 320(%rsp),%rbx | |
| 1424 leaq 64(%rsp),%rdi | |
| 1425 call __ecp_nistz256_sub_fromq | |
| 1426 | |
| 1427 movq 384(%rsp),%rax | |
| 1428 leaq 384(%rsp),%rbx | |
| 1429 movq 0+32(%rsp),%r9 | |
| 1430 movq 8+32(%rsp),%r10 | |
| 1431 leaq 0+32(%rsp),%rsi | |
| 1432 movq 16+32(%rsp),%r11 | |
| 1433 movq 24+32(%rsp),%r12 | |
| 1434 leaq 32(%rsp),%rdi | |
| 1435 call __ecp_nistz256_mul_montq | |
| 1436 | |
| 1437 movq 384(%rsp),%rax | |
| 1438 leaq 384(%rsp),%rbx | |
| 1439 movq 0+64(%rsp),%r9 | |
| 1440 movq 8+64(%rsp),%r10 | |
| 1441 leaq 0+64(%rsp),%rsi | |
| 1442 movq 16+64(%rsp),%r11 | |
| 1443 movq 24+64(%rsp),%r12 | |
| 1444 leaq 288(%rsp),%rdi | |
| 1445 call __ecp_nistz256_mul_montq | |
| 1446 | |
| 1447 movq 448(%rsp),%rax | |
| 1448 leaq 448(%rsp),%rbx | |
| 1449 movq 0+32(%rsp),%r9 | |
| 1450 movq 8+32(%rsp),%r10 | |
| 1451 leaq 0+32(%rsp),%rsi | |
| 1452 movq 16+32(%rsp),%r11 | |
| 1453 movq 24+32(%rsp),%r12 | |
| 1454 leaq 32(%rsp),%rdi | |
| 1455 call __ecp_nistz256_mul_montq | |
| 1456 | |
| 1457 leaq 352(%rsp),%rbx | |
| 1458 leaq 96(%rsp),%rdi | |
| 1459 call __ecp_nistz256_sub_fromq | |
| 1460 | |
| 1461 movq 0+64(%rsp),%rax | |
| 1462 movq 8+64(%rsp),%r14 | |
| 1463 leaq 0+64(%rsp),%rsi | |
| 1464 movq 16+64(%rsp),%r15 | |
| 1465 movq 24+64(%rsp),%r8 | |
| 1466 leaq 128(%rsp),%rdi | |
| 1467 call __ecp_nistz256_sqr_montq | |
| 1468 | |
| 1469 movq 0+96(%rsp),%rax | |
| 1470 movq 8+96(%rsp),%r14 | |
| 1471 leaq 0+96(%rsp),%rsi | |
| 1472 movq 16+96(%rsp),%r15 | |
| 1473 movq 24+96(%rsp),%r8 | |
| 1474 leaq 192(%rsp),%rdi | |
| 1475 call __ecp_nistz256_sqr_montq | |
| 1476 | |
| 1477 movq 128(%rsp),%rax | |
| 1478 leaq 128(%rsp),%rbx | |
| 1479 movq 0+64(%rsp),%r9 | |
| 1480 movq 8+64(%rsp),%r10 | |
| 1481 leaq 0+64(%rsp),%rsi | |
| 1482 movq 16+64(%rsp),%r11 | |
| 1483 movq 24+64(%rsp),%r12 | |
| 1484 leaq 160(%rsp),%rdi | |
| 1485 call __ecp_nistz256_mul_montq | |
| 1486 | |
| 1487 movq 320(%rsp),%rax | |
| 1488 leaq 320(%rsp),%rbx | |
| 1489 movq 0+128(%rsp),%r9 | |
| 1490 movq 8+128(%rsp),%r10 | |
| 1491 leaq 0+128(%rsp),%rsi | |
| 1492 movq 16+128(%rsp),%r11 | |
| 1493 movq 24+128(%rsp),%r12 | |
| 1494 leaq 0(%rsp),%rdi | |
| 1495 call __ecp_nistz256_mul_montq | |
| 1496 | |
| 1497 | |
| 1498 | |
| 1499 | |
| 1500 xorq %r11,%r11 | |
| 1501 addq %r12,%r12 | |
| 1502 leaq 192(%rsp),%rsi | |
| 1503 adcq %r13,%r13 | |
| 1504 movq %r12,%rax | |
| 1505 adcq %r8,%r8 | |
| 1506 adcq %r9,%r9 | |
| 1507 movq %r13,%rbp | |
| 1508 adcq $0,%r11 | |
| 1509 | |
| 1510 subq $-1,%r12 | |
| 1511 movq %r8,%rcx | |
| 1512 sbbq %r14,%r13 | |
| 1513 sbbq $0,%r8 | |
| 1514 movq %r9,%r10 | |
| 1515 sbbq %r15,%r9 | |
| 1516 sbbq $0,%r11 | |
| 1517 | |
| 1518 cmovcq %rax,%r12 | |
| 1519 movq 0(%rsi),%rax | |
| 1520 cmovcq %rbp,%r13 | |
| 1521 movq 8(%rsi),%rbp | |
| 1522 cmovcq %rcx,%r8 | |
| 1523 movq 16(%rsi),%rcx | |
| 1524 cmovcq %r10,%r9 | |
| 1525 movq 24(%rsi),%r10 | |
| 1526 | |
| 1527 call __ecp_nistz256_subq | |
| 1528 | |
| 1529 leaq 160(%rsp),%rbx | |
| 1530 leaq 224(%rsp),%rdi | |
| 1531 call __ecp_nistz256_sub_fromq | |
| 1532 | |
| 1533 movq 0+0(%rsp),%rax | |
| 1534 movq 0+8(%rsp),%rbp | |
| 1535 movq 0+16(%rsp),%rcx | |
| 1536 movq 0+24(%rsp),%r10 | |
| 1537 leaq 64(%rsp),%rdi | |
| 1538 | |
| 1539 call __ecp_nistz256_subq | |
| 1540 | |
| 1541 movq %r12,0(%rdi) | |
| 1542 movq %r13,8(%rdi) | |
| 1543 movq %r8,16(%rdi) | |
| 1544 movq %r9,24(%rdi) | |
| 1545 movq 352(%rsp),%rax | |
| 1546 leaq 352(%rsp),%rbx | |
| 1547 movq 0+160(%rsp),%r9 | |
| 1548 movq 8+160(%rsp),%r10 | |
| 1549 leaq 0+160(%rsp),%rsi | |
| 1550 movq 16+160(%rsp),%r11 | |
| 1551 movq 24+160(%rsp),%r12 | |
| 1552 leaq 32(%rsp),%rdi | |
| 1553 call __ecp_nistz256_mul_montq | |
| 1554 | |
| 1555 movq 96(%rsp),%rax | |
| 1556 leaq 96(%rsp),%rbx | |
| 1557 movq 0+64(%rsp),%r9 | |
| 1558 movq 8+64(%rsp),%r10 | |
| 1559 leaq 0+64(%rsp),%rsi | |
| 1560 movq 16+64(%rsp),%r11 | |
| 1561 movq 24+64(%rsp),%r12 | |
| 1562 leaq 64(%rsp),%rdi | |
| 1563 call __ecp_nistz256_mul_montq | |
| 1564 | |
| 1565 leaq 32(%rsp),%rbx | |
| 1566 leaq 256(%rsp),%rdi | |
| 1567 call __ecp_nistz256_sub_fromq | |
| 1568 | |
| 1569 .byte 102,72,15,126,199 | |
| 1570 | |
| 1571 movdqa %xmm5,%xmm0 | |
| 1572 movdqa %xmm5,%xmm1 | |
| 1573 pandn 288(%rsp),%xmm0 | |
| 1574 movdqa %xmm5,%xmm2 | |
| 1575 pandn 288+16(%rsp),%xmm1 | |
| 1576 movdqa %xmm5,%xmm3 | |
| 1577 pand L$ONE_mont(%rip),%xmm2 | |
| 1578 pand L$ONE_mont+16(%rip),%xmm3 | |
| 1579 por %xmm0,%xmm2 | |
| 1580 por %xmm1,%xmm3 | |
| 1581 | |
| 1582 movdqa %xmm4,%xmm0 | |
| 1583 movdqa %xmm4,%xmm1 | |
| 1584 pandn %xmm2,%xmm0 | |
| 1585 movdqa %xmm4,%xmm2 | |
| 1586 pandn %xmm3,%xmm1 | |
| 1587 movdqa %xmm4,%xmm3 | |
| 1588 pand 384(%rsp),%xmm2 | |
| 1589 pand 384+16(%rsp),%xmm3 | |
| 1590 por %xmm0,%xmm2 | |
| 1591 por %xmm1,%xmm3 | |
| 1592 movdqu %xmm2,64(%rdi) | |
| 1593 movdqu %xmm3,80(%rdi) | |
| 1594 | |
| 1595 movdqa %xmm5,%xmm0 | |
| 1596 movdqa %xmm5,%xmm1 | |
| 1597 pandn 224(%rsp),%xmm0 | |
| 1598 movdqa %xmm5,%xmm2 | |
| 1599 pandn 224+16(%rsp),%xmm1 | |
| 1600 movdqa %xmm5,%xmm3 | |
| 1601 pand 416(%rsp),%xmm2 | |
| 1602 pand 416+16(%rsp),%xmm3 | |
| 1603 por %xmm0,%xmm2 | |
| 1604 por %xmm1,%xmm3 | |
| 1605 | |
| 1606 movdqa %xmm4,%xmm0 | |
| 1607 movdqa %xmm4,%xmm1 | |
| 1608 pandn %xmm2,%xmm0 | |
| 1609 movdqa %xmm4,%xmm2 | |
| 1610 pandn %xmm3,%xmm1 | |
| 1611 movdqa %xmm4,%xmm3 | |
| 1612 pand 320(%rsp),%xmm2 | |
| 1613 pand 320+16(%rsp),%xmm3 | |
| 1614 por %xmm0,%xmm2 | |
| 1615 por %xmm1,%xmm3 | |
| 1616 movdqu %xmm2,0(%rdi) | |
| 1617 movdqu %xmm3,16(%rdi) | |
| 1618 | |
| 1619 movdqa %xmm5,%xmm0 | |
| 1620 movdqa %xmm5,%xmm1 | |
| 1621 pandn 256(%rsp),%xmm0 | |
| 1622 movdqa %xmm5,%xmm2 | |
| 1623 pandn 256+16(%rsp),%xmm1 | |
| 1624 movdqa %xmm5,%xmm3 | |
| 1625 pand 448(%rsp),%xmm2 | |
| 1626 pand 448+16(%rsp),%xmm3 | |
| 1627 por %xmm0,%xmm2 | |
| 1628 por %xmm1,%xmm3 | |
| 1629 | |
| 1630 movdqa %xmm4,%xmm0 | |
| 1631 movdqa %xmm4,%xmm1 | |
| 1632 pandn %xmm2,%xmm0 | |
| 1633 movdqa %xmm4,%xmm2 | |
| 1634 pandn %xmm3,%xmm1 | |
| 1635 movdqa %xmm4,%xmm3 | |
| 1636 pand 352(%rsp),%xmm2 | |
| 1637 pand 352+16(%rsp),%xmm3 | |
| 1638 por %xmm0,%xmm2 | |
| 1639 por %xmm1,%xmm3 | |
| 1640 movdqu %xmm2,32(%rdi) | |
| 1641 movdqu %xmm3,48(%rdi) | |
| 1642 | |
| 1643 addq $480+8,%rsp | |
| 1644 popq %r15 | |
| 1645 popq %r14 | |
| 1646 popq %r13 | |
| 1647 popq %r12 | |
| 1648 popq %rbx | |
| 1649 popq %rbp | |
| 1650 .byte 0xf3,0xc3 | |
| 1651 | |
| 1652 #endif | |
| OLD | NEW |