1 .file "ghash-x86.s" 2 .text 3 .globl gcm_gmult_4bit_x86 4 .type gcm_gmult_4bit_x86,@function 5 .align 16 6 gcm_gmult_4bit_x86: 7 .L_gcm_gmult_4bit_x86_begin: 8 pushl %ebp 9 pushl %ebx 10 pushl %esi 11 pushl %edi 12 subl $84,%esp 13 movl 104(%esp),%edi 14 movl 108(%esp),%esi 15 movl (%edi),%ebp 16 movl 4(%edi),%edx 17 movl 8(%edi),%ecx 18 movl 12(%edi),%ebx 19 movl $0,16(%esp) 20 movl $471859200,20(%esp) 21 movl $943718400,24(%esp) 22 movl $610271232,28(%esp) 23 movl $1887436800,32(%esp) 24 movl $1822425088,36(%esp) 25 movl $1220542464,40(%esp) 26 movl $1423966208,44(%esp) 27 movl $3774873600,48(%esp) 28 movl $4246732800,52(%esp) 29 movl $3644850176,56(%esp) 30 movl $3311403008,60(%esp) 31 movl $2441084928,64(%esp) 32 movl $2376073216,68(%esp) 33 movl $2847932416,72(%esp) 34 movl $3051356160,76(%esp) 35 movl %ebp,(%esp) 36 movl %edx,4(%esp) 37 movl %ecx,8(%esp) 38 movl %ebx,12(%esp) 39 shrl $20,%ebx 40 andl $240,%ebx 41 movl 4(%esi,%ebx,1),%ebp 42 movl (%esi,%ebx,1),%edx 43 movl 12(%esi,%ebx,1),%ecx 44 movl 8(%esi,%ebx,1),%ebx 45 xorl %eax,%eax 46 movl $15,%edi 47 jmp .L000x86_loop 48 .align 16 49 .L000x86_loop: 50 movb %bl,%al 51 shrdl $4,%ecx,%ebx 52 andb $15,%al 53 shrdl $4,%edx,%ecx 54 shrdl $4,%ebp,%edx 55 shrl $4,%ebp 56 xorl 16(%esp,%eax,4),%ebp 57 movb (%esp,%edi,1),%al 58 andb $240,%al 59 xorl 8(%esi,%eax,1),%ebx 60 xorl 12(%esi,%eax,1),%ecx 61 xorl (%esi,%eax,1),%edx 62 xorl 4(%esi,%eax,1),%ebp 63 decl %edi 64 js .L001x86_break 65 movb %bl,%al 66 shrdl $4,%ecx,%ebx 67 andb $15,%al 68 shrdl $4,%edx,%ecx 69 shrdl $4,%ebp,%edx 70 shrl $4,%ebp 71 xorl 16(%esp,%eax,4),%ebp 72 movb (%esp,%edi,1),%al 73 shlb $4,%al 74 xorl 8(%esi,%eax,1),%ebx 75 xorl 12(%esi,%eax,1),%ecx 76 xorl (%esi,%eax,1),%edx 77 xorl 4(%esi,%eax,1),%ebp 78 jmp .L000x86_loop 79 .align 16 80 .L001x86_break: 81 bswap %ebx 82 bswap %ecx 83 bswap %edx 84 bswap %ebp 85 movl 104(%esp),%edi 86 movl %ebx,12(%edi) 87 movl %ecx,8(%edi) 88 movl %edx,4(%edi) 89 movl %ebp,(%edi) 90 addl $84,%esp 91 popl %edi 92 popl %esi 93 popl %ebx 94 popl %ebp 95 ret 96 .size gcm_gmult_4bit_x86,.-.L_gcm_gmult_4bit_x86_begin 97 .globl gcm_ghash_4bit_x86 98 .type gcm_ghash_4bit_x86,@function 99 .align 16 100 gcm_ghash_4bit_x86: 101 .L_gcm_ghash_4bit_x86_begin: 102 pushl %ebp 103 pushl %ebx 104 pushl %esi 105 pushl %edi 106 subl $84,%esp 107 movl 104(%esp),%ebx 108 movl 108(%esp),%esi 109 movl 112(%esp),%edi 110 movl 116(%esp),%ecx 111 addl %edi,%ecx 112 movl %ecx,116(%esp) 113 movl (%ebx),%ebp 114 movl 4(%ebx),%edx 115 movl 8(%ebx),%ecx 116 movl 12(%ebx),%ebx 117 movl $0,16(%esp) 118 movl $471859200,20(%esp) 119 movl $943718400,24(%esp) 120 movl $610271232,28(%esp) 121 movl $1887436800,32(%esp) 122 movl $1822425088,36(%esp) 123 movl $1220542464,40(%esp) 124 movl $1423966208,44(%esp) 125 movl $3774873600,48(%esp) 126 movl $4246732800,52(%esp) 127 movl $3644850176,56(%esp) 128 movl $3311403008,60(%esp) 129 movl $2441084928,64(%esp) 130 movl $2376073216,68(%esp) 131 movl $2847932416,72(%esp) 132 movl $3051356160,76(%esp) 133 .align 16 134 .L002x86_outer_loop: 135 xorl 12(%edi),%ebx 136 xorl 8(%edi),%ecx 137 xorl 4(%edi),%edx 138 xorl (%edi),%ebp 139 movl %ebx,12(%esp) 140 movl %ecx,8(%esp) 141 movl %edx,4(%esp) 142 movl %ebp,(%esp) 143 shrl $20,%ebx 144 andl $240,%ebx 145 movl 4(%esi,%ebx,1),%ebp 146 movl (%esi,%ebx,1),%edx 147 movl 12(%esi,%ebx,1),%ecx 148 movl 8(%esi,%ebx,1),%ebx 149 xorl %eax,%eax 150 movl $15,%edi 151 jmp .L003x86_loop 152 .align 16 153 .L003x86_loop: 154 movb %bl,%al 155 shrdl $4,%ecx,%ebx 156 andb $15,%al 157 shrdl $4,%edx,%ecx 158 shrdl $4,%ebp,%edx 159 shrl $4,%ebp 160 xorl 16(%esp,%eax,4),%ebp 161 movb (%esp,%edi,1),%al 162 andb $240,%al 163 xorl 8(%esi,%eax,1),%ebx 164 xorl 12(%esi,%eax,1),%ecx 165 xorl (%esi,%eax,1),%edx 166 xorl 4(%esi,%eax,1),%ebp 167 decl %edi 168 js .L004x86_break 169 movb %bl,%al 170 shrdl $4,%ecx,%ebx 171 andb $15,%al 172 shrdl $4,%edx,%ecx 173 shrdl $4,%ebp,%edx 174 shrl $4,%ebp 175 xorl 16(%esp,%eax,4),%ebp 176 movb (%esp,%edi,1),%al 177 shlb $4,%al 178 xorl 8(%esi,%eax,1),%ebx 179 xorl 12(%esi,%eax,1),%ecx 180 xorl (%esi,%eax,1),%edx 181 xorl 4(%esi,%eax,1),%ebp 182 jmp .L003x86_loop 183 .align 16 184 .L004x86_break: 185 bswap %ebx 186 bswap %ecx 187 bswap %edx 188 bswap %ebp 189 movl 112(%esp),%edi 190 leal 16(%edi),%edi 191 cmpl 116(%esp),%edi 192 movl %edi,112(%esp) 193 jb .L002x86_outer_loop 194 movl 104(%esp),%edi 195 movl %ebx,12(%edi) 196 movl %ecx,8(%edi) 197 movl %edx,4(%edi) 198 movl %ebp,(%edi) 199 addl $84,%esp 200 popl %edi 201 popl %esi 202 popl %ebx 203 popl %ebp 204 ret 205 .size gcm_ghash_4bit_x86,.-.L_gcm_ghash_4bit_x86_begin 206 .type _mmx_gmult_4bit_inner,@function 207 .align 16 208 _mmx_gmult_4bit_inner: 209 xorl %ecx,%ecx 210 movl %ebx,%edx 211 movb %dl,%cl 212 shlb $4,%cl 213 andl $240,%edx 214 movq 8(%esi,%ecx,1),%mm0 215 movq (%esi,%ecx,1),%mm1 216 movd %mm0,%ebp 217 psrlq $4,%mm0 218 movq %mm1,%mm2 219 psrlq $4,%mm1 220 pxor 8(%esi,%edx,1),%mm0 221 movb 14(%edi),%cl 222 psllq $60,%mm2 223 andl $15,%ebp 224 pxor (%esi,%edx,1),%mm1 225 movl %ecx,%edx 226 movd %mm0,%ebx 227 pxor %mm2,%mm0 228 shlb $4,%cl 229 psrlq $4,%mm0 230 movq %mm1,%mm2 231 psrlq $4,%mm1 232 pxor 8(%esi,%ecx,1),%mm0 233 psllq $60,%mm2 234 andl $240,%edx 235 pxor (%eax,%ebp,8),%mm1 236 andl $15,%ebx 237 pxor (%esi,%ecx,1),%mm1 238 movd %mm0,%ebp 239 pxor %mm2,%mm0 240 psrlq $4,%mm0 241 movq %mm1,%mm2 242 psrlq $4,%mm1 243 pxor 8(%esi,%edx,1),%mm0 244 movb 13(%edi),%cl 245 psllq $60,%mm2 246 pxor (%eax,%ebx,8),%mm1 247 andl $15,%ebp 248 pxor (%esi,%edx,1),%mm1 249 movl %ecx,%edx 250 movd %mm0,%ebx 251 pxor %mm2,%mm0 252 shlb $4,%cl 253 psrlq $4,%mm0 254 movq %mm1,%mm2 255 psrlq $4,%mm1 256 pxor 8(%esi,%ecx,1),%mm0 257 psllq $60,%mm2 258 andl $240,%edx 259 pxor (%eax,%ebp,8),%mm1 260 andl $15,%ebx 261 pxor (%esi,%ecx,1),%mm1 262 movd %mm0,%ebp 263 pxor %mm2,%mm0 264 psrlq $4,%mm0 265 movq %mm1,%mm2 266 psrlq $4,%mm1 267 pxor 8(%esi,%edx,1),%mm0 268 movb 12(%edi),%cl 269 psllq $60,%mm2 270 pxor (%eax,%ebx,8),%mm1 271 andl $15,%ebp 272 pxor (%esi,%edx,1),%mm1 273 movl %ecx,%edx 274 movd %mm0,%ebx 275 pxor %mm2,%mm0 276 shlb $4,%cl 277 psrlq $4,%mm0 278 movq %mm1,%mm2 279 psrlq $4,%mm1 280 pxor 8(%esi,%ecx,1),%mm0 281 psllq $60,%mm2 282 andl $240,%edx 283 pxor (%eax,%ebp,8),%mm1 284 andl $15,%ebx 285 pxor (%esi,%ecx,1),%mm1 286 movd %mm0,%ebp 287 pxor %mm2,%mm0 288 psrlq $4,%mm0 289 movq %mm1,%mm2 290 psrlq $4,%mm1 291 pxor 8(%esi,%edx,1),%mm0 292 movb 11(%edi),%cl 293 psllq $60,%mm2 294 pxor (%eax,%ebx,8),%mm1 295 andl $15,%ebp 296 pxor (%esi,%edx,1),%mm1 297 movl %ecx,%edx 298 movd %mm0,%ebx 299 pxor %mm2,%mm0 300 shlb $4,%cl 301 psrlq $4,%mm0 302 movq %mm1,%mm2 303 psrlq $4,%mm1 304 pxor 8(%esi,%ecx,1),%mm0 305 psllq $60,%mm2 306 andl $240,%edx 307 pxor (%eax,%ebp,8),%mm1 308 andl $15,%ebx 309 pxor (%esi,%ecx,1),%mm1 310 movd %mm0,%ebp 311 pxor %mm2,%mm0 312 psrlq $4,%mm0 313 movq %mm1,%mm2 314 psrlq $4,%mm1 315 pxor 8(%esi,%edx,1),%mm0 316 movb 10(%edi),%cl 317 psllq $60,%mm2 318 pxor (%eax,%ebx,8),%mm1 319 andl $15,%ebp 320 pxor (%esi,%edx,1),%mm1 321 movl %ecx,%edx 322 movd %mm0,%ebx 323 pxor %mm2,%mm0 324 shlb $4,%cl 325 psrlq $4,%mm0 326 movq %mm1,%mm2 327 psrlq $4,%mm1 328 pxor 8(%esi,%ecx,1),%mm0 329 psllq $60,%mm2 330 andl $240,%edx 331 pxor (%eax,%ebp,8),%mm1 332 andl $15,%ebx 333 pxor (%esi,%ecx,1),%mm1 334 movd %mm0,%ebp 335 pxor %mm2,%mm0 336 psrlq $4,%mm0 337 movq %mm1,%mm2 338 psrlq $4,%mm1 339 pxor 8(%esi,%edx,1),%mm0 340 movb 9(%edi),%cl 341 psllq $60,%mm2 342 pxor (%eax,%ebx,8),%mm1 343 andl $15,%ebp 344 pxor (%esi,%edx,1),%mm1 345 movl %ecx,%edx 346 movd %mm0,%ebx 347 pxor %mm2,%mm0 348 shlb $4,%cl 349 psrlq $4,%mm0 350 movq %mm1,%mm2 351 psrlq $4,%mm1 352 pxor 8(%esi,%ecx,1),%mm0 353 psllq $60,%mm2 354 andl $240,%edx 355 pxor (%eax,%ebp,8),%mm1 356 andl $15,%ebx 357 pxor (%esi,%ecx,1),%mm1 358 movd %mm0,%ebp 359 pxor %mm2,%mm0 360 psrlq $4,%mm0 361 movq %mm1,%mm2 362 psrlq $4,%mm1 363 pxor 8(%esi,%edx,1),%mm0 364 movb 8(%edi),%cl 365 psllq $60,%mm2 366 pxor (%eax,%ebx,8),%mm1 367 andl $15,%ebp 368 pxor (%esi,%edx,1),%mm1 369 movl %ecx,%edx 370 movd %mm0,%ebx 371 pxor %mm2,%mm0 372 shlb $4,%cl 373 psrlq $4,%mm0 374 movq %mm1,%mm2 375 psrlq $4,%mm1 376 pxor 8(%esi,%ecx,1),%mm0 377 psllq $60,%mm2 378 andl $240,%edx 379 pxor (%eax,%ebp,8),%mm1 380 andl $15,%ebx 381 pxor (%esi,%ecx,1),%mm1 382 movd %mm0,%ebp 383 pxor %mm2,%mm0 384 psrlq $4,%mm0 385 movq %mm1,%mm2 386 psrlq $4,%mm1 387 pxor 8(%esi,%edx,1),%mm0 388 movb 7(%edi),%cl 389 psllq $60,%mm2 390 pxor (%eax,%ebx,8),%mm1 391 andl $15,%ebp 392 pxor (%esi,%edx,1),%mm1 393 movl %ecx,%edx 394 movd %mm0,%ebx 395 pxor %mm2,%mm0 396 shlb $4,%cl 397 psrlq $4,%mm0 398 movq %mm1,%mm2 399 psrlq $4,%mm1 400 pxor 8(%esi,%ecx,1),%mm0 401 psllq $60,%mm2 402 andl $240,%edx 403 pxor (%eax,%ebp,8),%mm1 404 andl $15,%ebx 405 pxor (%esi,%ecx,1),%mm1 406 movd %mm0,%ebp 407 pxor %mm2,%mm0 408 psrlq $4,%mm0 409 movq %mm1,%mm2 410 psrlq $4,%mm1 411 pxor 8(%esi,%edx,1),%mm0 412 movb 6(%edi),%cl 413 psllq $60,%mm2 414 pxor (%eax,%ebx,8),%mm1 415 andl $15,%ebp 416 pxor (%esi,%edx,1),%mm1 417 movl %ecx,%edx 418 movd %mm0,%ebx 419 pxor %mm2,%mm0 420 shlb $4,%cl 421 psrlq $4,%mm0 422 movq %mm1,%mm2 423 psrlq $4,%mm1 424 pxor 8(%esi,%ecx,1),%mm0 425 psllq $60,%mm2 426 andl $240,%edx 427 pxor (%eax,%ebp,8),%mm1 428 andl $15,%ebx 429 pxor (%esi,%ecx,1),%mm1 430 movd %mm0,%ebp 431 pxor %mm2,%mm0 432 psrlq $4,%mm0 433 movq %mm1,%mm2 434 psrlq $4,%mm1 435 pxor 8(%esi,%edx,1),%mm0 436 movb 5(%edi),%cl 437 psllq $60,%mm2 438 pxor (%eax,%ebx,8),%mm1 439 andl $15,%ebp 440 pxor (%esi,%edx,1),%mm1 441 movl %ecx,%edx 442 movd %mm0,%ebx 443 pxor %mm2,%mm0 444 shlb $4,%cl 445 psrlq $4,%mm0 446 movq %mm1,%mm2 447 psrlq $4,%mm1 448 pxor 8(%esi,%ecx,1),%mm0 449 psllq $60,%mm2 450 andl $240,%edx 451 pxor (%eax,%ebp,8),%mm1 452 andl $15,%ebx 453 pxor (%esi,%ecx,1),%mm1 454 movd %mm0,%ebp 455 pxor %mm2,%mm0 456 psrlq $4,%mm0 457 movq %mm1,%mm2 458 psrlq $4,%mm1 459 pxor 8(%esi,%edx,1),%mm0 460 movb 4(%edi),%cl 461 psllq $60,%mm2 462 pxor (%eax,%ebx,8),%mm1 463 andl $15,%ebp 464 pxor (%esi,%edx,1),%mm1 465 movl %ecx,%edx 466 movd %mm0,%ebx 467 pxor %mm2,%mm0 468 shlb $4,%cl 469 psrlq $4,%mm0 470 movq %mm1,%mm2 471 psrlq $4,%mm1 472 pxor 8(%esi,%ecx,1),%mm0 473 psllq $60,%mm2 474 andl $240,%edx 475 pxor (%eax,%ebp,8),%mm1 476 andl $15,%ebx 477 pxor (%esi,%ecx,1),%mm1 478 movd %mm0,%ebp 479 pxor %mm2,%mm0 480 psrlq $4,%mm0 481 movq %mm1,%mm2 482 psrlq $4,%mm1 483 pxor 8(%esi,%edx,1),%mm0 484 movb 3(%edi),%cl 485 psllq $60,%mm2 486 pxor (%eax,%ebx,8),%mm1 487 andl $15,%ebp 488 pxor (%esi,%edx,1),%mm1 489 movl %ecx,%edx 490 movd %mm0,%ebx 491 pxor %mm2,%mm0 492 shlb $4,%cl 493 psrlq $4,%mm0 494 movq %mm1,%mm2 495 psrlq $4,%mm1 496 pxor 8(%esi,%ecx,1),%mm0 497 psllq $60,%mm2 498 andl $240,%edx 499 pxor (%eax,%ebp,8),%mm1 500 andl $15,%ebx 501 pxor (%esi,%ecx,1),%mm1 502 movd %mm0,%ebp 503 pxor %mm2,%mm0 504 psrlq $4,%mm0 505 movq %mm1,%mm2 506 psrlq $4,%mm1 507 pxor 8(%esi,%edx,1),%mm0 508 movb 2(%edi),%cl 509 psllq $60,%mm2 510 pxor (%eax,%ebx,8),%mm1 511 andl $15,%ebp 512 pxor (%esi,%edx,1),%mm1 513 movl %ecx,%edx 514 movd %mm0,%ebx 515 pxor %mm2,%mm0 516 shlb $4,%cl 517 psrlq $4,%mm0 518 movq %mm1,%mm2 519 psrlq $4,%mm1 520 pxor 8(%esi,%ecx,1),%mm0 521 psllq $60,%mm2 522 andl $240,%edx 523 pxor (%eax,%ebp,8),%mm1 524 andl $15,%ebx 525 pxor (%esi,%ecx,1),%mm1 526 movd %mm0,%ebp 527 pxor %mm2,%mm0 528 psrlq $4,%mm0 529 movq %mm1,%mm2 530 psrlq $4,%mm1 531 pxor 8(%esi,%edx,1),%mm0 532 movb 1(%edi),%cl 533 psllq $60,%mm2 534 pxor (%eax,%ebx,8),%mm1 535 andl $15,%ebp 536 pxor (%esi,%edx,1),%mm1 537 movl %ecx,%edx 538 movd %mm0,%ebx 539 pxor %mm2,%mm0 540 shlb $4,%cl 541 psrlq $4,%mm0 542 movq %mm1,%mm2 543 psrlq $4,%mm1 544 pxor 8(%esi,%ecx,1),%mm0 545 psllq $60,%mm2 546 andl $240,%edx 547 pxor (%eax,%ebp,8),%mm1 548 andl $15,%ebx 549 pxor (%esi,%ecx,1),%mm1 550 movd %mm0,%ebp 551 pxor %mm2,%mm0 552 psrlq $4,%mm0 553 movq %mm1,%mm2 554 psrlq $4,%mm1 555 pxor 8(%esi,%edx,1),%mm0 556 movb (%edi),%cl 557 psllq $60,%mm2 558 pxor (%eax,%ebx,8),%mm1 559 andl $15,%ebp 560 pxor (%esi,%edx,1),%mm1 561 movl %ecx,%edx 562 movd %mm0,%ebx 563 pxor %mm2,%mm0 564 shlb $4,%cl 565 psrlq $4,%mm0 566 movq %mm1,%mm2 567 psrlq $4,%mm1 568 pxor 8(%esi,%ecx,1),%mm0 569 psllq $60,%mm2 570 andl $240,%edx 571 pxor (%eax,%ebp,8),%mm1 572 andl $15,%ebx 573 pxor (%esi,%ecx,1),%mm1 574 movd %mm0,%ebp 575 pxor %mm2,%mm0 576 psrlq $4,%mm0 577 movq %mm1,%mm2 578 psrlq $4,%mm1 579 pxor 8(%esi,%edx,1),%mm0 580 psllq $60,%mm2 581 pxor (%eax,%ebx,8),%mm1 582 andl $15,%ebp 583 pxor (%esi,%edx,1),%mm1 584 movd %mm0,%ebx 585 pxor %mm2,%mm0 586 movl 4(%eax,%ebp,8),%edi 587 psrlq $32,%mm0 588 movd %mm1,%edx 589 psrlq $32,%mm1 590 movd %mm0,%ecx 591 movd %mm1,%ebp 592 shll $4,%edi 593 bswap %ebx 594 bswap %edx 595 bswap %ecx 596 xorl %edi,%ebp 597 bswap %ebp 598 ret 599 .size _mmx_gmult_4bit_inner,.-_mmx_gmult_4bit_inner 600 .globl gcm_gmult_4bit_mmx 601 .type gcm_gmult_4bit_mmx,@function 602 .align 16 603 gcm_gmult_4bit_mmx: 604 .L_gcm_gmult_4bit_mmx_begin: 605 pushl %ebp 606 pushl %ebx 607 pushl %esi 608 pushl %edi 609 movl 20(%esp),%edi 610 movl 24(%esp),%esi 611 call .L005pic_point 612 .L005pic_point: 613 popl %eax 614 leal .Lrem_4bit-.L005pic_point(%eax),%eax 615 movzbl 15(%edi),%ebx 616 call _mmx_gmult_4bit_inner 617 movl 20(%esp),%edi 618 emms 619 movl %ebx,12(%edi) 620 movl %edx,4(%edi) 621 movl %ecx,8(%edi) 622 movl %ebp,(%edi) 623 popl %edi 624 popl %esi 625 popl %ebx 626 popl %ebp 627 ret 628 .size gcm_gmult_4bit_mmx,.-.L_gcm_gmult_4bit_mmx_begin 629 .globl gcm_ghash_4bit_mmx 630 .type gcm_ghash_4bit_mmx,@function 631 .align 16 632 gcm_ghash_4bit_mmx: 633 .L_gcm_ghash_4bit_mmx_begin: 634 pushl %ebp 635 pushl %ebx 636 pushl %esi 637 pushl %edi 638 movl 20(%esp),%ebp 639 movl 24(%esp),%esi 640 movl 28(%esp),%edi 641 movl 32(%esp),%ecx 642 call .L006pic_point 643 .L006pic_point: 644 popl %eax 645 leal .Lrem_4bit-.L006pic_point(%eax),%eax 646 addl %edi,%ecx 647 movl %ecx,32(%esp) 648 subl $20,%esp 649 movl 12(%ebp),%ebx 650 movl 4(%ebp),%edx 651 movl 8(%ebp),%ecx 652 movl (%ebp),%ebp 653 jmp .L007mmx_outer_loop 654 .align 16 655 .L007mmx_outer_loop: 656 xorl 12(%edi),%ebx 657 xorl 4(%edi),%edx 658 xorl 8(%edi),%ecx 659 xorl (%edi),%ebp 660 movl %edi,48(%esp) 661 movl %ebx,12(%esp) 662 movl %edx,4(%esp) 663 movl %ecx,8(%esp) 664 movl %ebp,(%esp) 665 movl %esp,%edi 666 shrl $24,%ebx 667 call _mmx_gmult_4bit_inner 668 movl 48(%esp),%edi 669 leal 16(%edi),%edi 670 cmpl 52(%esp),%edi 671 jb .L007mmx_outer_loop 672 movl 40(%esp),%edi 673 emms 674 movl %ebx,12(%edi) 675 movl %edx,4(%edi) 676 movl %ecx,8(%edi) 677 movl %ebp,(%edi) 678 addl $20,%esp 679 popl %edi 680 popl %esi 681 popl %ebx 682 popl %ebp 683 ret 684 .size gcm_ghash_4bit_mmx,.-.L_gcm_ghash_4bit_mmx_begin 685 .align 64 686 .Lrem_4bit: 687 .long 0,0,0,29491200,0,58982400,0,38141952 688 .long 0,117964800,0,113901568,0,76283904,0,88997888 689 .long 0,235929600,0,265420800,0,227803136,0,206962688 690 .long 0,152567808,0,148504576,0,177995776,0,190709760 691 .align 64 692 .L008rem_8bit: 693 .value 0,450,900,582,1800,1738,1164,1358 694 .value 3600,4050,3476,3158,2328,2266,2716,2910 695 .value 7200,7650,8100,7782,6952,6890,6316,6510 696 .value 4656,5106,4532,4214,5432,5370,5820,6014 697 .value 14400,14722,15300,14854,16200,16010,15564,15630 698 .value 13904,14226,13780,13334,12632,12442,13020,13086 699 .value 9312,9634,10212,9766,9064,8874,8428,8494 700 .value 10864,11186,10740,10294,11640,11450,12028,12094 701 .value 28800,28994,29444,29382,30600,30282,29708,30158 702 .value 32400,32594,32020,31958,31128,30810,31260,31710 703 .value 27808,28002,28452,28390,27560,27242,26668,27118 704 .value 25264,25458,24884,24822,26040,25722,26172,26622 705 .value 18624,18690,19268,19078,20424,19978,19532,19854 706 .value 18128,18194,17748,17558,16856,16410,16988,17310 707 .value 21728,21794,22372,22182,21480,21034,20588,20910 708 .value 23280,23346,22900,22710,24056,23610,24188,24510 709 .value 57600,57538,57988,58182,58888,59338,58764,58446 710 .value 61200,61138,60564,60758,59416,59866,60316,59998 711 .value 64800,64738,65188,65382,64040,64490,63916,63598 712 .value 62256,62194,61620,61814,62520,62970,63420,63102 713 .value 55616,55426,56004,56070,56904,57226,56780,56334 714 .value 55120,54930,54484,54550,53336,53658,54236,53790 715 .value 50528,50338,50916,50982,49768,50090,49644,49198 716 .value 52080,51890,51444,51510,52344,52666,53244,52798 717 .value 37248,36930,37380,37830,38536,38730,38156,38094 718 .value 40848,40530,39956,40406,39064,39258,39708,39646 719 .value 36256,35938,36388,36838,35496,35690,35116,35054 720 .value 33712,33394,32820,33270,33976,34170,34620,34558 721 .value 43456,43010,43588,43910,44744,44810,44364,44174 722 .value 42960,42514,42068,42390,41176,41242,41820,41630 723 .value 46560,46114,46692,47014,45800,45866,45420,45230 724 .value 48112,47666,47220,47542,48376,48442,49020,48830 725 .byte 71,72,65,83,72,32,102,111,114,32,120,56,54,44,32,67 726 .byte 82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112 727 .byte 112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62 728 .byte 0 729
